2026-08-28

TL;DR - WikiSkill is a framework that consolidates agents’ execution experience into a persistent wiki used to evolve reusable skills. It improves agent performance and enables skills to transfer across models and model families.

• Separates raw execution histories, accumulated knowledge, and executable skills while continuously updating the shared wiki.
• Consistently outperforms prior skill-evolution methods across diverse benchmarks and models, and beats no-skill baselines in most tested settings.
• Evolved skills complement model scaling: larger models generally gain more, while smaller skill-equipped models can outperform substantially larger models without skills.
• Ablations identify persistent knowledge accumulation as critical; skills can also transfer effectively and sometimes outperform skills evolved by the target model itself.

TL;DR - RedEvoAgent is a black-box red-teaming agent that converts prior jailbreak trajectories into compact, interpretable attack skills that evolve with experience. It aims to expose dangerous tool use and persistent state changes in deployed LLM agents more efficiently and reliably.

• Profiles tool effectiveness and uses Deciding-Tool Attribution to determine which tools deserve credit for successful attacks.
• Applies a validation ratchet that retains skill updates only when they improve validation performance.
• Avoids the context overhead, retrieval bias, and limited interpretability associated with reusing full attack trajectories.
• Outperforms fixed and agentic baselines across multiple benchmarks, models, and execution harnesses, while improving tool efficiency and transferability.

TL;DR - A surgically implanted, pulsating muscle graft reportedly mimicked some benefits of exercise in mice. The animals developed greater muscle mass and strength as well as increased bone density, suggesting potential therapeutic value for patients unable to exercise.

• The study used a muscle graft designed to contract or pulsate after implantation.
• Implanted mice showed improvements in muscle mass and physical strength.
• Bone density also increased, indicating effects beyond the grafted muscle.
• The provided summary does not specify the graft’s mechanism, effect sizes, durability, or relevance to humans.

TL;DR - A large study reports an association between receiving the recombinant shingles vaccine and a lower risk of cardiovascular problems. The finding adds to evidence of a possible heart-health benefit, but the provided summary does not establish causality.

• The study examined the recombinant vaccine used to prevent shingles, a viral condition.
• Vaccinated people had a lower observed risk of cardiovascular problems.
• The result suggests vaccination could have benefits beyond preventing shingles.
• Effect sizes, study design, and potential mechanisms are not included in the provided content.

TL;DR - Rankovic and colleagues present a method for training large language models as uncertainty-calibrated optimizers for experimental molecular discovery. Accounting for uncertainty in the underlying data could make model-guided design decisions more reliable.

• Targets a key limitation of language models in molecular design: poorly calibrated uncertainty.
• Incorporates data uncertainty directly into language-model training.
• Frames language models as optimizers for selecting or proposing experimental discoveries.
• The provided summary does not specify benchmarks, molecular tasks, or quantitative results.

TL;DR - This Nature Methods Comment describes early efforts to build bioAFM-DB, a high-quality databank for biological atomic force microscopy data. Applying FAIR data principles could improve data reuse, comparison, and computational analysis across bioAFM research.

• The proposed resource focuses on organizing and preserving biological AFM datasets.
• The authors assess the initiative’s current state and identify challenges in establishing the databank.
• The Comment outlines potential courses of action rather than reporting experimental results.
• Details on specific standards, infrastructure, or benchmark results are not provided in the excerpt.

TL;DR - This is an author correction to the Nature Methods article on OrthoFinder, a method for scalable phylogenetic orthology inference. The provided content does not specify what was corrected or whether the changes affect the original findings.

• Published online in Nature Methods on 26 August 2026.
• The corrected article concerns improved accuracy and scalability in orthology inference.
• Orthology inference supports comparative genomics by identifying genes related through speciation.
• No technical details, corrected values, or revised results are included in the provided excerpt.

TL;DR - TTPO is a label-free test-time policy optimization method that combines self-distillation on majority-agreeing rollouts with grouped reinforcement learning on disagreeing ones. It matches label-supervised OPSD across five competition-level mathematical reasoning benchmarks while improving Qwen3-1.7B test-time accuracy from 38.0% to 45.2%.

• Uses an asymmetric objective designed to remain useful even when majority-vote pseudo-labels are incorrect.
• Distills agreeing rollouts via on-policy self-distillation and penalizes disagreeing rollouts through grouped reinforcement learning.
• Applies token-level selection to down-weight converged positions and penalize only confident errors.
• Reports strong label-free gains, including 25.2%–36.4% improvements without thinking and cross-task generalization.

TL;DR - CritICL is an inference-time framework that uses structured failure patterns from smaller models as critique-based in-context guidance for stronger models. It improves reasoning with fewer generations and lower token costs than conventional test-time scaling approaches.

