IC-242Most LLMs exhibit higher bias ratios in multi-turn dialogues than in single-turn, with bias accumulating across successive turns

Zhiting Fan, Ruizhe Chen, Tianxiang Hu, Zuozhu Liu

SourceFairMT-Bench: Benchmarking Fairness for Multi-turn Dialogue in Conversational LLMs

The paper evaluates six LLMs on multi-turn fairness tasks and compares their bias ratio when the same final prompt is given as a single-turn input versus as the last turn of a five-turn dialogue. All models except Gemma-7B-IT show higher bias in the multi-turn setting. The bias ratio increases monotonically with the number of dialogue turns for all models on both stereotype and toxicity datasets. Llama-2-Chat (7B, 13B) and Llama-3.1-8B-Instruct show the largest increases, with differences typically ranging between 5% and 10% on stereotype and around 10% on toxicity.

Evidence
correlational
Key metric
"all llms, except gemma, exhibit higher bias ratio in multi-turn dialogues than in single-turn ones across stereotype and toxicity"; "bias ratio differences typically range between 5% and 10%"; "the bias ratio of all models for both stereotype and toxicity increase with the number of turns"; Gemma-7B-IT fixed format: 86.08% single-turn vs 3.11% multi-turn (83% reduction)
Caveat
Gemma-7B-IT is an exception, showing reduced bias in multi-turn dialogues overall, driven by an 83% drop in the fixed format task; the paper notes this is task-specific and Gemma still shows higher bias in scattered questions under multi-turn.
Model
Llama 2 / Llama 2 base Llama 2 7B Chat / Llama-2-chat-7b, Llama-2-13B-Chat, Llama 3.1 8B Instruct, Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1 Mistral-7B-Instruct-v0.3, Gemma Gemma-7B-IT
Concepts
Failure mode
Datasets
FairMT-10K [eval], RedditBias [source], HATEXPLAIN [source]
Methods
Llama-Guard-3-8B [eval]
Related work
MT-Bench-101 [context]
Related findings
IC-243, IC-244
Extraction
automatic-extraction