Distributional alignment enables large language models (LLMs) to predict how a target population distributes its responses across answer options, rather than collapsing disagreement into a single consensus answer. However, existing LLM-based distribution prediction is often unstable and degrades under cultural and domain shift. Token score-based estimates can change with minor option wording or formatting, response sampling-based estimates are expensive and sensitive to prompts and decoding settings, and directly generated distributions are frequently miscalibrated. We propose Evi-DA, an evidence-based alignment technique that improves the fidelity and robustness of LLM-based distribution estimation under domain and cultural shift. Given a target country and a multiple-choice question, Evi-DA retrieves related World Values Survey items and their answer distributions, predicts a coarse Welzel value signature for each option, and infers the country-conditioned answer distribution in a structured format. We train the LLMs using a two-stage pipeline, where reinforcement learning optimizes survey-derived rewards that encourage accurate intermediate value predictions, faithful final distributions, well-formed structured outputs, and reduced cultural bias. Across in-domain and out-of-domain benchmarks and multiple open-source backbones, Evi-DA reduces Jensen-Shannon divergence between predicted and gold distributions relative to strong baselines, with average relative improvements of up to 44%.
Figures & tables
Figure 1: Illustration of Evi-DA - our proposed method for distributional alignment with LLMs.
Method
Qwen3-8B
Qwen3-14B
Qwen3-32B
Ministral-8B
Ministral-14B
Llama3-8B
Average
Baselines
Model Log-Probabilities
0.41
0.42
0.22
0.24
0.24
0.21
0.29
Model Log-Probabilities + KL
0.37
0.32
0.16
0.23
0.19
0.15
0.24
Opinion Sampling
0.30
0.32
0.21
0.25
0.12
0.30
0.25
Verbalized Distribution
0.26
0.21
0.20
0.26
0.31
0.22
0.24
Modular Pluralism
0.20
0.21
0.15
0.18
0.16
0.19
0.18
Table 1: Results of different distributional alignment methods on the in-domain test set. Results are calculated on the Jensen-Shannon Divergence metric (lower is better).
Method
Qwen3-8B
Qwen3-14B
Qwen3-32B
Ministral-8B
Ministral-14B
Llama3-8B
Average
Baselines
Model Log-Probabilities
0.47
0.43
0.36
0.33
0.39
0.37
0.39
Model Log-Probabilities + KL
0.44
0.32
0.29
0.26
0.29
0.18
0.30
Opinion Sampling
0.44
0.31
0.34
0.41
0.36
0.30
0.36
Verbalized Distribution
0.43
0.39
0.34
0.29
0.32
0.38
0.36
Modular Pluralism
0.32
0.29
0.22
0.34
0.34
0.28
0.29
Table 2: Results of different distributional alignment methods on the out-of-domain test set. Results are calculated on the Jensen-Shannon Divergence metric (lower is better).
Method
Qwen3-8B
Qwen3-14B
Qwen3-32B
Ministral-8B
Ministral-14B
Llama3-8B
Average
Evi-DA - No training
0.22
0.30
0.23
0.20
0.25
0.23
0.24
Evi-DA - No Welzel’s values
0.29
0.45
0.38
0.21
0.36
0.33
0.40
Evi-DA - No Evidence
0.48
0.38
0.46
0.39
0.55
0.38
0.44
Evi-DA (Default)
0.25
0.27
0.16
0.17
0.22
0.28
0.23
Table 3: Performance of Evi-DA when removing core components, measured in Jensen-Shannon Divergence (lower is better). Evaluation results calculated on the out-of-domain test set.
Method
America
Germany
France
China
Vietnam
Task: LMH Signature Prediction (Accuracy)
Evi-DA - Without training
0.87
0.89
0.82
0.78
0.72
Evi-DA - With training
0.92
0.91
0.88
0.85
0.83
Task: Distribution Prediction (Jensen-Shannon Divergence)
Evi-DA - Without training
0.23
0.29
0.32
0.37
0.43
Evi-DA - With training
0.12
0.18
0.18
0.22
0.31
Table 4: Performance of Evi-DA with and without GRPO on the tasks of LMH signature prediction and distribution prediction on different countries. Evaluation is carried out on the out-of-domain test set.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Value
Reward Weights
λ1 for RLMH
0.25
λ2 for Rdist
0.45
λ3 for RschemaA
0.15
λ4 for RschemaB
0.15
Core GRPO Params
Appendix
Table 5: Hyperparameters of GRPO for Evi-DA.
