A 1:4 debias-to-reasoning training mix under GRPO reinforcement learning yields the best bias-reasoning trade-off in small MLLMs, cutting measured stereotype scores by about 10% while retaining about 88% of reasoning accuracy.
A prompt array keeps the bias away: Debiasing vision-language models with ad- versarial learning
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
FairReason: Balancing Reasoning and Social Bias in MLLMs
A 1:4 debias-to-reasoning training mix under GRPO reinforcement learning yields the best bias-reasoning trade-off in small MLLMs, cutting measured stereotype scores by about 10% while retaining about 88% of reasoning accuracy.