‹ 2026-08-08

2026-08-09

2026-08-10 › Today

72 works · 3 categories · 35 topics · twitter 10 wechat 30 blog 7 arxiv 30 generated 2026-08-10 02:59:38 UTC

Top highlights — Research

Science重磅：AI首次生成完整噬菌体基因组，可存活、能抑菌 is the headline result: Stanford's Evo 1/Evo 2 genome language models designed complete phage genomes from scratch, yielding 16 viable synthetic phages that lyse E. coli C — the first demonstration that generative AI can produce functional life systems at genome scale, not just single genes or proteins.
Transformers Struggle to Use Their Emergent World Models: Revisiting the Tower of Hanoi, and the Illusion of Thinking sharpens the reasoning debate: both small and frontier models learn a linearly decodable, geometrically faithful Sierpinski-triangle representation of the puzzle, but large models lose it during long chain-of-thought — recasting the "illusion of thinking" collapse as a maintenance failure, not missing world models.
上海AI Lab、浙大、NUS团队：世界模型，何去何从？ argues the field should pivot from "predicting the world" to an agent-centric "World Proxy" that returns cheap, controllable feedback for planning and self-improvement, replacing visual-realism benchmarks with actionable information gain — a useful counterpoint to the Hanoi findings.
Explanation Stability of Test-Time Adaptation in Computational Pathology: A Large-Scale Benchmark is the practical deployment warning: across 2,958 adaptation runs, models adapting to scanner/staining/cohort shift can silently lose attribution fidelity while accuracy metrics stay flat, making explanation stability its own reliability axis.

LLM Agents 8

SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent

arXiv cs.AI

Rank 78 · Content 80 · Popularity 73

TL;DR - SkillProx is a proximal-gradient-inspired framework for evolving an LLM agent's textual "skills" (reusable procedural notes loaded into context, no weight updates), pairing closed-loop diagnostic edits with a utility-aware pruning stage. It matters because it treats skill deletion/consolidation as a first-class operation rather than a generic edit, addressing context bloat in self-improving agents.

Forward stage: re-executes diagnosis-driven edits on the same task batch, rolls back regressions, and feeds measured outcomes back into later diagnoses — creating explicit diagnosis→outcome feedback missing in prior text-gradient methods.
Backward (proximal) stage: decomposes a skill into auditable knowledge units, scores each via a frozen leave-one-out utility audit, then applies validation-gated consolidation, demotion, or removal.
Framing is a composite objective trading off task loss against skill complexity, mirroring proximal gradient descent in text space.
Reported +3.0 percentage points average accuracy over the strongest gradient-based baseline across in-distribution and OOD benchmarks and multiple backbone LLMs; ablations show closed-loop diagnosis and proximal refinement contribute complementarily.

details →

Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory

arXiv cs.AI

Rank 77 · Content 80 · Popularity 71

TL;DR - A training-free framework, Agent Memory Distillation (AMD), transfers structured memory from a large teacher agent to small (4B–8B) student LLM agents, substantially improving tool-use accuracy without any weight updates. It matters because small agents typically can't bootstrap their own successful trajectories for memory-based self-improvement.

Builds three complementary memory types from successful teacher trajectories: Workflow (task-level strategies), Subtask (intermediate-granularity behavioral examples), and Function (per-function calling conventions and pitfalls).
Injection is hybrid: Workflow and Subtask memories are added proactively at task start, while Function memory is retrieved reactively only when a tool call errors.
With GPT-5-mini as teacher, average accuracy gains were 27.2 percentage points on AppWorld, 11.2p on BFCL V3, and 3.4p on ToolSandbox across four student models, beating existing memory-based baselines.
Ablations attribute the largest share of gains to Subtask memory; effectiveness depends on both teacher capability and teacher–student compatibility, with 4B students benefiting most.

details →

Fisher-R1: Training LLM Agents for Reliable Hypothesis Testing

arXiv cs.AI

Rank 75 · Content 80 · Popularity 65

TL;DR - An arXiv paper showing that LLM agents automating statistical hypothesis testing often reach wrong conclusions through subtle inferential errors even when code executes correctly, and introducing Fisher-R1, an open-weight agent trained via RL to do this reliably. It matters because agentic "AI scientist" pipelines are increasingly trusted to produce empirical claims that current benchmarks don't validate.

P-Bench: 425 open-ended, realistic hypothesis-testing tasks across economics, biology, and medicine; each requires selecting a statistical method, computing a p-value, and drawing a conclusion from only a hypothesis plus a dataset.
Existing benchmarks miss this failure mode because they rarely check whether a reported p-value is statistically valid given the assumptions underlying the data — correct execution ≠ correct inference.
Fisher-R1 is trained on synthetic tasks with reinforcement learning using a verified statistical reward signal; the 14B model beats its own backbone and strong proprietary/open baselines.
Reported gains: ~21% average relative improvement in single-trial success over DeepSeek-V4-Pro, up to 26% on the hardest tasks, also outperforming GPT-5.4.

details →

DiDPO: Diff-in-Diff Policy Optimization for Coding Agent Training

arXiv cs.AI

Rank 69 · Content 80 · Popularity 45

TL;DR - DiDPO is a critic-free reinforcement learning method that derives fine-grained credit units directly from the structure of code diffs, addressing the fact that a single coding action bundles multiple independent edits whose individual contributions outcome- or step-level RLVR rewards cannot separate.

Organizes multi-turn coding interactions into thought–action steps, then discovers and compares code diffs across sampled trajectories rather than scoring whole trajectories or whole steps.
Splits each whole diff into sub-diffs using a "groupability score" that balances an anchor's semantic scope against the group mass it can form; highly similar sub-diffs are aggregated into anchors.
Anchors form advantage groups, and the diff-level advantage is projected back onto individual response tokens for training.
Reported to beat strong agentic RL baselines on long-horizon coding and reasoning benchmarks, exceeding comparable methods by over 10% on Qwen2.5-7B-Coder; the authors also open-source verl-code, an agentic RL codebase.

details →

Strategy-first synthesis planning for complex natural products

arXiv cs.MA

Rank 69 · Content 80 · Popularity 42

TL;DR - SynthEx is an LLM-based agentic framework that plans retrosynthetic routes to complex natural products that conventional, catalogue-driven design algorithms cannot handle, and expert chemists rated its key steps comparable to published human syntheses in blinded assessment.

Frames the gap in existing tools: they are tuned to catalogued reaction benchmarks and report near-complete success there, but falter on densely functionalized, polycyclic natural products where inventive chemistry is least represented in the record.
The agent proposes competing strategies, composes routine plus key steps into a cohesive route, then critiques and revises its own designs — a strategy-first rather than step-by-step search approach.
Reported route characteristics: more convergent than existing tools' output, and covering a region of reaction space catalogue-based methods cannot reach.
Releases SynthAtlas, an open interactive database of routes to 1,000+ natural products, targeting molecules lacking existing literature routes.

details →

上海AI Lab、浙大、NUS团队：世界模型，何去何从？

WeChat: 学术头条

Rank 64 · Content 75 · Popularity 40

TL;DR - A position/survey paper from Shanghai AI Lab, Zhejiang University, and NUS (arXiv:2608.02713) proposes shifting world models from "predicting the world" to an agent-centric "World Proxy" that sits between an agent and the real environment, returning low-cost, controllable feedback for planning, learning, and self-improvement. It reframes the evaluation target from visual realism/state-prediction accuracy to actionable information gain for agents.

Three requirements for a World Proxy: closed-loop agent-initiated queries/actions/interventions; environment grounding learned from real data, rules, trajectories, or interaction evidence; and optimization for actionable information gain rather than photorealism.
Three intervention levels: L1 inference-time guidance (memory/skill retrieval, execution simulation, verification injected into context, no parameter changes); L2 training-time optimization (scoring rollouts, failure diagnosis, synthetic trajectories/preference pairs feeding SFT, DPO, PPO, GRPO); L3 agent-proxy co-evolution (real trajectories flow back to update the proxy, which then retrains the agent).
Six functional forms by modeled information transfer: dynamics (next state/reward), spatial (observations from new viewpoints/poses), execution (code, commands, web clicks, API/tool results including stdout/stderr/tests), memory/experience (retrieved past failures and constraints), skill (reusable tools and action priors), and reward/verification (critique, preferences, safety checks).
Open problems named: limited fidelity and error accumulation over long trajectories, no reliable signal for when an agent should distrust the proxy and fall back to the real environment, reward hacking when the proxy acts as verifier, and the absence of agent-centric benchmarks that measure whether feedback actually improves the agent.

details →

Benchmarking the Benchmarks: Evaluating Benchmarks for Conversational Agents

arXiv cs.CL

Rank 58 · Content 65 · Popularity 42

TL;DR - An arXiv cs.CL paper proposing a reference-free, LLM-judge framework to audit the quality of benchmarks used to evaluate task-oriented conversational agents. It matters because unreliable benchmarks silently corrupt agent evaluation, and this gives a way to score the scorers.

Scores benchmarks along three axes — consistency, complexity, and policy coverage — and emits actionable diagnostics pinpointing specific weaknesses, without needing reference/gold benchmarks.
Validated three ways: agreement with independent human annotations, discrimination between benchmarks generated by LLMs of differing capability, and sensitivity to controlled quality-degrading perturbations.
Metrics reportedly separate quality levels consistently across multiple domains and across different judge models, suggesting the approach isn't tied to one judge.
Applies to both synthetic (LLM-generated) and manually curated benchmarks; the abstract reports qualitative consistency rather than specific numeric scores.

details →

ICML 2026 | 会查表却不会预测未来：南大TopBench测出大模型数据盲区

WeChat: PaperWeekly

Rank 53 · Content 60 · Popularity 38

TL;DR - TopBench (Nanjing University, ICML 2026) is a benchmark for "implicit predictive reasoning" in tabular QA, where the answer isn't in the table and the model must first infer that a natural-language request is actually a prediction task, then build a model to answer it. It matters because current LLMs handle lookup/aggregation well but frequently fail to even enter prediction mode, a prerequisite for autonomous data-analysis agents.

Setup: 35 real tables (Kaggle-sourced; healthcare, finance, daily consulting; <1K to >6M rows) and 779 queries across four task types — Single-Point Prediction (274), Decision Making (186), Treatment Effect Analysis (105), Ranking & Filtering (214); 384 regression vs 395 classification targets. Queries use dual personas (lay user vs data owner) and logic-driven sampling with hard-negative pairs and noisy candidate pools.
Evaluation: LLM-as-a-Judge with string/fuzzy/NLI re-anchoring to the model's original text to curb judge hallucination; regression scored via a composite of point error, interval coverage and an over-wide-interval penalty; Ranking & Filtering scored deterministically on generated CSVs via F1, Set Recall, NDCG and batch NMAE.
Results: Most models score below 0.60; the best reaches only ~0.65 on basic single-point prediction. Code execution isn't a fix — GPT-5.2 gains on Treatment Effect (0.51→0.65) but Qwen3-Instruct drops on single-point (0.57→0.43) by writing pandas filters instead of training a predictor, and Qwen3-Thinking loops through rows searching for an exact match ("Exhaustive Retrieval Loop") until context is exhausted.
Two bottlenecks isolated: supplying target column/task type/feature descriptions lifts scores (Qwen3-Instruct 0.43→0.56; DeepSeek-V3.2 0.57→0.68), showing intent alignment is one blocker; a predict-only ensemble given gold task structure still beats the best agentic E2E run (0.66→0.76 single-point), showing tabular modeling quality is a second, independent blocker.

details →

Medical/Healthcare AI 7

ResidencyRL: Reinforcement Learning in Simulated Clinical Environments

arXiv cs.AI

Rank 75 · Content 80 · Popularity 65

TL;DR - ResidencyRL is a multi-turn reinforcement learning method that trains clinical LLM agents in simulated patient encounters, mimicking medical residency. It shows that sequential clinical decision-making — not just static benchmark QA — can be learned in simulation and transfers to unseen evaluations.

Trains a policy agent against LLM patient/environment simulators capable of complex, adversarial behavior, over trajectories of up to 60 dialogue turns and 8 tool calls.
Uses a structured reward spanning diagnostic accuracy, management quality, communication, documentation, and safety.
Reported gains: +7.0% diagnostic accuracy under adversarial conditions (88.0% vs. 81.0%), 31% fewer missed red flags, and blinded clinician preference in 87.6% of side-by-side comparisons.
Transfers to held-out benchmarks: beats the base model on all six clinical axes of AMIE multi-visit, with directional improvements on AgentClinic and CRAFT-MD; authors note prospective real-world validation is still needed.

details →

Explanation Stability of Test-Time Adaptation in Computational Pathology: A Large-Scale Benchmark

arXiv cs.CV

Rank 72 · Content 85 · Popularity 42

TL;DR - A large-scale benchmark (2,958 adaptation runs) measuring how much test-time adaptation shifts model explanations in computational pathology, arguing explanation stability is a reliability axis distinct from accuracy. It matters because clinically deployed models adapting to staining/scanner/cohort shifts can silently lose attribution fidelity while accuracy metrics look fine.

