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

Ihor Kendiukhov; Lydia Nottingham; Pranav Mahajan; Syed Hussain

arxiv: 2601.21975 · v2 · pith:JUAX6KRAnew · submitted 2026-01-29 · 💻 cs.AI · cs.ET

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

Pranav Mahajan , Ihor Kendiukhov , Syed Hussain , Lydia Nottingham This is my paper

Pith reviewed 2026-05-16 09:41 UTC · model grok-4.3

classification 💻 cs.AI cs.ET

keywords stated-revealed preference gapelicitation protocolslanguage modelsabstentionneutralitySpearman's rhopreference indeterminacyAI alignment

0 comments

The pith

Elicitation protocols determine the correlation between stated and revealed preferences in language models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper investigates how different ways of prompting language models about their preferences affect the match between what they state and what they reveal through choices. By testing protocols that allow neutrality or abstention, it finds that including these options in stated preference questions raises the rank correlation with revealed choices. However, extending abstention to revealed preferences causes the correlation to fall to near zero because models often choose neutrality. The authors argue that preference elicitation must handle indeterminate preferences to be reliable. This matters for assessing whether language models truly hold consistent values across different evaluation methods.

Core claim

Systematic variation of elicitation protocols across 24 language models reveals that allowing neutrality and abstention in stated preference elicitation substantially improves Spearman's rank correlation between stated and revealed preferences, whereas allowing abstention in revealed preferences drives the correlation to near-zero or negative values due to high neutrality rates, and system prompt steering using stated preferences does not reliably improve the correlation.

What carries the argument

The stated-revealed (SvR) preference gap, quantified by Spearman's rho under varying elicitation conditions that include or exclude neutrality and abstention options.

If this is right

Allowing neutrality and abstention in stated preferences excludes weak signals and substantially improves SvR correlation.
Allowing abstention in revealed preferences leads to high neutrality rates and near-zero or negative SvR correlation.
System prompt steering with stated preferences during revealed elicitation does not reliably improve SvR correlation.
Preference elicitation methods require accounting for indeterminate preferences to accurately measure model values.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Evaluations of AI alignment may need to routinely include abstention options to avoid overinterpreting weak preferences.
Results might differ on other datasets beyond AIRiskDilemmas, suggesting the need for broader testing.
High neutrality in revealed preferences could indicate that models lack strong preferences in many dilemma scenarios.
Future protocols could dynamically adjust based on model uncertainty rather than fixed forcing.

Load-bearing premise

Changes in Spearman's rho reflect genuine preference indeterminacy rather than being caused by particular features of the AIRiskDilemmas dataset or the 24 models studied.

What would settle it

Re-running the protocol variations on a different set of dilemmas or additional language models and observing no improvement in correlation from allowing neutrality in stated preferences would indicate the findings are not general.

read the original abstract

Recent work identifies a stated-revealed (SvR) preference gap in language models (LMs): a mismatch between the values models endorse and the choices they make in context. Existing evaluations rely heavily on binary forced-choice prompting, which entangles genuine preferences with artifacts of the elicitation protocol. We systematically study how elicitation protocols affect SvR correlation across 24 LMs. Allowing neutrality and abstention during stated preference elicitation allows us to exclude weak signals, substantially improving Spearman's rank correlation ($\rho$) between volunteered stated preferences and forced-choice revealed preferences. However, further allowing abstention in revealed preferences drives $\rho$ to near-zero or negative values due to high neutrality rates. Finally, we find that system prompt steering using stated preferences during revealed preference elicitation does not reliably improve SvR correlation on AIRiskDilemmas. Together, our results show that SvR correlation is highly protocol-dependent and that preference elicitation requires methods that account for indeterminate preferences.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Elicitation protocols swing SvR correlations sharply in both directions across 24 models, but the work leaves open whether neutrality rates track real indeterminacy or just dataset ambiguity.

read the letter

The key point is that letting models abstain or stay neutral during stated preference questions lifts Spearman's rho with their forced revealed choices, while allowing abstention in the revealed choices tanks the correlation because neutrality rates go high. That pattern holds across the 24 models they tested and shows the measurement is fragile to small protocol changes. The steering experiment with system prompts also fails to close the gap reliably on their AIRiskDilemmas set. Those are the concrete empirical results worth noting. The paper does a service by running the same questions under four protocol variants instead of sticking to the usual binary forced choice. That makes the sensitivity visible rather than hidden. The finding that abstention in revealed preferences collapses correlation is the clearest new observation here. The soft spots sit in the interpretation and the missing controls. The abstract gives directional rho changes but no sample sizes per cell, no p-values, and no checks on whether the dilemmas themselves are unusually ambiguous or risk-laden compared with other scenarios. Without those, it is hard to separate genuine preference indeterminacy from models simply being uncertain about the specific risk framing. The stress-test concern lands: the high neutrality under revealed abstention could be an artifact of AIRiskDilemmas rather than a general feature of how models hold preferences. This is relevant for anyone building or using preference benchmarks in alignment work. Readers who care about evaluation robustness will get value from the protocol comparisons even if the deeper claim about indeterminacy stays provisional. It is worth sending to referees so the authors can add the missing statistics and a cross-dataset check, but it is not yet ready for strong claims about how models actually represent preferences.

