AI Weekly Brief

GPT-6 and Intelligent UI for everyone

GPT‑6 is rolling out globally in ChatGPT with Intelligent UI, delivering faster responses with visuals and interactive experiences you can explore and use directly.

Sharing AI progress in mathematics

OpenAI publishes new results on open problems in mathematics from an internal frontier model and shares Lean proof formalizations and research details on GitHub.

Claude Haiku 5.5

As previously promised , here's Anthropic's new fast, low cost model: Introducing Claude Haiku 5.5 . The previous Haiku, 4.5, was very much showing its age. It came out almost a year ago , and was priced at $1/million input and $5/million output - relatively expensive even back then, and a full 10x the price of OpenAI's GPT-6 Luna , released last month. The new Haiku exactly matches the price of GPT-6 Luna - $0.10/$0.50 - up to 100,000 tokens. Beyond 100,000 tokens the price increases 5x to $0.50/$2.50. Luna itself has a price increase at 272,000 tokens but only to $0.20/$0.75. Haiku 5.5 also uses a new, less generous tokenizer. My Claude Token Counter tool shows that the same long prompt uses around 1.25x as many tokens with Haiku 5.5 compared to Haiku 4.5, so there's a hidden price increase there. If your workloads fit in 100,000 tokens, Haiku is the same price as Luna and reports higher benchmark scores. Above 100,000 tokens, Luna looks like a much better deal. The most recent release of llm-anthropic finally fixed it so I don't need to ship a new version of that plugin for every new model. I tested the new model like this: llm install -U llm-anthropic llm anthropic refresh llm

Sophos cuts threat investigation time by 96% with OpenAI Daybreak

Discover how Sophos uses OpenAI’s Daybreak to cut cyber-threat investigation time by 96% and automate 52% of MDR cases while preserving human oversight.

Asana cuts model costs 76x in browser tests with GPT-6.1 Sol

Using GPT-6 Astra in Codex, Asana made its browser agent 76x cheaper and 5x faster in tests to offer customers more capable models.

How Oracle turns days of work into minutes with ChatGPT and Codex

Across recruiting, engineering, and operations, Oracle turns specialist knowledge into fast, repeatable workflows with ChatGPT Work and Codex.

Pollo AI turns creative ideas into campaigns with OpenAI

With GPT-5.6, GPT-6 Astra, and GPT‑Image‑2.5, Pollo AI helps creators turn bold ideas into detailed images and cinematic video ads.

LegalOn halves Codex costs while maintaining development speed

LegalOn cut estimated daily Codex costs by 65% while maintaining development speed. It matched Astra, Sol, and Luna to tasks and managed budgets strategically.

Disrupting AI-enabled “false front” operations

OpenAI disrupted two AI-enabled influence operations that used false-front journalists and a think tank to spread geopolitical messaging.

Helping teens learn, plan, and shape the future of AI

College Planner is coming to ChatGPT for Teens to help students manage college applications, alongside new flashcards, quizzes, and a teen AI council.

Radisson Hotel Group brings hotel discovery into ChatGPT

Radisson partnered with Accenture to build a ChatGPT plugin using OpenAI technology, helping travelers find, compare, and book hotels while planning their trips.

How Jump Trading is scaling quant research with ChatGPT

Jump Trading uses OpenAI to expand quantitative research. See how longer-running AI workflows combine multiple data sources with human review.

Advancing computer use with Ironclad

Learn how OpenAI and Ironclad are training and evaluating AI agents on complex contracting workflows to advance computer use for professional work.

Atlassian and OpenAI expand partnership to turn enterprise knowledge into action

Atlassian and OpenAI are expanding their partnership to connect frontier models with enterprise knowledge and help teams plan, build, and deliver work.

Ultrafast is rolling out today for GPT-6.1 Sol in the API, Codex, and ChatGPT Work. Near-Astra intelligence at up to 8x faster speeds than Sol Standard, so you can build as fast as the ideas come.

GPT‑6 with Intelligent UI rolls out globally to Plus, Pro, Business, and Enterprise users today and will expand to Free and Go users starting tomorrow. GPT‑6 in ChatGPT is powered by GPT‑6 Sol for Plus, Pro, Business, and Enterprise tiers, and GPT‑6 Luna for Free and Go tiers. Both are tuned for eve

We’re releasing a broad range of new mathematical results produced by an internal frontier model. We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to

We’re releasing a broad range of new mathematical results produced by an internal frontier model. We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to

SynthID Detector is now available to everyone. 🌐 Check whether online content was generated using @GoogleAI, or with tools from our industry partners – including @OpenAI, @NVIDIA, Kakao and coming soon, @Apple. Try it out → https://synthid.com

Meet EmbeddingGemma 2, our first natively multimodal open model for on-device embeddings. It expands beyond text to unify code, images, audio, and video in a shared space. 🧵

AgentGarten: Code Worlds for Evolving Agents

Interactive virtual worlds allow agents to learn through exploration and interaction. What agents can learn is bounded by the environments they practice in, which must be faithful, with consistent state, rules, and dynamics, and realistic, with observations that follow the real-world visual distributions. Achieving both across diverse worlds remains a bottleneck. We introduce AgentGarten, a framework that couples simulators and game engines with a shared neural renderer to build real-time interactive environments. Its simulation backends maintain persistent world state and execute program-defined interaction rules, while the renderer generates visual observations from structured conditions exported through a common interface. To build the neural renderer, we adapt a pretrained video model to geometry conditions, distill it with our proposed Adversarial Forcing, and optimize inference for real-time interaction. Adversarial Forcing makes history prefilling differentiable through exact replay, so that losses on later predictions update how the renderer encodes prior observations, and adds real-data adversarial supervision to improve its visual quality. In AgentGarten, agents perceive

Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight

Reinforcement learning with verifiable rewards (RLVR) turns agent experience into learning signals primarily through scalar outcome rewards after interaction. For group-relative objectives, however, this signal vanishes when all rollouts receive the same reward, even though their trajectories may reveal useful information about what the task requires and how the agent fails. We ask a complementary question: can hindsight teach an agent what it could have anticipated before acting? We introduce prospective learning, which uses post-hoc experience to supervise foresight predictions from the pre-interaction view, and instantiate it with Self-Retrospection Distillation (SRD). Intuitively, a completed trajectory reveals knowledge that would have been useful and pitfalls that should be avoided; SRD distills this privileged hindsight into trajectory-blind foresight of the same policy. Foresight serves only as a training target and need not be explicitly generated at inference time. Across 10 tool-integrated reasoning and long-horizon agentic tasks, SRD complements RLVR and self-distillation baselines with gains of up to 24.2 pp. Its advantage is especially pronounced when reward contrast

Learn2Play Bench: How Well Do LLM Agents Learn from Experience in Unfamiliar Environments?

