acceptodds
Under review as a conference paper at ICLR 2027

Investigating Tool-Memory Conflicts in Large Language Models

Abstract

Tool use is a defining capability of LLM-based agents, enabling them to act beyond their parametric knowledge and interact with external environment. Yet this reliance on external tools introduces a new vulnerability: the knowledge returned by tools may directly contradict an agent's internal memory. In this paper, we identify and systematically study this phenomenon, termed Tool-Memory Conflict (TMC), a previously unexplored class of knowledge conflict for LLMs. We find that even state-of-the-art LLMs struggle under TMC, with notable performance degradation on STEM-related tasks. We further reveal that LLMs prioritize tool knowledge and parametric knowledge differently depending on factors such as task domain and conflict severity. We evaluate existing conflict-resolution techniques, including prompting-based and retrieval-augmented methods, and show that none effectively resolve tool-memory conflicts, highlighting TMC as an open and critical challenge for building reliable LLM-based agents.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Knowledge Conflict, LLM

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

This paper studies what the authors call tool-memory conflict (TMC): cases where a tool-augmented LLM gives one answer when relying only on its parametric knowledge and a different answer when it is forced to use tools. The authors measure how often such conflicts occur across a set of proprietary and open-source models on math, medical, factual QA and general knowledge benchmarks. They then test whether several existing knowledge-conflict resolution methods can pick the right answer once a conflict is exposed, and probe whether models systematically lean toward memory or toward the tool. The main takeaways are that conflicts are common and vary a lot by model and domain, that conflict frequency and resolution ability are largely unrelated, that giving the model the full tool trace helps more than longer reasoning, and that most models show a mild preference for their own memory.

Critical strengths
  1. The problem is timely and practically relevant. As more agents call tools autonomously, how a model arbitrates between its own knowledge and tool outputs is a real question, and it has received surprisingly little direct attention.
  2. The distinction from context-memory conflict in Appendix C is well thought out. The point that tool evidence is produced by the model’s own actions, and therefore carries its own error chain and provenance, is a useful conceptual contribution and deserves a place in the main text.
  3. The finding that low conflict rate does not imply good conflict resolution is interesting. The observation that the strongest model rarely conflicts but resolves conflicts poorly, largely by sticking with its own answer, raises a real question about self-trust calibration in strong models.
Critical weaknesses
  1. The text and the tables are inconsistent in many places, to the point where I’m not sure which results are current. Section 5.3 refers to thirteen models, including one that never appears in any table, and nearly all of the bias numbers quoted there differ from Table 4. Several other statements also contradict the tables: the conflict rate given for Claude Opus 4.7 in Section 5.1, and the claim that Full Trace is the best method for every model, which does not hold for GPT-5.5 or Qwen3 8B. Together with the unresolved “Table ??” / “Fig. ??” references, this makes it hard to trust the analysis as written.
  2. The way conflicts are counted may not match the definition in Eq. 3. On SimpleQA, the “both wrong” share of conflicts is exactly zero for every model, while “both wrong” cases among non-conflicts are not. For free-form answers it is very unlikely that two wrong answers always coincide, which suggests conflicts were determined from correctness labels rather than from whether the two answers actually differ. If so, the conflict statistics need to be recomputed.
  3. The resolution results lack simple reference points. Since memory-wrong/tool-correct cases make up roughly half of all conflicts, a trivial “always pick the tool answer” rule would land around 50% accuracy, which already matches or beats several of the evaluated methods. Table 5 also shows that Full Trace has a strong tool bias, so part of its advantage may come from deferring to the tool rather than from better reasoning. Always-tool, always-memory and oracle baselines are needed before any claims about method quality can be made.
  4. The experimental setup doesn’t really test the scenarios that motivate the paper. The introduction emphasizes stale knowledge, code execution and unreliable tools, but the experiments use static benchmarks where many conflicts come from weak tool use (e.g., smaller models losing a lot of accuracy on SimpleQA once they search). This makes TMC easy to read as a tool-reliability problem that better engineering could fix, rather than an arbitration problem. Controlled experiments that vary the tool error rate, or use oracle and corrupted tool outputs, would separate these much more clearly.
Questions for authors
  1. How are instances where both answers are wrong handled in resolution accuracy? Under Eq. 5 they can never count as appropriate, so the achievable ceiling differs by model. What do the resolution numbers look like with these cases excluded?
  2. In tool-only mode, how often does the model’s final answer actually differ from what the tool returned? It would help to know how much of a_t reflects tool output versus the model’s own interpretation layered on top. A breakdown of tool-side errors (wrong tool, wrong arguments, wrong result, misreading) would also be informative.
  3. Do you have any explanation for GPT-5.5’s strong memory bias? For example, is it more likely to override the tool when its memory answer is confident or consistent across samples, and does this hold across domains?
Reviewer_31eSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces Tool-Memory Conflict (TMC), which occurs when an LLM's internal knowledge disagrees with its tool-assisted answer. It proposes an evaluation framework to measure conflict frequency, source bias, and resolution accuracy. Through experiments across multiple LLMs and QA benchmarks, the authors demonstrate that TMC is model- and task-dependent, and that existing conflict-resolution methods remain unreliable. The results also suggest that providing full tool traces improves conflict resolution. However, the evaluation primarily focuses on single-turn QA and post-hoc resolution, leaving open questions about whether it captures genuine source arbitration in realistic agent workflows.

