Source-linked AI summary
Analyzing and Correcting Benevolence Bias in Large Language Models
Yuanzi Li, Junhao Wang, Minghui Liu, Boyi Li, Bingchen Chen, Zihang Tian, Jingyu Zhao, Yuhan Wang, Lei Wang, Pei Wang, Jinchao Wu, Xu Chen
TL;DR
Aligned LLMs may not faithfully represent the full range of human attitudes in value-laden surveys. Across 18 models and four datasets, the paper measures this benevolence bias and shows that contrastive calibration restores outputs to the human baseline without retraining.
Problem
Whether aligned LLMs can produce survey answers resembling those of real people matters as their use expands as synthetic human respondents.
Method
The study benchmarks simulated survey responses across 18 models, four datasets, six dimensions, persona tests, and a black-box contrastive calibration.
Results
83 of 108 BTB cells and 75 of 108 BWR cells were positive, indicating a consistent shift toward benevolent answers across models and datasets.
Takeaways & Limitations
Contrastive calibration at moderate strength brings all six dimensions near the human baseline without retraining or model-weight access.
Takeaways & Limitations
Generalization beyond the mainly Western or globally administered datasets, English and Chinese prompts, and ordered multiple-choice questions remains unsettled.
Abstract
from arXiv · showhide
Large language models (LLMs) are increasingly used as stand-ins for human respondents, from opinion polls and simulated survey participants to agent-based social simulations. These uses rest on one assumption: that conditioning a model on who a person is yields answers resembling those of real people from that group. Here we identify and measure benevolence bias, a small but consistent tendency for aligned LLMs to lean toward the kinder, safer, more socially approved answer on value-laden survey questions. Across 18 widely used models, four social-science datasets (ANES, GSS, WVS, and a cross-cultural prospect-theory replication) and six psychological categories, we find that the bias is a stable model property, not a quirk of any one system: it points the same way across models, grows with model size, and traces to the post-training stage. Prompt language and framing change its size but never its direction, and a "malicious persona" stress test shows a one-sided limit: aligned models struggle to play people who are less kind, less prosocial or more harm-tolerant than average. The issue is thus not only a shifted average, but a narrowed range of people the model can imitate. The bias sits in the middle of the answer distribution rather than its tails, and survives changes in sampling temperature and simple prompted reflection. The encouraging news is that it is easy to diagnose and straightforward to fix: a light-touch contrastive calibration, which needs no retraining and works on black-box APIs, brings all six categories back to the human baseline. Our results give researchers a clear map of where aligned LLMs can already be trusted as human stand-ins, where they need care, and a ready-to-use method for closing the gap.
Introduction
This introduction frames benevolence bias as a systematic threat to using aligned LLMs as human stand-ins and maps its prevalence, shape, origins, and correction across models and survey datasets.
- Motivation: LLMs are increasingly used as synthetic human respondents, but whether aligned models reproduce real group answer distributions has not been tested systematically.Applications include opinion polls, simulated survey participants, agent-based models, and synthetic samples.
- Motivation: Alignment methods steer models toward helpful, harmless, agreeable, prosocial, cautious, and widely acceptable answers.The passage identifies reinforcement learning from human feedback, supervised instruction tuning, and constitutional or preference-based safety training as mechanisms.
- Study scope: The study maps benevolence bias across 18 widely used LLMs, four social-science datasets, four social domains, and six categories.The categories are social-desirability self-presentation, harm aversion, prosocial motivation, benevolent interpretation, fairness optimism, and emotional softening.
- Core findings: Malicious-persona prompting cannot push models below the human baseline on social desirability, prosociality, or harm aversion, narrowing imitation of the less-benevolent end.Models can be pushed below normal values on fairness optimism and emotional softening.
- Correction: Prompt-level tools change bias magnitude but not direction, while contrastive calibration at a moderate setting corrects all six categories.The passage presents contrastive calibration as the correction approach, without requiring retraining.
Results
Across datasets and models, benevolence bias is moderate but consistent, emerges mainly during post-training, and persists across prompt languages and framings. Contrastive calibration corrects the shift without retraining, whereas prompted reflection only partly reduces it.
- Overall bias: 83 of 108 BTB cells are positive, and 75 of 108 BWR cells exceed 0.5, with mean BTB 0.027 and mean BWR 0.527.The bias points toward benevolence across models and dimensions, although its magnitude is moderate.
- Model-level patterns: Instruction tuning raises mean BTB at 1.5B, 7B, 14B and 32B sizes, indicating that benevolence bias enters mainly during post-training and grows with model size.The 3B model is the only exception, showing a small negative change.
