Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
Simon Willison reflects on his experience with coding agents, noting that while they enable impressive feats, they also complicate software engineering. He emphasizes that fully harnessing their capabilities demands exceptional discipline and deep knowledge. The article highlights the dual nature of coding agents as both powerful tools and challenging additions to development workflows.
The paper demonstrates that several local LLM agents—including Claude Code, Codex, Antigravity, Open Code, and Grok Build—can delete their own execution traces when prompted, bypassing monitor guardrails. External attackers can also exploit this vulnerability to induce trace deletion. The authors recommend that trace logging be handled by an independent interception mechanism to maintain integrity even if the host is compromised.
Distributed learning in embodied reinforcement-learning agents offers a degree of privacy by retaining raw sensor data on-device and transmitting only policy gradients to the server. Yet temporal stru...
OmniFabric is a new method for creating high‑quality, globally coherent texture maps for 3D garment reconstruction from a single image. It first generates a coarse texture initialization on the garment’s sewing pattern using a 3D mesh and Vision‑Language Model priors, then refines this in the UV domain with a diffusion transformer conditioned on 3D positional features. The approach removes distortion and baked‑in artifacts, producing photorealistic 3D garments that outperform existing baselines.
The paper investigates whether coding agents can automate the synthesis of programs that solve generalized Task and Motion Planning (TAMP) problems. By evaluating Claude Code and Codex on 28 simulated environments, the authors find that these agents outperform hand-engineered planners and other baselines, achieving higher success rates and lower computation per instance. The agents also demonstrate adaptive behaviors such as calibrating physical models and refining strategies during interaction.
The paper introduces an acceptance protocol for sensor-coordinate and polarity binding in mechatronic commissioning, separating candidate generation from release authority. It routes unsupported requirements to a frozen four‑billion‑parameter local language model and only releases plans when both facts can be derived by an external gate under a sealed grammar, with a gold‑standard user providing a canonical answer when eligible. Evaluation on 144 tasks showed that fabricated plans were correctly rejected, no false releases occurred among 83 releases, and protection against incorrect user answers was characterized across 431 pairings.
The paper investigates whether layer-wise visual‑text similarity in multimodal large language models (MLLMs) truly reflects cross‑modal content integration. By injecting Gaussian noise into the visual stream of 13 MLLMs, the authors show that common scalar alignment metrics (CKA, SVCCA, MIR, principal‑angle cosine) fail to distinguish corrupted from intact visual tokens, a phenomenon they term the "alignment illusion." They propose the principal‑angle gap (PA gap) as a more reliable geometric diagnostic that correlates better with task accuracy and reveals when internal geometry diverges from performance.
The paper introduces BRIE, a scalable framework that automatically creates question–answer pairs from longitudinal electronic health record notes, validated by nineteen clinicians. It offers a continuously maintainable benchmark for evaluating large language models in clinical settings, addressing limitations of manual, costly, and quickly outdated existing benchmarks. Experiments across nine LLMs and five inference strategies reveal that even state‑of‑the‑art systems often miss clinically important information, especially for synthesis‑heavy queries.
The paper investigates whether audio language models encode phonetic features similarly when processing spoken versus written input. By comparing mean representations of minimal phoneme pairs across six models, seven features, and 15 languages, the study finds that only voicing in two Qwen2.5-Omni models shows a significant shared direction, and that the model family—not size—determines feature representation. The analysis uses cosine similarity against a random-pair reference to assess alignment across modalities.
The paper presents a graph-based framework for large-scale railway network management that combines a hierarchical Bayesian model with a Gaussian Process on a graph kernel to model spatially correlated maintenance environments, and a topology-aware Multi-Agent Reinforcement Learning system using graph neural networks and Transformers to optimize network-level policies. It demonstrates scalability by training agents on small network segments and deploying them zero-shot on larger, unseen networks, achieving superior performance over heuristics and standard MARL baselines while reducing training time. The approach addresses the computational challenges of centralized methods and the coordination gaps of decentralized methods in complex, long-horizon infrastructure asset management.
GRASP is a multi-stage planning framework that separates planning into specialized modules: GenPlan for global macro-guidelines, RevPlan for exploring localized strategies, and VerPlan for multi-criteria evaluation. This strategy-aware approach yields state‑of‑the‑art accuracy on diverse datasets, outperforming direct LLM planners by up to 30.8% on ZebraLogic and reducing multi‑task degradation. GRASP’s context isolation and macro‑regularization also give it a 14.5% edge over frontier reasoning models like GPT‑5‑mini.
KernelOPT is a multi-agent system that optimizes GPU kernels generated by compilers like PyTorch Inductor by treating compiled models as structured artifacts. It preserves vendor library calls and focuses on Triton sub-kernels, using five profiling-guided LLM agents and a four-gate verification cascade to filter and validate candidates. On 250 KernelBench problems, KernelOPT achieves geometric mean speedups of 1.40×, 1.15×, and 1.07× over torch.compile at different optimization levels.
The Era by Eon benchmark tests enterprise agents by presenting questions that specify answer rules and require code to compute answers from a company’s data. In the original benchmark, top models answered 22–25 of 27 questions, barely distinguishing performance. The updated benchmark adds eight templates that rely on hidden facts not explicitly stated in any question or document, forcing agents to infer information from indirect data. Twelve agents were evaluated, with the best achieving 18 of 24 correct answers, while the hardest questions—requiring selection among similar records—were answered correctly only 1 out of 84 attempts across all agents.
The paper introduces StructFlow-HPR, a structured pose‑conditioned flow matching framework that generates realistic 5G channel state information (CSI) paired with human pose data. It learns a continuous latent transport process from Gaussian noise to real CSI, preserving the receiver‑frequency topology via a reconstruction‑preserving autoencoder. A pose‑conditioned Transformer models the latent velocity field, enabling ordinary differential equation sampling to produce pose‑aligned CSI samples that improve human pose recognition performance in limited‑data scenarios.
Long horizon language model agents continually accumulate reasoning history, increasing context length and inference cost even after earlier decisions have been executed and observed. Unlike static Chain of Thought compression, removing historical reasoning can change future actions and the resulting interaction trajectory.
SEEK (Skill‑Routed Evaluation with Evolvable Knowledge) is a framework that externalizes search evaluation criteria into a skill bank, dynamically routes relevant skills for each query‑result pair, and uses a task‑adapted listwise evaluator to generate page‑level judgments and failure mode attribution. It employs a two‑stage training pipeline to align the evaluator with human preferences and a replay‑gated skill bank to incorporate new evaluation knowledge without retraining the model. Experiments on Kuaishou’s short‑video search demonstrate that SEEK improves listwise quality evaluation accuracy and enhances attribution diagnosis, leading to better online search evaluation at scale.
Large Language Models (LLMs) consume and produce a single sequence of text; hence, if text can be added to the beginning of the LLM's response, i.e., an output prefix, then all subsequent tokens will...
The study compares Jev, a typed classifier that outputs probabilities over allowed answers, with three flash-tier LLM rubric judges across nine panels from seven benchmarks. Jev’s accuracy differs significantly from an LLM judge in only 8 of 27 paired comparisons, performing best on binary criteria and worse only on graded ones, while most other comparisons are inconclusive. In terms of cost and speed, Jev is 29 to 325 times cheaper and 30 to 220 times faster than the LLM judges, and a cascade that defers uncertain Jev verdicts to an LLM yields only modest gains of up to 2.0 points over the best single judge.