acceptodds
Under review as a conference paper at ICLR 2027

LLM Human Response Alignment: A Multi-Sample Debiasing Framework

Abstract

Large language models (LLMs) offer scalable synthetic respondents, but their usefulness as human proxies depends on alignment with observed responses. We distinguish two sources of misalignment. Model-side bias arises when LLM and human response distributions differ under the same observed inputs. Latent-context omission limits prediction because relevant respondent information is missing. We propose a multi-sample post-hoc debiasing framework that keeps the LLMs frozen and learns an external correction from human-labeled data. The correction retains a response vector, preserving distributional variation and source identity that a single response can discard. The sampling design adapts to the task: fixed model sources for individual prediction and sampled personas for population-mean prediction. Under squared loss, the vector weakly reduces Bayes-optimal risk relative to either scalar summary, with strict gains when compression loses predictive information. Across three broad benchmarks spanning diverse domains, our method attains the best mean in every main comparison, reducing MAE relative to uncorrected LLM predictions by up to 47% at the population level and 54% at the individual level. We further show that our method preserves respondent rankings and improves revenue prediction.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Digital Twins, Large Language Models, Alignment, Multi-Sample Debiasing

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_HBW5OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a multi-sample LLM debiasing framework. By combining the outputs of multiple LLMs with predefined features, it trains a lightweight calibration network to improve the model's prediction accuracy for real human responses. The problem addressed in this paper holds certain research significance; the method is simple and practical, and the experiments are somewhat convincing. However, the overall approach is quite similar to traditional supervised classification, regression, and ensemble learning, with relatively limited innovation. Furthermore, the existing experiments have not yet fully demonstrated the necessity of introducing LLM outputs, nor do they include an in-depth analysis of individual and group prediction changes during calibration. Therefore, I believe this paper still needs to further clarify its methodological contributions and verify its effectiveness and reliability through more comprehensive experiments.

Critical strengths
  1. The research idea of this paper is commendable, as it focuses on the bias between LLM-simulated human responses and real human behavior, which holds certain research significance and practical application value.
  2. The method proposed in this paper is simple and easy to implement. Without requiring additional fine-tuning of the LLM, it merely employs a lightweight calibration network to improve prediction results, demonstrating a degree of practicality.
  3. The experiments in this paper are fairly convincing. The authors conducted evaluations on multiple datasets and considered prediction tasks at both individual and group levels, while also providing ablation studies and downstream application analyses, which to some extent support the effectiveness of the method.
Critical weaknesses
  1. First, the task in this paper is essentially still a classification or regression task, which predicts human responses by training a simple MLP that combines LLM outputs with predefined features. Although this method follows some basic paradigms of existing LLM calibration, I find its application scenario rather narrow and lacking in novel methodological design. Essentially, it merely applies existing supervised learning methods to a new domain, resulting in relatively limited overall innovation.

  2. Additionally, I noticed that in some tasks, LLM outputs seem unnecessary, as good predictive performance can be achieved using only predefined features. For example, in Table 3, w/o LLM even outperforms the method incorporating LLM outputs on certain metrics (e.g., Acc). This suggests that the model might fit the ground-truth answers well using only existing features. Then, is this paper calibrating LLM prediction bias, or directly training a new prediction model? I believe the authors need to further explain the necessity of introducing LLMs and their practical advantages over direct supervised learning.

  3. Finally, although the authors evaluated predictive performance at both the individual and group levels and conducted some analysis on individual ranking, I consider these experiments still insufficient. Improvements in group average predictions do not necessarily mean that individual predictions have been effectively calibrated. For instance, drastic changes in some individual predictions might improve the group mean while simultaneously harming the prediction accuracy for other individuals. Therefore, I suggest that the authors evaluate both individual and group performance on the same cohort of respondents and further analyze whether improvements in group predictions come at the cost of sacrificing individual accuracy or response diversity.

Questions for authors

No.

Reviewer_R9H4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method to improve LLM predictions of human responses. It keeps multiple LLM responses as a vector and uses human data to train a calibration model. The authors show that the full vector can retain useful information lost by averaging. Experiments on three datasets compare the method with single-response and averaged-response baselines. The gains are larger for population averages than for individual responses. Additional tests examine respondent rankings and revenue prediction.

