acceptodds
Under review as a conference paper at ICLR 2027

Before Answering: Evidence Sufficiency under Size-Matched Memory Construction

20 views

Abstract

Agents that answer questions from compressed or retrieved memory must recognize when the evidence a query needs is no longer in memory. Benchmarks for this task usually create insufficient-evidence examples by deleting supporting passages. We show that this construction leaks the label through memory size: on MuSiQue, a classifier that only counts paragraphs reaches an area under the ROC curve (AUROC) of for detecting unsafe memory, higher than the lexical estimator we initially evaluated. We propose a size-matched construction that provably removes this shortcut, and use it to study MemSafe, an estimator that cross-encodes the query with each memory unit and aggregates the units with a set transformer. Across three multi-hop question answering datasets and five seeds, MemSafe reaches and AUROC on MuSiQue and HotpotQA, to above a lexical baseline, while the third dataset, 2WikiMultiHopQA, is saturated. A frozen pretrained cross-encoder with a logistic head already closes of the MuSiQue gap between the lexical baseline and . At the same time, MemSafe degrades more than a weak baseline on the unanswerable questions released with MuSiQue, reaches only AUROC on SQuAD 2.0, and needs several thousand clinical training examples before it outperforms a feature-based estimator. Used as a gate for a 7B reader, it reduces the error rate on answered questions from to at coverage, outperforming both reader confidence and, on average, the ground-truth integrity label, although a 7B LLM judge is the better gate at coverage. These results indicate that the way insufficient evidence is constructed matters as much as the estimator that detects it.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Uncertainty Estimation, Trustworthy AI, retrieval-augmented generation, memory sufficiency, compressed-memory failure detection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ew2cOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies evidence sufficiency estimation before answering. Its main contribution is identifying a paragraph-count shortcut in benchmarks constructed by deleting supporting passages, proposing a count-matched construction, and introducing MEMSAFE, which combines a cross-encoder with a set transformer. Experiments cover multi-hop QA, cross-construction transfer, clinical records, and downstream answer gating.

Critical strengths
  1. The evaluation artifact is clear and consequential: paragraph count alone achieves 0.979 AUROC, showing that deletion-based construction can substantially inflate performance.
  2. The comparisons and ablations are reasonably comprehensive, including matched mean pooling, a long-context encoder, and an LLM judge.
Critical weaknesses
  1. The constructed task remains distant from realistic applications. Missing-evidence examples primarily rely on annotated supporting passages, whose absence does not necessarily make a question unanswerable. Evaluation on memories produced by actual retrieval or compression systems is needed to establish broader validity.
  2. Generalization and downstream utility remain limited. AUROC drops to approximately 0.639 on SQuAD 2.0, and transferred risk thresholds fail. Although answer gating reduces errors, the 7B reader still has approximately 63% error at only 5% coverage, providing limited evidence of practical reliability.
Questions for authors

Can the authors validate the approach on memories produced by actual retrieval or compression, and audit cases where questions remain answerable despite missing annotated evidence? How much coverage does set-level interaction gain over mean pooling at matched test risk?

Reviewer_e8TiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether an agent can determine before answering if its retained memory contains sufficient evidence. It identifies a paragraph-count shortcut in deletion-based evaluation and proposes a size-matched construction. This provides a formal characterization of size-only predictors. The study introduces MemSafe, a method that integrates query–passage cross-encoding with set-level aggregation. The evaluation covers three multi-hop question answering datasets, component ablations, transfer tests, clinical records, and downstream selective answering. MemSafe shows strong performance on the primary construction. Most improvements over lexical baselines come from the encoder and its fine-tuning. However, transfer and downstream experiments reveal substantial limitations.

Critical strengths
  1. Using paragraph count as a metric yields an AUROC of 0.979 on the original MuSiQue construction, but performance drops to near chance after size matching. Integrating formal argumentation and empirical controls provides strong evidence that the construction can artificially inflate sufficiency-detection performance.
  2. Matched-encoder baselines, ablation studies, transfer experiments, and clinical learning curves clarify the origins and boundaries of these observed gains. Explicitly reporting negative results enhances the interpretability of the empirical findings.
Critical weaknesses
  1. The authors correct a specific evaluation artifact through size matching, but obtain most performance gains from the cross-encoder and its fine-tuning. They gain only an additional 0.020 AUROC on MuSiQue and 0.007 on HotpotQA from set-level attention over matched mean pooling.