Scope: two histopathology benchmarks (Camelyon17, NCT CRC-HE), five architectures (CNNs through vision transformers and a pathology foundation model), seventeen TTA methods, and four attribution families.
Frozen-backbone TTA methods leave attributions nearly unchanged; continual methods like CoTTA and RoTTA cause the largest explanation drift.
Sensitivity is architecture-dependent — CNNs drift substantially more than transformer and foundation-model backbones — and drift grows with adaptation strength but is largely insensitive to batch size.
Explanation stability correlates only weakly with adaptation quality: some methods preserve attributions while degrading calibration or accuracy, yielding "silent failures" missed by accuracy-only or explanation-only evaluation. Metric, protocol, and full benchmark released as code.

details →

EliSeg: Verified Target Construction for Report-Grounded Abnormality Segmentation

arXiv cs.CV

Rank 69 · Content 80 · Popularity 42

TL;DR - EliSeg is an actor–verify–revise framework for segmenting abnormalities in chest X-rays directly from raw radiology reports, without any target identity, point, or box prompt. It matters because it removes the "hidden target oracle" assumption in prior segmentation work, forcing the model to decide which findings are actually eligible targets and how many masks to produce.

Defines report-grounded abnormality segmentation: the model must resolve target eligibility, cardinality, and finding-to-mask correspondence from an unfiltered report that may contain negated, prior, uncertain, or irrelevant findings.
Architecture: a grammar-constrained Actor proposes target slots and masks; an independent text-only Verifier reconstructs the eligible finding inventory; Revision re-executes the shared Actor only when the two target structures disagree.
On MIMIC-CXR-ILS it reportedly outperforms both direct segmentation methods and extract-then-segment cascades across findings, while suppressing masks for ineligible report mentions; CheXlocalize is used to show external transfer.
Ablations are reported as confirming that verification and revision play complementary roles; code is released at github.com/Maybach-dream/EliSeg.

details →

Representation-driven Endoscopic Visual Embedding Alignment for Latent Generation

arXiv cs.CV

Rank 69 · Content 80 · Popularity 42

TL;DR - REVEAL is a diffusion-transformer generative foundation model for endoscopy, trained on GastroNet-5M (5M multicenter frames), that aligns diffusion latents with in-domain endoscopic encoders instead of out-of-domain priors. It matters because it cuts the compute barrier to building specialized clinical imaging tools while doubling as a strong feature extractor.

Uses representation alignment with encoders pretrained on the endoscopic distribution, arguing out-of-domain priors fail to preserve fine textures and anatomical structure in this specialized image space.
Claimed to be the largest endoscopy generative foundation model to date; trained on the multicenter GN-5M dataset of 5 million frames.
As a feature extractor it is competitive with, and sometimes exceeds, classification-tuned endoscopic foundation models (EndoViT, Endo-FM), with reported robustness under realistic imaging corruptions.
Supports high-fidelity synthesis plus latent-space edits (inpainting, outpainting) with structural coherence, and is positioned as a backbone for conditional synthesis, segmentation, and OOD detection.

details →

ECHO: A Locally-Deployable Agentic Health Assistant with Temporal Memory, Safety Guardrails, and Speech Assessment

arXiv cs.AI

Rank 65 · Content 75 · Popularity 42

TL;DR - ECHO is a locally-deployable agentic health assistant for chronic care that combines a LangGraph/ReAct chatbot with persistent temporal memory, a two-stage safety layer, and multimodal speech assessment. It matters because it shows a full privacy-preserving clinical agent stack running on consumer hardware without sending patient data off-device.

Core agent uses a ReAct loop over LangGraph with 17 clinical tools and a temporal knowledge graph for cross-session memory; 94.9% tool-execution pass rate on a 59-scenario benchmark with GPT-5 Mini.
Hybrid safety layer: sub-1ms rule-based filtering for explicit crisis/jailbreak signals, plus a signed GNN with APPNP-style propagation for boundary cases — 88.8% accuracy and 90.6% unsafe recall on 2,537 annotated Turkish health queries, beating zero-shot LLM baselines including Llama 3.3 70B.
Speech module fuses Whisper acoustic and BERT text encodings via cross-attention to estimate emotion, depression, and pain, at 0.652 mean macro F1.
Delivered as a web app running entirely on consumer hardware with no external data transmission, targeting GDPR and KVKK compliance.

details →

Bayesian adaptively-weighted ensembles for few-shot abdominal segmentation

arXiv cs.CV

Rank 62 · Content 70 · Popularity 42

TL;DR - An arXiv preprint proposing a Bayesian adaptively-weighted ensemble that combines multiple few-shot segmentation algorithms for abdominal/pelvic anatomy, tuning ensemble weights per target domain. It matters because it addresses the practical bottleneck of deploying segmentation models at new clinical sites with scarce labels and institutional domain shift.

Multiple few-shot segmentation algorithms are first adapted on a small labelled support set; Bayesian optimisation then searches ensemble weights that maximise performance on a target-domain validation set.
Learned weights are frozen and applied to unseen query images from the same target domain, unlike fixed-weight ensembles that cannot adapt model contributions across anatomies or institutions.
Evaluated on the Cross-institution Male Pelvic Structures dataset using held-out anatomical structures and held-out institutions to jointly simulate label scarcity and domain shift.
Reported statistically significant gains over individual few-shot learners, fixed-weight ensembles, train-from-scratch baselines, and recent state-of-the-art ensembling methods (no numeric results given in the provided abstract).

details →

Is Self-Pretraining really useful to improve diagnosis in medical Time Series?

arXiv cs.LG

Rank 58 · Content 65 · Popularity 42

TL;DR - An arXiv study testing whether Self-PreTraining (SPT) with masking objectives helps transformers on medical time-series classification, finding consistent but modest gains across three clinical datasets. It matters because SPT is an architecture-agnostic way to squeeze more accuracy out of data-limited clinical settings.

Evaluated transformers on three tasks: rehabilitation robotics (Camargo), stress detection (Non-EEG Stress), and Parkinson's disease detection (Gait PD), training either from scratch or via SPT.
Four masking-based pre-training objectives targeting temporal and cross-modal representation learning yielded 0–6 percentage point accuracy improvements, varying by masking strategy, dataset, and architecture.
Gains held not only for multimodal/multivariate inputs but also for simple univariate inputs, suggesting the benefit is not purely cross-modal.
Model depth was varied systematically; deeper models benefited more, indicating capacity is needed to exploit the pre-trained temporal representations.

details →

Bioinformatics AI 4

Representative image for Science重磅：AI首次生成完整噬菌体基因组，可存活、能抑菌

Science重磅：AI首次生成完整噬菌体基因组，可存活、能抑菌 🔗 2 sources

WeChat: DrugAI WeChat: 学术头条

Rank 85 · Content 95 · Popularity 60

TL;DR — 斯坦福团队（King et al., Science, 2026）用基因组语言模型 Evo 1/Evo 2 从头生成完整噬菌体基因组，获得 16 株可存活、能裂解大肠杆菌 C 的合成噬菌体，首次证明生成式 AI 能设计基因组尺度的功能性生命系统，而不止于单基因、单蛋白或基因线路。

模型与生成流程：Evo 1/Evo 2 在超 200 万条噬菌体基因组上预训练，再用约 15,000 条微小噬菌体科（Microviridae）序列微调，以 ΦX174（5.4 kb 单链 DNA、11 个基因）起始区共识序列作提示（约 4–9 nt，采样温度 0.7–0.9），Evo 2 的基础生成能力更强。
多层过滤与产出：按长度、GC 含量、编码密度、ΦX174 式基因架构、宿主嗜性（主刺突蛋白相似度）及与野生型的进化距离筛选；302 条设计合成后 285 条组装成功，16 株（Evo-Φ）可抑制/裂解 E. coli C，且不抑制另外 6 株菌，宿主特异性保留；存活率与同天然基因组的相似度正相关。
新颖序列组合：设计中出现基因增删、基因顺序改变和非编码调控区变化，13 株可存活基因组带有无法由任何单一天然序列解释的突变；复制起点却高度保守，说明模型隐式学到了区域性约束。Evo-Φ36 将 DNA 包装蛋白 J 换成远缘噬菌体 G4 的较短同源物（此前文献报道该组合不可行），冷冻电镜确认衣壳相互作用兼容、颗粒稳定。
功能与抗性突破：生成噬菌体的裂解动力学跨度大于天然 ΦX174 类噬菌体；其鸡尾酒组合在 1–2 轮传代内攻克 ΦX174 抗性菌株 CR1/CR2，而天然噬菌体鸡尾酒 5 轮后仍失败，突破株源自 2–3 个生成基因组之间的重组——提示噬菌体疗法价值。
局限与治理：目前仅限小基因组，受训练数据偏倚限制；作者明确提示生物安全与生物安保风险，呼吁安全专家介入设计全生命周期并加强审查。

不同来源侧重：学术头条更强调抗性突破细节与生物安全治理呼吁，DrugAI 更强调模型机制、过滤条件与"基因组尺度生成"的里程碑定位。

details →

Nat. Rev. Drug Discov. | 药物发现中的人工智能：内涵、现状与未来路径

WeChat: DrugAI

Rank 68 · Content 75 · Popularity 53

TL;DR - A Nature Reviews Drug Discovery perspective (Bender, Thomas, Scannell et al., 2026) critically reviews AI in drug discovery, arguing that technical capability has advanced far faster than clinical impact, and proposes shifting evaluation from model validation to whether models actually improve project decisions.

Impact is misallocated: most AI work targets preclinical stages (hit finding) where labeled data is easy, while the biggest lever on cost-per-approved-drug is Phase II success — e.g. biomarker-based patient stratification roughly halves capitalized cost per launched drug ("looking under the streetlight").
Data is the bottleneck, not modeling: biological labels are scarce, high-dimensional, confounded, and "epistemically opaque" — whether a compound "hits a target" depends on assay type and ATP concentration; proxy readouts like liver-organoid cytotoxicity correlate poorly with clinical DILI, and thermal-shift assays only marginally track enzyme inhibition.
Chemical space is vast and local: real datasets cover a tiny biased fraction; the same functional group has different effects on different scaffolds, and four common ADME datasets share very few compounds or scaffolds, so applicability domains don't align for multi-objective optimization.
Validation must be use-case-specific: selection (hit ranking), exclusion (tox/clearance triage), and quantitative prediction (dose) demand different metrics — two models with similar AUC can behave oppositely. AlphaFold is cited as model-validation success that did not automatically translate into drug-discovery-process success; the proposed path forward is human-relevant data generation (iPSC models, organ-on-chip, Cell Painting, perturbation atlases), preclinical–clinical data feedback loops including negative outcomes, and consortium-scale purpose-built datasets.

details →

EpiBench | 大语言模型能否真正理解抗体表位？

WeChat: DrugAI

Rank 64 · Content 80 · Popularity 26

TL;DR - EpiBench is a new zero-shot benchmark (1,609 curated samples, 5 linked tasks) testing whether general LLMs can reason about antibody epitopes directly from antigen/antibody sequences rather than recite biomedical facts; results show current models capture some epitope-like sequence signal but fail at antibody-specific binding assignment.

Five chained tasks span targetable-region discovery, antibody-specific epitope identification, epitope binning, functional-epitope assessment (blocking/protection/ADCC), and mutation-driven escape prediction; labels derive from antibody–antigen complex structures, functional B-cell assays, and deep mutational scanning.
Nine general LLMs were evaluated zero-shot; they do reasonably on generic "which region looks epitope-like" but degrade sharply on antibody-specific localization and escape calls, and task-specific epitope predictors still outperform them where directly comparable.
Performance drops as antigen sequence length grows — long-context capability does not equal accurate residue-level tracking in long proteins.
Chain-of-thought gives inconsistent gains; error analysis shows models lean on motifs, physicochemical properties, CDR similarity, and memorized antibody knowledge. One SARS-CoV-2 RBD case applied correct escape knowledge at mis-mapped coordinates, yielding an inverted prediction — positioning EpiBench as a diagnostic tool for sequence-grounding gaps.

details →

CohortHijack: Robustness of Single Cell Annotation to Companion Cell Removal

arXiv cs.LG

Rank 62 · Content 70 · Popularity 42

TL;DR - An arXiv preprint introducing CohortHijack, a robustness audit showing that single-cell annotation tools which refine labels via neighboring cells or cluster voting can be flipped simply by removing other (non-target) cells from the query cohort. It matters because it exposes query cohort composition itself as a target-preserving attack surface in single-cell pipelines.