Learning from experience is essential for LLM agents to adapt to unfamiliar and dynmaic environments. Evaluating this ability is therefore important for understanding how effectively agents acquire and use new knowledge. Existing benchmarks have sought to evaluate this ability, but they primarily evaluate tasks whose rules are provided in the instructions or already familiar to pretrained models, making it difficult to distinguish learning from interactions from reasoning with existing knowledge. To address this, we introduce Learn2Play Bench, a benchmark of newly designed text-based games, whose rules are novel or counterintuitive, requiring agents to acquire knowledge through interaction rather than rely solely on pretrained knowledge. These games provide reproducible feedback and automatic scoring, enabling controlled evaluation of learning across repeated attempts. We also vary game instances to test whether agents can apply what they have learned to new situations. Therefore, we evaluate how backbone models, self-evolving methods, and agent harnesses affect agents' learning ability, revealing three findings: (1) Experience retention: Retaining complete records of actions and f

From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation

Realistic environment replicas are increasingly valuable for training and evaluating LLM agents, yet the original systems may be inaccessible or impractical to reproduce. We explore agentic language world modeling: rather than rebuilding an executable environment, a world model agent serves as the environment for a task agent and supports faithful and stateful simulation. We instantiate this paradigm with Trace2Env, a learning-free framework for settings where the original system is unavailable but historical interaction traces remain accessible. Trace2Env reconstructs these traces into a reusable environment worldbook containing environment schemas, grounded evidence, and induced behavioral knowledge. At runtime, the world model agent actively consults the worldbook together with persistent episodic state to infer each action's observation and lasting state effects. Across nine environments, Trace2Env improves both next-observation fidelity and long-horizon interaction consistency over conventional prompt-based LWMs. In multi-turn interaction, task agent actions generated against Trace2Env remain valid more often when replayed in the real environment, indicating that its simulated

TokenRouter: Efficient Serving System for Token-Level LLM Routing

Large language model (LLM) routing distributes inference work across different models, advancing the cost-quality Pareto frontier of LLM serving. While coarse-grained routing at the session or query level has been widely adopted in production systems, recent algorithmic work shows that fine-grained token-level routing can yield substantial efficiency and quality gains. However, efficiently serving token-level routed inference poses significant challenges to existing systems. Built on single-LLM assumptions, current systems suffer from severe step desynchronization and frequent batch admission delays under token-level routing, and they also impose high implementation complexity on developers. To address these challenges, we design TokenRouter, an efficient and developer-friendly serving system for token-level routed LLM inference. TokenRouter follows the principle of request-centric programming, model-centric execution: developers describe routing logic from the perspective of a single request, while the runtime launches a subserver for each LLM and dispatches requests asynchronously. Each subserver employs a delayed-batching scheduler, whose optimal hyperparameters are derived from

SuperNav: An Agentic Navigation System for Any Task in Any Scene

General-purpose service robots need navigation systems that can handle diverse human requests in unfamiliar environments, combining task generality with scene generality. Some existing methods fine-tune multimodal large language models (MLLMs) to predict navigation actions, making their behavior dependent on the coverage of navigation training data and potentially limiting generalization to new requests and environments. Our key insight is to let the MLLM focus on interpreting requests, understanding scenes, and making decisions while preserving its general-purpose capabilities and delegating motion execution to navigation tools. To realize this idea, we introduce SuperNav, which equips a pretrained MLLM with a specialized agent harness without navigation-specific fine-tuning of the MLLM. Our harness supports these decisions with Navigation Skills, agent-oriented Tools for physical interaction, and task-progress and context management. A unified visual-point interface connects decision-making to motion by allowing the model to specify destinations directly in images and revise its decisions from execution feedback. Together, these components support sustained navigation across diff

OpenTPU – An open-source AI accelerator, developed by AI

MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement

Reinforcement learning (RL) is the central training paradigm for advancing large foundation models towards self-improvement. This report introduces the MiMo-V2.6 series, an omni-modal family that pushes the frontier of model intelligence by scaling RL compute. Prior to RL, we conduct mid-training on a broad multimodal corpus to provide ample exploration space, and build a solid infrastructure on the pretrained hybrid-SWA architecture to support subsequent scale-up. We scale RL compute along three dimensions: (1) larger batches and higher throughput, with an asynchronous training that consumes 1,568 samples and 2.7-3.7B tokens per step at context lengths of up to 1M; (2) more diverse and complex environments, spanning code, general, visual, and cyber domains under a mixture of agent harnesses; and (3) more grader compute, via groupwise agentic grading that yields more accurate reward signals for long-horizon tasks and steers the model towards shorter, more token-efficient solutions. To keep training stable at scale, we freeze the MoE router and establish a multi-layer defense against reward hacking. We further build infrastructure for mixed-task agentic RL, including a unified traje

Dust: Pretraining Transformers Without Backpropagation

Impactful scheduling for GPU clusters

Introducing Falcon ASR

One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO

Multi-Agent Egocentric World Model with Fine-Grained Embodied Interaction

Egocentric world models predict first-person observations conditioned on an agent's actions, but most focus on a single agent. Real embodied settings often involve multiple agents that act and interact within a shared environment. Existing multi-agent world models rely on coarse actions like locomotion, camera control, or discrete commands, leaving fine-grained embodied interactions underexplored. We formulate multi-agent egocentric world modeling as synchronized ego-stream generation for multiple agents interacting through fine-grained actions in a shared world. This requires cross-view action consistency, shared-environment consistency, and consistent propagation of interaction-induced state updates. We propose Multi-agent Egocentric World Model (ME-World), which jointly denoises multiple ego streams in a shared token sequence, conditions each stream on all agents' target-view poses, and grounds generation with shared environment memory. We train and evaluate on real and synthetic multi-agent data and introduce shared-world consistency metrics for environment, update, and identity consistency. Experiments show ME-World improves shared-world consistency, action control, identity p

OuroWorld: Bringing Any 3D World Alive as Diverse, Endlessly Looping 3D Cinemagraphs

Recent 3D world models generate photorealistic, explorable scenes that remain frozen in time. OuroWorld is a mask-free framework that turns any static 3D Gaussian Splatting scene into a 3D cinemagraph: a dynamic scene with vivid, diverse motion looping seamlessly from any viewpoint. A vision-language model infers plausible dynamics and guides a video model to synthesize a reference video, which we lift and complete into multi-view videos. To learn from this imperfect supervision, we propose Inconsistency-Robust Periodic 4DGS: a Fourier-series deformation field guarantees looping by construction, while a Grounded Drift Field anchored at the reference view absorbs cross-view inconsistency. Unlike prior Eulerian methods limited to fluid-like motion, we capture general deformation, object motion, and illumination change. We introduce a ground-truth-free evaluation covering vividness, naturalness, loop seam coherence, and scene quality. On 39 reconstructed and generated scenes, OuroWorld outperforms all baselines and wins 70.8%-99.0% of user-study comparisons. Project page: https://ouroworld.userwei.com

DreamTrue: Action-Faithful Robot World Model with Counterfactual Post-Training

We present DreamTrue, a multi-view, cross-embodiment robot world model for action-faithful and physically plausible video prediction. Training such a model on existing robot datasets faces two obstacles: imprecise calibration can impair action following, while limited coverage of unsuccessful interactions can bias predictions toward successful outcomes. To improve action following across embodiments, we render action trajectories into image-space conditions and introduce offline geometric calibration to align these conditions with the target videos. To broaden interaction coverage, we introduce counterfactual post-training, modifying recorded action trajectories and generating future videos under a wider range of actions and contact configurations. To provide feedback on these predictions without paired ground-truth futures, we construct a human-annotated video dataset covering robot, object, and interaction defects and use it to train an embodied video reward model. Its scores guide reinforcement-learning post-training toward more physically plausible interaction outcomes. On AgiBot, DreamTrue attains state-of-the-art action following, while reducing the human-assessed interaction

Beyond Spatio-Temporal Priors: A Generalizable Approach for Dense Correspondence Matching

Dense correspondence matching has historically been bounded by simplifying spatio-temporal priors, such as smooth motion and rigid geometry. While effective for classical tasks, these assumptions break down in image editing and reference-guided generation (IEG), where transformations can preserve visual identity while breaking physical continuity. To establish identity-preserving correspondence across such transformations, we introduce FreeMatching, a generalizable framework combining generative and semantic foundation representations with heterogeneous supervision from classical datasets, tracked videos, and synthetic scenes. Teacher-guided iterative refinement further improves correspondence in IEG without dense correspondence annotations. Experimentally, a single FreeMatching model substantially improves correspondence quality on challenging IEG image pairs while retaining competitive performance on classical benchmarks. Furthermore, we demonstrate its utility as a quantitative metric for evaluating identity preservation, with scores that correlate with human judgment. The code is available at https://github.com/luping-liu/FreeMatching.

