2026-08-26

TL;DR - StarHarness evolves environment-specific agent scaffolding—such as prompts, tools, skills, subagents, and loop settings—without changing model weights. It improves enterprise benchmark performance by 20–35 percentage points and generalizes across held-out tasks and model families.

• Uses stratified task sampling based on baseline failures, with separate search, selection, and held-out evaluation sets.
• Achieves full-benchmark gains after only 4–12 accepted harness changes per environment.
• Transfers without re-evolution across GPT and Qwen model families.
• Improvements stem from repaired interfaces, encoded environment conventions, and operational knowledge that reduces false diagnoses and shortens some trajectories.

TL;DR - This paper identifies a “handoff tax” when one coding-agent model inherits another model’s trajectory. Escalating from a weaker model to a stronger one recovers less than half the quality gap at substantial added cost, while downshifting offers a better cost-quality trade-off.

• Experiments span Claude and GPT model pairs, varying handoff direction, timing, and inherited context.
• Transfer interfaces include full trajectories, compacted trajectories, and no trajectory beyond the preserved repository state.
• Escalation improves when less of the weaker model’s trajectory is retained.
• Downshifting behaves oppositely: removing the stronger model’s trajectory reduces quality.

TL;DR - Recuris is a recursive memory architecture that helps long-horizon agents track task progress, select relevant skills, and improve those skills through validation-gated updates. It increased task success across nearly all completed model-benchmark evaluations, with larger gains on longer tasks.

• Working Memory represents current progress and guides skill retrieval from Experiential Memory instead of relying on the full interaction history.
• Execution evidence localizes failures to memory components, enabling a fixed Meta-Agent to make bounded, validated updates to Skill Memory.
• Recuris improved 35 of 37 completed model-benchmark pairs across four benchmarks and ten models.
• Reported gains include +17.8 points for GPT-5.6 Sol and +15.6 for Claude Opus 5 on tau-bench, with improvements reaching +32.2 points on the longest tasks.

TL;DR - A multi-institution research survey proposes Graph Engineering as a foundation for next-generation LLM agent systems, explicitly representing tasks, agent coordination, and runtime state as evolving graphs. This could help complex multi-agent systems become more scalable, auditable, and recoverable.

• Task graphs encode goal decomposition, dependencies, parallel branches, executable workflows, and validation steps.
• Agent graphs model capabilities, permissions, delegation, communication, resource access, and human oversight.
• Runtime-state graphs track versioned changes and dependencies to support diagnosis, replay, rollback, and recovery from failures.
• Open challenges include unified capability graphs, validated self-evolution, graph-native agent operating systems, and privacy-preserving governance.

TL;DR - CAFE jointly improves an agentic search policy and its feedback critic by training them in alternating, coupled roles. Across seven in-domain and six out-of-domain benchmarks, this approach outperformed evaluated RL-based search agents on average and reduced answer-level hallucinations.

• A shared-parameter model alternates between acting as the search agent and providing corrective, in-trajectory feedback.
• Online RL rewards useful feedback requests using a call-versus-skip success gap and reshapes token-level advantages around feedback.
• Offline preference optimization trains feedback from matched successful and unsuccessful rollout trajectories.
• Ablations found that improving only the agent or critic eventually plateaued, while alternating updates continued to improve performance.

TL;DR - PeakBench evaluates whether LLM agents can correctly parallelize multi-tool workflows while respecting resource limits. It matters because sound dependency planning alone does not ensure safe, low-latency execution.

• Provides executable workflows with execution-grounded dependencies and measured resource profiles.
• Separately evaluates logical dependency planning and physical resource-constrained scheduling.
• Finds that strong planners can still execute inefficiently or trigger avoidable resource overflows.
• Shows that exposing resource information can reduce overflows and improve utilization.

TL;DR - SkillForge is a reinforcement-learning framework that lets LLM agents accumulate reusable skills while continuously verifying and refining them through environment interaction. It improves over append-only skill banks by maintaining skill quality as agents learn across episodes.