The attack removes selected companion cells while leaving the target cell's expression profile, base prediction, and trained model unchanged; removal strategies include random and structured selection with greedy, multi-start, and beam search.
On PBMC3K and Paul15 with logistic regression and calibrated linear SVM, multi-start search flipped 24.33% of linear-SVM targets and 19.67% of logistic-regression targets, removing only a small cohort fraction and keeping mean collateral changes under 0.4%.
Structured removal consistently beat random removal on Paul15, and ablations showed the effect vanished when neighborhood refinement was disabled — pinning the vulnerability on the refinement step, not the classifier.
With CellTypist, independent per-cell predictions never changed, but majority-voting refined labels did shift after small companion-cell removals.

details →

LLMs & Foundation Models 5

CreativeInstruct: Scalably Teaching LLMs to Balance Quality, Creativity, and Diversity

arXiv cs.CL

Rank 69 · Content 80 · Popularity 42

TL;DR - CreativeInstruct is an instruction-tuning method that teaches LLMs to emit special [StartCreativity] spans, letting a single post-trained model recover base-model-like creativity and diversity without losing output quality. It matters because post-training's diversity collapse hurts both creative writing and downstream RL exploration.

Injects learned [StartCreativity] control spans during instruction tuning to bias generation toward creative modes, avoiding the need for multiple models at inference time.
Introduces a structural diversity metric based on graph edit distance, aimed at capturing narrative-level variation that lexical and semantic metrics miss.
On narrative generation it matches or exceeds diversity of multi-model baselines and their distilled variants at equal quality; human annotators rated its outputs more creative than post-trained LLM outputs in 70.3% of cases.
Creativity transfers to RL: GRPO on a CreativeInstruct checkpoint gains ~4% on AMC and ~5 points on MATH versus the same training from the standard post-trained checkpoint.

details →

Why Knowing Both Hops Is Not Enough: Understanding Two-Hop Generalization in Language Models

arXiv cs.CL

Rank 69 · Content 80 · Popularity 42

TL;DR - A mechanistic study of why LLMs that know both individual hops of a two-hop query still fail to chain them, tracing the failure to a layer-wise mismatch and proposing a recurrent-style training fix. It matters because it explains a concrete compositional-reasoning gap and offers a training remedy.

Transformers trained from scratch in a controlled symbolic setting generalize on two-hop queries when the second hop matches the training distribution, but consistently fail when it is out-of-distribution.
Successful generalization coincides with the emergence of consistent intermediate representations for the same entities across contexts.
Failures stem from a cross-layer mismatch: lower layers build correct intermediate representations, while upper layers—though trained on the corresponding atomic facts—mainly learn to map them to outputs rather than reason over them.
A proposed recurrent-style training strategy lets transformers reuse reasoning circuitry across input forms, substantially improving OOD two-hop generalization.

details →

RP-OPSD: Reasoning-Pivot-Guided On-Policy Self-Distillation for Multilingual Reasoning Transfer

arXiv cs.CL

Rank 66 · Content 75 · Popularity 45

TL;DR - An arXiv preprint proposing RP-OPSD, an on-policy self-distillation method that focuses privileged supervision on "reasoning pivot" tokens to transfer LLM reasoning ability from English into other languages. It matters because it targets the specific tokens that drive cross-lingual reasoning transfer rather than treating all tokens uniformly.

Frames target-language reasoning as a mix of surface text generation and "reasoning pivots" — decisions that advance or redirect the reasoning chain — and argues distillation should concentrate on the latter.
Uses the distributional shift between matched teacher views with and without an English reference solution as an operational proxy to identify pivots, guiding privileged distillation and reference anchoring.
Reports gains over strong multilingual reasoning baselines and other OPSD variants on math reasoning benchmarks spanning 17 languages and multiple difficulty levels.
Analysis indicates the method upweights reasoning-control and problem-conditioned state-update tokens while downweighting surface-realization tokens; code is released at github.com/NJUNLP/RP-OPSD.

details →

On-Policy Delta Distillation for Multilingual Math Reasoning

arXiv cs.CL

Rank 63 · Content 60 · Popularity 70

TL;DR - An arXiv study of On-Policy Distillation (OPD) for multilingual math reasoning, showing that a variant using the teacher-vs-base probability gap (OPD²) yields stronger gains in Korean and Japanese than standard OPD. It matters because it offers an RL-alternative post-training recipe that reduces cross-lingual reasoning gaps.

OPD² replaces the plain teacher signal with the probability delta between a post-trained teacher and its base model, and consistently beats vanilla OPD in experiments on Qwen3.
Gains are largest for Korean and Japanese, and the approach generally narrows the English–Korean performance gap.
English-only OPD can still lift Korean/Japanese math performance, but tends to shift model responses toward English.
Implication: multilingual distillation data is needed to preserve target-language outputs, not just to improve accuracy.

details →

Toward Skill-Native LLMs Skill Entropy for Benchmarking and Training Long-Horizon Reasoning paper…

@_akhaliq

Rank 47 · Content 45 · Popularity N/A

TL;DR - A shared paper announcement for "Toward Skill-Native LLMs: Skill Entropy for Benchmarking and Training Long-Horizon Reasoning," which proposes a skill-entropy metric for evaluating and training LLMs on long-horizon reasoning. Content is limited to the title and a Hugging Face papers link, so the details below are inferred from the title.

Frames LLM capability in terms of "skills" as native units, arguing models should be skill-native rather than only token- or task-level optimized.
Introduces "skill entropy" as a measurable quantity, positioned for dual use: benchmarking existing models and serving as a training signal.
Targets long-horizon reasoning — multi-step tasks where single-shot accuracy metrics poorly capture the diversity or distribution of skills a model applies.
No results, datasets, or baselines are available in the provided content; the post is a link-share with no reported numbers.

details →

Multimodal & Generative 3

The Low Frequency Trap: Video Language Models Fail at Simple Event Bookkeeping

arXiv cs.AI

Rank 64 · Content 75 · Popularity 40

TL;DR - An arXiv study introduces "trace-grounded parametric profiling," a controlled video benchmark that audits which events a video-language model reports against executable ground-truth traces, not just the final count. It shows current VLMs collapse at simple event bookkeeping as event count and frequency rise, exposing temporal reasoning failures that aggregate accuracy hides.

Benchmark spans 2,190 synthetic videos across three tasks (bouncing-ball wall contacts, visual blinks, categorical state transitions), varying event count N and frequency F while holding rendering fixed, with per-video executable event traces enabling timestamp-level evaluation.
Failure is staged and representation-dependent: at an 80% reliability threshold, Gemini 3.6 Flash counts persistent state transitions up to 12 events at 0.5–1.0 Hz, but has no reliable positive-count region for transient blinks.
In the high-count/high-frequency regime only 0.2% of final counts are correct and just 18.1% of true events are recovered; higher sampling raises Bounce Ball accuracy 19.6% → 29.3% yet the reported sequence matches ground truth only 3.7% of the time.
Extra frames and alternative prompting strategies inflate scores without faithful event recovery; real-world video evaluations show the same concentration of success at low event counts.

details →

EmoWorld: A Decoupled Affective Field for Controllable Emotional Video Generation

arXiv cs.CV

Rank 58 · Content 65 · Popularity 42

TL;DR - EmoWorld is a training-free framework that decouples emotional atmosphere, semantic cues, and temporal progression in a frozen flow-matching Video DiT, enabling controllable emotional video generation. It matters because current generators collapse all affective factors into one text condition, leaving emotion unsteerable at inference.

Uses a one-time preparation stage that extracts layer-specific affect directions and a reusable cue library from geometry-preserving neutral vs. emotion-edited panoramas.
Three inference-time steering mechanisms: VAS (injects atmosphere directions into hidden states), SAS (separately scalable prompt residual for semantic cues), and TAS (interpolates endpoint residual fields across denoising and video time).
On Wan2.2: VAS gives +19% target-emotion alignment and −48% on a temporal-fluctuation proxy; SAS gives +37% alignment and +36% detected affect-bearing cues; TAS improves transition monotonicity by 15% over the strongest baseline.
Evaluated over 27 emotion categories in T2V and I2V, portable across multiple Video-DiT backbones, and supports camera-conditioned composition without updating generator parameters.

details →

iARCS: Iterative Agentic RL for Controllable 3D Scene Generation

arXiv cs.AI

Rank 58 · Content 65 · Popularity 42

TL;DR - iARCS is an iterative agentic RL framework that adapts a pretrained 3D scene generator to natural-language task requirements, so synthetic scenes satisfy functional constraints (walkability, reachability, clearance) rather than just looking realistic. It matters because constraint-faithful synthetic scenes are more useful as training data for computer vision and embodied AI.

Two-stage training: universal-reward pretraining for physical plausibility and layout quality, then task-specific fine-tuning against the stated natural-language requirements.
Reward functions are LLM-generated reward programs that are iteratively refined using feedback from training, avoiding hand-crafted per-task reward engineering.
Reported gains in constraint fidelity on walkability, reachability, and clearance tasks, with competitive scene diversity retained.
Data produced by iARCS improves the base generator, positioning it as a synthetic data generation pipeline rather than only a controllable scene editing tool.

details →

Efficiency & Systems 3

CoinRAG: Contextualized Information Nugget KV Cache Reuse for Long-Context RAG

arXiv cs.CL

Rank 76 · Content 80 · Popularity 68

TL;DR - CoinRAG is an arXiv preprint that speeds up long-context RAG by reusing precomputed KV caches at the level of fine-grained "information nuggets" instead of whole retrieved chunks, cutting prefill cost while improving answer quality. It matters because chunk-level KV reuse — the current efficiency trick for RAG — still feeds models redundant, noisy context.

Motivation: existing chunk-level KV cache reuse avoids re-encoding long retrieved contexts, but coarse chunks carry substantial redundancy and noise.
Method: a two-stage retrieval identifies query-relevant semantic units ("nuggets") inside retrieved chunks, then compositionally assembles their sliced, offline-computed KV representations together with a chunk-level context into a learned contextual representation.
Framing: the goal is an improved accuracy/latency Pareto frontier under low prefill-latency constraints, rather than accuracy alone.
Results: on LongBench multi-hop QA, CoinRAG reports a new Pareto frontier, reduced operational cost, and an average 5.3% relative F1 gain over baselines under a standard fast-prefill latency budget.

details →

Every Cache Entry Earns Its Place: Global Allocation of Resolution and Coverage for KV Cache Compression

arXiv cs.LG

Rank 69 · Content 80 · Popularity 42

TL;DR - GraceKV reformulates KV cache compression as a global resource-allocation problem under a fixed budget, rather than applying fixed token-eviction or merging rules, so cache capacity can flow across layers, heads, and context slots. It matters because KV cache storage and access are a primary bottleneck for long-context LLM inference.

Each layer–KV-head–slot combination is treated as an atomic unit represented by a prototype tree: leaves are token-level KV entries, and internal nodes compress their children into a single prototype; a set of non-overlapping nodes forms the unit's representation.
Two competing actions trade off globally for the shared budget — adding a new tree root expands information coverage, while splitting a node improves local resolution — letting the method jointly balance coverage and resolution instead of fixing one policy.
Training-free, with compression and inference running entirely on GPU.
Reported to rank first in 24 of 32 evaluated settings across long-context tasks and compression ratios, remaining robust up to 128× compression.

details →

LLM Inference Under Bursty Workload Distribution: Modifying the WAIT Algorithm

arXiv cs.LG

Rank 58 · Content 65 · Popularity 42

TL;DR - An arXiv preprint extending the WAIT LLM-serving scheduler to handle bursty, time-varying request traffic instead of the usual constant-rate Poisson assumption. It matters because real inference workloads are dynamic, and scheduling tuned for a fixed arrival rate leaves throughput on the table.

Argues prior LLM scheduling work assumes stationary Poisson arrivals, which misrepresents real-world bursty traffic patterns.
Adds a lightweight extension to the state-of-the-art WAIT algorithm that performs online estimation of request intensity from observed interarrival times, requiring no prior traffic knowledge.
Evaluated in simulation using Markov Modulated Poisson Process (MMPP) synthetic workloads with diverse request types.
Reports higher throughput than Sarathi-Serve, ORCA, and vLLM in the evaluated low arrival-rate shift scenarios, with comparable latency; results are simulation-only, not measured on a real serving stack.

details →

AI Safety & Alignment 1

Diffusion LLMs as Targets and Adversaries: Mechanistic Safety Exploits

arXiv cs.LG

Rank 69 · Content 80 · Popularity 42

TL;DR - An arXiv preprint showing that safety alignment in diffusion LLMs (DLLMs) is sparse, localized in identifiable "safety neurons," and inherited from the autoregressive models they were initialized from — making them both easy to jailbreak and effective as attack engines against other models.

