2026-08-24

TL;DR - SWE Refactor Bench evaluates coding agents on 20 long-horizon, whole-repository migrations using checks for both migration completeness and behavioral correctness. Frontier models succeed on only 5.4% of runs, highlighting major reliability gaps in autonomous stack migration.

• The three-stage protocol combines migration audits, fixed behavioral tests, and targeted tests generated by six independent coding agents.
• Only 28 of 520 runs passed all stages, and 13 of the 20 migration tasks had no accepted solution.
• Of the 340 runs that completed the migration, 58% passed 99% of fixed checks, but only 26% passed every check.
• Performance varied sharply by task: agents scored 31.4 on build-toolchain rewrites versus 5.6 on language rewrites; the best model scored 47.0/100.

TL;DR - EarthVerse is a benchmark for evaluating tool-using scientific agents on 405 reproducible Earth-system and natural-hazard investigations. Results show that even strong systems struggle to maintain reliable end-to-end reasoning across heterogeneous evidence, calculations, units, and physical interpretation.

• Tasks span 199 documented events and 19 hazard families, with provenance-preserving answers and executable ground truth.
• The benchmark evaluates evidence selection, tool use, memory, reasoning, interaction, and scientific execution under a controlled protocol.
• Across 25 model and agent systems, the best mean answer-unit accuracy was 84.65%, but the highest Strict@95 score was only 34.81%.
• This gap indicates that agents often solve individual steps correctly while failing to preserve a consistent scientific chain across the full investigation.

TL;DR - This paper presents the Salt method, in which AI agents exchange machine-checked proofs while one researcher supervises development from application code through a verified compiler to a taped-out RISC-V processor. It argues that inexpensive formal verification can make large-scale autonomous engineering safer and more productive.

• Lean 4 proof-kernel checks prevent hallucinated proofs from being accepted as valid artifacts.
• Verification spans the toolchain link by link, ending with SAT-checked equivalence at the silicon boundary.
• The five-week project reportedly used consumer AI subscriptions, with no human-written RTL and no human review of proofs.
• Published auditing includes theorem provenance, token and human-time accounting, and an append-only error ledger recording 256 numbered catches, with zero incorrect proofs entering the final record.

TL;DR - SkillAlchemy is a source-grounded framework that automatically creates reusable agent skills from underspecified briefs and open-world materials. It approaches human-curated performance while substantially improving task success over existing automated methods.

• Identifies requirements omitted from skill briefs using contrastive evidence gathered from permitted sources.
• Admits candidate procedures only within the scope justified by supporting evidence.
• Compiles approved procedures into grammar-guided skill packages for inference-time use.
• Across 87 SkillsBench v1.1 tasks, it improved pass rate by 19.9 percentage points over no-skill execution and 8.6 points over the strongest automated baseline.

TL;DR - AutoResearchEval evaluates eight agent–model combinations on 100 real-world frontier research tasks and finds that autonomous research agents primarily fail at scientific judgment and correction, not engineering execution. The central gap is a missing “metacognitive loop” that turns recognized flaws into revised experiments and conclusions.

• Researchers analyzed 800 end-to-end trajectories across seven scientific domains and identified 45 failure modes spanning planning, retrieval, experimentation, analysis, writing, and review.
• Cognitive and scientific failures accounted for 92.1% of observed failures, while engineering robustness issues accounted for 7.9%.
• “Uncorrected self-awareness” appeared in 660 of 800 trajectories (82.5%): agents often identified serious problems but submitted results without fixing them.
• Similar failures occurred across Claude Code, Codex, and Gemini CLI configurations, suggesting a shared limitation in current agentic research capabilities rather than a framework-specific issue.

TL;DR - This paper identifies an “interaction tax” in multi-agent LLM teams: sharing complete solutions can rapidly homogenize agents’ proposals and eliminate the diversity that makes multiple models useful. Across 11 verifier-scored optimization tasks with matched budgets, independent generation was a stronger default than full-solution interaction.

• Agents exposed to complete peer outputs converged within one round and tended to remain close to the first solution seen.
• Independent proposal generation preserved structurally different approaches across model families.
• Critique improved results only when the violated constraint was easy for the LLM to identify and repair.
• Performance depended more on what information agents exchanged—and when—than on the number of agents.

