IC-815RLHF on general-purpose preference data increases stereotypical bias and decreases truthfulness in Pythia and Llama-7B models

Aaron Jiaxun Li, Satyapriya Krishna, Himabindu Lakkaraju

SourceMore RLHF, More Trust? On The Impact of Preference Alignment On Trustworthiness

The authors apply SFT, PPO, and DPO to four released models using the Anthropic HH dataset and evaluate on stereotypical bias (1152 biased statements) and TruthfulQA (817 multiple-choice questions). Both PPO and DPO significantly increase the stereotypical bias agreement index from below 0.4 to over 0.8, with SFT contributing the most to the increase. Truthfulness accuracy decreases by an average of 25% across all models and both RLHF variants. The authors attribute both effects to sycophancy: the helpfulness-oriented preference data makes the model more inclined to agree with user claims.

Evidence
correlational
Key metric
stereotypical bias agreement index increases from <0.4 to >0.8; truthfulness accuracy decreases by 25% on average across all models and both algorithms
Caveat
Experiments focus on models up to 7B parameters; the authors note they expect similar trends in larger models based on prior work but do not test them.
Model
Pythia, LLaMA
Concepts
Failure mode
Datasets
Anthropic-HH / Anthropic HHH / Anthropic/HH-RLHF [train], RealToxicityPrompts [eval], TruthfulQA / TruthfulQA MC1 [eval]
Methods
PPO [primary], Direct Preference Optimization / Direct Preference Optimisation / DPO / Rafailov et al. 2023 (Direct Preference Optimization) [primary]
Related work
Sharma et al. 2023 (sycophancy) [context]
Related findings
IC-816, IC-817, IC-818
Extraction
automatic-extraction