Safety neuron pruning breaks alignment: self-pruning raises attack success rate from 2.6%→73.8% on LLaDA and 1.9%→86.6% on Dream; transfer pruning using neurons mapped from Qwen2.5 reaches 73.2% on Dream and 86.3% on Fast-dLLM, evidencing a shared mechanistic safety footprint across architectures.
The proposed SN-Guided Diffusion is a fully offline, black-box jailbreak that steers iterative denoising away from safety-triggering regions via a weighted safety-neuron loss, with reported AUROC = 1.0 separating benign from jailbreak prompts.
Attacks transfer to third-party targets: up to 77.1% ASR on Llama-3-8B-Instruct, 86.9% on Qwen2.5-7B-Instruct, and 74.3% on Gemini-2.5-Flash-Lite, using only 20 generation episodes per prompt — orders of magnitude cheaper than prior jailbreak frameworks.
Practical implication: initializing DLLMs from aligned autoregressive checkpoints imports their vulnerabilities, so alignment likely needs to be redone (not inherited) for diffusion-based decoding; code is released publicly.

details →

AI Weather Forecasting 1

Timestep-Conditioned Transformers for Global Weather Forecasting

arXiv cs.LG

Rank 58 · Content 65 · Popularity 42

TL;DR - GEM-3 is a ~134M-parameter probabilistic global weather model whose autoregressive timestep can be chosen at inference time from a single set of weights, removing the usual fixed-timestep trade-off between sub-daily detail and error accumulation. It matters because one model can serve both short-range and extended-range forecasting without retraining specialists.

Fixed timesteps force a trade-off: short steps (1–6h) resolve diurnal dynamics but accumulate more error over a horizon, while 24h steps reduce accumulation but lose sub-daily usability.
Explicit multi-timestep inference lets users configure the step at run time; mixed-timestep training also consistently improved rollout stability versus timestep-specialist models.
Architecture is a lightweight neighborhood-attention transformer on an equirectangular grid, extending the earlier GEM-2 with further architectural changes.
Claimed outcome is near-SOTA medium-range probabilistic skill plus stable extended-range rollouts, efficient training/inference, and decision-relevant diagnostics; no specific metrics are given in the provided abstract.

details →

AI for Chemistry 1

Commun. Chem.｜从"副反应"到"主反应"：计算化学网络编辑策略解锁隐藏合成路径

WeChat: DrugAI

Rank 59 · Content 70 · Popularity 35

TL;DR - Hokkaido University's Maeda and Mita groups (Communications Chemistry, July 2026) introduce "network editing," a computational strategy that deletes dominant pathways from automatically generated reaction networks to expose hidden, synthetically useful side channels — then experimentally validated it with a CO₂ radical anion (CO₂•⁻)-mediated arylcarboxylation building 5- and 6-membered N-heterocycles.

Method: SC-AFIR builds the full reaction-path network; RCMC kinetic simulation identifies dominant product-forming routes; those routes are artificially removed (simulating a protected/deactivated functional group); re-running kinetics reveals which previously low-probability paths gain yield. Selection is kinetics-based rather than chemist-intuition-based.
Compute: A neural network potential combined with xTB, NNP(+xTB), uses Δ-learning on the DFT–xTB energy gap to approach DFT accuracy at near-xTB speed. The 1a + CO₂•⁻ search yielded 20,831 equilibrium structures and 32,362 path-top structures in ~9 days on 128 CPU cores.
Experimental payoff: For N,N-diallylaniline, blocking one allyl group with electron-withdrawing protecting groups (Ac, Boc, Bz) lowered cyclization barriers; Ac performed best, giving indoline scaffolds in moderate-to-good yields across substituted arenes, while Bz unexpectedly produced a six-membered lactam via a non-classical pathway.
Limits stated by the authors: network editing can only re-rank paths already present in the computed network — it cannot predict genuinely new reaction modes requiring different reagents, catalysts, or conditions, and substituent design still relies on chemist judgment.

details →

AI for Math 1

GPT-5.6和Fable联手，解决了一道悬了25年的数学难题

量子位

Rank 64 · Content 70 · Popularity N/A

TL;DR - Microsoft Research principal researcher Dimitris Papailiopoulos used GPT-5.6 and Fable 5 to produce a proof that a simple polynomial-time algorithm achieves exact MIMO detection right at the maximum-likelihood threshold (SNR = 2logN), closing a gap open since ~2001. It matters as a concrete case of frontier LLMs contributing the core proof strategy to a long-standing open problem in information theory.

The problem: recovering N transmitted bits from a noisy N×N channel. ML detection is exhaustive (2^N) and NP-hard in the worst case (Verdú, 1989); prior methods (sphere decoding — shown exponential by Jaldén & Ottersten 2005, SDP relaxation, bit-flipping, AMP, statistical physics) never matched the 2logN threshold, with box relaxation stuck at 4logN.
The algorithm is two steps and O(N³): sign-rounded LMMSE estimation (proved to land within o(N) Hamming distance of the true bits), then greedy single-bit flipping, shown to converge in O(NlogN) steps.
Proof structure: within a neighborhood of the start, every incorrect point has a flip that strictly decreases the cost by a non-vanishing amount, and the cost grows with Hamming distance — forming a barrier so greedy search can only halt at the true bit string. A converse is also proved: slightly below 2logN, even ML detection fails.
Process notes: GPT-5.6 proposed an AMP-based route, Fable 5 proposed "sign LMMSE + greedy bit flipping"; the author picked Fable's path, had GPT patch gaps, then iterated for ~7 days having the models mutually simplify until he could hand-check line by line. He declined Lean formalization because he doesn't know Lean. Claims are from the author's X post, not (per the content) peer review.

details →

Agent Evaluation Methods 1

AV-AIVAT: 74x Cheaper Agent Evaluation with Certified Anytime-Valid Stopping in Imperfect-Information Games

arXiv cs.GT

Rank 70 · Content 70 · Popularity 70

TL;DR - An arXiv preprint combining AIVAT variance reduction with anytime-valid confidence sequences so head-to-head agent evaluations in imperfect-information games can stop as soon as evidence suffices, cutting required hands by a median 74x. It matters because benchmarking LLM agents is expensive, and naive early stopping silently breaks stated confidence levels.

AIVAT's conditional mean-zero corrections cut variance by a median 54x across 15 LLM agent configurations over 71,439 paired Heads-Up No-Limit Hold'em hands, but provides no stopping rule on its own.
AV-AIVAT pairs those corrections with continuously monitored confidence sequences; the online value model trains only on past games so no hand scores its own correction, preserving validity under optional stopping.
At 95% nominal level and ±1 Big Blind precision, raw outcomes need a median 74x more hands than AIVAT-corrected ones under the Asymptotic CS (AsympCS).
Exact finite-sample certification uses the Empirical-Bernstein CS, requiring an independently justified payoff bound (established structurally for Leduc hold'em); a width floor from the bet cap and bound limits gains, with a median 1.37x stopping-time ratio in descriptive HUNL runs.

details →

Embodied Vision & Touch 1

HOPE: Hand-Object Pressure Estimation from Monocular Videos

arXiv cs.CV

Rank 58 · Content 65 · Popularity 42

TL;DR - HOPE is a monocular-video method that predicts per-vertex normal pressure and contact directly on a hand mesh, reframing vision-based pressure estimation as hand-centric video prediction instead of planar, single-image inference. It matters because contact-rich hand-object interaction understanding (robotics, AR/VR, manipulation learning) needs pressure signals without instrumented surfaces or gloves at test time.

Output space is the hand mesh itself, making predictions independent of object shape and sensor layout, and enabling dynamic multi-object scenarios that planar-surface baselines cannot handle.
Heterogeneous supervision is unified by lifting tactile-glove pressure, planar-sensor pressure, and distance-based contact annotations into a shared hand vertex space, so bare-hand contact data regularizes learning where metric pressure labels are missing.
Architecture is a vertex-anchored video transformer treating each vertex as a persistent token that aggregates visual features and hand pose over time, with a contact-gated pressure head enforcing zero pressure without contact.
Evaluated on OpenTouch, PressureVisionDB, and hand-object contact benchmarks across object-pressure, surface-pressure, and contact-supervised HOI settings; the authors report generalization from gloved-hand training video to bare-hand egocentric and in-the-wild video.

details →

Infrared Target Detection 1

ECCV 2026｜国防科大提出SPIRE：红外小目标检测新突破

WeChat: CVer

Rank 42 · Content 40 · Popularity 48

TL;DR — NUDT's Prof. Wei An group has an ECCV 2026 paper, SPIRE, that reframes infrared small target detection (IRSTD) from pixel-level segmentation to single-point-supervised probabilistic centroid regression, cutting annotation, compute, and post-processing cost while staying competitive on target-level accuracy.

Pipeline: centroid point labels → Point-Response Prior Supervision (PRPS) expands each point into a smooth local probability response modulated by infrared radiation peak and local contrast → High-Resolution Probabilistic Encoder (HRPE), a single-branch encoder-only net with no decoder/skip connections → local peak extraction + sub-pixel refinement gives coordinates directly (no mask, no connected-component clustering).
Efficiency: 0.29M params, 7.68G FLOPs at 640×640, 261.2 FPS — vs SCTransNet, ~87.9% fewer FLOPs and ~97.4% fewer params, and reported 3.2–5.6× faster than three high-accuracy baselines.
Accuracy under a unified centroid protocol (hit if predicted center within 5 px): SIRST-UAVB Precision 99.82%, Fa 1.02×10⁻⁸ (both best), F1 97.05% (0.04 pt off top), Recall 94.44% (not top); SIRST4 best on all four metrics (P 95.00%, R 94.21%, F1 94.60%, Fa 28.53×10⁻⁸), and highest F1 under stricter δ=3.
Ablations: PRPS beats single-point impulse supervision by +2.78 F1 and unconstrained Gaussian by +0.77 F1; enlarging output stride drops F1 by >10 pts while further upsampling costs ~3.5× FLOPs with no F1 gain, and removing channel shuffle/reweighting raises false alarms. Code: NIRIXIANG/SPIRE-IRSTD; paper listed as arXiv:2604.05363.

details →

Interpretability & World Models 1

Transformers Struggle to Use Their Emergent World Models: Revisiting the Tower of Hanoi, and the Illusion of Thinking

arXiv cs.AI

Rank 72 · Content 85 · Popularity 42

TL;DR - An interpretability study of the Tower of Hanoi shows that both small in-house Transformers and frontier reasoning models build a linearly decodable, geometrically faithful world model of the puzzle (a Sierpinski triangle), but large models lose that representation during extended chain-of-thought — reframing the reported "illusion of thinking" collapse as a maintenance failure rather than an absence of world models.

Small Transformers trained from scratch on precomputed solution traces develop an emergent, causally involved representation of the puzzle state space matching the Sierpinski triangle structure.
Qwen3.6-27B and DeepSeek-R1-Distill-Qwen-32B encode this same world model near-perfectly at the end of the prompt, yet fail on the majority of flat-to-flat variants with more than 3 rings.
Probing across planning stages localizes the failure to decaying world-model representations during generation; injecting the prompt-time representation at inference improves performance, establishing causality and partial recoverability.
The flat-to-flat variant (arbitrary initial and goal configurations, not all rings on one peg) is the discriminating task where current LRMs still break down.

details →

Learning Theory 1

An Optimal Agnostic PAC Algorithm

arXiv cs.LG

Rank 58 · Content 65 · Popularity 42

TL;DR - A theoretical paper constructing a learner for agnostic PAC learning over hypothesis classes of finite VC dimension d that attains the statistically optimal excess-risk bound, closing the long-standing gap between upper and lower bounds up to universal constants.

For a class $H\subseteq{-1,+1}^X$ with VC dimension $d\ge1$, the learner achieves, with probability $\ge 1-\delta$, $L(\widehat h)\le L^+7\cdot10^8\big(\sqrt{L^(d+\log(1/\delta))/n}+(d+\log(1/\delta))/n\big)$ from an i.i.d. sample of size $n$.
The bound is "first-order"/optimistic: it interpolates between the fast $O((d+\log(1/\delta))/n)$ rate in the realizable case ($L^=0$) and the slow $\sqrt{\cdot/n}$ rate as $L^$ grows.
The result settles agnostic PAC sample complexity up to universal constants at every fixed $L^*$, matching the lower bounds of Devroye, Györfi, and Lugosi (1996).
The constant $7\cdot10^8$ is explicitly large, so the contribution is asymptotic/constant-factor optimality rather than practical tightness; no empirical results are claimed in the provided abstract.

details →

Structural Cell Biology 1

Cell Reports | 朱金伟/符传孩/林霖团队合作揭示线粒体动力学调控新机制：钙调蛋白变构激活INF2驱动线粒体分裂

WeChat: BioArt

Rank 61 · Content 65 · Popularity N/A

TL;DR - A Cell Reports paper from Shanghai Jiao Tong (Zhu Jinwei/Lin Lin), Soochow (Zhang Meng), and USTC (Fu Chuanhai) teams shows that Ca²⁺-bound calmodulin allosterically activates the formin INF2 to drive actin-dependent mitochondrial fission. It defines a CaM–INF2 signaling axis linking local calcium signals to organelle dynamics and to INF2-linked neuropathy (CMT) and kidney disease (FSGS). Note: this is a molecular/structural biology study, not an AI advance.