Critical strengths
  1. The Mean and Vector comparison uses the same samples and training settings. Vector reduces population-level MAE across eight LLM sources. This supports the value of keeping the full response vector.
  2. The method keeps LLMs frozen. It needs no extra human labels compared with the calibration baselines. The theory also explains why repeated sampling cannot recover missing personal information.
Critical weaknesses
  1. EEDI averages numerical codes for answer choices A to D. These choices have no natural order. Relabeling them can change the results. This weakens the evaluation.
  2. Tests mainly compare Vector with One, not Mean. The held-out-question experiment also omits Mean. The benefit over averaging remains unclear.
Questions for authors
  1. Do the EEDI gains hold when predicting answer-choice proportions instead of average numerical codes?
  2. Is Vector significantly better than Mean at both population and individual levels? Please report paired comparisons with appropriate uncertainty estimates. How does Mean perform in the held-out-question experiment?
Reviewer_zsJEOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a multi-sample framework for correcting LLM-generated samples. It uses a correction model that takes as inputs multiple LLM-generated samples and question and/or profile embeddings, and fits the model to the observed human responses. Through theoretical analysis and experiments on survey datasets, it demonstrates the improvement of multi-sample correction over the approaches that input only one LLM-generated sample or the mean of multiple LLM-generated samples to the correction model.

Critical strengths

The idea of preserving synthetic samples' distributional variation when learning the correction is interesting. The paper achieves this by inputting multiple synthetic responses to a correction model, and experiments demonstrate its gain over the single-response and the mean-response approaches.

Critical weaknesses
  1. While the multi-sample idea is interesting, the proposed method is not very flexible. (i) The trained correction model takes a fixed number KK of synthetic responses as input. If, for a new response simulation task, fewer than KK synthetic samples are available, or if one of the LLM sources becomes unavailable, then the trained correction model cannot be used directly and may require retraining. If more synthetic responses are available, the trained model cannot directly exploit the additional data. (ii) The paper does not offer guidance on how one should choose KK in practice, either. (iii) Concatenating KK responses as input may also substantially increase the input dimension. This is not a big issue for the scalar-response setting studied in the paper, but can be problematic for multi-dimensional responses (e.g., a 20-question response vector).
  2. In addition, the paper does not give much investigation of where the multi-sample approach's advantage comes from. The paper partly motivates it by saying that it preserves distribution variation. However, the mechanisms behind are not investigated in greater depths. Particularly, for population-level prediction, if one wants to preserve distributional information, then another natural approach is to input the response histogram, which can be much more efficient when there are much fewer response categories than synthetic response samples. Comparison with this alternative approach would help clarify whether inputting all the synthetic responses is truly necessary. The theoretical analysis is also straightforward and relies on simple filtration inclusion arguments, and offers limited insights into the gains from multi-sample framework.
Questions for authors
  1. What is the purpose of introducing latent-context omission, and what insights does it give? The paper introduces as a source of misalignment, and explicitly states that it cannot be resolved by the proposed multi-sample correction framework. But beyond this, what additional methodological or empirical insights does the notion of latent-context omission provide?
  2. The paper treats the individual-level and population-level problem under the same multi-sample methodological framework. Would such a unified treatment ignore some finer structures in the two different problems?

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers, the (S)ACs, and the PCs

We sincerely thank the three reviewers, the (S)ACs, and the PCs for their time and careful engagement. The reviews were constructive and have substantially strengthened the paper.

Strengths noted by the reviewers. The problem has research significance and practical value, and the method is simple and keeps the LLMs frozen (HBW5, R9H4). Preserving the distributional variation of synthetic responses is interesting (zsJE), and the controlled Mean vs. Vector comparison supports keeping the full response vector (R9H4).

Key new evidence.

  • Vector vs. Mean (R9H4, HBW5). On held-out questions, Vector improves on Mean in all 24 dataset-metric pairs, significantly (p < .05) in 23 of them (Table R1).
  • Label-free EEDI (R9H4). With answer-choice targets instead of averaged codes, Vector remains best and beats Mean at p < .01 at both levels (Tables R5 and R6).
  • Source of the gain (zsJE). Vector beats Histogram and Moments encodings of the same draws in all six settings, because keeping every response lets the correction weight unreliable surrogates implicitly (Table R3).
  • Flexibility and necessity of LLM outputs (zsJE, HBW5). One padded model handles a varying number of responses without retraining (Tables R1 and R2), and Vector outperforms Qwen3-4B fine-tuned on the same human labels on 10 of 12 pairs, with under 1% of its parameters (Table R4).

Revisions. We will report the label-free EEDI results in Sec. 5.2, add guidance on choosing K in Sec. 3.3, and add the new comparisons to the appendix.

We thank the reviewers again for their thoughtful feedback, and the (S)ACs and PCs for overseeing a careful review process.

Best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.