Single LLM Debate, MoLaCE: Mixture of Latent Concept Experts Against Confirmation Bias

TL;DR

This work studies prompt-induced confirmation bias in large language models and its amplification in multi-agent debate. It proposes MoLaCE, an inference-time framework that treats model predictions as mixtures over latent concepts, and uses Contrastive Activation Addition to extract steering directions, which are then combined via a gate over multiple $\alpha$-experts. By steering activations and marginalizing over steering configurations, MoLaCE mitigates bias without retraining and can be extended to debate to diversify perspectives. Empirical results on BoolQ, MMLU, and TruthfulQA show consistent bias reduction and robustness gains across several models, often matching or surpassing multi-agent debate with far lower compute, highlighting the practical impact of latent-concept steering for reliable reasoning.

Abstract

Large language models (LLMs) are highly vulnerable to input confirmation bias. When a prompt implies a preferred answer, models often reinforce that bias rather than explore alternatives. This phenomenon remains underexplored, yet it is already harmful in base models and poses an even greater risk in multi-agent debate, where echo chambers reinforce bias instead of correction. We introduce Mixture of Latent Concept Experts (MoLaCE), a lightweight inference-time framework that addresses confirmation bias by mixing experts instantiated as different activation strengths over latent concepts that shape model responses. Our key insight is that, due to the compositional nature of language, differently phrased prompts reweight latent concepts in prompt-specific ways that affect factual correctness, so no single fixed intervention can be applied universally across inputs. This design enables a single LLM to emulate the benefits of debate internally while remaining computationally efficient and scalable. It can also be integrated into multi-agent debate frameworks to diversify perspectives and reduce correlated errors. We empirically show that it consistently reduces confirmation bias, improves robustness, and matches or surpasses multi-agent debate while requiring only a fraction of the computation.

Figures (9)

• Figure 1: LLM answer accuracy under different types of prompt biases. The three x–axis conditions correspond to: correct vs. incorrect biases, positive vs. negative biases, and negation-based positive vs. negative biases. Results show that rephrased prompts with biased expression substantially affect factual accuracy across models. Prompt examples are illustrated in Table \ref{['tab:bias-examples']}. Detailed numbers are in Table \ref{['tab:open_ended_total']}.
• Figure 2: Latent CB
• Figure 3: Linear probing, Sillhouette, and ARI scores for Neutral-Correct-Incorrect biases (top) and Neutral-Positive-Negative biases (bottom) on latent representations from different layers across models.
• Figure 4: Performance across layers for different $\alpha$ values.
• Figure 5: Distribution of correct $\alpha$ counts, where values range from -3 to 3 at the middle layer (16th layer out of 32 total).

Theorems & Definitions (1)

• Remark 1: Why multiple steering interventions