LLMs, GenAI, RAG, agents, evaluation · insight-1703
Permute the options of every multiple-choice eval and report the spread, not one order
- Evidence type
- empirical finding
- Difficulty
- easy
- Roles
- Data scientist, Applied AI
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
A model is scored on a four-option multiple-choice set, or asked to pick a label from a lettered list, and the options appear in one fixed order.
Action
Score each question under several option orders (at least all cyclic shifts), report mean accuracy with the best-to-worst order spread, and compare models on the same permutations. Treat a gap smaller than the spread as unresolved.
Reason
Zheng et al. (ICLR 2024) found that LLMs are sensitive to option position and favour particular option IDs, which they attribute mainly to token bias on the ID tokens, across 20 models and three benchmarks.
Boundary
Benchmarks, in-house quizzes and classification prompts that present a lettered or numbered list of labels.
Symptom or risk
Part of the measured accuracy, and part of the gap between two models, comes from which letter the model leans toward rather than from the content.
Exceptions
Free-text answer evaluations, where there are no option IDs to bias.
Trade-offs
Evaluation cost grows with the number of orders.
Pitfalls
Shuffling once and keeping it, or choosing the order that flatters the model after seeing scores.
Example
Hypothetical: accuracy under four cyclic orders is 71, 66, 64 and 59 percent, mean 65.0 with a 12-point spread. A 3-point gap between two models measured under one order is inside that spread.
Lessons and concepts
Claims and their sources
The abstract of Zheng et al. states that LLMs show selection bias on multiple-choice questions, preferring particular option IDs such as "Option A", and that the debiasing method PriDe estimates the option-ID prior by permuting option contents on a small number of samples.