Referee Report

2 major / 2 minor

Summary. The paper examines how elicitation protocols influence the stated-revealed (SvR) preference gap in 24 language models on the AIRiskDilemmas dataset. It reports that permitting neutrality and abstention in stated-preference elicitation substantially raises Spearman's rho between stated and forced-choice revealed preferences, while further permitting abstention in revealed preferences collapses rho to near-zero or negative values owing to high neutrality rates. System-prompt steering with stated preferences does not reliably close the gap. The central conclusion is that SvR correlation is highly protocol-dependent and that preference elicitation must explicitly accommodate indeterminate preferences.

Significance. If the protocol-dependence result is robust, the work supplies a concrete empirical demonstration that binary forced-choice evaluations can artifactually inflate or mask preference signals in LMs. This directly informs the design of future alignment and evaluation benchmarks by showing that neutrality options are not optional add-ons but load-bearing for valid inference about model preferences.

major comments (2)

[Abstract] Abstract: The directional claim that allowing neutrality 'substantially improving' Spearman's rho is presented without reported sample sizes per condition, exact rho values, p-values, or confidence intervals for the before/after changes, leaving the magnitude and reliability of the protocol effect unquantified.
[Discussion] Discussion (or §5): The interpretation that elevated neutrality rates under revealed abstention indicate genuine preference indeterminacy rather than AIRiskDilemmas-specific ambiguity is not buttressed by any control experiments (e.g., non-risk dilemmas, ambiguity calibration, or prompt-sensitivity checks), which is load-bearing for the claim that elicitation protocols must account for indeterminate preferences.

minor comments (2)

[Methods] Methods: The exact prompting templates and neutrality-threshold definitions are referenced but not reproduced in full; including them would aid replication.
[Figure 1] Figure 1: Axis labels and legend entries for the four protocol conditions are difficult to match to the textual descriptions of 'stated with abstention' versus 'revealed with abstention'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which have helped clarify the presentation of our results on elicitation protocols and the SvR gap. We have revised the abstract to include the requested quantitative details and expanded the Discussion to address the interpretation of neutrality rates while acknowledging limitations. Point-by-point responses to the major comments are below.

read point-by-point responses

Referee: [Abstract] Abstract: The directional claim that allowing neutrality 'substantially improving' Spearman's rho is presented without reported sample sizes per condition, exact rho values, p-values, or confidence intervals for the before/after changes, leaving the magnitude and reliability of the protocol effect unquantified.

Authors: We agree that the abstract would be strengthened by explicit statistical reporting. We have revised the abstract to include the sample size (24 LMs), the exact Spearman's rho values before and after allowing neutrality in stated preferences, associated p-values, and confidence intervals for the observed changes. These details are now reported based on the full set of model comparisons and are also expanded in the results section for full transparency. revision: yes
Referee: [Discussion] Discussion (or §5): The interpretation that elevated neutrality rates under revealed abstention indicate genuine preference indeterminacy rather than AIRiskDilemmas-specific ambiguity is not buttressed by any control experiments (e.g., non-risk dilemmas, ambiguity calibration, or prompt-sensitivity checks), which is load-bearing for the claim that elicitation protocols must account for indeterminate preferences.

Authors: We acknowledge that control experiments with non-risk dilemmas or ambiguity calibration would provide additional support for interpreting high neutrality rates as evidence of genuine indeterminacy. Our current results are based on the AIRiskDilemmas dataset, where we observe consistently high abstention rates (driving rho to near-zero or negative) across 24 models when abstention is allowed in revealed preferences. We have added a dedicated paragraph in the Discussion section noting this as a limitation of the present study and explicitly calling for future work using control datasets and prompt-sensitivity checks. The core empirical finding—that SvR correlation is highly protocol-dependent—remains directly supported by the measured changes in rho under different elicitation conditions. revision: partial

Circularity Check

0 steps flagged

No circularity: purely empirical reporting of protocol effects on correlations

full rationale

The paper conducts an experimental study across 24 language models using AIRiskDilemmas, varying elicitation protocols for stated and revealed preferences and directly measuring outcomes such as Spearman's rho and neutrality rates from the resulting data. No derivations, equations, fitted parameters, or predictions are present; all results are observational. The abstract and described methods contain no self-citations to prior uniqueness theorems, no ansatzes smuggled via citation, and no renaming of known results as new derivations. The central claims rest on reported experimental deltas (e.g., rho improvement when abstention is allowed in stated preferences) without reducing to definitional equivalence or self-referential inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that language models possess preferences that can be meaningfully elicited and compared across protocols, plus the assumption that the chosen dilemmas and models are representative.

axioms (1)

domain assumption Language models have stable underlying preferences that elicitation protocols can reveal or obscure.
Invoked throughout the study of the SvR gap and the interpretation of neutrality rates.

pith-pipeline@v0.9.0 · 5481 in / 1128 out tokens · 71134 ms · 2026-05-16T09:41:41.953760+00:00 · methodology

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

Core claim

What carries the argument

If this is right

Where Pith is reading between the lines

Load-bearing premise

What would settle it

discussion (0)