U-Space: Uncovering When and Why Uncertainty Arises in Language Models

Large language models are informing decisions with ever-higher stakes. As the consequences of their errors grow, a central question becomes harder to ignore: how much can we trust an individual answer? Yet recognizing when to defer remains difficult because language models can present incorrect conclusions with fluent explanations and an authoritative tone. Uncertainty quantification seeks to address this disconnect by estimating the reliability of individual predictions. However, many existing methods require repeated generations or separately trained components, and their scalar estimates do not reveal where uncertainty arises or how it evolves during reasoning. Recent work has also shown that generation length can be strongly associated with uncertainty estimates and correctness, raising the question of how much of an estimator's predictive power comes from uncertainty-specific information rather than output length alone. Mechanistic interpretability offers a way to address these limitations by connecting human-interpretable concepts to intermediate model states. Building on this capability, we introduce the U-Space, a low-dimensional subspace that makes a model's evolving uncer

MC-Sparse: Deconstructing and Closing the Dense-Sparse Attention Gap in Diffusion Transformers

Sparse attention is a primary approach to reducing the latency of diffusion transformers in long-sequence generation tasks, such as video and high-resolution 3D asset generation. However, existing methods can degrade generation quality and fidelity at high sparsity levels. Through controlled oracle comparisons, we trace this degradation to three sources: constraints imposed by token grouping, inaccurate interaction selection, and the attention contributions lost when tokens are discarded. Guided by this analysis, we propose Meta-Cached Sparse Attention (MC-Sparse), a training-free framework that selects individual key-value (KV) tokens while organizing similar queries into tile-aligned groups for efficient GPU execution. MC-Sparse caches metadata comprising query groups, KV indices selected using exact attention probabilities, and residuals between dense and sparse attention outputs, and reuses them across subsequent denoising steps. Across video and 3D generation models, MC-Sparse achieves higher fidelity to dense-attention outputs and larger denoising speedups than existing sparse-attention baselines, without visible quality degradation. Relative to dense attention, it delivers a

TestPrism: Rethinking Test Evaluation Beyond a Single Reference

Large language model (LLM) coding agents have advanced test generation across diverse programming tasks. However, the common practice of evaluating tests against a single reference solution overlooks alternative valid implementations and can overstate test quality. We introduce TestPrism, comprising 300 test tasks from 17 sources and 3000 candidate implementations, evenly split between valid and invalid solutions. Its primary metric, Joint Success Function, requires the generated tests to fail on the initial program state, accept every valid candidate, and reject every invalid candidate. Across fourteen baseline coding agent configurations, Joint Success Function reaches only 28.00%, whereas single reference success reaches 59.67%. Our analysis reveals missed behaviors, unsupported assertions, and faulty test construction. To address these weaknesses, we introduce TestHelix, which combines heterogeneous synthesis of test and repair pairs with peer cross validation and recursive self improvement (RSI). Across two models, TestHelix improves Joint Success Function by 8.67 to 9.00 percentage points over the native harness comparators in the TestHelix evaluation

Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks

Learning to act in unfamiliar environments requires agents to infer how the world works and revise that understanding as new evidence arrives. Yet limited observations can support multiple world models that explain past interactions but predict different outcomes in unseen states. We introduce Memento 3, building on the Memento series to enable frozen LLM agents to continually learn explicit world models through external memory. The agent maintains a natural-language rulebook as persistent semantic memory, recording revisable hypotheses about environment dynamics while leaving unknown aspects underspecified. It compiles this rulebook into executable code for prediction and planning. Through a continual loop of observation, reflection, rule revision, compilation, and verification, the agent uses prediction errors to refine both the rulebook and its code. Updated code is accepted only when the LLM judges it faithful to the rulebook and cell-exact replay reproduces the observed transitions. We investigate this process as a model-based route to recursive self-improvement (RSI): the agent autonomously explores the environment, revises its world model, and uses verified updates to guide

SparseDecoding: Decoding-Aware Pruning for Accurate and Efficient LLM Inference

The memory-bound nature of the decoding stage of large language model (LLM) inference incurs significant latency. Layer-wise training-free network pruning approaches guided by the Hessian have been a prominent solution to this problem, as pruning reduces the number of nonzero parameters read from memory during decoding. Nevertheless, typical methods in this line compute the Hessian using pre-collected natural sequences, whereas the model is fed self-generated tokens during decoding, creating a distribution shift between the two sequences. The Hessian calculated on the natural sequence is different from that calculated on the generated sequence. We observe that this discrepancy causes the activation distribution during generation to deviate from that used for pruning, further hurting the pruned model performance. Moreover, most existing LLM pruning methods that bring actual speedup primarily target the sparse matrix-matrix (SpMM) multiplication, providing limited support for the sparse matrix-vector (SpMV) operations, which dominate decoding. To solve these problems, we introduce SparseDecoding, a principled decoding-aware pruning framework tailored for accurate and efficient LLM de

EmbeddingGemma 2: an open, lightweight multimodal embedding model

We're sharing progress on ChatGPT for Teens, our ChatGPT experience for people under 18, alongside a preview of College Planner, new study tools, and support for college advisers and teen voices. ChatGPT for Teens applies automatically to accounts identified as belonging to someone under 18, with pr

When possible, we want to give people a choice—and be clear about what watermarking can and can’t do. A text watermark helps answer one question: “Was this likely generated by an OpenAI model?” It works by embedding an invisible statistical signal in text as it’s generated. A detector can look for t

Building predictive biology models is one of the most important scientific challenges of our time, but solving this requires new experimental biological data at an unprecedented scale. I’m pleased to announce that @GoogleDeepmind is partnering with @ChanZuckerberg’s Virtual Biology Initiative to cre

SynthID Detector can quickly identify our watermarks in video, audio and images – even after edits, like adding filters. ✂️ You can also use existing AI identification features found in @GoogleChrome, @Google Search, and the @GeminiApp.

LEGO: A Lifting-Free Approach for Exocentric-to-Egocentric Video Generation

Generating an egocentric video from a single exocentric recording is a challenging case of novel view synthesis, as the two cameras share little overlap and much of the target view is unobserved. Current state-of-the-art methods reconstruct the scene explicitly by estimating depth, lifting the video into a point cloud, and re-rendering it from the egocentric camera to condition a video diffusion model. This deterministic mapping assigns each pixel to a single reprojected location, which preserves texture but translates depth errors into misplaced content. We ask what a video diffusion model should receive as its condition and propose a lifting-free answer: a learned view synthesizer, an LVSM-style transformer fine-tuned to render the egocentric view directly without depth, point clouds, or reprojection, resolving cross-view correspondence internally. In contrast, its probabilistic mapping averages each region over candidate source locations according to a learned correspondence distribution, preserving structure while fine texture is averaged away. We argue that this trade-off suits a diffusion generator, whose denoising training excels at restoring detail, so an effective conditio

MIMESIS: Learning User Simulators as Training Environments for Interactive Agents

Training and evaluating interactive language agents typically requires rich user interactions, yet collecting human feedback is expensive and difficult to scale. Simulated users offer a scalable alternative, but they must both resemble real user behavior and provide useful learning experiences for agents. In contrast, most agent-training frameworks rely on off-the-shelf assistant LLMs, whose helpfulness can make them overly cooperative, explicit, and behaviorally homogeneous compared with real users. We introduce MIMESIS, a purpose-built user simulator trained on human conversations with explicit reasoning supervision and 13 realistic behavioral patterns derived from real user interactions. Empirically, our 9B model achieves a SOUL-Index of 65.7, surpassing the strongest frontier model. Compared with Claude-Opus-5, the strongest baseline on RealUserSim and SimulatorArena, MIMESIS improves behavioral fidelity by 13.4 points and reduces Turing distance by 3.6 points, respectively. We then freeze the simulator and train an agent by interacting with the frozen simulator using multi-turn reinforcement learning. Across eight environments, training with MIMESIS yields better agent perform

