acceptodds
Under review as a conference paper at ICLR 2027

Evaluating Alignment of Behavioral Dispositions in LLMs

Abstract

As people turn to LLMs for social advice, understanding their behavior in such contexts becomes essential. In this work, we focus on behavioral dispositions: the underlying tendencies that shape responses in social contexts. We introduce STAR, a framework for studying how closely the dispositions expressed by LLMs align with those of humans. STAR builds on established psychological questionnaires, adapting their items into realistic advice-seeking scenarios, as self-report may not transfer to actual advisory behavior. Using STAR, we construct a dataset of 23k scenarios, each validated by 3 raters and annotated with preferences from 10 participants. Across 25 LLMs, we find that (1) when human consensus is high, frontier models can fail to reflect it in 15–20% of cases, and smaller models fail at substantially higher rates; (2) when humans disagree, LLM recommendations are substantially less diverse than human choices, both within individual models and even across models from different providers, potentially narrowing the range of options users are guided toward; (3) LLMs' self-reported values are poor predictors of their recommendations. To support future research we make our dataset and code publicly available.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Behavioral Alignment, Situational Judgment Tests, Benchmarks, Computational Social Science

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_DjSTOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper releases a corpus of 2k scenarios developed from existing psychological questionnaires that focus on four behavioral dispositions, allowing comparison between the self-reported and true dispositions expressed by LLMs. They get these scenarios annotated by 10 people each, which provides a baseline of how real users would respond and allowing for direct comparisons against model advice. They find that models are misaligned with users, especially when there is greater disagreement between users, and even when there is relative consensus amongst users, models demonstrate less consensus. In my opinion, the biggest contribution is the dataset.

Critical strengths

The dataset is genuinely very useful and impressive in its scale - having 23k annotations is an expensive and difficult task and I appreciate that this will enable a lot of future research. Their analysis also very sound, barring a few concerns that are more rooted in the dataset construction than the analysis itself. The writing is good, the process is clear, and the dataset and code are easy to follow.

Critical weaknesses
  1. The core argument of your paper rests on the idea that high human consensus should imply high model consensus and low human consensus should imply low model consensus. I'm not sure if that presupposition is necessarily true. For one, your self-described limitation of focusing on US and UK users heavily skews the results since these social norms are highly contextual. While these models might perform with less consensus, that is likely intentional and due to years of work encouraging model creators to be less western-centric in the values. I think it's slightly concerning to suggest that more consensus in these cases is preferable. I also wonder whether the static wording of the prompt increases consensus in other cases. Perhaps the background of the user, induced via history, memory, or even linguistic signals (which are now more and more common) would better align the systems to specific users.
  2. I see that you chose 4 traits to focus on, which is understandable. I am confused why you picked these four when the taxonomy already has 4 second-order groups, of which these four only cover 3. Emotional regulation and impulsiveness are both part of the "self-control" group while "well-being" is omitted entirely. If this is intentional, I would like to know why and have that reflected in the paper.
Questions for authors
  1. In the abstract you claim "23k scenarios" but my reading suggests you have actually about 2.3k scenarios with 10 annotations each. Can you clarify?
  2. How were annotators recruited? How were they paid? What platform did they work on? How were they filtered? How many samples did each user do?
  3. Similar questions ^ for the human validation of the SJT samples.
1 comment

I thank the authors for their responses! I have updated my rating.

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

This paper explains a framework for comparing distributions of traits found in LLM and human advice. The key insight is to reframe surveys (typically preference statements which reveal traits) as characteristics of the advice given in response to a scenario. This more closely matches typical LLM outputs in advice-seeking dialogue, allowing for the measurement of LLM behavior rather than self-reporting. The survey questions come from established questionnaires measuring empathy, emotional regulation, assertiveness, and impulsiveness. Each survey statement is transformed by an LLM into 16 situational judgment tests consisting of a scenario and two alternative actions which support or oppose the preference indicated in the original statement. The test texts are validated by 3 human annotators; only those with unanimous approval are retained. 10 annotators provide human judgments for each test to measure the distribution of human behavior against which LLM responses are compared. Analysis shows correlation between LLM and human distributions when the entropy of human distribution is low. When the human entropy is high, the LLM output distributions are not. This is true even when aggregating across model families. Additional analysis shows LLM self-reported traits and situational advice is misaligned. Case studies provide qualitative insights.

Critical strengths

The dataset seems very rigorously constructed. The paper demonstrates the proper level of human validation for data collection. This is laudable given how few datasets do. The reuse of existing psychometric instruments grounds the work in the learnings and applications of personality psychology.

I think the transformation of questionnaires from self-descriptive claims to situational judgments is a very clever solution to the problem of measuring model behavior in a naturalistic setting.

Critical weaknesses

It’s a shame that the preference statement modification and situational judgment tests are LLM generated. Even with all the validation steps, the use of model generated texts may introduce unseen confounds which then impact LLM advice. It seems that Gemini advice shows better directional alignment than comparable models, possibly an advantage provided by the provenance of the generated situations. This could be partially addressed by comparing distributions over situations generated by other models with these generated by Gemini, or better still by real situations from existing chat interaction datasets.

A full table of results broken down by model and trait is missing.

Questions for authors

Suggestion for appendix: It would be nice to know some more details about the underlying psychometric instruments, including who uses them, how they were constructed, and if they are validated against e.g. life outcomes or other measures.

Reviewer_X9QvOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

STAR evaluates the alignment of LLM dispositions with human behavior. Grounded in psychological questionnaires, it shifts the focus from self-reporting to revealed behavior. The paper introduces SJTs, which present actions supporting or opposing a psychological preference, to make LLMs choose between different actions. Only unanimous SJTs are preserved. The work focuses on emotional intelligence traits and introduces the Trait-Positive Rate (TPR), which measures the percentage of human annotators who prefer an action and the percentage of LLM responses recommending that action. The results show that LLM outputs are far less diverse than human outputs, even when human preferences vary, and that repeated sampling and using different models do not help. Even when humans disagree, mean within-model agreement remains high. Smaller models show substantially lower agreement. Finally, the paper introduces metrics for directional alignment and consensus, showing that smaller models struggle to understand whether to suppress or express a trait even when human consensus is high.

Critical strengths

-The paper shows that self-reports do not transfer to advisory behavior.

-The methodology and dataset construction is strong and the paper has several important takeaways

-The approach is tested across various model families and sizes.

-The human annotation agreement is strong, for instance only unanimous SJTs being preserved.

-Smaller models show greater disagreement with humans, even when humans strongly agree. Overall, the methodology is very strong, but the evaluation metrics are weak.

Critical weaknesses

-The prompt requires the LLM to choose exactly one of two actions which constrains the model output behavior.

-The Distributional Gap metric overstates what it actually measures and the authors should soften the claim. The difference between two aggregate proportions does not fully capture the difference between the underlying human and model response distributions.

-The claim about capturing distributional shifts in human opinion is too strong, since the measurement is limited to binary outcomes.

-We do not want models to simply replicate human consensus, the authors should provide more support for this statement. They should explain why we want human consensus on an action to become LLM consensus.

-Were any of the 10 human annotators psychology experts or did they have any understanding of human responses?

Questions for authors

See weakness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.