2026-09-12

TL;DR - RAG-Safety-Bench evaluates how retrieval-augmented generation affects LLM safety by isolating retrieval context from retriever quality. Results across five open-source LLMs indicate that existing safety guardrails may not reliably carry over to RAG deployments.

• Tests four settings: no RAG, oracle harmful-answer documents, related documents without answers, and random safe documents.
• Finds an inverse relationship between benign capability and unsafe capability.
• Shows that baseline model safeguards do not guarantee safe outputs after retrieval.
• Finds model-specific evidence that even benign retrieved documents can increase unsafe generation.

TL;DR - 2AM separates memory from robotic control: a multimodal agent retains interaction history and steers an episodically stateless, RGB-only vision-language-action model through language and optional 2D spatial hints. On LIBERO-Mem, it achieves 76.3% average completion, suggesting that richer agent-policy interfaces can enable long-horizon manipulation without embedding memory in the action model or relying on depth and geometric planning.

• The multimodal agent is the sole holder of task memory, while one action model executes all task-relevant motion.
• The agent translates history into subtask instructions plus optional 2D grasp, placement, and movement hints at multiple time scales.
• Training uses structured hint labels, condition dropout, spatial noise, and temporal jitter to make execution robust to imperfect guidance.
• 2AM reports 76.3% average completion, 63.0% relaxed success, and 11.8% strict success, versus 14.8% completion for the strongest reported baseline.

TL;DR - GRIPNet is a real-time CT detector that incorporates the Gaussian-like radial intensity profile of pulmonary nodules into its architecture. It substantially improves small-nodule detection and localization across three public benchmarks, potentially strengthening early lung-cancer screening.

• Analysis of 18,218 annotated lesions found that nodule intensity typically peaks centrally and decays radially, with mean R² above 0.86 across datasets and size groups.
• Pinwheel convolutions model radial gradients, while a dual-frequency module separates boundary details from broader structural context.
• Dilated masked attention captures the spatial extent of intensity decay, and an adaptive loss emphasizes samples based on conspicuity.
• GRIPNet achieved mAP@0.5 of 95.3% on KanserSet, 91.6% on LUNA16, and 97.9% on Lung-PET-CT-Dx, with improved high-IoU localization at real-time speed.

TL;DR - This study frames cross-lingual clinical annotation projection as constrained text generation, using LLMs to insert entity tags without altering source documents. The approach outperformed prior projection methods across six languages while enabling deterministic validation and character-offset reconstruction.

• GLM 5.2 achieved a mean strict-match F1 of 0.9201 across 18 language–entity combinations; locally deployable Gemma4:31B reached 0.9133.
• The best LLM configuration surpassed the previous state of the art in all 18 settings, improving strict F1 by 0.0564–0.1512.
• The workflow projected Spanish Disease, Symptom, and Procedure annotations and produced 55,416 grounded mentions with reconstructed offsets.
• Local inference and deterministic validation could reduce expert effort and costs when building clinical NLP datasets for lower-resource languages.

TL;DR - Brain-PACE is a Siamese deep-learning framework that directly estimates structural brain-age acceleration from paired T1-weighted MRI. Its stronger associations with cognitive impairment and regional tau burden suggest it may provide a useful longitudinal imaging marker for early neurodegeneration.

• Brain-PACE detected accelerated ageing in 42.6% of participants with mild cognitive impairment.
• Faster estimated ageing correlated with worse functional and cognitive scores, including FAQ ($r=0.35$), ADAS13 ($r=0.30$), and CDR-SB ($r=0.32$).
• It was associated with tau burden in the posterior cingulate ($r=0.59$), precuneus ($r=0.47$), and entorhinal cortex ($r=0.37$).
• The framework extends LILAC with spatial attention, soft label distributions, and a Cramér-distance objective to reduce bias and provide predictive uncertainty.

TL;DR - ActMap estimates an LLM answer’s correctness from a single generation by compactly encoding hidden-state activity across every layer and generated token. It enables low-overhead uncertainty scoring for abstention, routing, and selective verification.

