acceptodds
Withdrawn from ICLR 2027

How Interaction History Shapes Evidence Use In Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

LLM agents rely on interaction history to support multi-step tasks, yet this history can change not only what information they access, but also how they use it in decisions. We investigate this process in MemSyco-Bench, where historical preferences conflict with task evidence. Our circuit-level analysis of LLM evidence use, combining input edits and activation interventions, yields three findings. First, preference and correct evidence retain opposing influences in later layer answer competition; preference’s larger state-change amplitude helps it prevail, rather than evidence being entirely suppressed. Second, repeated preference and middle-layer attention mass contribute to this amplitude advantage, linking historical reading to downstream decision effects. Third, early preference-content states weaken later scope-sensitive reading: reducing these states makes option queries more responsive to applicability, but does not strengthen scope’s absolute correction of the answer. Thus, content and scope interact during reading without ensuring appropriately constrained decisions. Matched tests show that early preference-state and scope-related state interventions shift answers in the same broad directions in Qwen3.5-4B and Qwen3.5-27B, without establishing an identical circuit. In a 300-instance behavioral screen, among cases made robustly correct by preference deletion, the fraction reversed by full history is lower at 27B than at 4B (18.3% vs. 24.1%), but remains nonzero. The tested 27B model is therefore not immune to this history-induced failure; the comparison does not isolate model scale as its cause.

Primary area:
interpretability and explainable AI
Keywords:
large language models, agent memory, mechanistic interpretability, activation patching

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_rx7hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies why LLMs follow a historical user preference over stronger task evidence. The authors select the Memory-Evidence Conflict task of MemSyco-Bench in an A/B format. The paper select 100 tasks where Qwen 2.5 choose the preferred option but answers correctly once the preference turns are deleted. Then stop the model from reading the preference turns or the correct evidence at middle layers, fixing the input to models. And then measure the changes of answer margins and explore where the changes causes in later layer states. The experimental results show that correct evidence affects the final asnwer of LLMs, but preference has much stonger effect on the final answer. The paper finds that the repeated preference and larger attention mass on the preference turns cause that.

Critical strengths
  1. The causal analysis is careful. Sites are selected on discovery scenarios and tested on held out scenarios, also consider the option order. Key effects are also compared with matched historical windows and equal-norm random directions, which makes the causal claims more convincing.
Critical weaknesses
  1. Limited novelty and contribution. The methods are standard activation patching and attention ablation. The main message, that a misleading context wins in later layers while the correct signal stays active, is close to prior findings on false demonstrations and sycophancy (Halawi et al., 2024; Wang et al., 2026). The paper cites these works only to say that the history setting is underexplored, and does not show how its mechanism differs from theirs.
Questions for authors

See weaknesses.

Reviewer_cNZ2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates how interaction history leads language models to favor user preferences over task evidence. It evaluates three Qwen models on 300 MemSyco-Bench instances spanning 50 scenarios, supplemented by mechanistic analyses on smaller subsets and targeted Gemma4 checks.

Critical strengths

The paper focues on an interesting question about how historical preferences interfere with evidence-based decisions in language models, which is meaningful to understand how context influence language models.

Critical weaknesses
  1. Answer-format dependence of the mechanistic explanation. The identified reading pathway relies on candidate-option positions introduced by the A/B reformulation, which are absent from the original free-form task. Behavioral agreement across formats does not establish mechanistic equivalence, leaving the explanation of the original failure insufficiently validated.

  2. Insufficient quantitative reporting and fragmented presentation. Quantitative results across experimental dimensions are insufficiently reported, and the fragmented organization makes the main findings difficult to assess.

Questions for authors

Please find the weakness above.

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

The paper investigates the competitive effects of user preferences and factual evidence on an LLM-based agent's recommendations. It reports that the preference effect is stronger, that that advantage is due to the repetition and the attentional weight the agent gives to preferences, and that helping the agent notice when a preference doesn’t apply doesn’t necessarily make it ignore that preference in decision-making. It identifies specific layers and modules involved.

Critical strengths

The methodology is careful about establishing causality. Negative results are reported honestly.

Critical weaknesses

The text is very difficult to follow.

It's not clear that the preference advantage comes from targeted attention rather than simply greater token length.

Most of the mechanistic analysis rests on few scenarios.

The broader significance is not apparent.

Questions for authors

Does the preference advantage remain when the preference and evidence regions are matched for token count and repetition?

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

Related papers

Discussion (0)

Sign in to comment.