Paper of the Day · November 19, 2026

LMLineage: Content Lineage and Context Governance for LLM Agents

LMLineage traces each sentence an SI agent writes back to the input passages that support it, without opening the model, and uses that trail to trim stale context.

What they found

The authors built LMLineage, a system that sits outside the model. It links every output sentence of an agent session to the input passages that support it, with confidence scores. It needs no access to model internals and no extra model runs. They tested it on 71 hand-labeled sentences from 10 Chinese web question-answering cases, plus 90 turns of multi-turn sessions across three generator models. Matching small windows of text by embedding similarity worked far better than matching whole documents or exact words. A fast greedy-window method was faster and more accurate on average than exhaustive line matching. Each turn took about 1s on a single GPU, even as context grew to 40 turns. Their drift detector, which finds input documents the agent has stopped using so they can be compacted, made no false compactions (0%). Its recall was limited by how accurate the attribution was.

Why it matters

SI agents pull in many documents and instructions over long sessions, and it is hard to tell which source shaped which statement. A cheap, model-agnostic way to trace this could help people audit agent answers. It could also help keep context short without summarizing everything, since user instructions are kept word for word. The test sets are small, and the abstract does not say how well this holds up outside them.

How it changes the race

This is an incremental, practical tooling advance. It could lower the cost of running long agent sessions and of auditing them, because it works on closed models without extra inference. It does not change model capability itself.

Safety and control

The system gives a way to trace which inputs support an agent's output, which could help with incident tracing and cleaning memory, as the authors' case study describes. Attribution shows which text supports a sentence, not why the model said it, and the abstract reports that drift detection recall depends on attribution accuracy.

The abstract, by the authors

Interpretability for AI agents should not be confined to understanding the model's internal weights. It must also cover the agent's reasoning and decision-making process over multi-source context. We present LMLineage, an out-of-band system that attributes every output sentence of an agent session to its supporting input spans with calibrated confidences, without requiring white-box access to the model or any additional inference passes. On top of the lineage metadata, we build concept-drift micro compaction: instead of waiting for the context-length limit and summarizing everything, the system detects when previously used input documents have stopped being referenced ("drifted") and surgically compacts only those, preserving user instructions verbatim. We evaluate on a multi-tier corpus: 71 gold annotated sentences from 10 Chinese web-QA cases, plus 90 turns of independently collected multi-turn sessions across three generator models (gpt-4o-mini, qwen3.8-27b, deepseek-v4-flash). Results show that (i) window-level embedding similarity is far better aligned with LLM paraphrasing than whole document or lexical matching; (ii) the greedy-window fast path is both faster and more accurate on average than exhaustive line-matching (a multiple-comparisons bias in the latter inflates web recall at the cost of overall accuracy); (iii) per-turn attribution completes in 1s on a single GPU, remaining stable as context accumulates to 40 turns; and (iv) concept-drift detection is conservative (0% false compaction) with recall bounded by attribution accuracy. We further describe the concept-drift micro compaction design and an interactive incident-tracing/memory-cleaning case study. A companion patent application covers the lineage construction method.

Quoted from arXiv:2611.22511. The research and its words belong to its authors.

The authors

Wei Hou, Jingxin Geng, Xin Chen, Zheqing Jiang

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.