OneSearch-VL: Unified Multimodal Deep Research Agent for Image and Video

Single-image, multi-image, and video deep research require different visual operations but share a workflow of visual grounding, external retrieval, and fact composition. A key challenge is to preserve the dependencies linking localized visual anchors, entity relations, source-supported facts, and answer-producing operations. We introduce OneSearch-VL, a unified agent centered on the Visually Grounded Evidence Graph (VGEG), which encodes these dependencies as a shared task-level reference for data construction, process supervision, and operation-level evaluation. Our VGEG-based data engine constructs and verifies multi-image and video questions and filters expert trajectories. Using these data, we assemble OneSearch-VL-SFT-110K and OneSearch-VL-RL-10K for SFT and RL, respectively. We further derive the Evidence-aware Visual-Grounded Rubric reward (EVGR) from VGEG annotations to supervise evidence traceability and visual grounding during RL. For fine-grained evaluation, we construct OneSearch-MI-Bench and OneSearch-Video-Bench, organizing questions by the research operations encoded in their VGEGs. Experiments show that OneSearch-VL-8B improves over Qwen3-VL-8B with tool access by 2

Q-Learning with Scalar Adjoint Matching

Flow policies capture rich and diverse action distributions, and fine-tuning them with off-policy RL to improve beyond the demonstrations has drawn growing interest. However, fine-tuning a flow policy against a learned value function is not trivial, because the policy generates its action over many flow steps. Adjoint matching offers a principled way to update the flow model itself by propagating value information from the final action back to each flow step, but it requires a vector--Jacobian product through the policy at every step, a cost that grows with the number of flow steps and the policy size. We observe that the batch-averaged velocity Jacobian of pretrained flow policies concentrates on its diagonal. Motivated by this finding, we derive a closed-form scalar adjoint that scales the value gradient at the final action by the flow time, eliminating the per-step vector--Jacobian products. We further find that controlling the critic's value at policy-generated actions is particularly important under the scalar adjoint. Based on these findings, we propose Q-learning with Scalar Adjoint Matching (SQAM), which combines the scalar adjoint with a value penalty at those actions. SQA

Reasoning-Informed Visual Editing

Large Multi-modality Models (LMMs) have made significant progress in visual understanding and generation, but still face challenges in visual editing, particularly in following complex instructions, preserving appearance consistency, and supporting flexible input formats. To study this gap, we introduce RISEBench, the first benchmark for evaluating Reasoning-Informed viSual Editing (RISE), and extend it to RISEBench++, a more comprehensive and fine-grained benchmark for this emerging task. RISEBench++ extends the taxonomy into a hierarchical scheme spanning six reasoning dimensions: Temporal, Causal, Spatial, Logical, and Counterfactual Reasoning, together with Hybrid Reasoning integrating multiple reasoning types across multi-turn edits. These dimensions are further decomposed into 12 subcategories and 65 fine-grained task types. We expand input formats to include multi-image conditioning and scale the benchmark to 1000 human-annotated test cases, released in English and Chinese. We also improve our evaluation framework, assessing Instruction Reasoning, Appearance Consistency, and Visual Plausibility with human judges and an LMM-as-a-judge approach for more reliable and calibrated

OpenAI “rogue” agent activities found on Wikimedia projects

OpenAI “rogue” agent activities found on Wikimedia projects Given how tempting a target wikis are for rogue agent swarms, it's not a huge surprise that Wikipedia found evidence of that activity once they went looking: The Wikimedia Foundation conducted its own investigation to see whether Wikimedia websites had been similarly affected by AI agents, focusing on those operated by OpenAI. We can confirm that we have discovered some activity by these “rogue” OpenAI agents on Wikimedia platforms. The unauthorized bot activities included edits to our wikis, some unsuccessful attempts to exploit a public note-taking tool we host, and heavy traffic, which are described more below. They found evidence of agents editing sandbox pages, trying to use pieces of infrastructure such as Etherpad to help proxy content from elsewhere, and saw widespread crawling and "hundreds of thousands of data queries" to their Wikidata Query Service. My best guess is that most of this was a similar (or the same) swarm of agents as those that defaced that German wiki while training for research tasks. The Wikipedia sandbox wiki edits appear to have started on May 12th, and the initial test edits to the UseModWiki

llm-openai-decisions 0.1a0

Release: llm-openai-decisions 0.1a0 OpenAI released their new Jev-style Decisions API , as previously announced at last week's DevDay. Since I already have an llm-typesafe plugin for talking to Jev, I had GPT-6 Astra read the new OpenAI API documentation and build an llm-openai-decisions plugin inspired by llm-typesafe . Unlike Jev, the new gpt-6-luna decision model supports image input in addition to text. Both models charge for input it and not for output: OpenAI's is 10 cents per million input tokens, Jev's is 4.2 cents per million. Otherwise the API shape is very similar to Jev, at least conceptually. Jev supports three question types for yes/no, choices, or scores. OpenAI Decisions supports the same three types. Install the plugin like this: llm install llm-openai-decisions Here's an example query against an image attachment: llm -m openai-decisions/gpt-6-luna \ -a https://static.simonwillison.net/static/2025/two-pelicans.jpg \ -s ' Does this image contain any mammals? ' And example output: { "type" : " predicate " , "name" : " evaluation " , "probability" : 0.0 } Consult the README for full details of how to run the other types of questions. Tags: openai , llm , coding-ag

llm-mistral 0.16

Release: llm-mistral 0.16 Adds support for reasoning models, such as the newly released Mistral Large 4 . Tags: llm , mistral , llm-reasoning

A self-learning scientific agent for X-ray diffraction

A central challenge for scientific agents is to turn analytical experience into reusable expertise grounded in physical evidence. Here we introduce Gan Jiang, a self-learning agent for powder X-ray diffraction built on a diffraction-analysis ecosystem we developed: XMatcher, XQueryer, XDecomposer and WPEM. Together, these engines span phase identification, multiphase decomposition and physics-constrained whole-pattern modelling. Gan Jiang converts analytical experience into executable skills by diagnosing failures, revising skill instructions and code, and validating revisions before reuse, without retraining the language model or changing the underlying physical models. Skills selected using development data and frozen before held-out evaluation achieve higher refinement scores than the original expert-designed skills across FullProf, GSAS-II and PyWPEM. The agent resolves strongly overlapping reflections, quantifies a five-phase ancient Egyptian cosmetic, tracks lattice evolution in an operating battery and compares atomic configurations in a disordered oxide catalyst. On DeltaXRDbench, it leads the evaluated methods in single- and multiphase identification across simulated and e

Pumpire: Unified Benchmark for Metric Distance Estimation

We present Pumpire, a unified benchmark for evaluating metric point-pair distance estimation capability of both image- and video-level 3D foundation models, with or without depth priors. In contrast to previous approaches that normally evaluate depth and camera intrinsics separately or evaluate point-clouds with geometric similarity metrics, which cannot directly reflect models' point-to-point distance estimation capability, Pumpire directly assesses point-to-point distances from the reconstructed geometry. To this end, we collect a large-scale and diverse dataset (pumpire-6k) comprising 100 real-world scenes, each annotated with physically measured point-pair distances and containing 64 frames, for a total of 6,400 frames. Building on this dataset, we establish a holistic evaluation protocol that covers both image- and video-level 3D foundation models and enables direct assessment of point-pair distance errors and cross-setting comparison. We conduct extensive experiments across 29 baseline configurations of representative 3D foundation models and provide a comprehensive analysis of the results. By offering this benchmark, we target the more fundamental ability to perceive and est

What Did the Agent Actually Do? Evidence-Grounded Oversight for Long-Horizon Agents

As agents take on long-horizon tasks, users shift from making individual decisions to overseeing autonomous execution. Yet the volume of agent activity and the fragmentation of supporting evidence make it difficult to determine which decisions warrant user verification. We study monitors that identify consequential decisions and locate evidence to help users assess their implications. We introduce AgentMonBench, a software-engineering benchmark comprising three subsets that cover two complementary dimensions: alignment between requirements and behavior, and awareness of consequential autonomous decisions for verification. To support these judgments, we propose the Evidence-Grounded Behavior Graph (EBG), a training-free method that groups source-linked evidence into behaviors and organizes their relationships into a graph. EBG presents task-oriented views of this graph to help monitors interpret behavior in context. Experiments across eight models show that EBG improves decision identification and evidence localization in most settings compared with direct access to the original context. Further experiments show that EBG's evidence-localization gains persist across input scales and

Building advertising for the way people use AI

OpenAI introduces a new visual ad format in ChatGPT and expands measurement tools, attribution partnerships, and brand suitability for advertisers.

