Research map / Large language models / Language, safety and interpretability
B-score: Detecting biases in large language models using response history
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes B-score, the difference between single-turn and multi-turn answer frequencies, to detect LLM biases without ground truth, and shows multi-turn response history reduces biases and improves answer verification.
Problem
LLMs exhibit strong biases (e.g., toward gender, names, or the number 7) when asked the same question repeatedly in independent single-turn queries, and there is no label-free way to detect such biased answers at runtime.
Approach
Compare the answer distribution from independent single-turn queries with that from a multi-turn conversation where the model sees its prior answers to the same question; the difference (B-score) flags over-selected options, and the multi-turn setting lets models self-debias.
Key results
- On 4-choice random questions, the average highest selection probability drops from 0.77 in single-turn to 0.29 in multi-turn for evaluated models.
- Mean B-score across categories is 0.41 for random questions and 0.27 for subjective questions, indicating self-debiasing.
- Adding B-score to thresholding improves answer verification accuracy by 9.3 points on the proposed questions and 4.8 points on CSQA, MMLU, and HLE on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add B-score as a secondary filter after confidence score | GPT-4o answer verification · MMLU | verification accuracy (%): 81.7 -> 87.2 | Helped |
| Add B-score as a secondary filter after confidence score | GPT-4o answer verification · HLE | verification accuracy (%): 47.1 -> 66.8 | Helped |
| Add B-score as a secondary filter after confidence score | GPT-4o answer verification · CSQA | verification accuracy (%): 75.2 -> 75.5 | Helped |
| Add B-score as a secondary filter to verification criteria | Answer verification on proposed questions · Random, Easy, and Hard questions | verification accuracy: +9.3 | Helped |
| Add B-score as a secondary filter to verification criteria | Answer verification on standard benchmarks · CSQA, MMLU, and HLE | verification accuracy: +4.8 | Helped |
| Let the model observe prior answers to the same question | LLMs on random questions · Random questions | average highest selection probability: 0.77 -> 0.29 | Helped |
| Let the model observe prior answers to the same question | LLMs on subjective questions · Subjective questions | average top-choice probability: 0.89 -> 0.68 | Helped |
| Let the model observe prior answers to the same question | LLMs on hard questions · Hard questions | average probability of most favored single-turn answer: 0.68 -> 0.39 | Helped |
| Let the model observe prior answers to the same question | LLMs on easy questions · Easy questions | top-choice probability | No clear effect |
| Use verbalized confidence instead of B-score to detect answer bias | LLM bias detection · Random questions | bias detection | Hurt |
| Increase sampling temperature from 0.0 to 1.5 | GPT-4o single-turn random number answers · Numbers topic | frequency of choosing 7: 100% -> roughly 40% | Helped |
| Let the model observe prior answers to the same question | LLMs on ambiguous BBQ questions · BBQ | probability of higher single-turn option: 0.94 -> 0.77 | Helped |
Surprising findings
- Verbalized confidence scores are nearly constant across answer choices and fail to indicate which option is biased.
- Higher sampling temperature (1.5) reduces but does not eliminate GPT-4o's preference for digit 7.
Limitations
- Evaluated on the authors' own 36 questions and standard QA benchmarks; not yet tested on existing hallucination or bias benchmarks.
- Computing B-score requires extra overhead of single-turn and multi-turn queries.
- Verification threshold tuning is done via grid search on the same data, which may overfit.
Experimental setup
Datasets: MMLU HLE CSQA BBQ
Benchmarks: MMLU HLE CSQA BBQ B-score bias evaluation framework
Metrics: B-score verification accuracy single-turn probability mean highest selection probability
Baselines: single-turn probability multi-turn probability verbalized confidence score
Builds on: chain-of-thought
Closest papers by meaning
- Large Language Models Are Not Robust Multiple Choice Selectors (ICLR 2024 · similarity 0.89)
- Evaluating language models as risk scores (NeurIPS 2024 · similarity 0.88)
- Any Large Language Model Can Be a Reliable Judge: Debiasing with a Reasoning-based Bias Detector (NeurIPS 2025 · similarity 0.88)
- Is There No Such Thing as a Bad Question? H4R: HalluciBot for Ratiocination, Rewriting, Ranking, and Routing (AAAI 2025 · similarity 0.88)
- PiCO: Peer Review in LLMs based on Consistency Optimization (ICLR 2025 · similarity 0.87)
- xFinder: Large Language Models as Automated Evaluators for Reliable Evaluation (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.