Mind the Gap: How Elicitation Protocols Shape the Stated-Revealed Preference Gap in Language Models

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A recent study, "Mind the Gap: How Elicitation Protocols Shape the Stated-Revealed Preference Gap in Language Models," investigates the discrepancy between language models' stated values and their contextual choices. This research systematically examines how various elicitation protocols influence the stated-revealed (SvR) correlation across 24 LMs. Findings indicate that allowing neutrality and abstention during stated preference elicitation significantly improves Spearman's rank correlation (ρ) between volunteered stated preferences and forced-choice revealed preferences. Conversely, enabling abstention in revealed preferences leads to near-zero or negative ρ values due to high neutrality rates. The study also found that system prompt steering using stated preferences did not reliably enhance SvR correlation on AIRiskDilemmas, underscoring that SvR correlation is highly protocol-dependent and requires methods accounting for indeterminate preferences.

Key takeaway

For NLP Engineers designing or evaluating language model preference elicitation, recognize that protocol choices critically impact stated-revealed correlation. You should incorporate options for neutrality and abstention in stated preference elicitation to improve signal quality. Be cautious with abstention in revealed preferences, as it can obscure genuine preferences. Design methods that explicitly account for indeterminate LM preferences to avoid misleading evaluations.

Key insights

Elicitation protocols critically shape the stated-revealed preference gap in language models.

Principles

• SvR correlation is highly protocol-dependent.
• Binary forced-choice entangles preferences with protocol artifacts.
• Preference elicitation needs methods for indeterminate preferences.

Method

Systematically studied elicitation protocols' effect on SvR correlation across 24 LMs, varying neutrality/abstention and system prompt steering.

In practice

• Allowing neutrality and abstention improves SvR correlation.
• Abstention in revealed preferences can degrade correlation.

Topics

• Language Models
• Preference Elicitation
• Stated-Revealed Gap
• Evaluation Protocols
• AI Ethics
• Model Alignment

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Cross-Lingual Response Consistency in Large Language Models: An ILR-Informed Evaluation of Claude Across Six Languages — ILR-informed expert judgment reveals interpretable, domain-dependent cross-lingual variation in LLM outputs.
• Prompt Engineering: Saying what matters most — Effective prompts are critical for guiding LLMs, defining behavior, controlling output quality, and managing operational costs.
• Self-Preference Is Weak or Absent in Verifiable Instruction-Following Revision: A Four-Model Test Under Genuine Authorship — LLMs show no detectable self-preference when revising their own verifiable instruction-following drafts, accepting machine-verified fixes at similar rates to fresh models.
• When AI Says It Feels — Encouraging LLMs to express feelings via self-rewarded RL can enhance robustness but degrade truthfulness.
• From Prompt Risk to Response Risk: Paired Analysis of Safety Behavior of Large Language Model — LLM safety analysis benefits from paired prompt-to-response severity transitions and joint relevance assessment.
• Structured Intent as a Protocol-Like Communication Layer: Cross-Model Robustness, Framework Comparison, and the Weak-Model Compensation Effect — Structured intent representations significantly enhance goal alignment and reduce variance across diverse AI models and languages.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.