acceptodds
Under review as a conference paper at ICLR 2027

First Impressions, Lasting Investments: Social Impressions Shape Dynamic Trust Adaptation in LLMs

Abstract

As large language models (LLMs) increasingly engage in sustained social interactions, their impressions of others can shape consequential decisions. Existing work has examined social reasoning and interactive behavior, yet how LLMs translate social impressions into investment decisions and adapt these decisions as new feedback accumulates remains unclear. We introduce a two-stage experimental paradigm that manipulates partner honesty to induce social impressions, then tracks investment decisions in repeated trust games. Across model families and scales, medium-to-large models (14B-32B Qwen models and DeepSeek V4 Pro) initially invest less in less honest partners. As favorable feedback accumulates, investments generally rise and stabilize, while honesty-related differences persist. Generating and retaining an explicit partner evaluation can amplify this initial honesty penalty and slow the convergence of investment in dishonest partners toward its asymptote. However, the two smallest models (4B and 9B) show no reliable positive initial penalty even after an explicit partner evaluation. To investigate this missing penalty, we apply linear probes at four points across the paradigm to assess impression formation and retention. Impression-related information is already decodable at the end of Stage 1 and remains so before the first investment. Prompt cues that retrieve the prior evaluation and clarify its relevance to investment can induce the honesty penalty, supporting a gap between available social impressions and their spontaneous use in investment. These findings identify the translation of social impressions into adaptive behavior as a useful lens for understanding LLM social intelligence and provide a dynamic framework for studying social decision-making in sustained interactions.

Primary area:
interpretability and explainable AI
Keywords:
Large Language Models, Social Intelligence, Social Impressions, Social Decision-Making, Adaptive Behavior

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_fDB5OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how LLMs use their impressions of others when making decisions over time. The authors introduce a two-stage experiment where models first observe a partner's honesty and then play 20 rounds of a trust game with that partner. They find that larger models initially invest less in dishonest partners, but increase their investments after receiving positive feedback. Explicitly asking models to evaluate their partners makes these initial impressions more influential. For smaller models, the authors show that information about honesty is present in their internal representations but does not always affect their decisions. Additional prompts can help these models use this information.

Critical strengths
  1. The two-stage experimental design separates impression formation from later decision-making, which helps study how past social information affects future behavior.
  2. The paper studies trust as a changing process rather than relying on a single decision, which I found to be interesting.
  3. The comparison between explicit evaluation, no evaluation, and a color control is well designed and helpful.
  4. The combination of behavioral experiments, internal representation analysis, and prompt interventions is very helpful.
Critical weaknesses
  1. The model comparison is heavily focused on Qwen models. Six of the seven models belong to the Qwen family, making it difficult to know whether the findings hold across different model families. Testing additional model families would strengthen the general claims.
  2. The honesty penalty is not necessarily a measure of better trust decisions. Since partners return the same favorable amount regardless of their honesty, investing less in dishonest partners may actually reduce the model's payoff. The paper should analyze the trade-off between using past impressions and maximizing rewards.
  3. The paper does not clearly separate learning from reaching the maximum investment. Some models quickly reach the 10-token limit, which can hide differences in how they update their decisions. An additional analysis that accounts for this ceiling would make the results clearer.
  4. The probing results may partly reflect information about honesty levels rather than a learned social impression. The probes predict a score that is itself closely related to the observed honesty history. Comparing against a simple honesty-count baseline would help establish what additional information the probes capture.
  5. The prompt interventions may directly encourage cautious investment rather than simply help models use existing impressions. A control prompt that mentions caution without referring to the partner's previous behavior would help separate these effects.
  6. The paper reports average investment patterns but gives less attention to differences across individual trials. Analyzing how often models actually change their investment strategy, rather than only comparing average trajectories, would help establish how consistent the findings are.
Questions for authors
  1. Does the order of honest and dishonest reports in Stage 1 affect the initial investment? For example, do models give more importance to recent behavior?
  2. Have the authors examined whether the explicit partner evaluations are consistent across trials? Could differences in the words models choose help explain the variation in investment behavior?
Reviewer_BXC6OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines if LLMs use information about the trustworthiness of their partner when making decisions regarding investments. They find that larger models attenuate their investments with less trustworthy partners, but smaller models do not. They find that, despite the fact that the small models do not seem to respond to trustworthiness, their partner’s trustworthiness is decodable from their activations, suggesting that the information is read but not used. Finally, they find that additional prompt cues can induce the small models to attenuate investments with untrustworthy partners.

