AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Latent Debate: A Surrogate Framework for Interpreting LLM Thinking towards Binary Decisions

Lihu Chen, Xiang Yin, Francesca Toni

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Reasoning mechanisms and failure modes · 5 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces latent debate, a structured surrogate framework that interprets LLM True/False predictions via hidden supporting and attacking latent arguments and their interactions.

Problem

Understanding the internal thinking process of LLMs and the causes of hallucinations on binary True/False prediction tasks remains a key challenge.

Approach

Treats an LLM's internal representations as a debate among latent arguments (supporting and attacking) and uses a structured surrogate model to interpret how these interactions produce the binary output, unlike human debates that are explicit.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Add token-wise semantic weights to initial strengths (ablation)Latent Debate · Prediction consistency, Llama-2-13Bconsistency (%): 95.0 -> 96.6Improved
Add token-wise semantic weights (ablation)Latent Debate · Prediction consistency, Llama-3.1-8Bconsistency (%): 80.1 -> 83.6Improved
Use quadratic connections instead of simple topology (ablation)Latent Debate · Prediction consistency, Llama-2-13Bconsistency (%): 96.6 -> 79.6Hurt
Use richer graph topologies (adjacent grid, quadratic, fully quadratic) (ablation)Latent Debate · Consistency, Llama-3.1-8B avg of 6consistency (%): 84.27 -> 78.50 / 58.60 / 47.80Hurt
Replace Logit Lens with Tuned Lens as argument interpreter (ablation)Latent Debate · Consistency, 3 datasetsconsistency (%): 86.60 -> 93.20Improved
Replace Logit Lens with Tuned Lens as argument interpreter (ablation)Latent Debate Detector · Hallucination detection, 3 datasetsAUROC: 86.54 -> 67.03Hurt
Use latent debate aggregation over simple average/majority voteLatent Debate · Prediction consistency, Llama-2-13Bconsistency (%): 90.0 -> 96.6Improved
Use latent debate with Mistral-7B vs majority votingLatent Debate · Prediction consistency, Mistral-7Bconsistency (%): 92.3 -> 92.4No clear effect
Use rightmost top argument instead of full debate (ablation)Single-argument baseline · Consistency, Llama-3.1-8Bconsistency (%): 83.57 -> 87.03Improved
Use debate features instead of single top-right argument for hallucination detectionLatent Debate Detector · Hallucination detection, Llama-3.1-8BAUROC: 0.58 -> 0.79Improved
Remove lower layers (first 3 to 12) from latent debate (ablation)Latent Debate Detector · Hallucination detectionAUROC: CounterFact 72.44 -> down 2.55 (first 3); TruthfulQA down 2.07 (first 12)Mixed
Use latent debate features instead of SAPLMA across tasksLatent Debate Detector · Cross-task detection (trained on Common Claim)AUROC: 68.89 -> 75.26Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy