acceptodds
Withdrawn from ICLR 2027

Learning to Retrieve Missing Evidence for Long-Term Memory QA

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-term memory enables language models to use past interactions in future conversations. However, evidence needed to answer a question may be scattered across distant turns, while the question itself omits clues needed to locate it. Retrieved facts can reveal these clues, motivating retrieval decisions conditioned on evidence already found. We introduce MERA (Missing-Evidence Retrieval Augmentation), which separates globally searchable memory from a question-specific evidence state. Verified evidence guides subsequent retrieval without restricting access to the global memory. We train a lightweight planner through reinforcement learning, rewarding queries that recover previously missing evidence. MERA achieves strong answer accuracy across Qwen3-30B and GPT-4o-mini backbones. With Qwen3-30B for evidence processing and answer generation, the trained 0.6B planner achieves 77.40% accuracy on LoCoMo and 71.29% on LongMemEval-S, exceeding a 30B planner without retrieval-grounded training by 4.10% and 3.96%, respectively. On LoCoMo, later retrieval rounds increase cumulative evidence recall from 55.5% to 80.5%.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Large language model, LLM memory, Reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ppxZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces MERA, a framework for answering questions over long conversation histories. It indexes short, self-contained facts extracted from the conversation where a lightweight planner generates retrieval queries using the question and accumulated evidence. The planner is trainable with GRPO using rewards based on newly recovered gold supporting turns. Experiments on LoCoMo and LongMemEval-S report improvements over searching recall and answer accuracy.

Critical strengths
  1. Well-motivated design that uses accumulated evidence to guide further searches while retaining access to the full memory.
Critical weaknesses
  1. The work seems incremental, which does not clearly establish a clear advance over closely related contextual and structured-memory approaches.

  2. Evaluation needs more justifications. It covers two datasets with dataset-specific training data, more specifications on the similarity between the training data and the evaluation data is needed. Generalization is insufficiently established, independent datasets outside the training distribution would strengthen the results. Table 1's results seem to come from different eval settings.

Questions for authors
  1. I would like to understand more on how similar the training data is as compared to the evaluation set in terms of conversation content, question types, and data generation procedures, and how potential overlaps are ruled out.
  2. To assess generalization, could the planner be trained using data from only one benchmark and evaluated on both? Excluding either LongMemEval-S or LoCoMo-derived training data would help distinguish cross-dataset transfer from gains due to benchmark-specific training.
  3. In Section 4.2, the footnote says "Results marked with ∗ are obtained from our own runs under the evaluation setup described above", which seems to come from a held-out test split of 101 questions based on Appendix B.1, are the results without * from evaluating on the same 101 questions or the 500 questions stated in 4.1? The accuracy numbers looks like they are evaluated on different settings.
Reviewer_AanROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper presents a method to retrieve missing memories in a retrieval setting in memory systems. Authors argue that in long conversations, memories created are scattered along different timelines and for a question, it may require to piece multiple memory evidence together to answer the question correctly. Retrieving all the related memories are difficult and hence the propose a solution for this makes sense. The paper presents an approach where they optimize the query planner with GRPO whereas the retriever, memory index remain unchanged.

While retrieving missing memories is undoubtedly useful in practice, it is not clear the proposed approach is making any contributions over existing approaches. There are many memory systems available in the literature that keep on improving the retrieval and answer generation accuracy. To make sense of the proposed approach, I believe this needs to be evaluated against more recent and advanced memory systems for accuracy. The proposed approach has to be comparable and competitive to warrant the claims that the authors are making. That is, in my opinion, a memory system may already show much higher accuracy on benchmark datasets (HyperMem, HyperGraphRAG, Memora, etc.). If such systems natively perform better, what value MERA brings in is not very clear.

Critical strengths

• The problem is real and has value. Paper is easy to follow.

Critical weaknesses

• I feel novelty of the entire approach is weak. There are plenty of iterative RAG retrieval methods that could apply for this task as well. So it is not clear why the proposed method is going to be different to those.

• The paper did not evaluate against more recent memory systems that have improved retrieval and memory representation techniques (e.g., memora, HyperMem, HyperGraphMem, etc.). I believe the proposed approach should be able to demonstrate accuracy over recent approaches because the MERA is directly competing against them. Or even apply on top of them to show value in finding missing evidence further.

Questions for authors

See weaknesses. Contribution of this work is not clear with respect to more recent memory systems. Compared baselines in the paper are methods that were introduced some time back.

Reviewer_WaU4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses question answering over long conversational histories, where relevant evidence may be scattered across multiple interactions or involve entities not explicitly mentioned in the query. Traditional memory systems focused on storing and retrieving information while the proposed MERA framework uses a graph-based memory representation and a lightweight RL-trained planner to iteratively retrieve missing evidence based on what has already been found.

The approach is evaluated on LoCoMo and LongMemEval using Qwen3 and GPT-4o-mini models, with accuracy assessed with LLMaaJ.

Critical strengths

The paper clearly formulates the problem statement and how it separates persistent memory from a question-specific evidence state that tracks what has been found and what is still missing. The proposed solution is interesting because although the method maintains a dedicated evidence graph, queries search the whole memory index, so entities outside the constructed graph remain reachable. The appendix, implementation details and diagnostic statistics are thorough and cover many aspects only briefly treated in the main text. The ablations are generally well designed.

Critical weaknesses

While the proposed solution is intersting, the paper has several weakness that do not allow me to recommend it for acceptance. First, the benchmarks contain relatively short conversation histories that largely fit within the context windows of modern LLMs, making the need for specialized retrieval insufficiently justified. The evaluation is also limited to two relatively small benchmarks, with no evidence of scalability or cross-dataset generalization, leaving open whether the planner learns dataset-specific patterns. Finally, the comparison lacks relevant question-driven retrieval baselines, and ablations on the reward design are missing. The iterative retrieval strategy may also propagate irreversible errors when relevant directions are incorrectly discarded. I also found some numerical inconsistencies (e.g., 75.73% vs. 77.40% on LoCoMo for apparently identical settings) and discrepancies between the main paper and Appendix F.

I'd recommend authors to clarify computational costs (offline vs. online), have up to date baselines,

Questions for authors

See weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.