• Compresses generation-time activations into a fixed 12 × 32 × 128 tensor occupying 96 KiB, with no measurable capture overhead.
• A lightweight classifier predicts correctness in under a millisecond; comparable MLP results suggest the representation, rather than classifier architecture, drives performance.
• Across QA, math, and summarization factuality tasks with three instruction-tuned 7–8B models, ActMap outperformed sampling, token-probability, attention, and embedding baselines.
• It matched the mean AUROC of ACT-ViT while using activation tensors 67× smaller and achieved lower calibration error on 10 of 12 evaluated pairs.

TL;DR - Vidu S2 introduces models for real-time 720p digital-character generation and live video editing, with an exploration of spatial video generation. It advances interactive generative video by allowing references and visual elements to be changed during a stream.

• Vidu S2-Avatar supports dynamic references, stronger instruction following, and real-time 720p generation.
• Vidu S2-Editing enables live style rendering and replacement of clothing, characters, or backgrounds.
• The authors explore real-time spatial video generation for both avatar generation and video editing.
• Reported experiments show Vidu S2 outperforming the evaluated baselines.

TL;DR - VWG-Bench evaluates whether video generators can follow symbolic rules, physical constraints, and intentional goals rather than merely produce visually convincing footage. It exposes persistent reasoning failures and introduces Vid-PRE, a model-agnostic prompt rewriter that improves constraint adherence without modifying generators.

• VWG-Bench covers 9 reasoning dimensions and 38 fine-grained tasks.
• A three-level VLM-as-Judge protocol separately assesses visual fluency, rule adherence, and goal realization.
• Leading models render well but struggle with logic-heavy and rule-constrained tasks.
• Vid-PRE uses reinforcement learning with text-only rewards to produce concise, constraint-aware prompts and improves reasoning across multiple generators.

TL;DR - This paper characterizes when external KV caching improves vLLM inference and introduces py-kvcache, an NVMe-backed connector that overlaps asynchronous disk reads with computation. It shows that caching should be enabled selectively based on workload, hardware, and prefix length rather than raw storage bandwidth alone.

• At 80K tokens, py-kvcache loads KV states from disk 2.0× faster than LMCache; scheduler-aware preloading contributes a 1.34× speedup.
• With GPU, CPU, and disk caching enabled, it is 1.23× faster than LMCache and within roughly 4% of native vLLM KV Offload.
• Performance depends on transfer granularity, staging-memory use, and scheduling, with benefits extending to irregular prefix chains and multi-turn workloads.
• Trace replays improved time to first token on a weaker GPU, but not on an H100 where requests were often below the caching break-even point and GPU memory retained enough prefixes.

TL;DR - Kashin-DCT is a language-model weight quantization method that replaces costly dense random transforms with sign-randomized DCTs, cutting each iteration from $\mathcal{O}(N^2)$ to $\mathcal{O}(N \log N)$. It enables efficient 4-bit-per-channel compression while improving numerical stability on configurations where competing methods fail.

• A greedy alternating-update algorithm produces the four-peak distributions needed to cluster each factor into stable 2-bit codes.
• Closed-form cluster-center initialization eliminates the multi-restart k-means bottleneck used by earlier approaches.
• Combined with OPTQ-style error compensation and QuIP-style incoherence preprocessing, the JAX implementation is competitive on OPT, Llama-2, and Pythia models with favorable wall-clock scaling.
• The bounded-$\ell_\infty$ factorization remains near FP16 perplexity in reported Pythia-6.9B and Mistral-7B stress tests where QuIP variants diverge or encounter NaNs.

TL;DR - This paper proposes the “convention gap,” a metric for quantifying how much cooperative success depends on implicit communication beyond literal messages. Across roughly 101,000 Hanabi actions, it better distinguished human-like convention use than game score and may help predict AI effectiveness with human partners.

• The convention gap compares failure probability predicted from literal communication with observed failure rates; Hanabi’s deterministic hint rules make it exactly computable.
• The gap was +26.2 percentage points for human-human pairs, −0.7 for AI-AI pairs, and +16.4 for human-AI pairs.
• Human implicit coordination was strongest for cards receiving no hints, where the human-pair gap reached approximately +46 points.
• In human-AI games, partners eliciting larger convention gaps produced fewer human failures despite providing similar literal information.