Ca²⁺-CaM binds the INF2 DID domain with nanomolar affinity; a crystal structure of the DID–Ca²⁺-CaM complex reveals an atypical dual-interface mode engaging both the αN helix and the armadillo repeat (ARR) domain, unlike canonical CaM targets.
Activation is allosteric rather than purely competitive: CaM induces DID rearrangement that disrupts DID–DAD autoinhibition, freeing the FH1-FH2 catalytic core for F-actin assembly — filling a gap since INF2 lacks classic Rho GTPase regulation.
In cells, boosting CaM signaling enhanced ER-localized INF2-driven actin remodeling, shortening mitochondria and increasing fission; CaM-binding-deficient INF2 mutants were unresponsive to calcium stimulation.
The CMT-associated mutation INF2 R91G strengthens Ca²⁺-CaM binding, causing excessive actin assembly and mitochondrial over-fission, suggesting a gain-of-function disease mechanism.

details →

Top highlights — Industry & News

Data-layer recursive self-improvement goes open-source: 当题库追不上模型，AI开始给自己出题：中国这支团队跑通了数据层RSI covers the release of BigBang-V1, a 35B MoE base model (~3B active, 262K context) trained with 100% AI-synthesized post-training data — an attack on the training-data bottleneck rather than compute scaling.
Small model, outsized benchmark claim: 35B模型科研能力超过 1T 级 DeepSeek V4 Pro：BigBang 跑通RSI数据闭环 reports the same SJTU/DP Technology/Shanghai Algorithm Innovation model outperforming the far larger DeepSeek V4 Pro Preview on hard scientific research benchmarks — the strongest evidence yet that an RSI data loop can substitute for parameter count.
Vendor-led serving optimization ahead of the next Qwen: RT by @NVIDIAAI: As we prepare for Qwen 3.8 drop, @NVIDIAAI team led the optimizations in vLLM for… reports 25K total tokens/s/GPU on GB200, a concrete datapoint on turning open-weight models into production throughput.
Orchestration decoupled from any single vendor: RT by @hardmaru: ベースモデルに依存しないオーケストレーションに向けてブログ：https://sakana.ai/fugu-gemma4/ Sakana… announces Sakana AI's Fugu conductor retrained on the open Gemma 4 base, moving routing intelligence toward swappable, sovereign agent stacks.

LLM Agents 4

35B模型科研能力超过 1T 级 DeepSeek V4 Pro：BigBang 跑通RSI数据闭环

WeChat: 深度学习自然语言处理

Rank 64 · Content 70 · Popularity N/A

TL;DR - Shanghai Jiao Tong AI Institute, DP Technology, and Shanghai Algorithm Innovation Institute released BigBang-V1, a 35B research-oriented model post-trained on 100% AI-synthesized data, which reportedly beats the much larger DeepSeek V4 Pro Preview on several hard scientific/research benchmarks. It matters as a claimed working demonstration of a recursive self-improvement (RSI) loop at the data layer rather than the model-scaling layer.

Reported scores: FrontierScience Research 46.2 vs 40.7, HLE 50.3 vs 48.2, BioMysteryBench Human-Difficult 15.7 vs 13.7, PaperBench 53.6 vs 50.4 (BigBang-V1 vs DeepSeek V4 Pro Preview); plus BrowseComp 76.5 and SWE-Bench Pro 54.2, suggesting transfer beyond science.
The pipeline pairs a Generator Agent (a code agent that edits/runs/debugs the data-synthesis program itself — choosing domains, reasoning-chain length, tools, and verification methods) with an adversarial Critic Agent doing two-tier format and value/difficulty screening.
An outer loop trains models on data from competing pipeline versions and evaluates on held-out real research tasks, recalibrating the Critic when its scores don't translate into measured capability gains — an explicit guard against reward hacking and Critic-pleasing.
Case studies: 3/3 stable identification of Norovirus GII.4 from noisy FASTQ data (comparison model 0/3), and a 26-circle packing solution of 2.635983 in ~3.02s vs 2.630957 in ~351.07s, attributed to exploiting hexagonal structural priors with SLSQP restarts.
Caveat: results are self-reported by the team via a WeChat announcement; humans still define objectives, budgets, held-out task sets, and acceptance criteria.

details →

RT by @hardmaru: ベースモデルに依存しないオーケストレーションに向けてブログ：https://sakana.ai/fugu-gemma4/ Sakana…

@hardmaru

Rank 61 · Content 65 · Popularity N/A

TL;DR - Sakana AI announced that Fugu, its multi-agent orchestration product exposed as a single foundation-model endpoint, now has a "conductor" model retrained on the open Gemma 4 base, aiming for base-model-independent orchestration. It matters because it decouples routing intelligence from any single vendor's model, a step toward sovereign, swappable agent stacks.

Fugu is a two-layer system: a swappable "model pool" that performs the actual work, and a "conductor model" that decides task allocation across those models.
Clients send one request to a single endpoint; Fugu decides how to process it and dispatches to appropriate models, reportedly exceeding single-model performance and cost-performance.
The model pool was designed to be interchangeable from the start; this release extends interchangeability to the conductor itself, now trained on Gemma 4.
On Sakana's internal evaluation set, the Gemma 4-based conductor matched prior performance while reducing cost; next steps target conductors built on Sakana's own in-house models for sovereignty requirements. No public benchmark numbers were given.

details →

RT by @_akhaliq: NVIDIA just released the NeMo Gym conversational tool-use assets on Hugging Face A…

@_akhaliq

Rank 50 · Content 50 · Popularity N/A

TL;DR - NVIDIA published a set of conversational tool-use assets for its NeMo Gym pipeline on Hugging Face, giving developers reference data for building and evaluating agentic tool-calling models. It matters because open reference datasets for multi-turn tool use are a bottleneck for training and benchmarking LLM agents.

Released as a bundle tied to NeMo Gym's conversational tool-use pipeline, distributed via Hugging Face.
Contents described as "golden" policy/tool reference pairs — i.e., known-good mappings between conversational policy behavior and tool invocations.
Also includes prompt histories, implying multi-turn dialogue context needed for realistic agent training/eval rather than single-shot function-call samples.
Content is thin (a short announcement post): no benchmark numbers, dataset size, license, or model results were provided, so scope and quality can't be assessed from this item alone.

details →

Opus 5狂烧6.9亿token做游戏，GPT-5.6用5美元复刻了

量子位

Rank 43 · Content 40 · Popularity N/A

TL;DR - A developer used Claude Opus 5 with a single ~2000-word prompt (690M tokens, $423) to one-shot a playable Three.js speedboat racing game, "INK TIDE"; another dev then reproduced a rougher version in Codex with GPT-5.6 Sol/Luna for roughly $5. It's a concrete datapoint on multi-agent coding workflows and the cost/quality tradeoff between frontier models.

The "single prompt" was effectively a requirements doc: hard constraints → visual standards → functional requirements → dev process → delivery criteria, with a dedicated section for the water system; built on Vite + TypeScript + Three.js with all models, textures, and audio generated procedurally (no external assets).
The prompt explicitly orchestrated sub-agents by domain (water, toon rendering, boat physics, AI opponents, character animation, audio/visual, performance), fixing shared architecture and per-agent file ownership first to fight "attention dilution" in long tasks.
A screenshot harness plus a visual-review sub-agent acted as QA: each subsystem was run, screenshotted, critiqued, and iterated until the reviewer had no complaints.
The GPT-5.6 replica (Sol Ultra as orchestrator, Luna Max as sub-agent, ~5 hours, 2 prompts) matched surface elements but was visibly coarser in water detail and visual coherence; the $5 figure is subscription-quota accounting (~20% of a weekly limit), not an API bill. Opponent AI in the original is hardcoded and deliberately weak (fastest NPC only 0.5% above optimal pace; another scripted to err 2–3 times per lap).

details →

Bioinformatics AI 3

谷歌掀起人事巨震！Demis Hassabis放权，all in AI制药

WeChat: 智药局

Rank 57 · Content 60 · Popularity N/A

TL;DR - Demis Hassabis has stepped down as Google DeepMind CEO to become board director and chief scientist, shifting his focus to AI-for-Science and a more active CEO role at AI drug-discovery firm Isomorphic Labs. It signals DeepMind reorienting from single-problem scientific research toward Gemini-driven commercialization and agents, while its founding figures double down on AI-driven drug discovery.

Broader DeepMind churn: AlphaFold lead John Jumper departed for Anthropic after ~9 years, the AlphaFold team was reportedly disbanded, and the AI4S group is pivoting from protein folding to a full Gemini-powered research system; Jeff Dean is leaving after 27 years to found DiscoveryLoop, an AI-for-science/engineering automation startup that may include drug R&D.
Isomorphic Labs (spun out of DeepMind in late 2021) aims to build a general-purpose drug design engine reusable across targets and disease areas, rather than per-target pipelines.
Technical lineage: AlphaFold3 (May 2024, with DeepMind) predicts structures and interactions of all life molecules; IsoDDE (Feb) is claimed to beat physics-based methods on small-molecule binding affinity and more than double AlphaFold3's protein–ligand structure prediction accuracy.
Commercial footing: $600M raise (Apr 2025, Thrive Capital leading, GV and Alphabet participating) and >$3B in research deals with Novartis and Eli Lilly; clinical entry is described as imminent but has slipped past the previously targeted end-2025, with no disclosed timing, target, or indication.

details →

双虚拟敲除 CellOracle + scTenifoldKnk 全流程复现（二）

WeChat: 生信技能树

Rank 47 · Content 45 · Popularity N/A

TL;DR - Part two of a WeChat tutorial series reproducing a 2026 Cell Prolif paper that used dual virtual knockout (CellOracle + scTenifoldKnk) to identify STAT3 as a dentinogenesis regulator; this installment covers the single-cell QC and cell-type annotation steps on GSE146123 in R/Seurat.

QC thresholds were chosen after an AI-assisted literature scan of dental pulp/odontogenic scRNA-seq papers (MT% cutoffs ranging 5%–40% across Pang 2023, Yang 2024, Zhang 2026, Liu 2026); the author settled on nFeature_RNA 200–6000, nCount_RNA 3–30000, percent_mito <10%, percent_hb <1%.
Standard Seurat pipeline: LogNormalize (scale factor 1e4), 2000 variable features, ScaleData, PCA, then Harmony integration on orig.ident to correct sample batch effects, with UMAP/neighbors on dims 1:20.
Clustering resolution was swept over 0.05–1 and inspected with clustree; resolution 0.5 was used for the working annotation.
Annotation targets the original paper's 9 major populations (endothelial 21.2%, mesenchymal 20.7%, perivascular 16.9%, glial, peri-odontoblastic layer, pulp, preodontoblasts, epithelial, immune), with marker-based distinctions noted (mesenchymal: COL1A1/DCN/LUM vs. pulp: VIM/BGN/POSTN).
Note: content is truncated mid-annotation, and the actual CellOracle/scTenifoldKnk perturbation steps are deferred to later installments.

details →

免费注册中 | 2026三维基因组与相分离高级研习班，从染色质空间结构到生物分子凝聚体

WeChat: BioArt

Rank 40 · Content 35 · Popularity N/A

TL;DR - An event announcement for the "2026 3D Genome and Phase Separation Advanced Workshop," a free one-day training session hosted by Guangzhou Epibiotek and Danaher Life Sciences at Zhejiang University on Aug 29, 2026. It is a commercial/academic outreach event rather than new research or AI advancement.