• Makes skill invocation explicit so RL jointly optimizes environment actions and decisions about when to use skills.
• Uses interaction evidence to verify and refine stored skills rather than assuming they remain effective.
• Supports multiple skill-induction pathways, enabling the skill bank to grow while controlling quality.
• Consistently outperforms SkillRL in experiments on ALFWorld, WebShop, and AppWorld.

TL;DR - oFM is a multimodal foundation model trained on longitudinal records from more than one million oncology patients, integrating clinical history, genomics, transcriptomics, and pathology images. Its frozen patient embeddings substantially improved prognostic and treatment-benefit predictions over expert-curated baseline features.

• The model encodes daily clinical and molecular episodes, combines them with H&E pathology images, and integrates observations over time into patient-state embeddings.
• Training used patient-level partitions from a real-world cohort of 1.67 million cancer patients.
• oFM improved overall-survival AUC from 0.563 to 0.774 and also improved treatment-response and progression-free-survival prediction.
• Across 11 comparative-treatment cohorts, its embeddings achieved three-fold higher pooled, scale-normalized treatment-benefit AUTOC and improved benefit ranking in 9 cohorts.

TL;DR - B-MIM is a self-supervised pretraining method that reduces global semantic alignment to make 3D CT encoders more sensitive to fine anatomical details. It improves cross-dataset segmentation of intricate structures while requiring only partial parameter updates during downstream training.

• Modifies the iBOT objective to prioritize local patch reconstruction and high-frequency morphology.
• Pretrains a 3D Swin Transformer on 9,955 abdominal CT studies curated from 17 public sources.
• Improves topological fidelity (clDice) for liver vessel segmentation across datasets.
• Achieves competitive tumor-segmentation Dice scores versus fully fine-tuned baselines while updating fewer parameters.

TL;DR - A Nature study shows that the Rb tumor suppressor has a dual role in HR+ breast cancer: CDK4/6 inhibition activates Rb to suppress cell-cycle genes but also drives estrogen-responsive, pro-proliferative transcription. This mechanism explains both the benefit of combining CDK4/6 inhibitors with anti-estrogen therapy and reduced efficacy in ESR1-mutant tumors.

• CUT&RUN mapping revealed that activated, hypophosphorylated Rb relocates to promoters and enhancers, joining ER-enriched 3D transcriptional hubs that activate genes such as CCND1.
• CRISPR disruption of Rb-bound enhancers and RB1 knockout eliminated drug-induced estrogen-response gene activation, supporting a direct regulatory role.
• Rb interacts with ER and KDM5A, counteracting KDM5A-mediated repression and expanding H3K4me3 across ER target genes.
• Fulvestrant blocked this proliferative feedback in endocrine-sensitive models; in ESR1-mutant models, degrading mutant ER restored stronger cell-cycle suppression.

TL;DR - A Nature Cell Biology study introduces ctEV-CLOCK, a method for time-resolved isolation of circulating tumor-derived extracellular vesicles, and shows that their secretion and tumor-promoting activity follow circadian rhythms. Synchronizing targeted therapies with peak vesicle-protein abundance substantially improved efficacy in mouse tumor models.

• ctEV-CLOCK combines metabolic glycan labeling with tumor-marker-targeted aptamer click chemistry to isolate newly secreted ctEVs at 12-hour resolution.
• Across several mouse tumor models, ctEV abundance and pro-migratory, pro-invasive activity peaked during the nocturnal active phase; host-clock disruption largely abolished this rhythm.
• Nighttime ctEVs were enriched in actionable proteins, including HPSE and CD47, linking vesicle rhythms to angiogenesis, invasion, and immune evasion.
• Administering an HPSE inhibitor or anti-CD47 antibody near the corresponding ctEV protein peaks improved antitumor responses; suppressing extracellular-vesicle release eliminated the time-of-day advantage.

TL;DR - A Protein & Cell study established a method to isolate human skeletal-muscle lipid droplets and mapped their proteome, revealing features that may explain why endurance athletes remain insulin-sensitive despite high muscular lipid stores. The findings offer molecular targets for studying diabetes and other metabolic diseases.

