Every processed story in chronological order, with the newest coverage first. Filter by tag, source, or score to drill in.
The benchmark demonstrates that adapter/harness design can swing Pass@1 by over 54 percentage points on the same model, showing that existing SWE-bench evaluations of general-purpose agents conflate harness quality with model capability — a gap Claw-SWE-Bench is designed to isolate.
AgentHarness introduces a concrete open-source pattern for separating verification from the main reasoning model in long-horizon agent loops, with purpose-built small weights that reportedly outperform much larger open-source models on BrowseComp benchmarks.
The findings demonstrate that how procedural knowledge is structured for LLM agents — not just what it contains — measurably changes agent search behavior and task outcomes, establishing Skill organization as a distinct design variable for agent systems.
North Mini Code 1.0 brings an Apache 2.0-licensed agentic coding model with a low active-parameter footprint (3B of 30B) to the open-source ecosystem, making it freely usable and modifiable for local and commercial deployments.
The paper demonstrates that replacing linear repository traversal with domain-scoped parallel agent spawning improves multi-file change localization for a small model, while also identifying that naive filesystem access and forced multi-agent consultation can actively harm performance or inflate costs.
The benchmark reveals that frontier coding agents can reliably execute computational social science workflows, while also exposing prompt-framing vulnerabilities that could introduce bias into AI-assisted scientific production.
The results show that targeted RL fine-tuning on high-quality, task-specific data can close — and reverse — a 231-billion-parameter gap in model size, at a training cost under $500, on a real financial reasoning benchmark.
AutoPDE's explicit strategy representation closes a key gap in LLM-based PDE solvers, where numerical decisions previously remained hidden in code and were difficult to inspect or correct when solves failed.
Fable 5 represents Anthropic's most capable generally available model to date, and the dual launch with Mythos 5 introduces a tiered access model that pairs broad public release with a restricted, safeguard-lifted variant for vetted cyberdefense use cases — a structure Anthropic describes as central to releasing powerful models both safely and quickly.
MemToolAgent demonstrates that structured memory management — without any LLM fine-tuning — can substantially improve tool-use accuracy, with an 80% relative gain on NESTFUL showing the approach's potential to close the gap between static LLM agents and agents that learn from experience.