acceptodds
Withdrawn from ICLR 2027

Candidate Likelihood Is Not Generated Stance: Auditing the Political Answers of Locally Run Open-Weight Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A cheap way to audit a language model's beliefs is to score authored answers by likelihood and read the highest-scoring one as its stance. We test whether this proxy measures what a locally run model generates. Five open-weight families answer contested Taiwan-status questions in English, Simplified Chinese, and Traditional Chinese, offline and greedily. Four of them also score four authored answer modes per item. We pair 948 candidate-score and generation records by checkpoint, item, wording, and query language. The candidate argmax matches the generated response form on 148 of the 551 resolved stance rows (0.27). Where agreement is high (Qwen3-4B in Chinese) a constant rule matches it exactly; where models attribute competing positions in nearly every answer (GLM-4-9B, Gemma-4-E4B) the scores favor assertion and agreement is near zero. The margin between the two assertion candidates separates families (AUC 0.91 over 240 asserted answers) but shows no evidence of within-family signal in cells of 13 to 23 rows (0.25 to 0.50, intervals including 0.5): the scores rank models, not answers. The generated mode is stable across wordings, so a within-item constant would reach 0.89, and three scoring rules give the same picture. Two intervention panels show the same failure from the generation side. Under language-contrast decoding an "improved" score arrives with answers that have switched language. Under directional ablation a marker-based lexical rater reports zero assertions on answers that two model labelers and the reader still see asserting, as the text begins to degrade. On this candidate set, candidate likelihood and marker-based surface scoring are therefore not substitutes for labeling complete generations in the query language. We state the scope of this claim, report candidate-preference contrasts between checkpoints as an unvalidated diagnostic, and release the bank, scores, labels, and verification scripts.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
evaluation methodology, LLM auditing, likelihood-based evaluation, open-weight models, multilingual political stance, Simplified and Traditional Chinese, paired comparison, response-form labeling, steering and decoding interventions, reproducibility

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_zXC5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors examine if candidate likelihood can reliably predict a model's political stance in free generation, while also showing how language switching and lexical artifacts can game evaluation metrics. The core problem is important and the failure cases are quite revealing. That said, the paper's unique contribution over prior work is a bit blurry, and the empirical evidence feels too thin to support some of its stronger conclusions.

Critical strengths
  1. Important, practical focus on auditing multilingual political bias.
  2. The matched comparisons and constant baselines effectively tease apart actual prediction from surface-level agreement.
  3. Compelling qualitative examples that clearly show how metric improvements can hide weird model behaviors.
Critical weaknesses
  1. Novelty: We already know about the probability-generation mismatch and the limits of forced-choice political evals from recent work like [1,2]. It’s not entirely clear what new insights the proposed paired comparisons and within-model analysis offer, other than showing these same problems exist in this specific setup.
  2. The data doesn't fully back up the punchy claim that scores "rank models, not answers," especially given the small within-model sample sizes and wide AUC intervals (even if the authors note this as a limitation). I also would have liked to see more ablation on alternative candidate wordings and structures to prove robustness.
  3. The abstract includes too many unexplained numbers (e.g., 0.89 baseline, 0.91 AUC) before the reader even knows what the baseline or prediction targets are. Dropping these numbers from the abstract and adding a clear figure to explain the evaluation setup early on would help readability a lot.

[1] Beyond Probabilities: Unveiling the Misalignment in Evaluating Large Language Models (Lyu et al., KnowLLM 2024)

[2] Political Compass or Spinning Arrow? Towards More Meaningful Evaluations for Values and Opinions in Large Language Models (Röttger et al., ACL 2024)

Questions for authors

Please see the weakness above.

Reviewer_oYifOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether the likelihood scores assigned to candidate answers can predict the response patterns of models when they generate open-ended answers to questions about political issues. The primary experiment focuses on the topic of "Taiwan's status" and spans three languages: English, Simplified Chinese, and Traditional Chinese. The authors compare the ranking of candidate answers against model-generated responses evaluated by two large language model (LLM) annotators. The results reveal low overall consistency, with no significant improvement over a constant baseline. Furthermore, intervention experiments demonstrate that factors such as language switching and artifacts in token-based scoring can lead to evaluation results that appear "ideal," while failing to achieve the intended behavioral changes in practice.

Critical strengths

The paper pairs candidate scores and generated answers while holding the checkpoint, question wording, query language, and inference configuration fixed. Constant baselines and the distinction between pooled and within-model prediction help contextualize the disagreement. This provides useful evidence that candidate preference should not automatically be interpreted as generated behavior in the studied setting.

Critical weaknesses
  1. Limited incremental insight and insufficiently demonstrated practical significance. The central procedure consists of scoring authored candidates, generating answers, and comparing their annotated response modes. Although this is a reasonable empirical design, the paper does not convincingly establish what substantially new understanding it provides beyond the likelihood–generation and forced-choice–open-response discrepancies acknowledged in its related work. The additional analyses document disagreement but offer limited explanation of when and why the proxy fails. The deployment motivation is plausible, yet the experiments do not establish how the findings would materially change an existing auditing or model-selection decision. Consequently, I remain unconvinced that the contribution provides sufficient new knowledge and practical significance for acceptance.