• Exploits the observation that failure modes recur across model scales within the same model family.
• CritICL-dynamic predicts input-specific failure modes and retrieves relevant critiques.
• CritICL-static applies a global failure-mode profile for consistent guidance.
• Experiments report gains over standard in-context learning and performance competitive with or better than test-time scaling methods.

TL;DR - Scaling model-generated distillation datasets can amplify subtle teacher-specific traits in students, even when the training examples are off-task and never explicitly mention those traits. This creates a latent behavior-transfer risk that conventional data screening may miss.

• Larger independent off-task datasets made an induced teacher trait more detectable in students compared with matched no-trait controls.
• Scaling either amplified an already favored target trait or shifted behavior from a related alternative toward the intended trait.
• Learned LoRA updates showed a parallel scaling trend, with effects observed across model families, trait types, multi-trait settings, and cross-model transfer.
• The findings motivate trait-aware curation and evaluation of synthetic distillation data, including data that appears benign.

TL;DR - This paper compares three paradigms for consolidating domain-specific RLVR capabilities into one language model: expert task-vector merging, mixed-domain RL, and multi-teacher on-policy distillation. Their average performance is similar, but substantial benchmark-level differences make the best choice dependent on cost, data balance, and capability-preservation priorities.

• Average performance differs by at most 1.4 points across paradigms, while individual benchmark gaps reach 8.6 points.
• Cross-domain relationships reflected in task-vector geometry help explain domain-level performance variation.
• Mix RL is sensitive to domain proportions, MOPD is bounded by teacher performance, and Merge compresses expert updates into a single model update.
• All three improve single-sample accuracy without measurable solution-coverage gains or degradation of held-out capabilities.

TL;DR - This paper decomposes block-drafting rejection into an unavoidable “information floor” caused by missing within-block context and a “model gap” caused by imperfect proposals. The results show that revealing just one preceding token removes most of the information floor, while current drafters still have substantial room for improvement.

• On Qwen3-4B, the all-parallel information floor reaches 0.286 at the final block position, capping optimal per-slot acceptance at 71%.
• Realizing one earlier token removes 86–100% of the floor, indicating that the missing information is highly local.
• Mutual-information analysis independently supports this short-range conditioning effect.
• Proposal quality remains a major bottleneck: the final-slot model gap causes 43–64% of DFlash rejection and 85–92% of DSpark’s oracle-conditioned rejection.

TL;DR - Researchers developed an enzymatic platform for constructing thiophosphate molecular motifs while avoiding large quantities of expensive reagents. The approach could enable more practical enzymatic synthesis cascades for drug development.

• Uses enzymes to construct thiophosphate-containing molecules.
• Reduces reliance on costly reagents required in large amounts.
• Establishes a foundation for multi-step enzymatic synthesis of drug-relevant thiophosphate motifs.

TL;DR - This Perspective proposes a framework for assessing reproducibility in hybrid Earth system models that integrate AI with traditional modelling. It matters because AI can improve prediction while introducing additional barriers to validating and reproducing model results.

• Identifies reproducibility as a growing challenge when AI components are incorporated into Earth system models.
• Introduces a framework for evaluating reproducibility in these hybrid modelling systems.
• Provides practical recommendations for strengthening reproducible research and modelling workflows.
• Focuses on methodological guidance rather than reporting new predictive results.

TL;DR - A pre-registered audit shows that difference-in-differences analyses on bounded rating scales can falsely suggest bias in LLM judges because censoring at scale limits creates spurious interactions. This calls into question preference effects reported without accounting for differential attenuation near rating floors or ceilings.

• The registered learner-profile effect on scaffolding preference was null: +0.085 points (95% BCa CI: −0.167 to +0.353; p = 0.684).
• A nominally significant +0.378 interaction (p = 0.002) was not identifiable as a genuine preference difference.
• A zero-differential-preference construction reproduced 79–85% of that interaction using the observed severity shift and rating-scale floor alone.
• The paper derives the censoring mechanism in closed form and shows its contribution can be estimated from an audit’s own ratings.

TL;DR - Structural studies of a key glycolipid-synthesis enzyme reveal an unexpected reaction mechanism and an intrinsic regulatory “brake” that arose during primate evolution. The findings connect enzyme architecture with both biochemical control and evolutionary history.

• The enzyme produces crucial sugar-carrying lipids.
• Its structures uncover a previously surprising catalytic mechanism.
• A built-in feature limits enzyme activity, functioning as a regulatory brake.
• The available summary attributes this brake to primate evolution but provides no experimental details.