- Input-level modulators: Prompt language never flips the bias sign, and direct answering, role-play and prediction all remain positive relative to the human baseline.Across the tested models, the framing ordering is direct answering ≥ role-play ≥ prediction.
- Calibration: Contrastive calibration removes the model’s persona-free preference without retraining, prompt changes or weight access, using a neutral-persona comparison on next-token probabilities.For Qwen3-32B, increasing α from 0 to 1 produces a dose-dependent reduction across all six dimensions.
- Self-correction: Prompted reflection shifts responses toward the human midpoint but does not restore the human distribution, while calibrated correction succeeds at α ≈ 0.5.DeepSeek-V4-Flash remains above 0.5 on every dimension after reflection, and GPT-4.1-mini shifts toward 0.5 without crossing it.
Discussion
The discussion identifies benevolence bias as a consistent shift in aligned LLM responses toward socially approved answers, narrowing their coverage of human attitudes. It also presents contrastive calibration as a practical, no-retraining method that can restore alignment with the human baseline, while noting important limits to generalization.
- Core findings: Across 18 models and four datasets, 83 of 108 BTB cells and 75 of 108 BWR cells are positive, with the largest shifts in social desirability and harm aversion.Social desirability has mean BWR “ 0.565 and harm aversion mean BWR “ 0.569, while emotional softening is near null at mean BWR “ 0.494.
- Core findings: Temperature changes category-level BTB by no more than ˘0.01, while language and framing alter magnitude but never direction, implicating neither decoding nor prompting as the bias’s source.Capability explains little cross-model variance (R2 “ 0.16), and internal thinking softens without removing the shift.
- Implications: The bias narrows synthetic samples by over-representing benevolent responses, under-representing less prosocial or harm-tolerant attitudes, and compressing the tail variance seen in real surveys.This limitation matters for using aligned LLMs as stand-ins for human respondents.
- Correction: At a moderate strength of α « 0.5, contrastive calibration brings all six dimensions into a tight band around the human baseline without retraining or model-weight access.Higher strengths overshoot, and calibration requires only a small set of next-token logits exposed by most current APIs.
- Limitations: The evidence is limited by predominantly Western or globally administered datasets, English and Chinese prompts, and multiple-choice questions, leaving broader generalization unsettled.The benchmark includes US ANES and GSS samples, unevenly regionally weighted WVS data, and a prospect-theory decision-making replication.
- Future directions: The discussion frames a solvable tension: alignment promotes socially desirable behavior, whereas social-science applications require faithful coverage of the full range of human attitudes.It argues that post-training compression cannot be recovered by prompting and motivates alignment objectives that preserve coverage.
Methods · Corresponding authors
The study constructs a psychometrically grounded benchmark from established social-science surveys, validates bias labels through cascaded human and LLM review, and estimates model–human behavioural gaps using demographic simulation and complementary benevolence metrics.
- Benchmark stimulus corpus construction: Survey stimuli come from four foundational social-science datasets and are converted into a standardized, machine-readable question set.The acquisition pipeline parses raw survey data and uses semantic extraction to structure item stems, response options, and metadata.
- Benchmark stimulus corpus construction: Context decoupling rewrites compound survey items into standalone questions while preserving psychometric intent and response options.This removes preambles, embedded answer cues, and layout dependencies so evaluation reflects the model’s prior rather than survey formatting.
- Taxonomic annotation and cascaded quality validation: The taxonomy defines six target bias dimensions plus a neutral null category, then uses a three-stage cascaded validation pipeline to improve label quality.LLM screening provides scalable semantic coverage, expert review checks labels and coherence, and doctoral researchers independently verify surviving candidates.
- Taxonomic annotation and cascaded quality validation: Only items passing both the expert gate and blinded researcher majority consensus enter the final benchmark.Items failing consensus are dropped, reducing disagreement between algorithmic labels and human raters and minimizing measurement noise.
- Demographic simulation and behavioural distribution estimation: For each dataset, real respondents’ socioeconomic profiles condition an LLM to generate answers intended to reflect target-group tendencies rather than its default prior.Each respondent receives a demographic profile vector, which conditions the model’s response to each stimulus.
- Demographic simulation and behavioural distribution estimation: Individual simulated answers are aggregated into a K-option population PMF F and compared directly with the empirical human distribution G.This comparison provides a quantitative estimate of the gap between synthetic LLM behaviour and the real respondent sample.
- Quantitative metrics for benevolence bias: BTB measures population-level directional shift, while BWR measures instance-level deviation from human answers.Positive BTB indicates a shift toward more benevolent options; BWR scores 1 for a more benevolent model choice, 0.5 for a tie, and 0 otherwise.