acceptodds
Under review as a conference paper at ICLR 2027

Behavior Underspecification in Large Human Language Models

Abstract

Large human language models (LHLMs), language models designed to simulate human responses, are increasingly used as scalable simulators of human behavior. Given a description of an individual, an LHLM generates responses from that person’s perspective, enabling applications in behavioral science, marketing, public policy, and beyond. But what kind of human behavior do they actually model? Behavioral science has long documented a consequential gap between what people say they do and what they actually do. We hypothesize that LHLMs may not reliably distinguish between these constructs because their training data disproportionately capture what people say, while their training objectives do not explicitly differentiate among them. We refer to this ambiguity as behavior underspecification. To study it, we evaluate ten contemporary LHLMs on a cross-domain dataset comprising 98 matched measures of what people said and eventually did, derived from more than 37K human responses. With experiments, we find that behavior underspecification is widespread, that LHLMs model what people say better than what people do, and predictions of what people do often reflect what people say. The effect of behavior underspecification persists even in otherwise accurate models. Fine-tuning on human-response data does not necessarily resolve this problem and can sometimes exacerbate it. Together, our findings show that evaluating LHLMs requires asking not only how accurately they simulate human responses, but precisely what aspects of human behavior they simulate.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
human behavior modeling, large human language models, say-do gap

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_Koy7OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores the discrepency between how well LLMs predict what people say versus what people do. It introduced the concepts of fidelity asymmetry and target ambiguity as indicators of "behavioral underspecification." It then conducted hypothesis tests on seven general-purpose models and three fine-tuned models with a constructed dataset. The results reveal that behavioral underspecifications as measured by both fidelity asymmetry and target ambiguity are common, and fine-tuned human-behavior models do not consistently reduce this phenamenon.

Critical strengths
  1. The research question is important in this domain, it also has many implications such as how to fine-tune LLMs to better distinguish what people say vs what people do, and how do we better frame behavioral tasks to the models.
  2. The problem is well formulated, with formal definitions and notations clearly given.
  3. The application of behavioral science theory, the theory of planned behavior, to alleviating the fidelity asymmetry seems to be inspiring and promising.
Critical weaknesses
  1. Evaluation data: Though the dataset contains 98 matched say-do pairs across multiple domains and over 37,000 human responses as described, it seems to be limited in its form. If I understand correctly, the "do" data here are all binary voting records. As voting is highly contextualized, and follows a very specific process, it may hardly cover the vast behavioral space and could introduce bias using it represent human behavior, as this paper scoped. I think the authors could consider narrowing this paper's scope into human's voting behavior or including more diverse datasets.
  2. Evaluation metrics: While the formula supports arbitrary distributional distance metrics, the experiments used only the TVD. It might be helpful to try another distance metric as a robustness check.
  3. The discussion of fine-tuned models lacks formal evidence support. The comparison between fine-tuned and general-purpose LLMs are informative, but with the fine-tuning data and process varying a lot, it is difficult to conclude the dynamics of behavior under specification and the impact of fine-tuning. I think the paper could benefit from having at least one controlled experiment, where an LLM is particularly fine-tuned by what people say vs what people do in the dataset under a controlled process. With this controlled experiment, it would be clearer how fine-tuning influences.
  4. There are also models particularly tuned on what people do. The authors may consider including them into the comparison:

[1] Lu et al. "Can LLM Agents Simulate Multi-Turn Human Behavior? Evidence from Real Online Customer Behavior Data." Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

[2] Xie et al. "Be. fm: Open foundation models for human behavior." arXiv preprint arXiv:2505.23058 (2025).

[3] Lei et al. "HumanLLM: Towards personalized understanding and simulation of human nature." Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 2026.

Questions for authors
  1. Since there are several categories of say-do pairs in the dataset, I'm curious how's the results differ among the categories? Is there a category where models tend to be more behaviorally underspecified?
  2. Page 3 remarks that this paper focuses on the behavioral underspecification at the population level. I wonder even though we don't consider the individual-level heterogenity at this stage, how will the individual-level results differ from what we have at the distributional level?
2 comments

I appreciate the authors' efforts in responding to my concerns and providing the additional results. However, I think my concerns on the evaluation data and metrics still remain. I will keep my score.