Opus 5.5 agents discover two room-temperature magnetic semiconductor candidates

Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses

Training AI agents with reinforcement learning can be challenging because their tools, context, and decision-making are managed by complex frameworks. Agent Lightning connects existing agents to RL training, making it easier to improve them without rebuilding them. The post Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses appeared first on Microsoft Research .

Can a Cloud-Native Harness Make Agents Reliable Beyond the Desktop?

Kubernetes co-creators Craig McLuckie and Joe Beda aim to bring agent harnesses fully into the cloud.

The Ballon d'Or, legal division: On Harvey's Legal Agent Benchmark (LAB), Mistral Large 4 is the #1 oss model.

Gmail. Google Sheets. Slack. Salesforce. i.e. the four horsemen of Monday. On AutomationBench, Mistral Large 4 took on 657 business workflows across simulated workplace apps, all four included, finishing ahead of Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro.

Mistral Large 4 runs general-purpose agents that gather information and produce finished deliverables across complex workflows.

Pinned: Meet Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier mo

Strands Decider 2B: a small, open-source, decision model

The model that didn't exist, so you made it yourself

Multimodal open d1 decision models for the edge

New in Managed Deep Agents 0.9: Reactions API for @SlackHQ Channels. When it comes to building effective user agent experiences, loading states are underrated. Now, you can build loading states and receipts using either heuristics, decision models, or custom logic.

Mold Linker Version 3.0.0 Release – Rewritten in Rust

DuckDB Ducklake

Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s

Port of the TypeScript compiler, checker and lsp to Rust, by LLM

Deno is joining Cloudflare

Deno is joining Cloudflare The Deno team released the first version of celld back in August - their open source implementation of the Durable Objects pattern from Cloudflare Workers. Today, Cloudflare are acquiring Deno outright, with the goal of building on celld to "make workerd self-hosting a first-class supported way to build and run apps using the Workers programming model" (see the Cloudflare blog .) The bad news is that Deno itself will not be maintained by Cloudflare beyond the next year: We will support the Deno runtime for another year with monthly releases containing bug fixes and security updates. After that year we will end our development of the Deno runtime. Deno will remain open source, and we welcome others who want to continue its development. Deno (and Node.js) creator Ryan Dahl explained that decision in a comment on Hacker News: It's a joint decision and I agree with it. I'm most invested in its success and have put the most work into it - and I no longer think it's where I can do the most important work. There are some good ideas in Deno and it's well engineered - but it ultimately is not solving big problems. It has been sucked into the gravity well of node c

ttok 1.0

Release: ttok 1.0 I released ttok 0.4 , ran uv tool upgrade ttok , piped a file into the new version... and realized that it was defaulting to the GPT-4 tokenizer when it should very clearly default to GPT-5/GPT-6 instead! I figured switching the default was a reasonable excuse to finally ship a 1.0. OpenAI haven't actually confirmed that GPT-6 uses the same tokenizer as the GPT-5 family yet - there's an angry issue about it - but I found this commit by William Liu which reports on an experiment he ran confirming that the tokenizers are likely the same: All seven GPT models (5.5, 5.6 Sol/Terra/Luna, 6 Astra/Sol/Luna) report 44,794 tokens and match each other on every one of the 31 fixtures. GPT-6 introduces no input-count change on this corpus. Tags: projects , ai , openai , generative-ai , llms , tokenization

ttok 0.4

Release: ttok 0.4 ttok is my CLI tool for counting tokens, using OpenAI's open source tiktoken library. It hasn't been in updated in a couple of years, but I finally fixed a Click warning, updated CI, and added a --list-models command to list available models. It works with uvx , so you can count tokens in anything like this: cat file.txt | uvx ttok Tags: projects , ai , openai , generative-ai , llms , tokenization

Anthropic can’t reliably control its AI agents. It’s cutting off its internal evals from the live internet instead

Anthropic said it "turned off live internet access" for "all our internal evaluations" until further notice.

An Anthropic AI model sent a false homicide tip to Philadelphia police

Anthropic did not discover this behavior until over two months after its AI submitted the false tip.

Aleph Alpha Kolibri: How the sovereign German LLM works

<a href="https:&#x2F;&#x2F;aleph-alpha.com&#x2F;en&#x2F;blog&#x2F;kolibri-has-landed-a-sovereign-open-weight-model&#x2F;" rel="nofollow">https:&#x2F;&#x2F;aleph-alpha.com&#x2F;en&#x2F;blog&#x2F;kolibri-has-landed-a-soverei...</a>

Malware reverse-engineering: solving an out-of-distribution investigation task. When faced with an unknown binary, Mistral Large 4 reverse-engineers it end-to-end. In this case, it concludes the sample is Cobalt Strike, extracts the IoCs and malware configuration, and writes a report with a YARA rul

Mistral Large 4 against the field: efficiently reasoning over diverse complex challenges. CTF speedrun: 19 challenges, with tool calls and solve times grounded in actual runs. It solved 18 of the 19.

For their Mistral Large 4 release, @MistralAI ran a blind Surge human eval, with our expert software engineers evaluating five frontier models on coding quality. Le Chonk ranked #1 among open-weight models and #2 overall, behind only Opus 5. Unit tests ask: does it work? Human code review asks: woul

HarnessSQL: Harness-Native Training for SQL Agents in Realistic Database Environments

Text-to-SQL models are commonly trained to map questions directly to static queries, whereas real-world database agents operate through stateful, multi-turn interaction with live databases -- inspecting schemas, executing probe queries, diagnosing errors, and revising hypotheses. This creates a critical train-deploy mismatch, as the execution harness that mediates this interaction is introduced only at inference time. To bridge this gap, we propose HarnessSQL, a harness-native post-training framework that preserves the full interaction structure throughout both supervised fine-tuning and reinforcement learning. HarnessSQL builds isolated, executable database environments paired with hidden execution oracles, rolls out teachers directly inside the target SQL harness, and retains only verified trajectories for full-sequence SFT, followed by execution-reward RL. Across Spider 2.0-SQLite, HarnessSQL dramatically boosts the execution accuracy of compact models, raising Qwen3-8B from 15.5% to 45.2% and Qwen3-14B from 22.2% to 54.8%, while transferring effectively to out-of-distribution interactive benchmarks such as BIRD-Interact and LiveSQLBench. Our findings demonstrate that training database agents directly within their execution harness is essential for mastering complex, long-horizon database workflows.