• PLIN4, rather than PLIN2 or PLIN5, was the most enriched perilipin coating protein on athletes’ skeletal-muscle lipid droplets.
• The study identified lipid droplet–anchored mitochondria enriched in respiratory-chain complex IV subunits COX6B1 and COX7C, suggesting enhanced fatty-acid oxidation.
• TPD52L2 was newly identified on lipid droplets; knockdown reduced triglycerides and droplet size and number, while overexpression promoted cell proliferation.
• Researchers plan to compare these “healthy” athlete lipid droplets with those from people with diabetes to investigate insulin resistance.

TL;DR - This research roundup covers new computational and single-cell methods for biomarker discovery, high-dimensional visualization, RNA-velocity quantification, and disease profiling. Together, they improve the reliability and clinical relevance of analyses across complex cellular datasets.

• Machine-learning models can identify multigene disease biomarkers from scRNA-seq data, but require patient-level validation, biological interpretation, and standardized workflows.
• Bonsai represents high-dimensional single-cell relationships as trees, preserving distances and developmental hierarchies more faithfully than conventional 2D embeddings such as UMAP or t-SNE.
• The tidesurf tool corrects strand-orientation errors in 10x 5′ RNA-velocity analysis that can misassign reads and reverse inferred cell-state transitions.
• Large multi-omic atlases—including a four-million-cell pancreas map—reveal disease-associated cell states, regulatory programs, and potential cellular plasticity relevant to diabetes, regeneration, and cancer.

TL;DR - Researchers used AlphaFold3-guided engineering to create Cas9d Ultra, a compact and highly specific genome-editing system suitable for single-AAV delivery. Its derived base editors achieved in vivo Pcsk9 editing and lowered LDL cholesterol in mice, supporting potential single-vector gene therapies.

• Cas9d Ultra combines a shortened, stabilized gRNA scaffold with protein mutations that strengthen interactions with gRNA and target DNA.
• The engineered system exceeded 60% cutting efficiency in mammalian cells—a 3.1-fold improvement over wild-type Cas9d—and showed high specificity in GUIDE-seq testing.
• Cas9d-derived adenine and cytosine base editors enabled embryo editing, including Tyr knockout and generation of albino mice.
• Single-AAV delivery produced 3.5%–11.9% liver editing with 9dABE and an average 15.7% with 9dCBE, reaching 28.9% and significantly reducing serum LDL-C.

TL;DR - OPDVR combines on-policy distillation’s dense token-level guidance with verifiable task-level rewards for LLM post-training, without introducing new hyperparameters. It aims to preserve teacher guidance while preventing incorrect trajectories from receiving positively aligned distillation signals.

• Reformulates sampled-token distillation rewards according to trajectory correctness.
• Uses ReLU gating so correct trajectories receive non-negative rewards and incorrect ones receive non-positive rewards.
• Converts sampled-token distillation into an RLVR-compatible objective that can be used with policy-gradient methods such as GRPO.
• Consistently outperforms standard on-policy distillation across six reasoning benchmarks.

TL;DR - TTTIR reframes image restoration as test-time state evolution, dynamically adapting lightweight operators to each degraded image. With 0.787M parameters, it reports state-of-the-art results across 12 low-light enhancement, deraining, raindrop removal, and dehazing benchmarks.

• A progressive state sequence generator builds coarse-to-fine restoration targets by combining spatial features with wavelet-based frequency states.
• A state transition evolution module performs a single-step fast-weight update during inference, creating an instance-specific restoration operator.
• On cross-dataset LOL-v2-Real evaluation, TTTIR reaches 30.78 dB PSNR and 0.9070 SSIM after training only on LOL-v1.
• Its depthwise-convolution inner loop keeps memory and latency growth roughly linear with image resolution; the implementation is open source.

TL;DR - The IJCAI 2026 paper introduces AC²-VLA, an action-context-aware framework that accelerates vision-language-action models by dynamically routing computation based on robot action state. On the SIMPLER benchmark, it reduced FLOPs to 29.4% of the CogACT baseline and achieved 1.79× faster inference without lowering task success.