We appreciate the reviewer's consideration of the additional results and explanations. It would be great if the reviewer could clarify what concerns still remain. To summarize, we:

•⁠ ⁠Clarified that “voting" is only one of the 98 say-do pairs we study

•⁠ ⁠⁠How TVD is the most appropriate, and widely used metric for this task (equivalent to Wasserstein in our formulation)

•⁠ ⁠⁠Identified ways in which fine-tuning can mitigate the problem

•⁠ ⁠⁠Discussed how behavioral underspecification is robust to the type of say-do pair

•⁠ ⁠⁠Highlighted that behavioral underspecification is even more prominent of models are evaluated at the individual response level

We are happy to address any questions that still remain unanswered.

Thanks again for taking the time to review our paper and go through our rebuttal.

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

This paper studies behavioral underspecification in large human language models, focusing on whether they distinguish between what people say they will do and what they actually do. It evaluates ten models on 98 matched say–do pairs from more than 37K human responses and finds that models generally predict stated behavior more faithfully than realized behavior.

Critical strengths
  1. The paper identifies an important and underexplored evaluation problem by separating stated and realized behavior rather than treating human-response simulation as a single target.
  2. The empirical study is broad and carefully constructed, covering ten models, multiple behavioral domains, and matched human say–do data with detailed robustness and subgroup analyses.
Critical weaknesses
  1. The definition of realized behavior is not always strong. Most do targets are self-reported actions, and only 6.1% use external records or incentivized measures. Moreover, 81.6% of the pairs measure say and do concurrently. This makes it unclear whether the benchmark consistently captures what people actually do rather than another form of self-report.
  2. The paper attributes the larger error on do questions to behavioral underspecification, but realized actions may simply require more information than stated intentions or preferences. The model is mainly given demographic information, while actions can depend on cost, opportunity, and other situational constraints. The large improvement after providing the person’s say response also suggests that missing information may explain part of the gap.
  3. The definition of target ambiguity may be affected by statistical power. A target is treated as ambiguous when the analysis fails to show that the prediction is significantly closer to its intended target. However, failure to establish a distinction does not necessarily mean that the model cannot distinguish the two targets. The paper itself shows that smaller subsets are more likely to be classified as ambiguous.
Questions for authors

See the weaknesses above.

Reviewer_5PwQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

Simulating human responses is a growing application domain for LLMs. Since human respondents show differences in what they say they do and what they actually do, this paper asks which of the two LLMs best represent. They build a dataset of 98 say-do pairs, each question binarized. They then compare population-level yes rates against 10 language models, introducing two new metrics characterizing 1) differences in error across “say” and “do” questions, 2) higher similarity of models’ “say” and “do” to the human equivalent rather than the other one. Authors conclude that models predict what people say more accurately than what they do, including those fine-tuned to simulate human responses.

Critical strengths

The motivation for the work and the exposition in the first 4 and half pages is clear and well-written.

Critical weaknesses

The results section (page 5 and a half onwards) lack the clarity in exposition of the earlier pages. Instead of dumping all results into a single table (Table 1) and very verbose results subsections, it would be better to guide the reader by having each Table/Figure have one key take away point.

The metrics considered are flawed in important ways. Authors take ΔFA<0 as “fidelity to say rather than do” but it can also be consistent with “always yes” behavior, e.g., for psay=0.7p_{\text{say}} = 0.7, pdo=0.5p_{\text{do}} = 0.5.

Modelpsay′p'_{\text{say}}pdo′p'_{\text{do}}ΔFA\Delta_{\text{FA}}ΔTAsay\Delta^{\text{say}}_{\text{TA}}ΔTAdo\Delta^{\text{do}}_{\text{TA}}
Say-copy0.70.70 − 0.2 = −0.20 − 0.2 = −0.20.2 − 0 = +0.2
Always yes1.01.00.3 − 0.5 = −0.20.3 − 0.5 = −0.20.5 − 0.3 = +0.2

Authors should at minimum include trivial baselines such as always-yes, say-copy, a 50/50 baseline, and a constant predictor at the pooled human mean. ΔTA is interpreted by authors as “can’t tell do from say” (L108-L109, L205-207, L212-215, L323-325, L328, L332, …), however, models which preserve the human do-say gap but with a constant shift may similarly be scored as one-sided TA:

Modelpsay′p'_{\text{say}}pdo′p'_{\text{do}}Model gapΔTAsay\Delta^{\text{say}}_{\text{TA}}ΔTAdo\Delta^{\text{do}}_{\text{TA}}Label
no say/do distinction0.80.800.1 − 0.3 = −0.20.3 − 0.1 = +0.2one-sided TA (do ambiguous)
correct gap, shifted up by 0.150.850.650.20.15 − 0.35 = −0.20.15 − 0.05 = +0.1one-sided TA (do ambiguous)

Authors could instead directly compare the models’ say-do gap with that of the human respondents. More broadly, the metrics in Section 4.3 add little, potentially being more indictive of test power than genuine differences (e.g., TA detected is just failure to reject!). Effect sizes with CIs are in my opinion more meaningful.