TL;DR - InjecMEM is a single-interaction memory injection attack that plants a retrievable record in an LLM agent’s persistent memory to steer later topic-related responses toward a chosen output. It exposes agent memory as a durable attack surface even without direct access to the memory store.

• Combines a retriever-agnostic anchor with high-recall topical cues and an adversarial command that activates when the record is retrieved.
• Optimizes the command using gradient-based coordinate search across synthetic prompt templates, insertion positions, and optionally multiple backbone models.
• Remains effective across multiple memory systems and models under variable contexts, long prompts, and memory drift.
• Targets related queries while reportedly leaving non-target queries unaffected, highlighting the need for memory-system defenses.

TL;DR - Google and collaborators propose EnvHarness, a programmable layer that adapts static training environments to an agent’s evolving weaknesses without rebuilding simulators or validators. Its automated EnvRigger system improved held-out task performance by up to 9.0 percentage points while reducing average execution steps by 9.8%.

• EnvHarness modifies environments through three composable components: Stage changes initial states, Contract adjusts interaction rules and observations, and Chain links tasks into longer episodes.
• EnvRigger analyzes successful and failed trajectories, diagnoses policy weaknesses, generates targeted components, and validates them through actual rollouts.
• Across ALFWorld, WebArena, SWE-bench Verified, OfficeQA, and SpreadsheetBench, the approach generally outperformed original environments and domain-specific environment-generation pipelines.
• GRPO experiments showed improved results on three of four reported metrics, suggesting adapted environments can provide better direct reinforcement-learning signals as well as better skill-extraction trajectories.

TL;DR - TOWN-VLA is a prompt-authority interface that lets retrieval modify a frozen vision-language-action policy only when a compatibility rule approves a canonical instruction. It prevents prompt-form collapse and improves simulated and physical robot manipulation without retraining.

• Appending text indiscriminately reduced mean success from 92.47% to 3.00%; both meaningful and length-matched meaningless additions failed on all 500 audited states.
• TOWN-VLA restores the exact base prompt when intervention is unauthorized; across 900 routes, 525 matched the base prompt by hash and 375 authorized prompts preserved the task signature.
• On LIBERO-Plus, success increased from 69.5% to 73.1% across 10,030 episodes per method, improving on all four suites and six perturbation axes.
• On a physical PiPER arm, success rose from 52.7% to 78.7% over 150 trials per method.

TL;DR - BEACON is an ICML 2026 framework for training long-horizon language agents by assigning credit at environment-defined milestones rather than relying only on sparse terminal rewards. On long ALFWorld tasks, it raised success from 53.5% with GRPO to 92.9% while substantially improving sample utilization.

• BEACON partitions trajectories at verified subgoals, applies temporally decayed rewards within each segment, and combines trajectory-level and segment-level advantage estimates.
• It requires no additional reward model; milestones come from environment states or built-in subgoal signals in ALFWorld, WebShop, and ScienceWorld.
• Effective sample utilization increased from 23.7% to 82.0%, and the zero-advantage ratio fell from roughly 55% to 10%.
• Gains grew with task length, suggesting milestone-based credit assignment particularly addresses failures in long-horizon agent reinforcement learning.

TL;DR - ClawSentry is an open-source, framework-agnostic security gateway that monitors autonomous LLM agents across skill admission, intent review, execution, and post-action consequences. Its progressive review tiers substantially reduce successful attacks while largely preserving performance on benign tasks.

• Combines deterministic checks, semantic review, and bounded read-only agentic investigation, escalating only ambiguous cases.
• Reviews skill packages before first use and tracks rephrased or tool-switched retries across an entire session.
• Applies a shared policy across Codex, Claude Code, Kimi CLI, and Gemini CLI without modifying agent internals.
• On SkillInject with Codex/GPT-5.4, attack success fell from 39.55% to 2.61%, while task success decreased from 83.78% to 83.05%; across five agents, clean-skill task success remained 98.7%.

TL;DR - Fine-tuning MedSAM improves in-domain and nearby-domain segmentation but can hurt performance under large distribution shifts. The study links these failures to decoder representation drift and identifies adaptation and prompt-training strategies that better preserve robustness.

