evidence-attribution

A curated list of research on LLM-agent traces: evidence tracing, execution provenance, failure attribution, observability, runtime safety, memory provenance, and learning from traces. Companion list of arXiv:2606.04990.

证伪优先的引用核验:论文存在性 / 元数据 / 论点支撑三重检查,可被 Claude Code、Codex、Cursor 以 MCP 调用 · Skeptical citation auditor for agent writing workflows · PyPI: citationguard

Evidence attribution in visual document understanding without coordinates or region labels: quote-and-retrieve interface + region-label-free GRPO

Measurement study of attention sinks that land on indexed image patches in late-interaction visual document retrieval (ColQwen2.5 / ViDoRe V3).

An evidence-annotated benchmark for RAG over university course syllabi — 80 questions with document, page and supporting-span labels, including unanswerable items, evaluated across closed-book, full-context, basic RAG and citation-constrained RAG.