It is unsurprising that fine-tuning on what people say need not help predict what they do. However, I imagine that a little bit of fine-tuning on both what they say and what they do could go a long way. Such fine-tuning experiments would strengthen the contributions of the paper (it need not be on a large model). Section 5.2.3 would be stronger if, rather than requiring actual ground truth, authors would first ask the model to predict “say” and then use that to predict “do”. They suggest this but do not run the experiment.

Questions for authors

Not questions but rather minor comments.

Minor: L78-L79 is stated without any evidence, even naively training on adequately worded survey questions would mean that training does differentiate between the two.

Minor: 98 say-do pairs are little, but I understand the challenges in sourcing such datasets.

L115-L188: What prevents authors from also studying the individual level behavior (e.g., evaluating the do-say accuracy gap).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear Area Chair,

Thanks for your time reviewing our work. During the rebuttal period, we posted detailed responses to all three reviewers addressing all their concerns, followed up during the discussion period, and checked in again on its final day. Reviewers 5PwQ and FgwS unfortunately have not responded. We believe we have addressed their concerns and questions and would like to summarize our rebuttal below.

Reviewer 5PwQ

  • Trivial baselines. We evaluated all four requested baselines (always yes, say-copy, 50/50 and the pooled human mean) with the same metrics and tests as Table 1. Our metrics correctly flag these baselines, and none of the ten models resembles any of them in its predicted "yes" rates or their variation across pairs.
  • Meaning of target ambiguity. We clarified that the metric describes where predictions fall, not whether a model understands the difference between say and do, and we will revise the wording throughout the paper accordingly.
  • Comparing the say-do gap directly. We showed that reproducing both rates of a pair is much harder than predicting each separately, and related this to our Joint paradigm, where models asked for both answers at once exaggerate the gap.
  • Effect sizes. As the reviewer recommended, we reported every contrast with 95% confidence intervals, which support the same conclusions as our tests.
  • Fine-tuning on both say and do data. We ran a new controlled fine-tuning experiment varying the share of say and do training data. Mixtures of the two reduce both errors, with do-heavy mixtures giving the best do predictions.
  • Conditioning on the model's own say prediction. We ran the requested experiment for five models. It helps, but much less than conditioning on the real answer, and we showed that errors in the say prediction carry over into the do prediction.
  • Individual-level analysis. We evaluated each person's own answers and found behavioral underspecification for all ten models, more consistently than at the population level.
  • Evidence for L78-L79. We added supporting literature and will present this explanation as a hypothesis.

Reviewer FgwS

  • Self-reported do measures. We explained that every do question asks factually about what the respondent has done, whereas every say question asks about values, attitudes, beliefs or intentions not tied to an actual event, and that all pairs were screened for this distinction with expert behavioral scientists.
  • Whether do questions need more information. We explained why conditioning on demographics is standard practice and why population-level evaluation reduces the influence of individual situational factors.
  • Target ambiguity and statistical power. As the reviewers recommended, we reported 95% confidence intervals for every contrast and committed to revising the wording.

Reviewer Koy7

  • Evaluation data. Only one of the 98 pairs concerns voting. The pairs cover seven domains and include self-reports, official records and incentivized tasks, and we gave examples from each domain.
  • Metrics. TVD is the natural distance for binary outcomes, is widely used in prior work, and is identical to the Wasserstein distance in our setting.
  • Controlled fine-tuning. We ran the requested controlled experiment on the balance of say and do training data.
  • Behavior-tuned models. We already include Centaur, and we reviewed the three suggested models. One targets a much narrower setting (online shopping actions), one has no publicly released weights and is trained mostly on stated responses, and we have requested access to the third.
  • Differences between categories. Appendix D shows that behavioral underspecification appears in every category of say-do pair, demographic group and domain we can test.
  • Individual-level results. Behavioral underspecification is even more consistent at the individual level.

We would be grateful if these responses could be taken into account in the final assessment. Thank you for your time and for handling our submission.

Sincerely, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.