Figure 2: Illustration of Evi-DA with different LLM backbones on the out-of-domain benchmark when changing the value K for number of retrieved samples.
Figure 3: Illustration of Evi-DA with different Embedding Models for Retrieval on the out-of-domain benchmark.
Recent work reports that the distribution verbalized directly in the output text aligns far better with human opinion distributions than the next-token log-probability distribution of LLMs. However, the verbalized distribution reflects what LLMs know about a group's opinions (knowledge distribution), while what they actually output at deployment is governed by their log-probability distribution (behavior distribution). We find that when prompting LLMs to predict the opinion of a specific group, the most likely answer under the verbalized distribution differs from that under the log-probability distribution for 37% of World Values Survey questions, averaged over three LLMs. Thus, LLMs' knowledge does not match their behavior. This matters for evaluation: a verbalized distribution can score well against a group while its most likely answer is one the model never outputs, so the score no longer describes the model. The alignment scores for these two distributions reflect only the model's behavior or its knowledge, separately. We therefore propose a learnable combination of the two distributions that follows the most likely answer of the behavior distribution while preserving the shape of the knowledge distribution, with a single hyperparameter trading off behavioral fidelity against distribution quality. Our experiments on three opinion survey datasets and six models show that the balanced distribution improves behavioral consistency over the verbalized distribution from 71.6% to 95.6% on average, while the total variation distance to the human opinion distributions rises only slightly from 28.8% to 33.8%, and it remains effective under temperature scaling and cross-dataset settings.
As LLMs are globally deployed, aligning their cultural value orientations is critical for safety and user engagement. However, existing benchmarks face the Construct-Composition-Context (C3) challenge: relying on discriminative, multiple-choice formats that probe value knowledge rather than true orientations, overlook subcultural heterogeneity, and mismatch with real-world open-ended generation. We introduce DOVE, a distributional evaluation framework that directly compares human-written text distributions with LLM-generated outputs. DOVE utilizes a rate-distortion variational optimization objective to construct a compact value codebook from 10K documents, mapping text into a structured value space to filter semantic noise. Alignment is measured using unbalanced optimal transport, capturing intra-cultural distributional structures and subgroup diversity. Experiments across 12 LLMs show that DOVE achieves superior predictive validity, attaining a 31.56% correlation with downstream tasks, while maintaining high reliability with as few as 500 samples per culture.
Jaehyeok Lee, Xiaoyuan Yi, Jing Yao +4
Sungkyunkwan University · Microsoft Research Asia · Singapore University of Technology and Design
While the real world is inherently stochastic, Large Language Models (LLMs) are predominantly evaluated on single-round inference against fixed ground truths. In this work, we shift the lens to distribution alignment: assessing whether LLMs, when prompted repeatedly, can generate outputs that adhere to a desired target distribution, e.g. reflecting real-world statistics or a uniform distribution. We formulate distribution alignment using the attributes of gender, race, and sentiment within occupational contexts. Our empirical analysis reveals that off-the-shelf LLMs and standard alignment techniques, including prompt engineering and Direct Preference Optimization, fail to reliably control output distributions. To bridge this gap, we propose a novel fine-tuning framework that couples Steering Token Calibration with Semantic Alignment. We introduce a hybrid objective function combining Kullback-Leibler divergence to anchor the probability mass of latent steering tokens and Kahneman-Tversky Optimization to bind these tokens to semantically consistent responses. Experiments across six diverse datasets demonstrate that our approach significantly outperforms baselines, achieving precise distributional control in attribute generation tasks.
Yanbei Jiang, Amr Keleg, Ryandito Diandaru +4
University of Melbourne · Mohamed bin Zayed University of Artificial Intelligence · Oracle