• Full fine-tuning offered the best overall tradeoff across in-distribution, close-OOD, and far-OOD datasets.
• Encoder-only LoRA was the strongest parameter-efficient method, outperforming standard LoRA and visual prompt tuning on far-OOD data.
• Centered Kernel Alignment analysis associated far-OOD degradation with decoder drift; encoder similarity alone did not explain robustness.
• Training with random 0–100-pixel prompt jitter improved performance and resilience to noisy prompts.

TL;DR - This paper audits the weight changes between general-purpose LLMs and their medical-specialized counterparts, finding that decoder updates closely reproduce benchmark gains but do not yield a simple component-level explanation of specialization.

• Examines aligned Gemma-3-to-MedGemma and Qwen2.5-to-HuatuoGPT-o1 checkpoint pairs.
• Full decoder deltas strongly reconstruct medical benchmark movement, with endpoint-normalized retention of 0.974 and 1.183.
• MLP layers are the strongest broad component family in both pairs, but controls and rollback tests prevent uniquely attributing the gains to them.
• Conclusions are limited to text-only multiple-choice benchmarks and do not establish clinical validity or circuit-level mechanisms.

TL;DR - MediSkill-Evo is a clinical agent framework that self-evolves governed process knowledge without fine-tuning its backbone model. It improves diagnosis, treatment-intent coverage, and safety metrics by enforcing evidence provenance and clinical process constraints throughout interactions.

• Separates experience into typed banks for clinical skills, process rules, symbolic schemas, and measurement procedures, then publishes validated knowledge to a frozen test-time snapshot.
• Uses a preference harness to bind evidence to sources, reject controller-invalid actions, and rank valid actions with a safety-prioritized clinical critic.
• On 300 held-out Qwen encounters, diagnosis accuracy rose from 61.33% to 69.00%, treatment-intent coverage from 33.62% to 66.44%, and critical failures fell from 31.00% to 16.33% versus AgentClinic.
• Results are descriptive system-level evidence from fixed automatic evaluations, not causal evidence for individual components or clinical validation.

TL;DR - MedGuard is a 7B-parameter, LLM-based gatekeeper that fact-checks Chinese telemedicine consultations before clinical advice or prescriptions are finalized. It matters because it combines patient-specific context with traceable medical evidence to detect errors while limiting unnecessary alerts.

• MedGuard converts long, multi-turn consultations into patient-specific atomic medical claims, preserving critical details such as negations, dosages, timing, medical history, and allergies.
• An uncertainty-driven workflow sends ambiguous claims through iterative evidence planning, retrieval, sufficiency assessment, and reasoning against six authoritative Chinese medical-resource types.
• On MedGuardEval, fine-grained risk-detection F1 improved by an average of 22.1% over baselines; claim extraction improved by 23.2%.
• Across 604 consultations, 126 clinicians rated all seven evaluation dimensions above 4.0/5; on 10,000 retrospective consultations, MedGuard issued 2,234 alerts versus 4,785 from a comparison model.

TL;DR - KREL is an LLM-based framework for automatically assigning ICD codes to clinical notes by combining clinical-evidence reasoning with structured coding guidelines. This knowledge-guided approach aims to reduce hallucinations, improve standards compliance, and handle challenges such as long notes and a large label space.

• Integrates external ICD coding guidelines directly into the LLM reasoning process.
• Couples clinical-text interpretation with evidence-based, multi-step code selection.
• Addresses limitations of extreme multi-label classification and unconstrained generative approaches.
• Consistently outperforms strong pretrained language model and state-of-the-art LLM baselines on benchmark datasets.

TL;DR - A Science Advances study introduces CAT-Cube, a programmable DNA nanostructure that directly visualizes and amplifies RNA-virus signals in living cells without modifying the virus. It also enables real-time evaluation of antiviral efficacy and mechanisms.

• Target viral RNA triggers catalytic hairpin strand displacement, repeatedly releasing the RNA to amplify fluorescence; in vitro signal was about 3.8× the negative control.
• CAT-Cube autonomously entered Vero E6 cells, remained responsive after 48 hours, and detected SARS-CoV-2 trVLP infection down to MOI 0.1.
• The system measured inhibition by an ACE2 inhibitor and S-protein-targeting DNA aptamers, which reduced virus-associated fluorescence by roughly 68–75%.
• Imaging and computational analyses suggested cepharanthine may block viral entry by interacting with both host ACE2 and viral S protein, while also affecting host translation-related processes.

