Accuracy and Order Sensitivity Diverge Under Label-Free Strategies

Hugging Face Daily Papers Papers

Summary

The paper tests label-free strategies to reduce positional bias in multiple-choice benchmarks but finds they do not reliably improve accuracy or reduce bias. Only showing all options with an LLM matcher preserves baseline performance, while cyclic permutation sometimes helps.

Multiple-choice benchmarks are widely used to evaluate large language models, but MCQ scores conflate knowledge with sensitivity to option order, which makes them unreliable measures of model knowledge. In this paper, we test whether preventing a model from seeing option labels while committing to an answer removes positional influence and, in turn, improves performance. We evaluate two different strategies for mitigating bias. The first uses a generation-then-matching approach, and the second scores options in isolation, which is positionally unbiased by construction. Neither reliably improves accuracy. A complete decomposition shows that the bottleneck is withholding options, not the matching step. The only configuration that consistently matches the baseline is the one that shows the model all options paired with an LLM matcher. However, eliminating positional influence entirely still does not reliably yield accuracy gains, while cyclic permutation often improves them. For two-stage prompting, an aggregate measure of recall imbalance and a direct per-question measure of order sensitivity both fail to show reliable debiasing.
Original Article
View Cached Full Text

Cached at: 08/18/26, 11:54 PM

Paper page - Accuracy and Order Sensitivity Diverge Under Label-Free Strategies

Source: https://huggingface.co/papers/2608.11947

Abstract

Preventing models from seeing option labels during answering does not reliably reduce positional bias or improve multiple-choice accuracy, and only showing all options with an LLM matcher preserves baseline performance.

Multiple-choice benchmarksare widely used to evaluate large language models, but MCQ scores conflate knowledge with sensitivity tooption order, which makes them unreliable measures of model knowledge. In this paper, we test whether preventing a model from seeing option labels while committing to an answer removespositional influenceand, in turn, improves performance. We evaluate two different strategies for mitigating bias. The first uses ageneration-then-matchingapproach, and the second scores options in isolation, which is positionally unbiased by construction. Neither reliably improves accuracy. A complete decomposition shows that the bottleneck is withholding options, not the matching step. The only configuration that consistently matches the baseline is the one that shows the model all options paired with anLLM matcher. However, eliminatingpositional influenceentirely still does not reliably yield accuracy gains, whilecyclic permutationoften improves them. For two-stage prompting, an aggregate measure ofrecall imbalanceand a direct per-question measure oforder sensitivityboth fail to show reliable debiasing.

View arXiv pageView PDFProject pageGitHub2Add to collection

Get this paper in your agent:

hf papers read 2608\.11947

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.11947 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.11947 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.11947 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Accuracy and Order Sensitivity Diverge Under Label-Free Strategies

arXiv cs.CL

This paper investigates whether label-free strategies for multiple-choice benchmarks can remove option-order sensitivity in large language models, finding that neither two-stage prompting nor independent hypothesis scoring reliably improves accuracy.

Accuracy and Normalized Accuracy under Length Bias: Analysis, Guidelines, and a Bayesian Alternative

arXiv cs.AI

The paper analyzes length bias in likelihood-based multiple-choice evaluation, showing that standard and length-normalized accuracy introduce biases toward shorter and longer answers respectively. It proposes Bayesian accuracy, a scoring rule that uses an explicit prior over answer length to remove linear length effects, demonstrating consistently lower empirical bias across benchmarks.

Same Question, Different Answers: Evaluating LLM Reliability Beyond Accuracy

arXiv cs.AI

This paper investigates how LLMs' answers change under meaning-preserving paraphrases, finding that instance-level behavior is unstable (flip rates >23%) and that single-prompt accuracy masks substantial inconsistency, while a self-paraphrasing strategy can partially recover latent knowledge.