From Reactive Containment to Proactive Assurance: Lessons from OpenAI, Anthropic, and Google Agent Security Incidents

In 2026, cybersecurity evaluations involving OpenAI, Anthropic, and Google agents reached real systems outside their authorized test scope. The paths were different. OpenAI agents exploited research infrastructure, coordinated across runs, and compromised parts of Hugging Face's production environment. Anthropic reported cases in which a misconfigured third-party environment exposed real systems to agents pursuing simulated cyber tasks. In a separately reported evaluation, Google's Gemini accessed three real organizations through an unintended internet route; Google stated that the model stopped in all three instances. Taken together, the cases show why an evaluation cannot rely on an assumed boundary. That boundary must be verified while the agent is operating. This comparative instrumental case study develops a Proactive Agent Security Assurance Cycle (PASAC) and a five-layer Boundary Assurance Stack. The framework combines risk-tiered task design, executable scope contracts, pre-run validation, least-capability access, independent egress enforcement, credential restrictions, cross-run monitoring, automatic stop conditions, and evidence-based reauthorization. A leading-indicator model, nine design propositions, and seven falsifiable hypotheses turn these lessons into a testable research program. Because the public Gemini record is limited to attributed statements and journalism, its detailed causal mechanism remains provisional. The central conclusion is straightforward: pr

Into the Omniverse: How Developers Turn Ideas Into Simulations With Frontier AI Agents

Turning a simulation idea into a working application means assembling assets, connecting physics and rendering, and checking that the scene behaves as intended. Developers are combining frontier AI models with NVIDIA Omniverse libraries to help carry out that work — building applications for exploring scenarios, investigating failures and improving designs. Developers direct AI agents through […]

NVIDIA, Microsoft Kick Off a New Beginning for Windows PCs With RTX Spark and AI Agents

At a Microsoft event in San Francisco on Wednesday, NVIDIA founder and CEO Jensen Huang and Microsoft CEO Satya Nadella outlined how NVIDIA and Microsoft are co-engineering hardware and software for AI agents to run on Windows PCs. NVIDIA was founded because of Windows, Huang said. Now AI agents are coming to Windows. “If you […]

Why Telecom Operators Are Building Their AI Strategy on Open Models

Telecom operators are increasingly building their AI strategies on open models — and the reasons go beyond mere cost.  Open models give telcos the ability to trust, control and customize AI across their most critical workloads — from autonomous networks to customer care.  NVIDIA’s latest State of AI in Telecommunications report reflects this shift, with […]

Docker Agent

The maker of non-text AI model Jev valued at $7.5B just weeks after launch

What has users and large corporations so excited about Jev is TypeSafe’s claim that it works significantly faster and uses far fewer tokens than LLMs.

Getting the most out of Opus 5.5 in Claude and Claude Code

FTL: A new operating system for clouds

<a href="https:&#x2F;&#x2F;github.com&#x2F;nuta&#x2F;ftl" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;nuta&#x2F;ftl</a>

FastBench: Can Streaming VLMs Perceive High-Dynamic Real-World Streams?

Streaming Video Large Language Models (VLMs) enable continuous video understanding, yet existing benchmarks focus on low-dynamic scenarios. Under bounded context budgets, models must balance temporal history, spatial resolution, and temporal granularity; sparse sampling at 1--2 FPS misses fast events. We introduce FastBench to evaluate high-dynamic perception in real-world video streams. Its trajectory-grounded pipeline combines QA generation from high-FPS clips, filtering of questions answerable at 2 FPS, answer verification using SAM3 and CoTracker3 trajectories, and three rounds of human inspection. FastBench contains 306 QA pairs across eight domains, six capabilities, and forward, instant, and backward temporal scopes, with human-annotated evidence intervals. We also present ProactiveFrame, a training-free baseline that adjusts incoming frame rates through text tokens. A dual-tier sliding window retains recent high-FPS observations while downsampling older ones into sparse history. Experiments reveal substantial limitations: the strongest model, Gemini-3.5-Flash, scores only 50.7%. Denser sampling improves Qwen3-VL-8B from 32.9% at 2 FPS to 44.6% at 24 FPS, but gains saturate as history is compressed. ProactiveFrame outperforms sparse uniform sampling by 5.4 and 1.5 percentage points, yet remains well below oracle-guided focusing, showing that current VLMs struggle to determine from the stream alone when finer temporal perception is needed. FastBench provides a testbed f

WOVEN: Weaving Visual World Modeling into Multimodal LLMs

Multimodal large language models (MLLMs) struggle with spatial, embodied, physical, and temporal reasoning. We hypothesize that these failures reflect a shared deficit in visual transition reasoning, and test whether this capability can serve as a shared training primitive, one that different models can learn from different supervision sources and reuse across different tasks, with a systematic training recipe. Existing benchmarks document these deficits separately but do not support controlled comparisons across scenes, actions, and reasoning operations. We therefore introduce WOVEN, a training source and benchmark for visual transition reasoning that organizes transition supervision by scene, action, and reasoning type, using diverse, realistic rollouts from video-pretrained generative models: 36,076 examples across 20 scene types, 5 action types, and 8 reasoning types. We first evaluate 38 frontier MLLMs (e.g., GPT-5.4 and Qwen3-VL-235B-A22B) and find a substantial and systematic deficit: even the strongest models fall far below humans, and the failures recur across model families and persist with scale. We then train MLLMs at multiple scales on WOVEN and find that they learn a shared capability that transfers broadly: training subsets of only about 2,000 items each collectively improve 22 of 26 external benchmarks by up to 27.3 percentage points, and WOVEN data can replace 30-50% of a task's own training data with comparable accuracy. Controlled comparisons further yield

Predicting Alignment Generalization with Value Representations

LLM developers post-train their models to exhibit prosocial values and behavioral traits, which are enumerated in an alignment target. However, while recent post-training developments have yielded models that score highly on alignment evaluations, training models on sets of narrow behaviors still influences their behavior across unseen contexts and environments in unexpected ways. In this paper, we establish the task of alignment generalization prediction, i.e., predicting how fine-tuning a model to follow a given value changes its behavior across a wide range of held-out values. We conduct a large-scale analysis of alignment generalization effects across 66 values found in modern alignment targets, and benchmark representational techniques on the alignment generalization prediction task. We find that representations based on model activations when applying values in context significantly outperform methods based on textual descriptions of the values. Specifically, the best activations-based methods achieve correlations of 0.45 with our generalization matrix, compared with 0.05 from description-based baselines. We then show the applicability of representations that predict alignment generalization toward downstream tasks by using them to measure how similar the values in a multi-value alignment target are, which we find is significantly correlated with model robustness. Finally, we show initial evidence towards a shared, model-independent value space, which we use to develop

ViSkill: Reinforcing VLM Agents with Evolving Visual-Native Skills

Skill-augmented agents improve sample efficiency by distilling successful trajectories into reusable strategies. Yet most existing approaches remain text-centric, linearizing spatial layouts and action-state correspondences into language that loses critical geometric structure. Recent efforts have begun incorporating visual evidence, but construct and update skills separately from policy optimization, leaving their mutual improvement underexplored. We propose ViSkill, a visual-native skill learning framework that encodes successful interactions as composite visual skill cards directly accessible to VLM agents. Retrieved skills guide both inference and reward shaping, while successful trajectories are distilled back into the library, forming a closed feedback loop in which skill accumulation and policy improvement reinforce each other. An optional cold-start mechanism further accelerates early-stage learning. Evaluated on Sokoban, FrozenLake, and PrimitiveSkill, ViSkill achieves an overall success rate of 0.89, rising to 0.91 with cold-start initialization, outperforming all evaluated proprietary and open-source baselines while converging faster than standard PPO. Our code is available at https://github.com/ZJU-REAL/ViSkill.

SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models

Existing spatial reasoning benchmarks mainly test spatial perception: reading off relations already visible in the input. Yet real-world spatial intelligence demands predictive spatial reasoning: constructing a scene from observations, anticipating how an intervention changes it, and reasoning about the unseen outcome. We introduce SpaceCast-Bench, the first benchmark to directly and diagnostically evaluate this capability. Built around an observe-transform-infer framework, its 3,862 questions from 182 real-world scenes span 16 task types at three levels: static perception, local prediction, and global prediction, progressively requiring scene understanding, spatial state updating, and relational inference over unobserved outcomes. Evaluating 21 models exposes a stark gap: the strongest model reaches only 58.0% against 87.2% human performance, while spatially specialized models remain near random chance. Controlled analyses further reveal that bridge views are critical for integrating distributed observations, and that explicit 3D evidence benefits models more reliably than generated outcome images or videos. Fine-tuning on our programmatically generated data lifts Qwen3-VL-4B from 34.0% to 65.7% with macro-average gains across six out-of-domain benchmarks.