  2. The authors define sufficiency through annotated supporting facts, but do not quantify how often alternative evidence remains sufficient after deletion. They achieve only 0.639 AUROC on SQuAD 2.0 and fail to maintain the target risk after transfer. They also report approximately 63% reader error at 5% coverage.

Questions for authors
  1. Have the authors audited how often annotation-defined unsafe memories remain answerable from the retained text? Quantifying this discrepancy and performance on audited examples would clarify the validity of the evaluation target.
  2. What explains the 7B reader’s 29.1% exact-match accuracy even with complete memory? Have answer-format and normalization issues been ruled out? What are the dominant remaining errors? This affects how much weight the downstream results should receive.
Reviewer_3MQyOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to determine, before answering a question, whether the current memory still retains the evidence required by the question. The authors find that directly deleting supporting paragraphs to construct insufficient-evidence examples creates a correlation between paragraph count and the integrity label, so a model may achieve high scores merely by counting paragraphs. They therefore propose a size-matched sample construction method and build MEMSAFE using a cross-encoder and a set transformer to predict evidence completeness and missing information. The method achieves 0.9682 and 0.9831 AUROC on MuSiQue and HotpotQA, respectively, but transfer performance across constructions and external datasets declines markedly. Downstream experiments show that its score can improve answer selection, but it has not yet reached a reliable risk level for answering.

Critical strengths
  1. The problem has practical significance. The authors show that a count shortcut in sample construction can seriously affect evaluation: on MuSiQue, a paragraph-count control reaches 0.979 AUROC, which drops to 0.5007 after size matching. This reminds us that a high integrity-detection score may come from construction cues rather than genuine evidence judgment.
  2. The controls and failure analysis are fairly thorough. The experiments include mean pooling over shared encoded representations, a joint encoder, lexical methods, and an LLM judge, and also examine cross-construction transfer, clinical data scale, and downstream answer filtering. The authors explicitly report results that do not support the preset hypotheses, which helps to understand the scope of the method.
Critical weaknesses
  1. I am still not sure how far the contribution of the size-matching method can generalize. The size shortcut was mainly identified in the authors' initial deletion-based construction, and it needs further explanation of how prevalent it is in existing evaluations or in real memory processing. Existing benchmarks and baselines can support detection and component analysis on controlled samples, but the main experiments are still based on curated multi-hop QA and artificially constructed memory, lacking samples produced by real search-agent retrieval, compression, or forgetting processes. On 2WikiMultiHopQA, several encoder methods already exceed 0.99 AUROC, while MEMSAFE achieves only 0.6393 when transferred to SQuAD 2.0, which also indicates that the current high scores are not yet sufficient to demonstrate broader applicability.
  2. The presentation of the method and experiments is relatively difficult to follow. The main text introduces multiple sets, integrity states, prediction targets, and risk metrics at once; although the appendix provides a notation table, reading still requires repeated back-referencing. Figure 1 uses Stage I/II and Stage A/B simultaneously, and the later text switches among evidence risk, reader risk, risk budget, and coverage, making it unintuitive what exactly different experiments control and measure. The paper needs a concrete example that runs through sample construction, model prediction, and answer filtering, rather than merely supplementing symbolic explanations.
Questions for authors
  1. How common is the construction that deletes supporting paragraphs without adding replacement paragraphs in existing sufficiency evaluations? Can the actual impact of quantity leakage in an existing protocol be demonstrated, to clarify the scope of this finding's contribution?
  2. Can the evidence sufficiency of memory after retrieval, compression, or forgetting be checked on a small batch of real search-agent trajectories, and the performance of different gates be compared? Would such results support the conclusions from the controlled experiments?
  3. Can a complete example be provided, showing the original memory, different integrity states, MEMSAFE's output, and the final decision to answer or refuse, while clearly distinguishing the 5% evidence-risk target, 5% coverage, and reader error?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.