LMMR trains small model agents to pick memory sources and reading depth
A research paper describes a learned retrieval method that lets compact language model agents decide which stored record to open next and how much of it fits a token budget.
In plain words
Small AI agents have little room to hold background facts when answering questions that need several steps of reasoning. The paper's method, LMMR, teaches the agent to choose both the next document to consult and the amount of it to read. In reported tests it answered more accurately while reading less text.
What was announced
The authors frame the problem around multi-hop questions, where needed facts sit in separate records and the surrounding text of each record fills the limited context a small model can handle. LMMR links paragraphs and sentence fragments back to their source documents, and during training it pairs views that supply the same missing support, while relevant distractors and repeated content serve as contrasting examples. At inference, a reranker scores what each candidate would add to an anchor passage, and a selector picks evidence that fits a token budget.
The reported results come from the official validation sets of HotpotQA, 2WikiMultihopQA, and MuSiQue, covering 22,398 questions. With a frozen 2B answer model and a 400-token budget, the paper reports F1 scores of 61.47, 43.26, and 30.58 respectively, beating query expansion by 3.06, 2.14, and 4.55 points while cutting memory input by 23.1 to 26.9 percent. Against a coarser retriever trained on the same questions, LMMR used 23.5 to 27.7 percent fewer input tokens. Compared with that retriever followed by LLMLingua-2 at similar input lengths, F1 gains reached 3.99 points, significant on HotpotQA and 2WikiMultihopQA after Holm correction. All eight small answer models tested on HotpotQA improved over query expansion and MMR in F1 and exact match.
What it does not show
- The source is an abstract only; full method details, experimental setup, and code availability cannot be verified from the supplied passages.
- The reported scores are the authors' own claims and have not been independently confirmed.
- The abstract does not state which eight small answer models were tested or their individual results.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.