acceptodds
Under review as a conference paper at ICLR 2027

Recursively Updated Summaries Can Amplify a Remembered Betrayal’s Effect on a Qwen3-8B Agent’s Stated Willingness

Abstract

Long-horizon language-model agents compress their interaction history into summaries. We ask whether a single remembered relational event changes an agent’s later willingness to rely on the offending party, and whether two common summarization procedures preserve or distort that change. In a scripted three-agent simulation on Qwen3-8B, with a scenario family (n = 30 lexical variants of one script) as the unit of inference and a single-candidate digit-logit rating as the instrument, a deliberate betrayal on day 4 lowers Alice’s willingness toward Bob by 0.46 on a 7-point scale; the gap between Bob and Charlie, the recipient of the disclosure, falls by 0.136 relative to neutral control (95% CI [−0.158, −0.117]; 30/30 families). An initial comparison suggested that both a recursively updated summary and a summary regenerated from the archive amplified this effect. A content audit revealed an instruction asymmetry between the summarizer prompts; in a preregistered replication with matched prompts the regenerated-summary result disappears on the checkpoint average, while a smaller recursive-summary amplification remains (−0.055, CI [−0.075, −0.037], 27/30 families negative; the originally secondary contrast). A preregistered sentence-swap study shows that the verbatim wording of the retained event sentence does not by itself produce this difference; the responsible feature of the surrounding text is not isolated, so we report a difference between recursively updated and regenerated memory states, not an isolated effect of recursion. Minimal edits to the memory text show the effect is bound to whichever name occupies the offender role and depends on the confidentiality-request clause; a per-candidate decomposition shows that compact “judgment” sentences widen the gap mainly by raising willingness toward the comparator, a caution for gap-based instruments. On the original corpus, explicit recall of the event stays intact. The instrument could not be validated on other model families in our calibration protocol, so all findings are limited to one model and lexical variants of one scenario.

Primary area:
interpretability and explainable AI
Keywords:
LLM agents, agent memory, memory compression, recursive summarization, relational memory, long-horizon agents, behavioral evaluation, counterfactual interventions

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_XcUsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studied a detailed problem that when LLMs agents compacting long horizon experiences repeatively, it can distort the behavioral influence of a remembered betrayal. They used Qwen3-8B as the base model to build a simulated scenario with 3 agents, together a 32-day interaction. The results show that the betrayal reduces willingness to rely on both participants in the disclosure

Critical strengths

The experiments is transparent. Analyses are preregistered, families are treated as the sole independent unit, and the authors report the collapse of their originally primary hypothesis rather than presenting only the surviving result.

Critical weaknesses

W1. The paper asks whether compression preserves or "distorts" a betrayal's influence, but never defines what an undistorted influence would be. Full history is treated as ground truth without justification. No human reference, task outcome, or normative account specifies how much an agent should lower its reliance after a deliberate breach. Without one, "amplification" is a label rather than a finding: a larger effect under compression could equally be read as better calibration.

H1 is also close to tautological. The model reads that Bob deliberately disclosed Alice's secret against her request, and then rates Bob lower. The interventions (Sec. 4.4) show the rating follows whichever name the sentence blames and whether the sentence mentions the confidentiality request. In other words, the model reads the sentence. The paper's own conclusion concedes that the result is "not ... evidence that recursion itself is the isolated causal variable" (Sec. 8). What remains of the title's claim is that two different texts produce different ratings, which is expected of any model that conditions on its input.

W2. There is one scenario. The 30 "families" are the same script with a topic noun substituted. Thirty-one of the 32 days are the identical sentence about a "routine, uneventful interaction." No character ever acts. The outcome is an expected digit computed from next-token logits, not a decision. The authors acknowledge that n=30n=30 "supports robustness across topics, not across scenario structures" (Sec. 3.1) and that the large dzd_z values "reflect how alike the families are" (Sec. 7). The p<10−6p<10^{-6} results therefore certify insensitivity to a noun swap: a single fixed stimulus treated as a sample [1].

The instrument does not extend beyond this one model either. It failed calibration on all three other models tried, and its redesign failed construct validity on Qwen3-8B itself (Sec. 6). By the authors' own account, a calibration failure is "the absence of a valid measurement" (Sec. 6). There is thus no valid measurement on any model except the one the instrument was developed on. First-token scores of this kind are also known to diverge from what instruction-tuned models actually generate [2], and the paper never checks a generated decision.

