AI Coding Live News

On November 2, Stanford researchers Owen Queen and James Zou published an interactive report, "The shelf life of LLM findings", retesting 50 widely cited claims about LLM behavior on 20 models released from July 2024 to September 2026 by eight developers. Verdict: 22 claims (44%) expired, 17 (34%) weakened, 8 (16%) persist and 3 (6%) strengthened; the median time from publication to expiration for expired claims was 14.7 months. Capability failures such as letter counting, lost-in-the-middle and fabricated citations have largely faded, while preference issues such as sycophancy and judges favoring their own answers persist.

Zoom

ControlNet creator Lvmin Zhang and researchers from Stanford and CUHK unveiled MTEdit (arXiv:2610.23751). To resolve the visual degradation and cumulative error drift inherent in conventional sequential rollout for multi-turn image editing, the team introduced Packed Trajectory Canvas (PTC). MTEdit constructs a landmark dataset of 500K trajectories and 2.69M editing turns (averaging 5.38 turns, up to 8 turns), demonstrating stable instruction following and ambient identity consistency across 8 consecutive edits.

Zoom

Researchers from Stanford University and UC Berkeley introduced Librarian Agents (arXiv:2610.22273). To overcome the bottleneck of reactive RAG methods struggling with complex cross-document reasoning at query time, the framework uses reinforcement learning to train LLMs to perform anticipatory directed curation on messy long-context workspaces into structured synthetic filesystems. On BrowseComp+, it hits 55.86% accuracy while reducing navigation token consumption by 33.8% to 51%.

Zoom

Robotic world models struggle to predict physics when key environmental properties (friction, mass, contact dynamics) cannot be directly seen. Researchers from Stanford, CMU, and Toyota Research Institute introduced T2-z (arXiv:2610.22364, GitHub: T2-z-latent-optimization/T2-z). By freezing the world model and optimizing a compact environment latent z_e at test time from as little as a single interaction trajectory, T2-z halves physical error and raises real-world robot task success from 33.73% to 72.96%.

Zoom

Tackling step-level verification in mathematical reasoning and in-context copying bottlenecks in Transformers, NYU/UCLA/Stanford researchers introduce LeanPRM (arXiv:2610.17697)—using the Lean 4 theorem prover to automatically label 152K steps across 10,634 natural-language proofs with zero human annotation, boosting Qwen3-4B step accuracy on LeanPRM-Bench from 48.2% to 74.7% (vs. 60.3% for PRM800K). Meanwhile, Cornell University's Copy-Paste Language Models (CPLMs, arXiv:2610.19247) add a lightweight Copy-Paste head (~3% compute overhead) that relieves the LM head gradient bottleneck, boosting LAMBADA accuracy by +2.5-4.0 points and allowing a w=128 sliding-window CPLM to match full-attention perplexity.

Zoom

Microsoft Research and Stanford University have unveiled AutoEnvScaling, an automated data flywheel that frames RL environment creation itself as a terminal execution task. A proposer agent builds diverse environments directly from solver rollouts, slashing cost per valid environment by 75% while achieving a 25x higher valid environment yield over prompt-only baselines. Crucially, unifying environment design and solving within a single model enables Recursive Self-Improvement (RSI), yielding +12.2 points on Terminal-Bench 2.1 and doubling scores on Terminal-Bench 4.0.

Zoom

Researchers from Stanford University and the University of Oxford have published an empirical study examining 'no-thinking' behavior in reasoning LLMs. While developers frequently prompt models to disable explicit thinking to reduce latency and token consumption, the authors discover pervasive 'Thinking Inertia': models persist in generating explicit inference traces even under strict negation instructions, revealing an inherent trade-off between answer-only compliance and task accuracy.

Zoom

Inverse graphics—interpreting raw visual inputs into executable generative code representing scene geometry and dynamics—constitutes a foundational cognitive frontier for physical world models and embodied agents. Existing code generation benchmarks remain confined to static 2D/3D geometries or conversational software development, failing to assess an agent's grasp of physical causality. Researchers from Stanford University, MIT, and Johns Hopkins introduce 4DCodeBench (arXiv:2610.03715), the first benchmark evaluating agents on 4D inverse graphics through executable code synthesis. Agents must parse dynamic videos and implement compact abstractions such as physics simulators to reproduce observed continuum behaviors, including fluid dynamics, non-rigid deformation, and structural fractures. Evaluating frontier models reveals a stark capability gap: models with strong static 3D spatial reconstruction fail to generate executable simulations for complex time-evolving dynamics. 4DCodeBench establishes an open-source testbed and dataset to measure how coding agents interpret the physical dynamics of the physical universe.