Scientific scope: chromatin compartments, TADs, chromatin loops, and enhancer–promoter interaction networks; biomolecular phase separation as a lens on transcriptional regulation, nuclear condensates, and disease mechanisms; plus RNA regulation and single-cell multi-omics.
Logistics: 120 seats, 8 invited speakers, free registration (lunch/tea included) with a refundable 300 RMB deposit; restricted to researchers at universities, hospitals, and research institutes. Co-organized by Zhejiang University's student BioLinkX bioinformatics association.
Sponsor context: Epibiotek provides epigenetics sequencing services (super-enhancers, 3D chromatin, epitranscriptome, spatial interactomes, RBPs, single-cell epigenomes, protein phase separation) and runs the "EpiCloud" analysis platform, with planned tumor exosome RNA marker and medical epigenomics databases.
Note: content is promotional — no speaker names, agenda details, or technical results are given, and there is no explicit AI/ML component beyond bioinformatics data analysis.

details →

LLMs & Foundation Models 3

当题库追不上模型，AI开始给自己出题：中国这支团队跑通了数据层RSI

量子位

Rank 68 · Content 75 · Popularity N/A

TL;DR - Chinese team "无尽前沿" (SJTU AI School, DP Technology, Shanghai Institute for Algorithm Innovation) released BigBang-V1, an open-sourced 35B MoE base model (~3B active, 262K context) claimed to be the first trained natively via recursive self-improving, with 100% AI-synthesized post-training data. It matters because it targets the training-data bottleneck rather than compute scaling.

Data-layer RSI loop: a Generator Agent (a code agent that edits/runs/debugs the data-synthesis program itself, picking domains, reasoning-chain length, and tools) pairs with an adversarial Critic Agent that checks format/execution/completeness, then correctness, verifiability, difficulty, diversity, and training value.
Two-level co-evolution: the fast inner Generator–Critic loop is guarded by an outer loop that trains models on competing pipeline variants and evaluates on held-out real research tasks, back-calibrating the Critic when high-scored data fails to improve models — the stated defense against reward hacking and synthetic-data collapse.
Task selection is grounded in "frontier + verifiable" science (formal methods, code execution, numerics, simulators, domain tools), so the task pool doesn't saturate like a static benchmark bank.
Reported results (per the team's tech report, not independently verified): 10 first-place finishes among 35B models; claimed to beat 1T-scale DeepSeek V4 Pro Preview on FrontierScience Research and PaperBench; BrowseComp 76.5, SWE-Bench Pro 54.2; case studies include a 47bp transposon insertion localization and a paper-reproduction run scoring 0.7657 (331/485 points). Weights on Hugging Face, code on GitHub.

details →

R to @OpenAI: Plus and Pro users can access the updated version of GPT‑5.6 Sol in ChatGPT along…

@OpenAI

Rank 54 · Content 55 · Popularity N/A

TL;DR - OpenAI is rolling out an updated GPT‑5.6 Sol to ChatGPT Plus and Pro users, alongside a new response-control slider, while expanding GPT‑5.6 Luna access for free users. It matters as a product-tier segmentation signal: the same model name now ships in different tuned variants per surface.

The updated GPT‑5.6 Sol is scoped to the Chat experience only; the Sol variant powering Work and Codex is unchanged by this release.
A new "slider" ships with it, implying user-adjustable behavior (per the linked post, aimed at everyday chat quality) rather than a separate model selection.
OpenAI's linked announcement claims improved accuracy and consistency for Sol, plus expanded/unlimited everyday chat access to GPT‑5.6 Luna for free users.
Content is a short announcement post with no benchmarks or eval details provided, so the accuracy/consistency claims are unquantified here.

details →

Pinned: We’re making better intelligence easier to access in ChatGPT for everyone: - GPT-5.6 Sol…

@OpenAI

Rank 54 · Content 55 · Popularity N/A

TL;DR - OpenAI announced a ChatGPT tier-wide model rollout: GPT-5.6 Sol becomes the backbone for both Instant and deep-reasoning modes on Plus/Pro, while Free and Go users get unlimited text chats on GPT-5.6 Luna. It matters as a signal of how frontier labs are consolidating model families and pushing reasoning-grade capability down the pricing ladder.

GPT-5.6 Sol now serves both fast "Instant" responses and extended reasoning for paid Plus/Pro tiers, unifying what were previously separate model routes.
OpenAI claims Sol delivers "more factual, focused" outputs — a positioning around hallucination reduction and response conciseness rather than raw capability jumps.
Free and Go tiers move to unlimited text chat on GPT-5.6 Luna (a lighter/cheaper sibling), effective the day after the post — implying meaningful inference-cost reductions.
Content is a short product announcement thread with no benchmarks, model cards, or architecture details; capability claims are unverified from this source alone.

details →

Efficiency & Systems 2

RT by @NVIDIAAI: As we prepare for Qwen 3.8 drop, @NVIDIAAI team led the optimizations in vLLM for…

@NVIDIAAI

Rank 64 · Content 70 · Popularity N/A

TL;DR - NVIDIA's AI team announced it led inference optimizations for Qwen 3.5 in vLLM, reporting 25K total tokens/s/GPU on a GB200 system, with a linked vLLM blog deep dive. It matters as a concrete datapoint on how vendor-led serving optimizations translate open-weight models into production-grade throughput ahead of the next Qwen release.

Claimed headline result: ~25K total tokens/s/GPU (combined prefill+decode, per the "total" framing) for Qwen 3.5 served on NVIDIA GB200 (Grace-Blackwell) hardware.
Optimization work was contributed upstream into vLLM rather than a proprietary stack, so the gains land in a widely used open-source serving engine.
Framed as groundwork for an upcoming "Qwen 3.8" release — i.e., the serving-path optimizations are expected to carry forward to the next model drop.
Content is a short promotional post; specific techniques (kernels, quantization, attention/MoE handling, batching, disaggregation) are not stated here and would require the referenced vLLM blog post (vllm.ai/blog/2026-08-06-qwen…) to verify.

details →

Anthropic回应造芯！327万年薪招揽芯片工程师

WeChat: 智东西

Rank 57 · Content 60 · Popularity N/A

TL;DR - Anthropic publicly confirmed for the first time that it is building an in-house silicon team to design custom AI chips for Claude, co-designing model and hardware for faster, more efficient large-scale serving. It signals frontier labs pushing vertical integration down to the hardware layer, following OpenAI's own chip effort.

Anthropic says it will keep a "multi-chip" strategy — AWS, Google, NVIDIA and AMD hardware remain core to its compute expansion; no architecture, use case, fab partner, or timeline disclosed.
Job postings in San Francisco, New York and Seattle span the full flow: front-end design, pre-silicon verification, physical design, DFT, analog/mixed-signal, process/foundry, design automation, and packaging/SI/PI, at $320K–$485K/year.
Hires will write chip specs and interface definitions, make build/buy/license/partner IP decisions, and co-design with Anthropic's inference, performance, kernel and infrastructure teams through bring-up and system integration; Anthropic also wants Claude used for RTL review, verification triage and spec checking.
Context per prior reporting: Reuters (April) flagged exploration of custom silicon amid tight advanced-chip supply, The Information (July) reported Samsung talks as a possible manufacturing partner; OpenAI announced its first in-house chip in June.

details →

3D CAD Generation 1

RT by @huggingface: every 3D mesh is kind of a compiled binary: printable, usable, but frozen. you…

@huggingface

Rank 54 · Content 55 · Popularity N/A

TL;DR - Hugging Face is highlighting Cadena, a tool that "decompiles" static 3D meshes into editable, step-by-step CAD programs, with a demo hosted on HF Spaces. It matters because mesh outputs from scanning or generative 3D are effectively frozen artifacts, and recovering parametric CAD history restores editability.

Frames the core problem as an analogy: a mesh is like a compiled binary — usable and printable, but not modifiable at the level of design intent (e.g. "make this hole 2mm wider").
Cadena's stated output is not just a cleaned mesh but a CAD program reconstructed step by step, implying reverse-engineering of parametric modeling operations rather than surface fitting alone.
Distribution is via a Hugging Face Space (huggingface.co/spaces/kulibi…), i.e. an interactive browser demo, with a video shown in the post.
Content is thin — a promotional post with no benchmarks, model architecture, dataset, or accuracy claims; technical details would require the linked Space or paper.

details →

AI Chips & Compute 1

摩尔线程：2026上半年营收17.36亿元，已超2025全年

雷峰网 (AI科技评论)

Rank 50 · Content 50 · Popularity N/A

TL;DR - Chinese GPU maker Moore Threads (摩尔线程, 688795.SH) reported H1 2026 revenue of ¥1.736B, up 147% YoY and already exceeding its full-year 2025 total, driven by scaled shipments of its MTT S5000 AI training/inference cards and KUAE compute clusters. It signals that a domestic full-function GPU vendor is reaching credible large-model training scale outside the CUDA/NVIDIA supply chain.

Financials: gross profit ¥989M (+103.8% YoY), net loss narrowed 95.7% (attributable) / 52.4% (ex-items); H1 R&D spend ¥769M (+38.2%), ~¥5.9B cumulative since 2022; 2,167 patent applications filed, 788 granted.
MTT S5000 clusters deployed in Beijing, Wuxi, Hangzhou with claimed ~95% linear scaling efficiency, training-loss curves matching mainstream international cards, checkpoint-resume support, and >90% effective training time.
Demonstrated workloads: a from-scratch MoE-236B model trained on >25T tokens, full-stack native training of the "EvoPhys-World" 5D world model (reported 37 straight days atop Stanford WorldScore's world-generation axis), and open-sourced MusaCoder code LLM.
Ecosystem: MUSA stack claims 100% core math-library CUDA compatibility, 3000+ PyTorch operators, 55 core AI operator classes, official vLLM/SGLang support, and 800k+ developers; product line spans cloud (S5000), edge (E300 module), and endpoint (AIBOOK, AICUBE).

details →

AI Cost Governance 1

180万刀，连亚马逊都烧不起Claude了

量子位

Rank 47 · Content 45 · Popularity N/A

TL;DR - 量子位报道亚马逊一项"为网站补全作者信息"的 Claude Sonnet 任务失控烧掉 180 万美元、超预算 860%、5 个月后才被发现且最终未部署上线，折射出企业级 Agent 用量与成本失控已成普遍问题。

事故根因是 Agent 失败后不知疲倦地反复重试，且缺乏用量告警与止损机制；按 Claude Sonnet 公开价（输入 $3/百万 token、输出 $15/百万 token）折算，180 万美元最多约对应 6000 亿 token。
同类失控并非孤例：Meta 内部排行榜 Claudeonomics 聚合 8.5 万员工数据，30 天消耗 73.7 万亿 token（按公开定价约合 2.21 亿美元/月），6 月起发备忘录限额并建 "AI Gateway" 中央平台监控预算；Uber 前四个月烧光全年 AI 编程预算后设定每人每工具每月 1500 美元上限。
排行榜式激励触发古德哈特定律——亚马逊 "KiroRank" 被关停后改用 "normalized deployments" 衡量产出；Uber COO 承认 token 支出与可衡量产出的关联尚未成型，调查显示仅 26% 企业对自身 AI 成本有全面可见性。
对比背景：亚马逊 2026 年资本开支预计约 2200 亿美元（同比增近六成），Q2 AWS 净销售额 422 亿美元、同比增 37%，贡献总营业利润约 60%；文章以 2012 年骑士资本 45 分钟亏损 4.4 亿美元的自动化事故作类比，强调自动化放大错误的速度与规模。

details →

AI Research Reproducibility 1

大厂不再迷信顶会：Auto Research时代，论文含金量正在缩水

WeChat: PaperWeekly

Rank 61 · Content 65 · Popularity N/A

TL;DR - SAI ran an AI-agent-based reproduction study over all 168 ICML 2026 Oral papers, fully reproducing 105, and found most accepted top-conference results do not hold up under actual re-execution. It matters because it quantifies the gap between peer-review acceptance and verifiable science in the emerging "Auto Research" era.

Of 92 papers with ≥5 verifiable claims, only 8 scored above 80% reproduction and 34 above 40%; the median reproduction score was 28% (rising to 42% among the 80 papers when counting only full, non-downscaled runs).
101 of 105 papers hit at least one obstacle: 58 had code that wouldn't run as released, 48 had numbers inconsistent with the paper, 42 lacked data, 38 shipped no runnable code, and 4 depended on retired/unavailable models.
SAI Review covered 78% of issues raised by ≥2 human reviewers and surfaced 903 code/reproducibility issues humans missed (vs. only 22 the humans caught and it missed); humans still outperformed on novelty and research positioning.
Full reproduction is expensive: median ~$8,900 per Oral paper at Google Cloud on-demand rates, 17 papers over $100K, top near $2.2M — a conservative estimate excluding salaries and 2–3x exploratory reruns.
Context: the study follows public criticism from OpenAI researcher Keller Jordan that headline-grabbing ICLR/ICML/NeurIPS papers are often oversold; SAI notes Review is still Beta and its findings need further verification.

details →

AI Research Roundup 1

2026必看AI干货!《大模型/AIGC/GPT-4/Transformer/DL/KG/NLP/CV AI+X》集合

WeChat: 专知

Rank 43 · Content 40 · Popularity N/A

TL;DR — A WeChat link-aggregation post from 专知 (Zhuanzhi) compiling hundreds of 2025–2026 AI survey papers, PhD theses, conference tutorials, and reports spanning LLMs, agents, multimodal models, and AI+X domains. It is a curated index rather than new technical work, useful mainly as a reading-list entry point.

