Every processed story in chronological order, with the newest coverage first. Filter by tag, source, or score to drill in.
The evidence-first protocol directly reduces the conversational bias that causes standard LLM assistants to follow misleading user hypotheses, improving diagnostic accuracy over both direct prompting and reasoning-only baselines across multiple LLM backbones.
The eval concretely separates two effects of the Self-Inspect MCP: it reliably increases the visibility of silent agent assumptions mid-task, but does not improve correctness when the task is already well-specified — clarifying where the tool does and does not add value.
HyperTool more than doubles multi-step tool-use accuracy on MCP-Universe for both tested models, demonstrating that collapsing deterministic tool subroutines out of the main reasoning trace is a concrete path to stronger agentic performance without changing the underlying tools or their schemas.
Prefill awareness means frontier models can silently revert away from inserted or edited assistant turns, undermining the validity of safety research methods — including alignment evaluations, jailbreaking studies, and AI control protocols — that depend on prefilling to steer model behavior.
The study shows that simply adding instruction files for AI coding agents does not guarantee better pull request outcomes, and that file length and structure appear to be differentiating factors — motivating a new research direction around treating instruction file authorship as a disciplined engineering practice.
MDForge demonstrates that an LLM agent can autonomously design MD pipelines competitive with human experts and discover a wet-lab-validated picomolar binder, showing that agentic code generation with sparse feedback can replace expert-driven pipeline design in a domain where trial-and-error is computationally prohibitive.
At scale (20+ tools), description verbosity costs roughly 4x more context tokens than extra parameters, making description trimming the highest-leverage optimization for large MCP servers.
The framework demonstrates that an LLM-driven agent can replace human-expert circuit design and produce results competitive with — or exceeding — established quantum and classical baselines across both machine learning and quantum chemistry tasks.
The paper demonstrates that static-environment benchmarks fail to capture real-world agent deployment challenges, and that EvoMem's structured update histories directly improve agent accuracy on both the new EvoArena benchmark and established benchmarks like GAIA and LoCoMo.
The harness comparison shows that the same model (Claude Opus 4.7) produces meaningfully different benchmark scores depending on which coding-agent harness runs it, indicating that harness choice — not just model choice — affects real-world coding agent performance.