Zoom

LLM-guided evolutionary algorithms (such as AlphaEvolve) have achieved breakthrough results in computational optimization, including circle packing and systems tuning. However, existing methods optimize strictly over iteration counts rather than fiscal expenditure, burning dozens of dollars per task. Researchers from NUS, Stanford, and HKUST present FrugalEvo (arXiv:2610.03675), a cost-aware program evolution framework maximizing gain per dollar. FrugalEvo implements an asymmetric dual-LLM division of labor: a powerful, higher-cost LLM explores high-level solution concepts, while a cheap, high-speed LLM generates and iteratively refines concrete implementations. Additionally, FrugalEvo engineers prompt harnesses to maximize prefix KV-cache reuse across evolutionary generations. Across 10 mathematical optimization problems and ALE-Bench-Lite suites, FrugalEvo matches or outperforms OpenEvolve, EvoX, and SwarmResearch. On circle packing, FrugalEvo achieves SOTA performance using GLM-5.3/Flash for only $0.55 and GPT-5.6 Terra/Luna for $1.68—slashing costs by up to 99% compared to conventional $50 multi-agent baselines.

Zoom

Reinforcement learning (RL) post-training is widely assumed to instill novel agentic reasoning capabilities into large language models. However, groundbreaking research from Stanford, UW-Madison, and Google DeepMind titled 'Sharpening Tax in Post-Training' (arXiv:2610.01509) challenges this paradigm. The authors demonstrate that RL post-training predominantly sharpens preexisting base-model behaviors, inflating single-shot Pass@1 accuracy at the steep cost of solution coverage (Pass@K) on multi-turn agentic tasks. Under sufficient test-time compute budgets, raw pretrained models wrapped in lightweight harnesses frequently solve a broader variety of challenging tasks than their aligned counterparts. The team formalizes the 'Sharpening Tax' diagnostic metric and introduces Posterior-Tempered Group Sampling (PTGS), a plug-and-play Bayesian temperature adaptation algorithm that minimizes the tax while simultaneously elevating both Pass@1 accuracy and diverse solution coverage.

Zoom

Modern unified multimodal models integrate visual generation and semantic understanding within a single parameter footprint, allowing them to critique their own generations. However, post-training pipelines conventionally depend on larger external teacher models. Researchers from Stanford, the University of Washington, and collaborating institutions (including Yejin Choi, Jure Leskovec, and Li Erran Li) introduce UniEvo-VL. The framework assigns a single model to act simultaneously as student (conditioned on raw prompts) and teacher (conditioned on privileged self-critiques), minimizing per-state divergence across denoising diffusion trajectories sampled on-policy. Built on Qwen-image-2512, UniEvo-VL elevates GenEval scores from 0.747 to 0.808 and GenEval2 Soft-TIFA from 32.97 to 35.53 without external supervision.

Zoom

Coding agents are moving beyond virtual codebases into the physical world of robotics. However, existing benchmarks primarily focus on isolated control policies, overlooking the broader engineering capabilities required of real-world robotics engineers. Researchers from Harvard, MIT, and Stanford introduce RLE-Bench, a comprehensive benchmark and qualifying exam for coding agents as Robot Learning Engineers. Spanning interactive control, policy learning, perception/estimation, and mechanical design, RLE-Bench evaluates agents across heterogeneous artifact synthesis, multimodal feedback reasoning, and hardware-constrained execution, introducing the aggregate RLE Index to benchmark frontier agent engineering capabilities.

Zoom

As autonomous software engineering agents increasingly build upon artifacts generated by peer agents, they exhibit a destructive tendency to reimplement capabilities from scratch rather than reuse existing modules, causing codebase bloat. Researchers from UW-Madison and Stanford introduce LibraryDesignBench, the first two-phase benchmark measuring how effectively agents architect reusable libraries for downstream agents. Across 242 expert-validated programming problems, 15 library tasks, and four languages, agent architects successfully reproduce human-engineered abstractions on 11 of 15 tasks. However, downstream consumer agents systematically underutilize these libraries, reimplementing existing routines. Detailed failure audits reveal that downstream agents reimplement functions not due to missing features, but because agent-written libraries are structurally rigid and difficult to invoke. Equipping agent designers with subagent-testing harnesses dramatically improves API ergonomics and downstream reuse.

Zoom