Heaviest concentration is on LLM agents: surveys on agent memory (evaluation taxonomies, SSGM controlled-memory framework), agentic RL, deep-research systems, multi-agent collaboration, agent communication protocols/security, and full-stack agent safety.
Multimodal and embodied AI are the second cluster: vision-language-action models, 3D/4D world modeling, UAV vision-language navigation, embodied edge foundation models, and the 67-page《具身智能发展报告（2025年）》.
Includes conference-cycle news items (CVPR 2026 awards with Google D4RT best paper, ResNet/YOLO test-of-time; ICML 2026 tutorials/spotlights), plus vendor items like GLM-5 "vibe coding → agentic engineering" and a DeepSeek open-sourced memory module paper credited to Liang Wenfeng.
Content is thin on technical detail — entries are titles/links only, so no results, benchmarks, or methods can be verified from this post itself.

details →

AI Talent & Org Shifts 2

Jeff Dean离职创业，对Gemini有什么影响？

WeChat: 机器之心

Rank 43 · Content 40 · Popularity N/A

TL;DR - Jeff Dean, Oriol Vinyals, Quoc Le, and Sanjay Ghemawat left Google to found Discovery Loop, an automated-ML-research startup, on the same day Google named Koray Kavukcuoglu SVP over Gemini 4 development; Alphabet stock fell over 5%. The piece argues the near-term hit to Gemini 4 is limited, but the structural loss of architecture-level judgment is real.

All three Gemini co-technical leads departed within seven weeks: Noam Shazeer (Transformer, sparse MoE, MQA) to OpenAI on June 17, then Dean and Vinyals on August 5.
Short-term impact is argued to be small: Gemini 4 pretraining has started, the training recipe is locked, execution staff remain, and Kavukcuoglu is a 13-year DeepMind insider (WaveNet, DQN). Gemini 3.5 Pro's repeated delays — Bloomberg attributes them to missing internal quality bars — predate the exits.
Three claimed longer-term losses: architecture-level intuition (Transformer, MoE, seq2seq, NAS), cross-layer TPU→compiler→Pathways→model co-design judgment, and organizational gravity pulling other senior researchers out.
Google hedged rather than blocked: Alphabet is a founding investor, Google Cloud supplies first-year compute plus a joint ML-systems research framework; seed led by Radical and Khosla. Discovery Loop aims to automate ML research itself, with Quoc Le suggesting they may find a non-Transformer architecture.

details →

爆料：哈萨比斯原本要和Jeff Dean一起走！

量子位

Rank 40 · Content 35 · Popularity N/A

TL;DR - 量子位援引 Pathfounders 的爆料称，Demis Hassabis 原本计划与 Jeff Dean 等人同期离开谷歌，被管理层以担忧股价为由劝留，改以"卸任 Google DeepMind CEO、升任 DeepMind 董事长兼 Alphabet 首席科学家"的方式过渡。若属实，这意味着谷歌最核心 AI 组织的领导层与决策重心正在发生结构性重排。

实权交接：原 CTO Koray Kavukcuoglu 出任 Google DeepMind 高级副总裁，直接向 Pichai 汇报，接管 Gemini 与消费产品的日常运营、资源配置与发布节奏；Hassabis 退向科研、长期战略与对外事务。
权力西移：Koray 与负责关键 AI 编程项目的 Sebastian Borgeaud 已从英国迁至加州，部分研究员离开伦敦总部，决策中心由 DeepMind London 收归 Alphabet 加州总部。
资源重配：据《金融时报》此前报道，独立的 AlphaFold 团队已被拆分，研究人员被调往 Gemini、AI 编程、基因组学等方向——研究导向让位于产品竞争。
去向猜测：文章基于 Hassabis 公开访谈中"用 AI 加速科学与医学、甚至先治愈癌症"的表述，推测其若离开将围绕 AI for Science 展开，最可能深耕 Isomorphic Labs；Pathfounders 预测其一年内彻底离职。以上均为未经谷歌证实的爆料与推测，非官方公告。

details →

AI Talent & Startups 1

Jeff Dean 创业路演 PPT，惊现 34 位创始人，谷歌系人才占领半壁江山

WeChat: 图灵人工智能

Rank 33 · Content 25 · Popularity N/A

TL;DR - Jeff Dean left Google after 27 years to co-found Discovery Loop with Sanjay Ghemawat, Oriol Vinyals and Quoc V. Le, and his pitch deck surfaced a list of 34 founders who came out of Google Brain — evidence that Brain has become the single largest talent pipeline feeding today's AI startup wave.

Discovery Loop targets AI for Science: full automation of machine learning, scientific discovery and engineering, framed around "recursive self-improvement" — packaging an "automated AI scientist" so a handful of researchers can outpace large institutions.
The four co-founders' Google track record spans the infra era (GFS, MapReduce, BigTable, Spanner/TrueTime, Borg, LevelDB, TCMalloc) and the deep learning era (Google Brain, the 2012 unsupervised "cat" experiment, TensorFlow, TPUs, Seq2Seq, AutoML/NAS, chain-of-thought, AlphaStar, Gemini co-lead and Gemini 3 "Deep Think" reasoning).
The alumni list includes Dario Amodei (Anthropic), Ilya Sutskever (SSI), Chris Olah (Anthropic) and Yang Zhilin (Moonshot AI, mentored by Quoc Le); the article credits the 2016 Brain Residency program, which recruited on quantitative aptitude, self-driven AI passion and cross-disciplinary background rather than AI PhD pedigree.
Article notes the departure (dated 2026-08-05) coincided with significant Alphabet stock volatility; per the piece, SSI is also expected to unveil its first model this month.

details →

AI for Science 1

刚刚！谷歌之神创业，带走最顶尖AI天团，杀入AI4Science

WeChat: 智药局

Rank 61 · Content 65 · Popularity N/A

TL;DR - Google chief scientist Jeff Dean is leaving after 27 years to co-found Discovery Loop, a public benefit corporation with three longtime Google AI/systems veterans, aimed at automating the scientific discovery loop. It signals that AI4Science is becoming the next strategic battleground for top AI talent after code.

Founding team: Jeff Dean (MapReduce, Google Search infra), Sanjay Ghemawat (GFS, systems), Oriol Vinyals (research, AlphaFold/Gemini lineage), and Quoc Le (automated ML) — roles largely mirror their Google responsibilities.
Core thesis: today's science is a slow, labor-intensive trial-and-error loop; Discovery Loop wants systems that autonomously generate hypotheses, design experiments, run them, and iterate — enabling thousands of parallel experiments.
First milestone is automating ML research and engineering itself; the longer-term target is NAE Grand Challenges (better drugs, health informatics, economical solar, clean water, cyberspace security, discovery tools).
Framing note per the article: Vinyals has publicly argued human scientists remain essential for asking the right questions and steering AI. Company details beyond the website mission (funding, timeline) are not given in the source.

details →

AI in Consumer/Family 1

奥特曼的ChatGPT育儿大法，捅了马蜂窝

量子位

Rank 33 · Content 25 · Popularity N/A

TL;DR - Sam Altman's X post promoting ChatGPT-generated "family podcasts" (built from household calendars and kids' interests for the school run) drew heavy backlash, with a critic's one-line reply ("just talk to your kids") outperforming his post ~18x in likes. The episode surfaces OpenAI's broader push to position ChatGPT as a household assistant.

The proposed use case requires feeding ChatGPT family schedules, children's interests, and activity data — raising retention, consent, child-data-deletion, and shared-memory privacy questions the article flags as unresolved.
OpenAI has been staffing for it: TechCrunch reported a July job posting for a product manager dedicated to family products (since removed).
Sensor Tower estimates cited by TechCrunch: in Q2 2026, ChatGPT users aged 35+ rose to 31% (from 26% a year earlier) while 18–24 fell from 34% to 29%; US smartphone-using parents' adoption reached ~24%, up from 16%.
Safety stakes are higher than in workplace deployments — AP reported seven families sued OpenAI in November 2025 (four alleging ChatGPT's role in suicides, three alleging reinforced harmful delusions), and OpenAI concedes guardrails can be circumvented.

details →

Autonomous Driving VLA 1

现在不做VLA和世界模型的公司，还有哪些？

WeChat: 自动驾驶之心

Rank 36 · Content 30 · Popularity N/A

TL;DR - A WeChat industry roundup (from 自动驾驶之心) arguing that VLA (vision-language-action) plus world models has become a mandatory technical route for essentially every major autonomous-driving player, followed by a promotion for the account's 14-week paid research mentorship program. Useful mainly as a market-landscape signal on where AD stacks are converging.

Claimed VLA adopters: Li Auto (Mind-VLA), XPeng (VLA 2.0), DeepRoute (元戎), Xiaomi, Tesla (world model + FSD fusion), Waymo, NVIDIA; BYD and Changan are described as recent entrants.
Claimed world-model-leaning players: NIO, Huawei, Momenta, Pony.ai, WeRide — with few public papers; Geely, GAC, SAIC, Chery are said to have no public progress, which the author attributes to production-timeline KPIs rather than absence of R&D.
The author's framing: direction is settled, differentiation lies in the entry point — how VLA reasoning couples with world-model future prediction, how action generation and physics modeling share representations, and on-vehicle stability.
The bulk of the post is course marketing: 14 weeks (12 research + 2 writing), baselines VAD/UniAD/DiffusionDrive/OpenDriveVLA/Senna, datasets nuScenes/Waymo/Argoverse, 6 seats, 4090-class inference and 4–8 GPU training required. Note: promotional content, no original results or benchmarks.

details →

Cell Death Immunology 1

议程公布 | SMART Symposium：细胞焦亡与Gasdermin生物学研讨会

WeChat: BioArt

Rank 36 · Content 30 · Popularity N/A

TL;DR - Shenzhen Medical Academy of Research and Translation (SMART) has published the agenda for a Keystone-style symposium devoted entirely to pyroptosis and Gasdermin biology, to be held in person in Shenzhen on September 9–11, 2026. It is billed as the first dedicated international meeting on the Gasdermin protein family, signalling growing institutional investment in this cell-death/immunology frontier.

Scientific scope: Gasdermin-mediated pore formation as a lytic programmed cell death mechanism, covering both inflammasome-dependent and inflammasome-independent functions across mammals, fungi, and bacteria.
Sessions target mechanistic and structural bases of Gasdermin autoinhibition, activation, and post-translational regulation, plus translational prospects for novel therapeutics.
Logistics: venue is Cloud Park International Conference Center, Guangming Science City, Shenzhen; abstract deadline August 21, 2026; registration deadline September 4, 2026; in-person only, with short talks (accommodation covered) and a judged poster session.
Note: this is a conference announcement — the agenda, organizing committee, speaker list, and fee schedule appear as images in the source, so specific names and prices could not be extracted; no research results are reported.

details →

Embodied AI & Robotics 1

李飞飞最新访谈：透露空间智能和具身智能

WeChat: CVer

Rank 40 · Content 35 · Popularity N/A

TL;DR - Fei-Fei Li's World Labs acquired SceniX and, in an a16z interview with co-founder Yunzhu Li, laid out a plan to build a scalable "digital training ground" for robots — model- and embodiment-agnostic simulation infrastructure for spatial intelligence.

SceniX brings a real-to-sim-to-real pipeline with dense reconstruction (appearance, geometry, dynamics) aligned to reality; World Labs contributes Marble, a generative world model that turns single/multiple images or text into geometrically consistent worlds plus sparse reconstruction.
Core bottleneck framed as robot data scarcity: unlike LLMs there is no internet-scale corpus, so simulation is used to unlock scaling laws via controllable randomization of lighting, geometry, friction, and physics parameters.
Evaluation is the pitch's centerpiece: real-world robot eval is orders of magnitude slower than LLM eval and is dangerous/expensive, so aligned digital environments are used to distinguish e.g. a 90% vs 92% checkpoint quickly.
Positioning is infrastructure, not hardware or "robot brains" — data can train models from scratch or fine-tune VLA/World Action models; they argue "models change, infrastructure doesn't," and target semi-structured environments (warehouses, restaurants) before unstructured homes.

details →

Embodied Robotics AI 1

全行业都在卷大模型，他们却用不确定微分几何给机器人做了个大脑

WeChat: 机器之心

Rank 43 · Content 40 · Popularity N/A

TL;DR - 机器之心 profiles 信度启源 (Xindu Qiyuan), a newly founded Tsinghua-affiliated startup led by mathematician 刘宝碇, building a robot "brain" on uncertainty theory and "uncertain differential geometry" instead of large data-driven models. It matters as a claimed third technical route for embodied AI — theory-driven and white-box — positioned against end-to-end VLA models and classical control+AI hybrids.