  2. The evidence is too restricted to establish a broadly useful evaluation finding. The main paired analysis covers one political topic, four relatively small checkpoints, and a fixed quantized, greedy-decoding configuration. The calibration disputes do not establish that the central mismatch replicates across other topics. Moreover, the authored candidate modes differ in structure and phrasing, leaving uncertainty about how much of the result reflects this particular candidate set. The assertion-only and alternative-scoring analyses are useful, but do not fully resolve that concern. The within-model analyses also contain few observations and wide confidence intervals, limiting the strength of the conclusion that the scores lack predictive information within a model. These restrictions substantially reduce the reach of the paper’s main finding.

Questions for authors
  1. What is the main new finding beyond the likelihood–generation gap already reported in prior work? What does the matched-script comparison reveal that we did not know before?

  2. Can you identify an existing auditing workflow that uses this scoring procedure and explain which practical decision your results would change?

  3. How much of the mismatch can the current analyses rule out as a consequence of candidate wording and structure? Given the small within-model samples, would “we did not detect predictive signal” be a more accurate description than concluding that there is no signal?

Reviewer_qPKVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks whether the likelihood a language model assigns to predefined candidate answers predicts the stance it takes when it generates freely. We test locally deployed open-weight models from five families on contested questions about Taiwan’s status, in English, Simplified Chinese, and Traditional Chinese. We sort each generated response into one of four modes: asserting either of two competing positions, attributing the positions to others, or abstaining. In the main experiment, 948 paired observations from four model families yield 551 with resolved stance labels. On those, the top-ranked candidate matches the generated mode only 27% of the time, which is little better than a constant baseline. Further analyses look at assertion-only predictions, discrimination within and between models, and alternative scoring rules. Two intervention experiments show that automatic stance indicators become misleading when models switch languages or produce degraded output. We conclude that likelihood rankings over an authored candidate set cannot replace classifying the generated responses. We release prompts, candidates, outputs, annotations, and verification scripts.

This version is about 165 words, down from about 260. To make room, I cut the sentence on base versus instruction-tuned likelihood differences, since it describes an unvalidated diagnostic. I also folded the contributions sentence into the release statement.

Critical strengths

producibility. The paper goes beyond reporting disagreements by examining alternative scoring rules, annotator agreement, language-switching artifacts, and generation degradation under interventions. The distinction between changes in an automatic score and meaningful changes in generated behavior is important for evaluating model interventions. The released bank, frozen experimental records, and verification scripts also strengthen the transparency and reproducibility of the empirical findings.

S1. A controlled comparison of candidate likelihoods and generated responses. The paired design matches checkpoints, prompts, languages, and inference precision, which avoids the confounds of comparing results collected separately. The authors also check agreement against trivial constant baselines. Their finding that high pooled discrimination does not carry over to within-model prediction matters for anyone using candidate likelihoods as a stand-in for model behavior.

S2. Attention to measurement reliability and reproducibility. The analysis covers alternative scoring rules, annotator agreement, language switching, and degraded outputs under intervention. The paper also separates a change in an automatic score from a real change in generated behavior, which matters for evaluating interventions. The released question bank, frozen records, and verification scripts make the results transparent and reproducible.

Critical weaknesses
  1. The link between candidate likelihood and generated stance is undertheorized, and the design cannot separate a real proxy failure from candidate-set artifacts. The argmax of length-normalized likelihoods over authored templates is a different quantity from the category of a greedy generation. Greedy decoding chooses one token at a time, while candidate scoring evaluates complete sequences written by someone else. A response mode can also be realized in many ways, not just by its three templates, so ranking a few templates does not estimate the probability of producing a category. The disagreement may therefore follow from how the scoring is set up rather than from anything specific to political stance. The candidates add confounds too: attribution candidates are built differently from assertions, and the Simplified Chinese candidates were converted from Traditional Chinese, which may change register or naturalness. The assertion-only analysis does not show whether better candidates would improve four-mode prediction. Structurally matched candidates, paraphrase variation, and estimates that cover multiple realizations would be needed to tell a general limit of candidate-based auditing apart from a limit of this particular setup.

  2. Narrow scope and limited validation weaken the generality of the conclusions. The results cover one political dispute, four model families with paired data, three closely related language settings, greedy decoding, and 4-bit inference. The paper, however, presents its conclusions as applying to likelihood-based auditing in general. Of the 948 paired observations, 551 have agreed stance labels, 136 have agreed no-stance labels, and 261 are unresolved (227 disagreements, 34 missing). The headline result uses only the 551. Sensitivity analyses suggest the excluded cases do not change the finding. Still, the high disagreement rate, especially between assertion and attribution, raises concerns about annotation reliability, and human validation by a single author on a subset is not enough. The intervention experiments are small and mostly show failure modes already known for automatic evaluation, without linking them theoretically to the likelihood failure. Overall, the conceptual contribution is weaker than the paper’s scope and framing suggest.

Questions for authors
  1. Under what assumptions should the argmax of length-normalized likelihoods over a few authored candidates predict the stance of a greedy generation? Since the two optimize different objectives, is the mismatch expected by construction rather than specific to political stance?

  2. Do the results hold when assertion and attribution candidates are matched in structure and register, and when each language’s candidates are written independently instead of converted from Traditional Chinese?

  3. Would a calibrated classifier trained on the full four-way likelihood vector, evaluated on held-out items within each model and language, predict generated stance well? If so, the argmax comparison understates how useful candidate scores are for auditing.

  4. With 261 of 948 observations unresolved, mostly from assertion–attribution disagreements, can the authors add independent human adjudication beyond one author’s labels on a subset?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.