W3. The preregistered primary contrast (fresh−full) went from −0.135 to −0.004 after one clause was added to each summarizer prompt (Table 1). The claim now rests on the originally secondary contrast. Its size is −0.055 on a 7-point scale, below the paper's own preregistered ±0.10 tie band, with 25/30 families counted as ties (Table 1; App. F). The recursive−fresh difference is −0.224 at day 4 and approximately zero at days 8, 16, and 32 (App. F). The authors cannot attribute it to length, position, fabricated content, or update history (Sec. 7). When one prompt clause erases the primary result, the object of study is that prompt pair, not "recursive summarization."

References [1] H. H. Clark. The language-as-fixed-effect fallacy: A critique of language statistics in psychological research. Journal of Verbal Learning and Verbal Behavior, 12(4):335–359, 1973. [2] X. Wang, B. Ma, C. Hu, L. Weber-Genzel, P. Röttger, F. Kreuter, D. Hovy, B. Plank. "My Answer is C": First-Token Probabilities Do Not Match Text Answers in Instruction-Tuned Language Models. Findings of ACL, 2024.

Questions for authors

Q1. What is the correct behavior, and why is full history the reference? The paper calls a larger effect under summarization "amplification" and implies distortion, but never states what a faithful carry-over of a betrayal into later willingness should look like. If an agent ought to become more wary after a deliberate breach of confidence, a larger effect under compression could just as well be read as an improvement. Without a normative target, such as human judgments on the same scenario or a task where reliance has consequences, the sign of H2 has no interpretation. What is the target, and how would one know that a memory procedure got it wrong?

Q2. What hypothesis was at risk in H1? The model rates Bob immediately after reading a sentence stating that Bob deliberately broke Alice's confidence. A lower rating seems to follow from basic reading comprehension. The interventions confirm the rating tracks what the sentence says: swapping actors redirects the effect, and deleting the confidentiality clause weakens it (Sec. 4.4). What result would have surprised you? What does H1 establish about agents beyond the fact that an LLM conditions on its context?

Q3. Is "recursive vs. fresh summarization" a well-defined treatment? Adding one clause to each summarizer prompt eliminated the originally primary result. It also changed the length ordering, verbatim retention, and fabrication rates (Sec. 4.2, Table 3). This suggests the prompt wording, not the procedure, dominates the measured quantity. A claim about recursive summarization would then require results across a range of reasonable summarizer prompts and models. Why should the outcome for one matched prompt pair be read as a property of recursive summarization rather than of these two prompts?

Q4. What deployment does this simulation stand in for? Alice never makes a decision and Bob and Charlie never act. Thirty-one of the 32 days are the same sentence about a "routine, uneventful interaction," and the outcome is a stated rating, not a choice. Real long-horizon agent memory involves heterogeneous events, conflicting evidence, repair, retrieval, and decisions with consequences. Which of these features does the setup preserve, and why would the results transfer? The answer determines whether the recommendations Sec. 8 addresses to builders of agent memory are supported.

Q5. What construct is being measured, and why is the Bob−Charlie gap its measure? Betrayal lowers willingness toward Charlie, who did nothing wrong, by 0.33, versus 0.46 for Bob (App. C). The largest intervention effects come mainly from changes in Charlie's score (Sec. 4.5). If the construct is reduced reliance on the offender, Charlie is not a neutral comparator. If it is general wariness, a difference score is the wrong estimand. Which construct is intended, and why measure it as a gap between two affected parties?

Q6. What effect size would matter? The surviving effect (−0.055 on a 7-point scale) is smaller than the ±0.10 band your own preregistration treats as a tie (Sec. 3.2). What smallest effect of interest motivated the study, and does the result meet it?

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

This paper investigates whether summarizing an interaction history changes a language model’s stated willingness to rely on someone who previously betrayed a confidence. Using Qwen3-8B and 30 topic substitutions into one scripted scenario, the authors compare full histories, recursively updated summaries, and summaries regenerated from the archive. After correcting an instruction asymmetry, recursively updated summaries produce a small additional betrayal-related willingness gap relative to full history (−0.055 on a seven-point scale); regenerated summaries show no checkpoint-averaged difference. Sentence interventions investigate role binding, confidentiality information, and event wording.