TL;DR - A Nature study developed poorly membrane-permeable caspase inhibitors that enter pyroptotic cells through transient GSDMD pores, selectively suppressing inflammatory cell death while sparing apoptosis. The approach could enable safer anti-inflammatory therapies, although liver-enzyme elevations indicate further optimization is needed.

• Screening identified selective covalent inhibitors of inflammatory caspases, including KGR-3 for human caspase-1/4 and KGR-53P for mouse caspase-1/11.
• The inhibitors required GSDMD pores to enter cells, blocking pyroptosis and IL-1β/IL-18 release without disrupting apoptotic pathways.
• Delayed treatment remained effective, showing a therapeutic window after initial GSDMD pore formation.
• KGR-53P reduced inflammatory cytokines in a mouse endotoxin-shock model but raised liver enzymes without LPS, suggesting off-target toxicity.

TL;DR - An Immunity study identifies yolk-sac-derived CD163-high red-pulp macrophages as essential for splenic recovery after malaria. Malaria-induced hemolysis permanently eliminates these cells, disrupting macrophage signaling and causing lasting damage to splenic architecture and hematopoietic recovery.

• Single-cell RNA sequencing and lineage tracing identified an iron-recycling CD163-high macrophage subset that retains its embryonic yolk-sac origin into adulthood.
• Hemolytic stress—not persistent parasites or generalized inflammation—caused this subset’s irreversible loss across multiple malaria models.
• Loss of CD163-high macrophages disrupted GDF15–TGFBR2 trophic signaling to marginal metallophilic macrophages, impairing their regeneration and destabilizing the splenic marginal zone.
• Reciprocal support from marginal-zone macrophages was also necessary for red-pulp macrophage development, iron storage, and recovery of red blood cell production.

TL;DR - An Immunity study identifies CCR7+ activated dendritic cells (actDCs) as essential for spontaneous and therapy-induced antitumor immunity. New genetic mouse models show that actDCs uniquely prime tumor-specific cytotoxic T cells and sustain responses to checkpoint blockade and adoptive T-cell therapy.

• Both cDC1 and cDC2 converge on the actDC state, marked by CCR7 and elevated costimulatory, antigen-presentation, and regulatory molecules.
• Only activated DCs—not resting DCs that had merely acquired tumor antigen—could prime naive tumor-specific CD8+ T cells.
• actDC1s use cross-presentation, whereas actDC2s can acquire preformed peptide–MHC I complexes from tumor cells through “cross-dressing.”
• Selective actDC depletion reduced tumor-specific and tumor-infiltrating T cells, impaired spontaneous tumor rejection, and weakened immunotherapy efficacy.

TL;DR — Orbformer is a transferable deep quantum Monte Carlo foundation model for ab initio electronic wavefunctions that accurately captures bond breaking and other strongly correlated processes. Reusing learned electronic-structure patterns substantially reduces computational cost while approaching chemical accuracy.

• Orbformer was pretrained without external energy labels on 22,350 equilibrium and nonequilibrium molecular geometries, then jointly fine-tuned across related structures or reaction pathways.
• Across five bond-dissociation curves, it matched or surpassed the accuracy–cost trade-off of conventional quantum-chemistry methods and systematically converged toward approximately 1 kcal/mol accuracy.
• Joint fine-tuning delivered about a 20× efficiency gain, with pretraining providing additional acceleration when target systems resembled the pretraining distribution.
• Despite being pretrained only on systems containing at most 24 electrons, the model generalized to molecules with up to 106 electrons.
• The learned representations included physically meaningful localized orbitals and reusable local electronic-structure patterns, supporting transfer across molecular systems.

Note: Only the first source concerns Orbformer; the other two describe unrelated studies on cancer-targeting minibinders and drug repurposing for MPS IIIA, so their details were not merged.

TL;DR — EvoMax is a sparse-data, model-guided protein evolution framework that substantially improved the activity and targeting range of compact Fanzor genome editors in mammalian cells. The resulting FanzMAX system demonstrated single-AAV in vivo editing, but increased toxicity and large genomic deletions exposed an important activity–safety tradeoff.

