2026-09-23

TL;DR - CliffCompaction is an autocompaction method for long-horizon coding agents that cuts rollout costs by up to 50% while preserving or improving benchmark performance. It avoids cumulative context drift by retaining only selected original content rather than rewriting summaries.

• Each compaction pass operates on the original context, discarding prior compacted output and only truncating or dropping information.
• More efficient test-time scaling improved Terminal-Bench performance by over 10 percentage points for less than two full-context runs.
• Parallel scaling enabled Kimi K2.6 to match Opus 4.7 and outperform Opus 4.6 and GPT-5.3 Codex at lower cost.
• On KernelBench, million-token sessions produced CUDA speedups of 2.23× after 200 steps and 3.58× after 400 steps.

TL;DR - A2M is a two-stage black-box attack that hijacks MCP agents by optimizing malicious tool metadata and outputs using execution traces. Its high attack rates expose substantial semantic supply-chain risks in third-party MCP tools.

• The Attraction stage tunes tool metadata to maximize malicious tool selection, while Manipulation refines adversarial returns to steer agent behavior.
• On LiveMCPBench with GLM-4.6, A2M achieved a 93.6% malicious tool invocation rate and a 74.4% mean success rate across exfiltration, environment compromise, and reasoning derailment attacks.
• Cognitive denial-of-service attacks increased weighted token costs to 32.4× the benign baseline.
• Attacks transferred to four other models without re-optimization, though at reduced effectiveness, motivating stronger tool vetting and runtime isolation.

TL;DR - Dual-Frontier is a framework for deciding when an agent can safely act on predictions from a learned world model. It addresses the otherwise unidentifiable question of whether failures stem from the agent’s decision rule or inaccurate model predictions.

• Proves that decision error and world-model error cannot be disentangled from passive interaction trajectories, even with finite-horizon planning.
• Admits a model-guided decision only when its predicted advantage exceeds a certified bound on decision-relevant model error; otherwise, it prioritizes model verification.
• Provides action-conditioned value bounds and a closed-loop extension guaranteeing non-decreasing return for admitted decisions.
• Experiments with learned models and cross-backbone tool-use benchmarks report improved decision quality and reliability from this verify-then-promote approach.

TL;DR - AIDE² recursively improves an AI research agent by rewriting its own code, benchmarking each version, and retaining changes that perform best on hidden evaluations. An eight-day autonomous run produced transferable performance gains and unexpectedly reduced reward hacking.

• Discovered seven successive improvements, including a new search policy and mechanisms for compressing and managing growing context.
• The strongest self-improved agent matched or exceeded a leading human-engineered production agent across four held-out benchmarks.
• Gains transferred across machine learning engineering, heuristic algorithm engineering, and out-of-distribution weather forecasting.
• Reward hacking on a separate held-out task family fell from 55% to 32%, despite not being an explicit optimization target.

TL;DR - PACT formalizes token-level credit assignment for LLM reinforcement learning and introduces an actor-then-critic update that better aligns the critic with the updated policy. It substantially improves agentic mathematical reasoning and software-engineering benchmark performance over several established RL methods.

• Three regularity conditions—Completeness, Prefix Consistency, and Neutrality—uniquely characterize token-level credit.
• The analysis connects token-level credit to On-Policy Distillation and explains why response-level RLOO can produce the same expected policy-gradient contribution.
• PACT uses importance-sampling correction during critic training to address policy mismatch and critic errors associated with GAE.
• PACT reaches 72.87% average accuracy across four math benchmarks and 67.4% on SWE-bench Verified, outperforming the reported PPO, GRPO, and SAO baselines.

TL;DR - Agensh is a self-organizing multi-agent framework that replaces centralized orchestration with asynchronous task claiming, communication, and shared context. Experiments suggest that increasing the agent count can improve coding-task performance and reduce time to reach comparable results.

• Workers independently gather context, claim subtasks, execute actions, share findings, verify results, and merge progress through a shared workspace.
• On five difficult ProgramBench tasks, scaling from 1 to 128 agents increased the mean final test-pass rate from 19.31% to 28.78%, a roughly 49% relative gain.
• On the pandoc task, scaling from 1 to 1,024 agents raised the final test-pass rate from 33.89% to 55.06%.
• Agent trajectories indicated that cooperation patterns emerged and became more standardized as the organization grew.