The paper’s strongest contribution is its careful demonstration of how prompt differences and comparator effects can complicate conclusions about agent memory. However, the surviving empirical finding is too narrow, and its interpretation remains insufficiently resolved, for acceptance at a top general AI conference.

Critical strengths

he analysis uses scenario families as the inferential unit rather than treating question templates, anchors, and checkpoints as independent samples. Reporting raw effects alongside standardized effects is particularly appropriate given the low variability among these lexical variants.

Critical weaknesses

Narrow generalization is the major weakness. All findings use one model and 30 lexical variants of one scenario. Statistical consistency across topics does not establish robustness across scenarios or models.

Moreover, the calibration gates address useful properties, including anchor agreement and basic ordering. They do not establish that a 0.055-point change in a digit-logit expectation corresponds to a meaningful change in reliance decisions. All principal outcomes remain prompted ratings, and the alternative choice instrument did not pass calibration.

Questions for authors
  1. How stable is the matched recursive–full result across individual templates and forward versus reverse anchors?

  2. DO the observations hold true for other models?

  3. Do recall, appraisal, and belief probes on the matched corpus support the proposed dissociation?

Reviewer_iH85OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how memory compression affects the downstream behavioral willingness of a Qwen3-8b model to rely on a party after a scripted relational betrayal. Using a controlled 3-agent simulation (Alice, Bob, Charlie) across 30 lexical scenario variants, the authors measure Alice's willingness to rely on the offender (Bob) versus the disclosure recipient (Charlie). The study initially finds that both recursive and fresh summarization amplify the negative effect of betrayal compared to full history. However, a rigorous content audit reveals a prompt asymmetry. In a pre-registered replication with matched prompts, the fresh-summary amplification disappears, while a smaller, statistically robust recursive-summary amplification remains (Δ=−0.055\Delta = -0.055). Further ablation studies (sentence swaps, role swaps) show the effect is bound to the offender’s role and the confidentiality clause, though the exact mechanism driving the recursive amplification remains unisolated.

Critical strengths

The preregistered replications, family-level statistical inference (avoiding pseudo-replication), and the honest reporting of the initial prompt asymmetry flaw are exemplary.

The decomposition of the "willingness gap" is a noticeable strength. Showing that "judgment-only" edits primarily affect the comparator (Charlie) rather than the target (Bob) is an interesting methodological contribution for anyone using gap-based metrics in LLM evaluation.

The deterministic text interventions (actor swap, facts-only, judgment-only) cleanly isolate that the effect is bound to the offender role and the confidentiality-request clause, ruling out simple "forgetting" as the cause.

The self-correction in Section 5 (excluding circular willingness scores and adjusting permutation baselines) shows a good understanding of the pitfalls of representation probing.

Critical weaknesses

These aren't critical but rather tied to the limited scope of the work. In essence, the findings are strictly limited to Qwen-3-8b. The authors' attempts to extend the instrument to Mistral-7B, Phi-4-mini, and Y-1.5-9B failed calibration. While honestly reported, it means the phenomenon's generalizability across architectures is currently unknown.

As the authors rightly note in the limitations, the operative variable driving the recursive/fresh difference is not isolated. The recursive summary is longer, fabricates filler events, and has a different update history. The paper identifies a difference between the states, but cannot pinpoint the causal feature of the recursive process.

The surviving recursive-full effect is −0.055-0.055 on a 7-point scale. While statistically robust across the 30 families, its practical significance for real-world, long-horizon behavior is debatable.

The 30 families are merely lexical variants of a single scripted scenario with identical routine filler. Real-world agent logs are highly heterogeneous, and a summarizer facing diverse, noisy logs may behave differently.

Questions for authors
  1. While the authors correctly state that verbatim retention does not by itself produce the difference, do the authors have a leading hypothesis for what does? For instance, does the compounding of the "fact preservation" instruction over multiple recursive updates cause the model to over-index on the negative social gist of the day-4 event compared to a single fresh pass?

Could you elaborate on why the digit-logit instrument failed for models like Mistral or Phi-4? Was it a tokenization issue (e.g., the digit tokens not being single, distinct tokens), a divergence in instruction-following for the strict "output only a digit" constraint, or something else? A brief diagnostic insight here would be highly valuable to the community.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.