Long Text to Predictive Features: LLM-Guided Blockwise Feature Engineering via Executable Program Search

Industrial risk-control systems typically rely on structured-data models for efficient prediction, yet substantial valuable information remains embedded in unstructured long text. Extracting this information through manual feature engineering is labor-intensive, while requiring a large language model (LLM) to process every real-time input may not meet practical deployment requirements. To address this challenge, we propose LLM-BlockFE, an LLM-guided offline feature construction framework that converts long text into executable feature programs, thereby avoiding LLM calls during online inference. LLM-BlockFE constructs feature programs by incrementally appending immutable code blocks and evaluates candidate features using a downstream model. To address the tendency of conventional greedy search to become trapped in suboptimal solutions, our method introduces a block-level rollback mechanism based on depth-calibrated credit allocation and advances multiple independent search trajectories in an interleaved manner, reducing redundant exploration by sharing fixed descriptions of each trajectory's exploration direction. After the search, the resulting programs are frozen and deployed to extract structured features for downstream prediction models. Across two public and two private datasets, LLM-BlockFE achieves absolute AUC improvements of 0.0069 to 0.0358 over the strongest baseline on each dataset in the full-dataset comparison. Post-launch monitoring across five deployed financi

Latent Core Tokenizer: Compress, but Meaningfully

Tokenizers are commonly optimized for compression, but a compact vocabulary does not necessarily distribute its capacity evenly across languages. We introduce the Latent Core Tokenizer (LCT), a language-agnostic approach that separates structural discovery from vocabulary construction. LCT uses Minimum Description Length, entropy-based boundary signals, and morphotactic constraints to identify reusable linguistic units before constructing a shared vocabulary. Across 104 languages with a 200K-token vocabulary, LCT achieves lower fertility and higher MorphScore than BPE, Unigram, and parity-aware BPE, while maintaining comparable cross-lingual disparity in tokenization cost. Across four multilingual downstream benchmarks, LCT improves aggregate score by 1.48, 1.83, and 2.00 points over BPE, Unigram, and parity-aware BPE, respectively. Our findings show that compression alone does not predict representation quality and highlight the importance of morphology-driven structural discovery and how frequency is used to allocate the final vocabulary across languages.

OnTrack: Real-Time Monitoring and Intervention in LLM Agent Trajectories via Streaming Structure-Aware Optimal Transport

Agents are deployed in applications from trip planners and stock trading to IT incident triage. In most cases, LLM agents work autonomously with minimal rule-based safeguarding, leading to cost and safety issues from irreversible actions. Recent works resolve this either by using a safeguard agent to monitor behavior or evaluating logs post-hoc. The first adds cost and latency to every step; the second delivers its verdict after the run, when tokens are burned and damage is done. To overcome this, we propose OnTrack, a streaming monitoring mechanism that compares an agent's steps and dependencies against recorded successful runs to alert users or block the agent in about a millisecond per step. We study this problem in three regimes of decreasing access: full reference access (historical runs and tool schemas), intermediate access (only tool schemas), and no prior knowledge (only step logs as generated). Expectation of OnTrack's monitoring capabilities reduces as data access drops, ranging from plan violation detection to identifying loops, stalls, and repeated tool calls. Finally, we evaluate OnTrack using SWE-bench trajectories. Based on the first 8 steps, our method ranks failing trajectories below succeeding ones better than content similarity approaches (+0.057 AUROC). With an abort policy, we save about 18% of compute that would be burned on failing runs, where 83% of interrupted runs were actually heading to failure (5 out of 6 aborts were correct).

Which Skill to Distill? SGUID: Selecting a Compact Skill Bank for Model-Skill Co-Evolution

Skills, reusable procedural guidance added at inference, can substantially improve LLM downstream performance (Li et al., 2026). Prior work retrieves skills from a bank by semantic relevance, then uses them as inference-time patches or for model distillation. The individual utility of each skill, however, is largely neglected. We first show that, in on-policy distillation where skill-conditioned policies serve as teachers, fewer than 25% of retrieved skills provide useful distillation signals. We then propose SGUID, a method for selecting a compact subset of skills for distillation. SGUID retains a skill only if it consistently yields effective learning signals during training. The selected skills are then distilled to produce a better model. Our results show that not all skills are worth distilling. Across four models from the Olmo and Qwen families, distilling 6 selected skills matches or exceeds full-bank distillation in mean avg@12 on three of the four models, and on all four after a second round that distills 3 newly selected skills, while the full banks are up to 11x larger. Importantly, SGUID supports stable model-skill co-evolution: after a distillation round, a new candidate bank is curated from the updated model's rollouts, and SGUID selects which skills to internalize next. In the second round, this loop selects 3 new skills and improves Qwen3-8B from 64.3% to 66.3%. The selection step is essential for stability: on Qwen3-4B, naively updating the model with unfilte

Cited but Not Consulted: A Counterfactual Audit of Legal Chain-of-Thought Faithfulness

Large language models increasingly justify legal decisions by naming the statute or precedent behind a verdict, treated as evidence that the decision follows from it. We test this directly: holding case facts fixed, we substitute the named legal authority for an unrelated one and decode a model's evolving verdict from its hidden states. Across seven open-weight models (8B-70B) and four benchmarks spanning judicial and contractual reasoning, when explicitly required to justify a verdict by naming the governing authority, models name the correct one in 66.7%-100% of generations, while the verdict changing when the authority changes is far less consistent: 0.0%-21.7% on CaseHOLD, 30.0%-76.7% on ECHR and SCOTUS, and 43.3%-50.0% on ContractNLI. Neither scale nor a purpose-built legal-reasoning model (a best-effort LoRA reproduction; Section 6) closes this gap. A red-teaming evaluation on five core models finds compliance with an adversarial instruction hidden in the case facts (73.3%-96.4%) exceeds verdict-swap sensitivity by a wide margin, holding without exception across model rankings. Naming a legal authority is thus a poor proxy for a verdict's dependence on it, while the same verdict remains separately vulnerable to adversarial manipulation. Both findings replicate across checks ruling out prompt-wording noise and confounded sampling, and bear directly on the use of generated legal explanations as compliance or audit artefacts.

Accurate but Not Humble: Evaluating Epistemic Humility in LLM Agents under Knowledge Conflict

When retrieved evidence contradicts an agent's prior beliefs, does it revise its answer, acknowledge uncertainty, or persist with an incorrect conclusion? Existing evaluations of agentic systems focus primarily on task success, offering limited insight into how agents handle such conflicts. We propose to evaluate agents on epistemic humility (EH): the agent's willingness to recognize, act on, and communicate uncertainty during task execution. We operationalize EH through three trajectory-level behavioral dimensions: Identify, Solve, and Escalate (ISE). Through knowledge conflict, situations where the backbone language model's parametric knowledge contradicts the evidence it encounters, or where two contextual sources disagree, we evaluate two conflict settings: (1) controlled conflict and (2) naturally occurring conflict during multi-step agentic execution, each paired with matched no-conflict controls. Evaluating four agents, we find that higher task accuracy does not necessarily correspond to greater epistemic humility: some high-accuracy configurations recognize conflicts during execution but do not communicate unresolved uncertainty in their incorrect final answers. Trajectory-level analysis further reveals that agents frequently detect conflicts in early steps of execution but fail to maintain or resolve them in later steps. Finally, we show that model-level interventions can improve EH, but often at the cost of task accuracy, suggesting that epistemic humility emerges f

