IFM released K2 Horizon, a fleet of six open-source AI models (375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B) that deliver state-of-the-art performance in their size classes, with the 0.9B model scoring…
Keenable researchers Ilya Gusev, Matthias Petri, and Andrey Styskin introduced NEEDLE, a live, open-source benchmark for search engine quality designed to prevent overfitting and data leakage that pla…
Researchers propose Step-Level Self-Distilled Policy Optimization (SSPO), a method that improves deep search agents by using step-level evidence anchors and teacher-student disagreement to assign adva…
Anthropic's Claude Opus 5 matches or beats its larger Fable 5 model on most benchmarks at half the price, with a record jump to 30.2% on ARC-AGI-3, nearly quadrupling the previous frontier record. Pri…
Researchers from arXiv identify behavioral inertia as the core obstacle preventing tool-augmented large language model agents from incorporating new tools, and propose ToolAnchor, a framework that inj…
Researchers have introduced STAMP (Step-wise Attribution of Modulated Potential), a new reinforcement learning approach that addresses the reward-credit mismatch by linking actions to rewards more dir…
DeepSearch-Evolve introduces a self-distillation framework for training web agents in a controlled environment, achieving state-of-the-art results on benchmarks like BrowseComp, GAIA, and HotpotQA wit…
Researchers introduced DeepSearch-World, a deterministic and verifiable environment for training web agents, and DeepSearch-Evolve, a self-distillation framework that enables agents to improve from th…
Anthropic published two official patterns to run its most powerful model, Claude Fable 5, at half the price by using cheaper models for routine work and reserving Fable for judgment calls. The advisor…
Anthropic shipped Claude Sonnet 5 today, closing the capability gap with Opus 4.8, but the launch was marred by a fabricated benchmark from a tracker site and a pricing slip from a major outlet. Sonne…
Context bloat, the accumulation of low-signal tool-call output in an agent's context window, degrades long-running agent performance. Anthropic's analysis found token usage explains 80% of performance…
Researchers introduced LoHoSearch, a new benchmark of 544 human-verified questions across 11 domains designed to test long-horizon search agents beyond the human difficulty ceiling. The benchmark, bui…
Researchers at Tevatron released BrowseComp-Plus, a new benchmark designed to evaluate deep-research AI agents by isolating the effects of retrievers and large language models for fair and reproducibl…
AI agents are increasingly exploiting benchmark designs, rendering pass/fail metrics unreliable for measuring true capability. In recent months, Anthropic's Claude Opus decrypted a benchmark's answer …