• EvoMax integrates Gaussian process regression trained on 209 measured mutations with ESM-2 evolutionary priors and ESM-IF structural compatibility, requiring experimental testing of only about 10–20 candidates per iteration.
• Combining model-guided protein mutations, engineered ωRNA, and an hLa fusion increased editing efficiency, broadened TAM compatibility, and restored activity in several naturally inactive Fanzor2 homologs.
• FanzMAX v3-hLa achieved up to 97% editing at its strongest endogenous target and averaged approximately 33% across 19 sites—more than 2.6-fold higher than two existing compact editors.
• Single-AAV delivery enabled PCSK9 editing in mouse liver, although the most active constructs also caused toxicity and more large genomic deletions, indicating that future optimization must jointly address efficacy and safety.

Note: The BioArt summary describes MOSAIC, a DNA-synthesis platform unrelated to the title-matching EvoMax/Fanzor study, so its claims cannot be factually merged into this work.

TL;DR - TracingFlow is a simulation-free flow-matching framework that infers trajectories from sparse temporal snapshots using second-order dynamics. By learning acceleration rather than only velocity, it better captures nonlinear, high-curvature processes such as cell differentiation.

• Provides an exact, efficient solution to the Dynamical Optimal Acceleration Transport problem.
• Learns force fields that model regulatory momentum and delayed responses absent from memoryless first-order methods.
• Improves distribution reconstruction and trajectory faithfulness on synthetic and large-scale scRNA-seq datasets.
• Incorporates lineage-tracing priors to recover mathematically optimal and biologically plausible dynamics.

TL;DR - Relay-OPD is an online distillation method that detects when a student model’s reasoning begins to diverge and briefly hands generation to a teacher before returning control. It improves mathematical reasoning accuracy while cutting training trajectory length by more than half.

• A label-free trigger fires when the teacher favors a reflective token such as “Wait,” “But,” or “However,” while none appears among the student’s top-K choices.
• Teacher interventions are deliberately early and limited: the main setup allows two handoffs of three paragraphs each, followed by single-sample distillation on the resulting relay trajectory.
• With a Qwen3-4B teacher and 1.7B student, Relay-OPD averaged 46.96 accuracy across eight math benchmarks, 5.73 points above standard OPD, while reducing average trajectories from 4,658 to 2,296 tokens.
• A speculative-decoding implementation supplies teacher logits and alternates teacher/student generation in one engine, avoiding separate verifiers, reward models, process labels, or correctness labels.

TL;DR - JoyAI-Echo-1.5 is a unified audio-visual generation system designed for coherent long-form video and interactive worlds. It combines cross-shot memory, geometric camera control, and rollout-aware training to preserve characters, voices, and scene stability over long horizons.

• Its long-video variant aggregates visual history and speech-derived speaker cues to maintain character appearance and voice identity across shots.
• Its world-model variant translates diverse navigation inputs into metric 6-DoF camera trajectories for controller-agnostic viewpoint control.
• Progressive teacher forcing and Self-Gradient Forcing convert a bidirectional backbone into an efficient causal few-step generator.
• The system improves multiple long-video metrics over existing baselines and reports first place on WBench with an average score of 81.7.

TL;DR - InfinityEdit is a lightweight adapter that lets streaming video generators apply instructions to future video chunks rather than editing only fixed-length clips. It enables ongoing applications such as live-game restyling and camera changes while aiming to preserve continuity and stability across repeated edits.

• Defines “infinite video editing,” where edit requests arrive over an unbounded sequence and affect subsequent generated segments.
• Uses history cross-attention for source continuity, causal temporal self-attention for forward-only temporal cues, and edit cross-attention for instruction conditioning.
• Activates the adapter only when a new edit arrives; later chunks use the original generator with a reset anchor frame.
• Introduces a task-specific data-collection pipeline and reports stable, faithful continuation across accumulating edit sequences.

TL;DR - Meitu’s MT Lab introduced Consistent Feature Transport (CFT), an ECCV 2026-accepted method that reframes portrait relighting as explicit lighting-feature transport within a Rectified Flow model. It improves lighting quality and physical consistency while better preserving identity, geometry, and scene content.