The premise: real-world environments violate probability theory's frequency-stability assumption, causing three failure modes in learned models — forced classification of unseen inputs, false confidence (e.g. Wiener-process motion predictions implying infinite path length), and data hunger requiring recalibration per scene.
The approach stacks solid geometry → differential geometry → uncertain differential geometry, modeling perturbations as "belief" (uncertain measure) to compute a trustworthy physical boundary, then solving for optimal control inside it — analogous to keeping a safe following distance rather than estimating lane-change probabilities.
A "七窍" (seven-module) framework mirrors human cognition: hearing/touch/vision for feature extraction, plus verification (vestibular-like millisecond safety cutoff), prediction (compensates sensor latency to infer the current true state), decision (belief maximization, no action sampling), and control — all non-neural and traceable end to end.
Claimed demo: a humanoid grasping a cup at an uncalibrated position at human speed, in an uncut ~10s video, with no task-specific pretraining, no GPU racks, and one month from robot procurement to full deployment. These are vendor/company claims with no benchmarks, ablations, or third-party evaluation provided; the article itself notes scaling to precise assembly, multi-robot, and open environments remains unvalidated. The team is currently raising a new funding round.

details →

Robotics Foundation Models 1

R to @GoogleDeepMind: FInd out more → https://goo.gle/3Tvi61l

@GoogleDeepMind

Rank 57 · Content 60 · Popularity N/A

TL;DR - Google DeepMind announced Gemini Robotics 2, a robotics model that extends Gemini's capabilities to "whole-body" robot control rather than just arm/hand manipulation. It matters because it signals a push from foundation-model-driven grasping toward full-body, multi-robot embodied autonomy.

Positioned as "whole body intelligence" — coordinated control spanning locomotion and manipulation ("from feet to fingertips"), not isolated end-effector control.
Emphasizes fine dexterity alongside coarse body motion, implying a single policy stack handling multiple control granularities.
Highlights robot "teamwork," suggesting multi-robot coordination on shared tasks.
Content is a promotional announcement post only: no benchmarks, hardware platforms, training details, or evaluation results are provided, so capability claims are unverified here.

details →

Robotics World Models 1

李飞飞最新访谈：人也不全靠现实数据学习啊

WeChat: 图灵人工智能

Rank 40 · Content 35 · Popularity N/A

TL;DR - Fei-Fei Li's World Labs acquired robotics simulation startup SceniX and, in an a16z interview with co-founder Yunzhu Li, laid out a plan to build a model-agnostic "digital training ground" where robots are trained and evaluated in simulation before real-world deployment. It matters because it frames scarce robot training/eval data — not hardware or policy architecture — as the field's core bottleneck.

SceniX builds a real-to-sim-to-real pipeline with dense reconstruction (appearance, geometry, dynamics) aligned to real environments; World Labs contributes Marble, a foundation model turning single/multiple images or text into geometrically consistent worlds, plus sparse reconstruction and generative 3D strengths.
The stated case against pure video-model routes is consistency (spatial, temporal, cross-view, cross-interaction) — video predictors can make pushed objects vanish, giving no usable training signal.
Simulation is pitched for two gains: reliability via systematic randomization of lighting, geometry, friction and physics parameters to cover the state space, and efficiency, including speeding up robot behavior beyond teleoperation collection rates; evaluation is called the bigger neglected problem, since distinguishing a 90% from a 92% checkpoint in the real world is orders of magnitude slower than LLM eval, plus dangerous and expensive.
Positioning is deliberately neutral — no hardware, no robot brain; data can train models from scratch or fine-tune VLA/World Action models, on the thesis that "models get replaced, infrastructure doesn't." Near-term focus is semi-structured environments (warehouses, restaurants) over unstructured homes; Li cites Waymo's billions of simulated hours and argues simulation enables counterfactual reasoning humans also perform.

details →

Top highlights — Opinions

Where AI still hits a wall in frontier math: 王虹与三维挂谷猜想：深度解析基础数学的AI能力边界 uses Wang Hong and Joshua Zahl's 127-page three-dimensional Kakeya proof as a case study, arguing AI is an "outstanding student" for delegated subtasks but not yet a paradigm-creating "master" — a boundary independently echoed by adjacent takes on Tao's ICM 2026 lecture and MLS-Bench.
Machines as the primary audience for research: 李德毅院士：当机器成为论文的第一读者——图灵方程将不再沉寂 argues scholarly impact will shift from human citation counts to whether work is retrievable and callable by AI systems, with Turing's long-dormant reaction-diffusion equations as the cautionary precedent.
The local/on-device inflection: RT by @huggingface: Many AI tasks can now run fully local or on-device. - Redact PII in the… contends the recent wave of 0.6B–3B open-weight models makes many practical tasks runnable in-browser, leaving model routing across a fragmented ecosystem as the open problem.
AI as macroeconomic general-purpose technology: RT by @ylecun: My Nobel lecture and the video of the Stockholm lecture are online. It covers what… surfaces a creative-destruction framing of AI's macro impact — notable less as a technical result than as economics-side signal amplified from within the AI research community.

AI & Economic Growth 1

RT by @ylecun: My Nobel lecture and the video of the Stockholm lecture are online. It covers what…

@ylecun

Rank 43 · Content 40 · Popularity N/A

TL;DR - A retweet by Yann LeCun of an economist's announcement that their Nobel lecture paper and Stockholm lecture video are now publicly available, covering creative destruction and framing AI as the next general-purpose technology (GPT) revolution. It matters as a high-profile economic framing of AI's macro impact rather than a technical result.

Content is an availability announcement (paper on AEA's site, video on nobelprize.org), not new empirical AI research; details beyond the abstract-level topic list are not provided.
Stated scope: what creative destruction implies for secular stagnation, firm dynamics, and the middle-income trap.
Positions AI as the next general-purpose technology revolution, aligning it with historical GPTs (steam, electricity, IT) in growth-theory terms.
Item is thin — only the announcement text and links were available, so no findings, magnitudes, or methodology can be summarized.

details →

AI & Research Impact 1

李德毅院士：当机器成为论文的第一读者——图灵方程将不再沉寂

WeChat: 图灵人工智能

Rank 33 · Content 25 · Popularity N/A

TL;DR — Chinese Academy of Engineering academician Li Deyi argues that in the AI era, machines—not humans—are becoming the first and primary readers of basic-research papers, so scholarly influence will shift from human citation counts to whether work is retrievable, callable, and reusable by AI systems ("the silicon-based circle"). He uses Turing's 20-year-dormant reaction-diffusion equations as the cautionary historical case that he believes will no longer repeat.

The Turing equation case study: Turing's 1952 The Chemical Basis of Morphogenesis lay ignored for ~20 years due to five compounding factors — conflict with the then-dominant gene-determinism paradigm (Watson/Crick DNA, Wolpert's positional-information gradients), cross-disciplinary math/biology barriers plus a journal audience mismatch, no experimental validation (no compute to simulate patterns, no assays to track morphogen diffusion), Turing's death two years post-publication leaving no advocate or successors, and the counterintuitive claim that diffusion creates order. It resurfaced only via Gierer–Meinhardt's 1972 computational simulations and Prigogine's dissipative-structures theory.
New impact metric: the traditional loop (humans read → cite → publish) is being displaced by machine retrieval and invocation. Explicit AI usage — in RAG pipelines and tool calls — is logged automatically, bypassing the gap between being depended on and being cited (Matplotlib: ~20k citations vs. ~300k dependent packages).
Non-paper artifacts count: GitHub code, a lemma in a formal math library, or an agent-callable tool may exert more influence inside AI systems than a top-journal paper, since dissemination through the silicon circle is faster and broader than through the human "carbon-based" circle.
Human role reframed: as AI drives the cost of routine problem-solving down, scarcity moves from "producing results" to "being worth using" — the researcher's value shifts to choosing fundamental problems, designing workflows, and verifying outputs, with imagination, consciousness, and pre-evidential insight held as irreplaceably human.

details →

AI for Mathematics 1

Representative image for 王虹与三维挂谷猜想：深度解析基础数学的AI能力边界

王虹与三维挂谷猜想：深度解析基础数学的AI能力边界 🔗 5 sources

WeChat: 图灵人工智能 WeChat: 智药局 WeChat: 专知 WeChat: 图灵人工智能 WeChat: 学术头条

Rank 41 · Content 30 · Popularity 65

TL;DR — A cluster of Chinese-language WeChat commentary centered on Wang Hong (王虹) and Joshua Zahl's 127-page proof of the three-dimensional Kakeya (挂谷) conjecture, used as a case study for where today's AI hits its ceiling in frontier research: AI is an "outstanding student" for delegated technical subtasks, not yet a "master" that creates paradigms. Adjacent pieces (Terence Tao's ICM 2026 lecture, the MLS-Bench benchmark, Daphne Koller on drug discovery) independently corroborate the same boundary from evaluation, workflow, and applied-science angles.

Why 3D Kakeya is the test case: unlike the flexible 2D setting, the 3D problem has rigid geometry, infinitely nested multi-scale/chaotic structure, "sticky" Kakeya configurations, and outruns existing Fourier/harmonic-analysis tooling. Wang and Zahl's contribution is framed as system-level innovation — abandoning prior research paths and building new machinery — spanning four branches of mathematics across 127 pages.
The four claimed AI limits: reuse of existing paradigms instead of inventing frameworks; no global research-strategy judgment; inability to sustain very long, multi-layer, cross-domain logical chains; and no mathematical intuition for structures without precedent or template. AI is positioned as an accelerator for verification, gap-hunting, symbolic simplification, and pruning dead ends, and as dominant on bounded, competition-style problems.
Empirical echo (MLS-Bench): 140 real research tasks across 12 ML areas, each with a real codebase, ≥3 transfer conditions, and ≥3 reproduced strong human methods (anchored scoring: weakest baseline = 0, strongest = 50, theoretical bound = 100). All five frontier models failed to beat the strongest human method; expert review found they mostly recombine losses/modules/tricks from provided baselines, "optimize/debug" prompts beat "discover a new method," and removing the parameter-count guard let models win purely by scaling — unisolated variables masquerading as discovery. With more compute freedom they performed worse, failing to allocate budget to high-information experiments. (Full run ≈700 H100-hours per candidate; a 30-task subset ≈100 H100-hours is used in official frontier-model release evals.)
The downstream bottleneck (Tao): AI plus formal tools like Lean accelerate only proof generation and verification, congesting the remaining stages — exposition, publication, digestion, canonization — creating "proof indigestion." Cited: First Proof reported 7 of 10 novel research-level problems solved to publishable quality at $10–$1,000 each, while erdosproblems.com accumulates AI-submitted proofs without enough qualified verifiers. Prescriptions: disclose AI assistance, reward refereeing/surveys/translation over priority races, and apply a "talk test."
Same ceiling in applied science (Koller): insitro's founder argues AI's drug-development bottleneck is identifying correct disease mechanisms, not generating molecules — >90% of trials fail on wrong mechanisms; 38 targets each carry 50+ programs while new targets/year fell from ~100 (2015) to ~30 (2024). Cell atlases cover a tiny slice of perturbation space and lack causal data; agentic self-driving labs need fast, cheap, objective feedback that human clinical trials cannot provide.

Emphasis differs by source: 图灵人工智能 argues the boundary conceptually (Kakeya) and procedurally (Tao's proof lifecycle), 学术头条 supplies the quantitative benchmark evidence, 智药局 transfers the argument to biomedicine, and the 专知 post is an unrelated content-marketing index of 100+ "AI + military" reports with no results or original analysis.

details →

On-Device Small Models 1

RT by @huggingface: Many AI tasks can now run fully local or on-device. - Redact PII in the…

@huggingface

Rank 54 · Content 55 · Popularity N/A

TL;DR - A Hugging Face-amplified take arguing that a wave of recent small open-weight models (0.6B–3B) now makes many practical AI tasks runnable fully locally or in-browser, with model routing across this fragmented ecosystem framed as the next open problem.

Task-specific small models cited: OpenAI privacy-filter (1.5B) for in-browser PII redaction, Nemotron-3.5 ASR (0.6B) for streaming speech-to-text, OvisOCR2 (0.9B) for document parsing, LFM2.5 (2.6B) for agent loops, Shieldstral (3B) for policy-based moderation, and Marlin-2B for video understanding.
All are described as open-weight releases from the last few months, implying local/on-device deployment is now viable without API calls — a privacy and latency argument as much as a capability one.
The parameter range (sub-3B) suggests specialization over generality: narrow models beat one large generalist for these workloads at edge-feasible sizes.
The author's stated bottleneck is routing — selecting the right specialist per request across a diverse model zoo — not raw model quality. This is an opinion/prediction, not a benchmarked result.

details →