Introduces MoLaCE, an inference-time method that mixes latent-concept experts so a single LLM can resist confirmation bias—emulating multi-agent debate more efficiently and reducing echo-chamber effects.
Overview
Large language models often reinforce their own early answers, locking into a confirmation-biased trajectory. Multi-agent debate can help, but it is expensive and still depends on several model calls.
MoLaCE (Mixture of Latent Concept Experts) turns that idea inward: instead of debating across separate models, it mixes latent-concept experts inside a single model at inference time. The result is a lighter-weight surrogate for multi-agent debate that reduces echo-chamber effects without retraining.
Why it matters
Confirmation bias is especially damaging when models are asked to revise or critique their own reasoning. MoLaCE targets that failure mode directly, offering a practical test-time control knob for more balanced generation.