The BaseResearcharXiv
LMLineage: Content Lineage and Context Governance for LLM Agents
According to its abstract, a preprint by Wei Hou, Jingxin Geng, Xin Chen and Zheqing Jiang presents LMLineage, an outside system that traces each sentence an agent writes back to the input passages that support it, without access to model internals, and uses this to compact context that is no longer referenced.
Why it matters
The same thing, for different readers.
For lawmakers
A tool that records which sources support each agent statement could aid auditing of AI agents, though the evidence here is a small test set.
For researchers
Hou and colleagues propose interpretability at the level of an agent's context rather than its weights, attributing output sentences to input spans with confidence scores.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- An out-of-band system attributing agent output sentences to input spans, evaluated on 71 gold-annotated sentences from 10 Chinese web question-answering cases and 90 multi-turn sessions across three generator models.
- What it found
- The authors report that window-level embedding similarity aligns with language-model paraphrasing far better than whole-document or lexical matching; other results were cut off in the excerpt.
- What it does not show
- A small evaluation corpus, mostly Chinese web question answering; the excerpt is truncated.
- Status
- A preprint: not yet peer reviewed
The record
- Questions
- Are AI companies putting stronger safeguards into practice?
- People
- Wei Hou (Wrote it) · Jingxin Geng (Wrote it) · Xin Chen (Wrote it) · Zheqing Jiang (Wrote it)
- What we read
- We read the abstract.
- Found by
- The Radar, run of November 18, 2026, via
arxiv-ai-safety-query(tier 1: the source’s own channel or an official record) - Cite this
- arXiv, “LMLineage: Content Lineage and Context Governance for LLM Agents”, November 17, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-11-17-lmlineage-content-lineage-agents