Critical strengths

1.The paper identifies Tool-Memory Conflict (TMC), an important yet underexplored challenge in tool-augmented LLM agents, where internal knowledge may conflict with external tool outputs.

2.The paper provides a systematic analysis of TMC across diverse models and tasks, revealing meaningful insights into source biases and the limitations of existing conflict-resolution methods.

Critical weaknesses
  1. The evaluation protocol leaves some uncertainty about answer matching and the distinction between tool-use errors and genuine tool–memory conflicts, making the results harder to interpret and reproduce.

  2. The single-turn setting limits conclusions about dynamic evidence arbitration, while the resolution gains may partly reflect an additional reasoning step rather than conflict handling specifically.

Questions for authors
  1. The paper defines memory as internal parametric knowledge. Could the authors evaluate how providing additional task-relevant context as in-context memory affects TMC rates and resolution performance?

  2. Why is the memory-only response treated as internal knowledge rather than reasoning ability? A tool-simulation baseline through CoT could help distinguish source bias from reasoning limitations.

3.Would TMC still occur if the resolution method of TMC were incorporated directly into the memory-only turn or tool-only turn?

4.The unresolved cross-references (“Table ??” and “Fig. ??”) in Line 262 should be corrected.

5.Does the evaluation distinguish errors in tool selection, argument generation, execution, and result interpretation? How are these factors accounted for when identifying TMC?

6.How are ama_m and ata_t compared? Are exact matching, semantic equivalence, or LLM-based judgments employed? How are semantically equivalent answers handled?

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

This paper investigates whether conflicts arise between tools and memory in large language models, i.e., cases where an LLM's answer based solely on its internal memory is inconsistent with the answer it derives with the help of external tools. The authors measure the conflict rate across 11 LLMs and 6 benchmarks, examine both memory bias and tool bias, and evaluate six existing resolution strategies, finding that providing the full tool execution trace (Full Trace) performs best.

Critical strengths

Valuable Empirical Insights: Evaluating how tool-augmented LLMs arbitrate between internal memory and external tool outputs reveals a critical failure mode, particularly the finding that low conflict frequency does not imply strong conflict resolution capability.

Critical weaknesses
  1. Inconsistencies Between Text and Tables: The manuscript appears incomplete and contains major discrepancies between the main text and experimental tables. In Section 5.3, the narrative claims to evaluate 13 models and discusses Qwen3-Coder 30B-A3B, whereas Table 4 only lists 11 models and omits Qwen3-Coder entirely. Moreover, the specific values analyzed in Section 5.3 do not match Table 4 (for instance, the text states that GPT-5.5 has a tool bias of 1.1%, memory bias of 38.3%, and bias gap of -37.2, while Table 4 reports 7.3%, 57.7%, and -50.5; similar mismatches occur for Claude Opus 4.7 and Llama-3.1-8B). Unresolved LaTeX references (Table ?? and Fig. ?? on pages 5 and 7) and typographical errors (agentic syg s.) further undermine the reliability of the work.
  2. Limited Methodological Contribution and Artificial Evaluation Setup: The paper benchmarks existing prompting strategies without proposing a new conflict resolution method, as the best-performing Full Trace baseline simply retains standard tool execution logs in the prompt. Furthermore, the evaluation protocol generates decoupled memory-only and tool-only answers independently and asks the model to choose between them side by side, which differs from realistic agent workflows where tool invocation and verification happen sequentially.
Questions for authors

See the weakness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.