Critical strengths

The paper asks an important question about the social intelligence of LLMs, their ability to effectively use information about their interlocutors in determining how to interact with them. The difference they find between large models and small models appears substantial and robust (although testing on a yet wider set of models would help confirm the small model vs large model difference).

Critical weaknesses

1.) I believe there is insufficient evidence that the linear probe is capturing information about the partner’s honesty. There are major lexical differences between the honest and dishonest contexts. Most notably, the word TRUE or FALSE each appear between 0 and 12 times depending on the level of honesty. The activations could just be tracking how many times the word FALSE appears in context – we do not know that this is mapping on to an assessment of the partner’s honesty.

2.) We do not know that the additional cues in the prompt are “promoting the behavioral use of social impressions." It may be that the prompt is having its own effect, independent of prior social impressions. This is especially true of the “impression relevance cue” which states “You may take your prior assessment of this same partner into account when deciding how cautiously to interact.” This is essentially providing instructions to the model to consider acting more cautiously.

3.) When activations are extracted after the adjective is generated, that generated adjective is in context (for T2 and T3). This is creating a confounding issue for the correlation between probe’s predicted score and the social impression score (Figure 4) because the social impression score is based on logits for adjective generation. The same tokens that signify a dishonest social impression are therefore in context for the probe. This means the probe may be acting as a “dishonesty word” detector rather than actually picking up on the honesty of the partner. Does the association weaken if you redact the adjectives when extracting activations (note that this only applies to T2 and T3).

Questions for authors

1.) The difference between large models and small models in their ability to make use of social impressions is based on a very small sample size of models. Testing more large/small model pairs would strengthen this claim.

2.) I would like to see the results for the linear probe if you removed lexical differences between honest and dishonest conditions. This means that the output could not print TRUE or FALSE, it would simply state what the partner reported and then reveal the actual choice. You would also need to be careful about balancing frequency of A, B, and C tokens between the honest and dishonest conditions so the probe would not pick up on those frequencies. Does this still capture differences between honest and dishonest partners?

Reviewer_s6cvOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates how social impressions influence trust-based decision-making and behavioral adaptation in large language models (LLMs). The authors introduce a two-stage experimental framework in which LLMs first observe partners with different honesty histories and then participate in repeated investment games with those partners under matched economic returns. Results show that medium-to-large models generally invest less in previously dishonest partners, gradually increase their investments following positive feedback, but often retain persistent honesty-related differences. Explicitly evaluating a partner strengthens initial impressions and can slow subsequent behavioral adaptation. Interestingly, smaller models exhibit an impression-to-action gap: although their internal representations retain information about partner honesty, they fail to consistently use it in investment decisions. Targeted prompts that retrieve prior impressions and emphasize their relevance can partially bridge this gap.

Critical strengths
  • I think the two-stage paradigm is a significant methodological strength. By systematically manipulating verified partner honesty while holding subsequent investment returns constant across conditions, the authors isolate the influence of prior social information on initial decisions and subsequent behavioral adaptation. The inclusion of both no-evaluation and color-control conditions further strengthens the interpretation of the explicit-evaluation effect. This careful design, together with paired comparisons and robustness analyses under variable returns, provides convincing evidence that social history can influence LLM investment behavior independently of economic feedback. It addresses an important confound in prior studies of LLM trust and strategic interaction.
  • The combination of behavioral experiments, hidden-state probing, and targeted prompt interventions is particularly valuable. The finding that smaller models retain highly decodable impression-related information immediately before investment, despite failing to exhibit corresponding honesty-sensitive behavior, provides an interesting perspective on LLM social reasoning. The recovery of honesty-sensitive investment through retrieval and relevance cues further supports the proposed impression-to-action gap.
Critical weaknesses
  • Regarding the impression-to-action gap, the current analyses do not fully establish its proposed interpretation. Because the complete honesty history remains available in context, successful linear probing may reflect decodability of the underlying factual history rather than retention of a distinct social impression representation. Furthermore, the independent probes across task positions demonstrate information availability but not its causal relationship to the investment policy. Similarly, retrieval and relevance cues may directly alter investment behavior through task framing or instruction effects rather than specifically activating previously formed impressions. This limitation is consequential because it leaves uncertainty about whether the experiments identify a distinct impression-to-action failure mechanism or a more general difficulty in using contextually available information
Questions for authors
  • Since the complete Stage 1 honesty history remains in context throughout Stage 2, how can the authors distinguish the decoding of a retained social impression from the decoding of factual honesty information directly available in the conversation? Are there additional controls or analyses that demonstrate that the decoded impression-related representation, rather than direct access to the history, is specifically implicated in the observed behavioral gap? Addressing this distinction could materially strengthen the paper's central mechanistic interpretation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.