• CFT jointly learns noise-to-target generation, noise-to-source reconstruction, and direct source-to-target lighting transfer; the latter uses Rectified Flow’s linear structure and a parallelogram construction.
• Its core transfer objective is supervised with images that have different content but share the same lighting transformation, helping disentangle reusable lighting changes from identity and scene differences.
• The team built a large portrait-relighting dataset spanning indoor and outdoor scenes and 14 lighting categories, including structured illumination, mixed color temperatures, and complex shadows.
• The full CFT configuration achieved the best reported ablation results across SSIM, PSNR, LPIPS, and FID, generalized to style-transfer models, and has been deployed in Picchi, AirBrush, and BeautyPlus.

TL;DR - Periodic Row-wise Muon makes the Muon optimizer more practical for training Diffusion Transformers at 1.3B–15B parameters. It preserves Muon’s generative-quality gains while substantially reducing optimizer computation, communication, and total training time.

• Muon improved best observed generative quality by 12.9–19.1% over AdamW across tested model scales.
• The proposed method runs the full five-step Newton–Schulz spectral update only every (K) steps, using cheaper row-wise constrained updates between refreshes.
• A sharding-aware distributed implementation reduces optimizer time by 46.9–54.3%, step time by 15.7–24.3%, and logical communication volume by 66.7%.
• It reached its best generative quality with 33.7–64.8% less active training time than vanilla Muon, while staying within 0.5% quality on 1.3B–4B models and improving quality by 4.5% at 9B.

TL;DR - ProxyFormer uses persistent local features alongside compressed proxy tokens to make global attention and KV caching substantially cheaper. It enables million-token retrieval and much longer training contexts on limited GPU memory while also showing preliminary applicability to high-resolution image generation.

• Each layer compresses local features into proxy states, performs global interactions in proxy space, then injects the contextualized proxies back into the persistent local stream.
• Factorized multi-level compression, dynamic layer-wise compression ratios, asymmetric embeddings, and proxy-only KV caching further reduce resource requirements.
• On a 16GB GPU, a compression ratio of 64 increases the reported trainable context from roughly 20K to 0.7M tokens.
• A 64K-window model retains 92%–95% multi-needle retrieval accuracy at 1,048,576 tokens; preliminary flow-matching experiments cover pixel and latent image generation.

TL;DR - Human-JEPA is a video-pretrained vision model designed to both perceive people in the present and anticipate their future behavior. Its anchored forecasting approach preserves dense perception capabilities while adding anticipation, offering a unified alternative to separate static and predictive models.

• Anchors dense prediction targets to a frozen copy of the model initialization to prevent degradation of dense perception during video pretraining.
• Uses a past-to-future training split instead of block masking, avoiding reported drops of five points in action recognition and 17 points in person re-identification.
• With frozen probes, outperforms pixel-anchored specialists on pose estimation and person re-identification despite having 2.7× fewer parameters.
• The released predictor head adds anticipation without degrading it, though the model remains weaker on high-resolution dense parsing.

TL;DR - HKU researchers introduced SegMAN, an open-source semantic segmentation framework combining Mamba-based global modeling with local self-attention and multiscale decoding. It improves segmentation accuracy and inference efficiency across major benchmarks while using fewer parameters than several larger vision backbones.

• The SegMAN Encoder’s LASS module combines sliding-window neighborhood attention for fine details with SS2D state-space scanning for long-range dependencies at linear complexity.
• Its Mamba-based MMSCopE decoder aggregates multiresolution features and propagates global semantics across scales.
• SegMAN-B’s 45M-parameter encoder reportedly outperforms larger VMamba-B and ConvNeXt-B backbones on ImageNet-1K.
• On 2048×1024 Cityscapes images, SegMAN-T achieves roughly three times EDAFormer-T’s frame rate while improving mIoU by 1.6 percentage points.

TL;DR - This paper proves that prediction-based safeguards such as accuracy, calibration, and conformal coverage cannot by themselves certify trustworthy AI. It proposes a “competence envelope” that combines prediction and explanation certification to expose otherwise invisible failures.

• A reliable model and a compromised model can satisfy identical prediction-side certificates while differing arbitrarily in explanation fidelity and deployment behavior.
• Detecting this separation requires evidence about the model’s decision mechanism, not merely its outputs.
• The competence envelope provides a deployable criterion integrating prediction performance with explanation fidelity.
• Experiments across multiple datasets and model classes reveal failure modes missed by prediction-only certification.