• A unified learned router coordinates cognition-cache reuse, visual-token pruning, and Transformer-layer skipping according to action context rather than visual complexity alone.
• Self-distillation trains the sparse model to match the dense model’s actions and internal features, avoiding reliance on sparse task-success signals.
• AC²-VLA averaged 76.8% success on four Google Robot Visual Matching tasks, compared with 74.8% for dense CogACT.
• Cache reuse can improve temporal consistency by suppressing frame-to-frame visual noise, though routing, memory movement, and cache lookup overhead limit realized speedups relative to FLOP reductions.

TL;DR - The IJCV paper introduces TEPR, an edge-prompted geometry-learning framework for multi-view 3D reconstruction from thermal imagery. It improves reconstruction in darkness, smoke, and low-light environments where RGB methods struggle and thermal images lack reliable texture.

• TEPR combines thermal appearance with learned edge prompts, modality-adaptive layers, and top-k sparse attention to improve cross-view matching and suppress noisy associations.
• Edge-enhanced prediction uses global multimodal features for camera parameters and dense edge features for depth boundaries, while hierarchical depth alignment enforces local-to-multiview geometric consistency.
• The authors built TI-FRD with 38,840 thermal frames across 30 scenes, including daytime, nighttime, and smoke sequences with multimodal sensor data and ground truth.
• Against VGGT on TI-FRD, TEPR reduced average reconstruction accuracy and completeness errors by about 38% and 33%, respectively, and trajectory error by about 44%, while running at 19.5 FPS.

TL;DR - A Molecular Cell study identifies DDIAS as the ssDNA-binding effector of the CIP2A–TOPBP1 complex, revealing a DNA synthesis-dependent pathway that repairs single-stranded DNA during mitosis. This mechanism helps preserve chromosome integrity and influences PARP-inhibitor sensitivity in BRCA1/2-deficient cells.

• DDIAS binds phosphorylated TOPBP1 through its BRCT7/8 domains and is recruited to mitotic DNA-damage sites.
• DDIAS’s ssDNA-binding activity is required for CIP2A–TOPBP1 function and mitotic genome stability.
• ssDNA accumulated in BRCA1/2-deficient cells during G2—especially after PARP inhibition—is largely removed by prometaphase through a DDIAS- and DNA synthesis-dependent process.
• The findings establish the CIP2A–TOPBP1–DDIAS axis as a mitotic ssDNA-repair system distinct from canonical DNA-damage responses.

TL;DR - UniGOOD is a unified framework for graph out-of-distribution generalization and detection under simultaneous covariate and semantic shifts. It matters because existing graph-learning methods typically address these two real-world distribution shifts separately.

• UniGOOD jointly optimizes a variational invariant-subgraph generator, cross-subgraph spectral contrastive learning, and a three-population invariance regularizer.
• Its contrastive objective clusters positive pairs to support generalization while separating negative pairs to improve semantic-shift detection.
• The framework uses invariant representations for label prediction and KNN distance for parameter-free out-of-distribution detection.
• It achieved the best reported results across the evaluated synthetic and real-world datasets, including a 9.53% relative AUROC gain on GD-Tox21-SIDER and a 9.21% relative FPR reduction on GD-HIV-ZINC.

TL;DR - WorldEcho reveals that robotic world models often fail to follow valid off-expert actions, limiting their reliability as policy-learning simulators. The proposed WorldSync training framework improves action-conditioned generation and supports more successful iterative policy improvement.

• WorldEcho evaluates action following beyond expert demonstrations using visual integrity and SE(3) trajectory alignment.
• Existing models handle expert actions reasonably but may ignore diverse off-expert commands or generate visually invalid rollouts.
• WorldSync targets distributional coverage, grounding video representations in robot dynamics, and alignment of predicted intervention effects with real futures.
• Experiments on RoboTwin and real robots show improved diagnostic metrics and higher policy success rates.