dsh-arxiv-ml-search

When an LLM answers machine-learning questions, it all too often relies on parametric memory — conclusions baked in during training that may be inaccurate or outdated. dsh-arxiv-ml-search lets the agent search arXiv for ML papers directly, verifying every claim against published abstracts and metadata instead of trusting its own weights.

The plugin ships two tools: arxiv_search, which supports keywords, authors, categories, date ranges, sorting, and paging while returning truncated abstracts to keep token costs low; and arxiv_get, which retrieves full metadata and complete abstracts for specific paper IDs, with optional segmentation for long text. A companion Skill file codifies the claim-checking workflow: restate the claim, run orthogonal queries, read the shortlist carefully, and report a verdict with paper IDs, flagging any approximate matches or thin evidence explicitly.

Built for engineers and researchers working on reinforcement learning, RLHF, calibration, reward shaping, and related topics who need to cite specific arXiv paper IDs as evidence. No Python runtime or build step is required — a single profile entry keeps the tools available in the toolset, and the agent automatically engages the verification workflow whenever an ML engineering question arises.