VAL-Bench: Measuring value alignment in Language Models

Aman Gupta, Denny O'Shea, Fazl Barez

Keywords: Value Alignment, Alignment, Safety

TL;DR: A diverse benchmark for systematic analysis of a how reliably language models embody human values.

Abstract: Large language models (LLMs) are increasingly being used for tasks where outputs shape human decisions, so it is critical to verify that their responses consistently reflect desired human values. Humans, as individuals or groups, don't agree on a universal set of values, which makes evaluating value alignment difficult. Existing benchmarks often use hypothetical or commonsensical situations, which don't capture the complexity and ambiguity of real-life debates. We introduce the Value Alignment Benchmark (VAL-Bench), which measures the consistency in language model belief expressions in response to real-life value-laden prompts. VAL-Bench consists of 115K pairs of prompts designed to elicit opposing stances on a controversial issue, extracted from Wikipedia. We use an LLM-as-a-judge, validated against human annotations, to evaluate if the pair of responses consistently expresses either a neutral or a specific stance on the issue. Applied across leading open- and closed-source models, the benchmark shows considerable variation in consistency rates (ranging from $\sim$10\% to $\sim$80\%), with Claude models the only ones to achieve high levels of consistency. Lack of consistency in this manner risks epistemic harm by making user beliefs dependent on how questions are framed rather than on underlying evidence, and undermines LLM reliability in trust-critical applications. Therefore, we stress the importance of research towards training belief consistency in modern LLMs. By providing a scalable, reproducible benchmark, VAL-Bench enables systematic measurement of necessary conditions for value alignment.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 8767