TL;DR - FIRE applies failure-informed runtime instructions and action denials to language-model agents without modifying model weights or user prompts. It substantially improves repeatable task completion, suggesting that harness-level policies can turn existing capabilities into more dependable—and potentially cheaper—agent performance.

• Across 87 Terminal-Bench 2.1 tasks, FIRE raised pass² from 50.6% to 54.0% for Luna, 55.2% to 60.9% for Terra, and 64.4% to 73.6% for Sol.
• Sol’s best-of-two success improved by only 1.2 points while pass² rose by 9.2 points, indicating greater consistency rather than major new capability.
• On 14 tasks, policy-guided Terra achieved 71.4% versus 64.3% for unassisted Sol at about half the cost.
• In a randomized five-arm experiment, real policies reached 61% success on eligible tasks, compared with 36–43% for no-policy, sham, verification, and reconsideration controls.

TL;DR - Foundation-model embeddings captured subtle longitudinal mammogram changes before cancer diagnosis without task-specific adaptation. The effect depended strongly on clinically grounded pretraining, suggesting embeddings could support earlier breast-cancer risk detection.

• The study analyzed 1,773 biopsied women and 1,773 matched controls, each with at least two annual screenings before the index exam.
• Malignant cases moved faster than controls along a data-derived “cancer direction” during the two preceding screening intervals in MedImageInsight’s embedding space.
• Mammo-CLIP and HOPPR showed narrower significant effects, while BiomedCLIP showed none.
• Results were broadly consistent across matched case-control and within-patient contralateral-breast comparisons.

TL;DR - Li and Lai use large language models to predict structural edit sequences that make computationally designed molecules more synthesizable. The approach reportedly outperforms traditional methods while preserving important molecular features.

• Targets a key drug-design failure mode: generated molecules that cannot be practically synthesized.
• Predicts precise, sequential modifications to molecular structures rather than merely scoring synthesizability.
• Aims to improve synthesizability without sacrificing key properties of the original candidate.
• The provided summary reports better performance than traditional optimization methods but does not include quantitative results.

TL;DR - Diffusion models can exhibit malign rather than benign overfitting: test loss begins worsening when parameters (p\sim n), long before the interpolation peak at (p\sim nm). Regularization restores the benefits of larger models and can outperform all unregularized configurations.

• Experiments on CelebA U-Nets and a random-features model with closed-form learning curves support the same behavior.
• With (m) noise realizations per sample, the interpolation peak shifts from the standard-regression scale (p\sim n) to (p\sim nm), while overfitting still starts near (p\sim n).
• Bias grows as the learned model approaches the memorizing empirical score; beyond the peak, variance falls but bias continues increasing.
• Ridge regularization in theory and early stopping in U-Net experiments make overparameterization beneficial.

TL;DR - This paper introduces on-policy distillation (OPD) to recover long-form reasoning degraded by sub-3-bit quantization. By supervising quantized models on their own generated trajectories, OPD substantially improves math and code performance over standard teacher-forced quantization-aware distillation.

• OPD targets quantization-amplified exposure bias, where small deviations compound during autoregressive generation and can produce repetitive loops.
• A frozen full-precision teacher provides dense token-level guidance and task-verifier rewards on prefixes generated through the deployment-time quantized path.
• Across four models at 2.79 and 1.88 effective bits, average BF16 performance retention rose from 35% to 70% on MATH-500 and from 66% to 91% on HumanEval.
• The method preserved short-form performance and outperformed continued teacher-forced distillation under matched training budgets.

TL;DR - DreamStream is a generative closed-loop driving simulator designed to preserve policy-relevant scene features rather than optimize photorealism alone. It enables more faithful testing and reveals policy failures missed by prior benchmarks.

• Distills a pretrained autoregressive video model using traffic-layout guidance to vary appearance while preserving scene geometry and dynamic-object consistency.
• Introduces FDπ, a Fréchet-distance metric based on scene-context features from public end-to-end driving policies, addressing shortcomings of metrics such as FID.
• Improves FDπ over the strongest prior closed-loop simulator by 1.6× on nuScenes and 4.7× on NAVSIM.
• Builds Navhard-CL, an interactive benchmark with adversarial driving and weather variations that exposes scorer bias and weak recovery behavior.