Overcoming Prior Barriers: Supervised Fine-Tuning under Long-Tail Distribution

Supervised fine-tuning (SFT) adapts pretrained large language models (LLMs) to downstream tasks, but the required concepts can receive substantially different levels of pretrained support. Frequent concepts are more likely to be well learned, whereas rare concepts may remain weakly represented. We introduce a novel notion named prior barrier to quantify how strongly the pretrained model supports competing concepts over the target concept. We observe that prior barriers follow a long-tail distribution, placing head and tail concepts at different starting points for SFT: head concepts face lower prior barriers, whereas tail concepts require additional instructions to overcome their higher prior barriers. Our theoretical analysis further derives a predictive risk bound for SFT under long-tail prior barriers, explicitly characterizing how the prior barrier and accumulated SFT evidence jointly determine predictive performance. Motivated by this prior barrier-dependent demand, we propose PASS, an adaptive SFT instruction selection method that constructs reference-derived concepts and estimates the distinguishing evidence provided by each instruction, and adaptively allocates the selection budget toward concepts that remain insufficiently covered under the current selection. In this way, PASS jointly considers which instructions can provide useful evidence and where additional supervision is needed under a limited budget. Experiments show that our method consistently outperforms sev

Can AI Agents Learn Their Way to the Top? Evaluating Heuristic Learning in a Long-Running Game Agent Competition

Adversarial games have driven advances from heuristic search to reinforcement learning, yet learning and adapting strategies from limited samples remain challenging. AI agents offer an alternative by turning game experience into revisions of executable policies. Building on heuristic learning (HL), we formalize Adversarial Heuristic Learning (AHL), a paradigm that uses AI agents as learning engines to refine game policies and supporting software while keeping model weights fixed. We introduce AAArena, a benchmark comprising 12 authentic adversarial games and 1,920 archived human programs, with an evaluation protocol modeled on real-world game competitions. Agents interpret rules, choose opponents, analyze replays, and revise game agents to achieve their highest ranking within fixed match and evaluation budgets. We evaluate \val{completedmodels} model and harness configurations: Opus5.5 with Claude Code earns 6 gold medals, while no evaluated configuration tops the remaining 6 human ladders. Performance is generally weaker in games with more complex rule specifications. Further experiments show that opponent selection and dense feedback support policy improvement, and that agents learn from both on-policy replays of their own matches and off-policy replays of other players' matches. These results highlight HL's potential in adversarial games and identify persistent challenges in game understanding, strategy implementation, and long-horizon policy development.

VFold: Symmetry-Aware Cross-Layer Value Cache Compression

While caching key-value (KV) states accelerates Large Language Model (LLM) decoding, this cache can dominate memory usage at long context lengths. One solution is to compress this memory by exploiting inter-layer cache similarities. However, most existing techniques necessitate architectural changes to LLMs and incur substantial overhead. In this work, we propose a symmetry-aware value cache merging strategy that reduces cache memory while avoiding both harmful performance degradation and architectural overhead during decoding. Furthermore, we show that this approach can be exploited alongside existing cache compression techniques, composing with high-ratio quantization or key cache pruning to reach compression ratios that neither method reaches alone, with minimal additional cost. Ultimately, our findings reveal a major source of underutilized capacity in the value cache, offering a simple yet highly effective direction for scaling context windows under memory constraints.

Verdict Without the Rule: Diagnosing and Auditing Regulatory Rule Sensitivity in LLM Compliance Systems

Large language model compliance systems are deployed on the assumption that a verdict depends on the regulatory rule it is given. We test this directly across five models and 20 regulatory and platform-policy domains: delete, swap, or negate the governing rule while holding the case fixed, and check whether the verdict changes (OCS) or the model's internal representation of compliance shifts at all (ICS-delta). Neither moves much: models' verdicts are often invariant to substantial perturbations of the supplied rule, and the guard model, evaluated here under a custom-rule adaptation of its native taxonomy, is the least rule-sensitive and least accurate of the five, barely above chance (51%, versus 90-92% for general-purpose models). This reflects easy cases more than blanket neglect: on cases where deleting the rule changes a previously correct model prediction, models do track it closely. Neither better prompting nor direct intervention on the model's internal representations closes this gap. Accuracy alone does not establish that a compliance verdict is grounded in the supplied rule.

EgoVoice: Proactive Spoken Assistance from Egocentric Multimodal Streams

Wearable augmented reality (AR) assistants are moving toward continuous real-world interaction, where they perceive the user's activity through first-person video and audio and provide timely spoken guidance without being explicitly asked. While proactive video assistants, spoken dialog systems, and egocentric task understanding have each advanced rapidly, existing systems do not address the joint problem of deciding when to speak and what to say from continuous first-person streams. We introduce EgoVoice, a framework for training and evaluating proactive egocentric spoken assistants. From HoloAssist video recordings of real human instructors, we construct clean audio streams through source separation and speech resynthesis, and convert each video session into a format where the model must decide at each moment whether to remain silent or provide spoken guidance. We fine-tune an omni-modal LLM with our data, and further improve its proactive intervention behavior with direct preference optimization. Experiments across closed and open-source models show that existing systems rarely produce well-timed, meaningful proactive interventions, while EgoVoice yields clear improvements in intervention timing, content relevance, and human preference over the zero-shot backbone.

NativeScope: Relation-Localized Retrieval over Native Topology with a Correct Anchor

Dense retrieval usually ranks text chunks by their semantic similarity to a question. This ignores structure that many data systems already store, including section membership, session boundaries, and native order. We propose NativeScope, a scope-then-rank method for queries with a known anchor and relation. It represents a query as q -> (A, r, B). The anchor A and relation r select native units through belonging, before, or after operators, and the target term B ranks only chunks that overlap the selected scope. An internal variant, NS-FullQ, ranks the same candidates with the full question. We evaluate both methods on 200 controlled document and memory records derived from QASPER and LongMemEval under a 1,024-token budget. NativeScope attains native-unit recall of 89.28 percent for documents and 72.50 percent for memories, improving over instance-wide Dense RAG by 42.75 and 22.00 percentage points. NS-FullQ reaches 87.78 percent and 68.50 percent; its differences from NativeScope are inconclusive, locating the primary gain in relational scoping rather than the shorter ranking query. With automatic Top-1 anchors, memory recall falls to 35.50 percent. NativeScope is therefore effective when anchor coordinates and native relations are reliable, but hard scoping inherits errors from the localization interface.

From Scan to Treatment Plan, AI Helps Close Breast Cancer’s Deadliest Gaps

Breast cancer is the most commonly diagnosed cancer among American women — yet the gaps in care are wide. A majority of women over age 40 skip the recommended annual screening. Radiologists are reading more mammograms with fewer colleagues. And when a diagnosis arrives, the tests that inform treatment can take weeks to return results.  […]

Amazon and others are done keeping data center deals secret. Is it enough to build trust?

Amazon says it will stop using NDAs when negotiating data center deals with local governments, following a similar move from Microsoft earlier this year. Secrecy has fueled community backlash against AI infrastructure, with opposition leading to hundreds of proposed and enacted moratoriums from New York to San Francisco. Meanwhile, a wave of startups is betting that consumers will hand AI agents access […]

The Agent Said It Was Done. The Database Disagreed.

On the estimation and validity of AI time horizons---a statistical look at the METR plot

METR's 50\% time horizon measures the human completion time of software tasks that an AI solves with 50\% probability, allowing AI capabilities to be expressed in interpretable units. On 228 tasks and 26 AIs, we recompute the time horizons using splines and item-response theory to relax the assumption that the AI difficulty of a task depends linearly on the log of human time. Our fitted spline can be interpreted as a function that \emph{converts} human time to AI difficulty; it is nearly flat in a region from 2--30 min but close to linear elsewhere. Hence, a time-horizon jump from 3 min to 30 min is much easier than one from 30 min to 5 hours despite the same multiplier of $10 \times$. Overall, we contribute time-horizon point estimates that perform better under a cross-validated suite of proper scoring rules, as well as diagnostic plots for assessing time horizons' construct validity. We suggest that time horizons be interpreted together with the diagnostic plots, especially as new time-horizon-based benchmarks are proposed or existing ones grow to include longer tasks.