Articles & Guides

OpenAI is now asking California to expand safeguards in SB 53, including monitoring frontier models during training and evaluation. The position marks a notable shift from the company’s earlier opposition to the bill.

A public-information review finds that leading AI labs have disclosed few concrete plans for isolating or shutting down a model that tries to evade human control. The gap is becoming more consequential as agentic systems gain access to real company infrastructure.

Cloudflare’s Agent Tracing adds agent-level observability to Workers, covering model calls, tool execution, approvals and subagents. Payload defaults differ across frameworks, while truncation and upcoming usage fees limit how the feature should be used.

After AI-related development spending grew roughly tenfold in six months, JetBrains chose not to restrict engineers to a small approved toolset. Instead, it built a shared access, routing, and accounting layer around those tools.

EnvHarness wraps existing environments with programmable plug-ins instead of rebuilding them from scratch. Its companion system, EnvRigger, studies black-box policy trajectories to create and validate environment changes aimed at specific weaknesses.

SWE-bench Science brings repository-level evaluation to scientific software engineering across 98 GitHub repositories and 20 domains. Its results show that coding agents can pass visible checks without preserving the scientific behavior that matters.

FlashPrefill V2 upgrades sparse attention for long-context LLM prefill with mean correction, GPU-oriented kernel engineering, and native support for paged KV caches and continuous batching. On NVIDIA H20, it reports up to 47.26× the speed of FlashAttention-2 at 128K context in FP8.

A study from Hume AI introduces a way to measure benchmark optimization in automatic speech recognition. It finds that some high-scoring open-source models may reproduce reference text from public benchmarks even when the audio is contradictory, masked, or ambiguous.

PolicyGuide compiles organizational policies into workflow graphs and proactively checks agent state at user-turn boundaries. The approach helps customer-service agents recover missing procedural steps instead of merely blocking risky actions.

τ₀-VLA introduces world-model-guided test-time computation for long-horizon robot manipulation, allowing a high-level policy to search through candidate subtasks when a decision is uncertain.

A broad benchmark finds that LLMs and dedicated embedding models are nearly tied overall, while their economics are dramatically different. The practical answer is task-based specialization, not wholesale replacement.

FACET introduces an environment-grounded framework for synthesizing terminal-agent tasks. It aims to preserve source intent while keeping instructions, solutions, environments, and verifiers consistent through execution-based validation and targeted repair.

Remembering more does not always help a language model solve the task in front of it. MemTrapBench measures two ways retrieved memories can distort reasoning and introduces AdaptiveMem as an inference-time mitigation.

SkillEvo argues that the main bottleneck in evolving agent skills is not merely the ability to edit them or the number of iterations, but whether evaluation keeps producing reliable directions for improvement. It turns multi-turn user simulation into a feedback generator and adds an independent governance layer to control factual and structural degradation.

ForgeWM introduces a progressive causal training recipe for converting a bidirectional action-conditioned video generator into interactive world models that run in one, two, or four denoising steps. The design targets low-latency gameplay while preserving alignment between controls and generated motion.

A Greek reasoning study finds that nearly unchanged accuracy can hide major behavioral gains. SFT teaches models to reason in the user's language, while RLVR repairs formatting and channel-control failures.

DeepMind introduces EXIMO, a three-stage method that combines VLM-guided exploration, imitation learning, and residual off-policy reinforcement learning to fine-tune vision-language-action policies for new tasks.

A new study proposes IAR, a staged post-training framework for answering questions about a fixed document collection without retrieved passages at inference time. It separates knowledge injection, QA accessibility, and recovery of general capabilities.

NARU is a benchmark for evaluating whether multimodal models can follow evolving narratives and interpret implicit cultural meaning in extremely long Japanese videos. Its results point to persistent weaknesses in long-range integration and culturally grounded reasoning.

A new study introduces Chain-of-Experience (CoE), a test-time framework in which language models retain traces from prior interactions and use feedback to refine later attempts. The work evaluates self-feedback and environmental signals across math, coding, and knowledge tasks.

QuoteBench argues that a coding agent’s result depends not only on the command it generates, but also on how that command is serialized, wrapped, and parsed before execution. A matched score can therefore conceal major failures introduced by the interface.