Archive
新模型、開放權重、多模態、語音、影像與模型架構。
推論系統
Based on host memory and PCIe bandwidth, FreeToken dynamically decides whether experts missing from the cache should be transferred to the GPU or computed directly on the CPU. The paper reports 1.3× to 2.1× higher decoding throughput than the strongest baseline across five consumer-grade systems, but the linked code repository is not currently publicly accessible.
代理工具
A new evaluation separates tool calls into state retrieval and task execution, finding that context compression can sharply increase an agent’s rechecking costs before success rates decline. Under fivefold compression, GPT-5.5’s completion rate changed from 80% to 85%, while its average retrieval calls rose from 21.0 to 63.9.
推論系統
DriveCache uses vehicle translation and rotation known before generation to schedule diffusion-model feature reuse across different driving scenarios. On Wan2.2 A14B, it delivers 1.276 dB higher PSNR than SeaCache at a nearly identical acceleration level, though the code has not yet been released.
AI 安全
A new study uses an additive log-odds model to measure subtle biases introduced by paraphrases, typos, animal names, and JSON fields, then combines cues that push in the same direction into a prompt. Some optimized prompts transfer across models, though the current experiments focus on artificially designed binary-choice questions.
代理工具
New research shows that even when the content being verified is identical, prior auditing and repair experience retained in context causes several open models to flag fewer errors. Signal detection analysis found no corresponding improvement in discriminability, suggesting that the agent pipeline changes the verification threshold rather than simply making the model more accurate.
代理工具
Rather than modifying agent framework internals, ClawGym II intercepts calls at the model-serving boundary and reconstructs multi-turn trajectories using a prefix tree. After training, Qwen3-30A3B’s Pass@1 on ClawGym-Bench increased by 9.98 percentage points with OpenClaw and 14.81 percentage points with Claude Code.
AI 安全
A new study finds that several guard models and activation probes show almost no change in judgment accuracy when rules are removed, shuffled, or replaced. In cross-rule tests designed to eliminate textual shortcuts, fast detectors perform near chance; only models using step-by-step reasoning clearly recover the ability to combine rules with scenarios.
代理工具
A Stanford team simulated more than 10,000 communities of language-model agents and found that, over multiple rounds of communication, initially weak opinions gradually develop into consensus or polarization. Using an extended Ising model, the researchers predicted individual trajectories on previously unseen communication graphs. Political questions, however, also produced a rightward shift, showing that collaboration does not necessarily eliminate model bias.
AI 安全
JailbreakSkill packages prompt rewriting, failure analysis, and skill evolution behind a unified interface, enabling red-team agents to generate new procedures from unsuccessful attack trajectories. Its second stage raises the cross-model average attack success rate on AdvBench and HarmBench by a further 17.5 and 13.4 percentage points, respectively, although evaluation relies on a GPT-4o judge.
推論系統
ParaTempo periodically probes the provisional answer distributions of individual reasoning paths, then independently prunes, retires early, or forks them. On four benchmarks, Qwen3.5-35B-A3B averaged 1.1 percentage points lower accuracy than fixed 16-path self-consistency, while reducing latency by 21.8% and generated tokens by 30.3%.
開發者平台
Anthropic ended access to the legacy Workbench on August 17. Saved prompts, versions, variables, and evals cannot be carried over to the new interface. The replacement is a Messages API playground with no server-side history, while the experimental endpoints for generating, improving, and templatizing prompts have also been retired.
推論系統
The new release stops overriding GPU memory information with the system’s `MemAvailable` value in Linux HIP builds and instead trusts `hipMemGetInfo`. The fix targets AMD APUs with configurable UMA carveouts and reduces the risk of the model loader overestimating capacity and failing later.