Research map / Large language models / Reasoning and chain of thought

Latent Veracity Inference for Identifying Errors in Stepwise Reasoning

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis

Read the paperCheck your idea against 108,967 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.

Proposes Veracity Search (a Metropolis search over per-step veracity bits using LM joint likelihood with the true answer as proxy reward) and Amortized Veracity Inference (SFT on VS pseudo-labels) to identify stepwise errors in chain-of-thought without step-level labels.

Problem

Identifying which steps of a chain-of-thought are incorrect without step-level supervision, since LMs often generate flawed reasoning steps that hurt performance and trustworthiness.

Approach

Treats each CoT step's correctness as a latent binary veracity variable in a latent-variable model, performs posterior inference via single-bit Metropolis search with simulated annealing using the LM joint likelihood P(v, y* | x, z) as reward, and amortizes this into a zero-shot veracity classifier (AVI) trained by SFT on the search outputs.

Key results

• VS reaches Hamming similarity 0.910-0.964 on ProntoQA across Qwen-4B, Qwen-8B, Llama-3B, Llama-8B, versus at most about 0.69 for the best baseline.
• AVI (Qwen3-8B) reaches Hamming similarity 0.955-0.967 on 3-5 hop ProntoQA, versus about 0.66-0.67 for Many2Many.
• On ProntoQA with synthetic noisy chains, AVI self-correction raises answer accuracy for Qwen3-8B to 0.87 (3-hop) versus 0.54 for no correction.

Surprising findings

• Few-shot Many2Many and CoT baselines often score near chance-level Hamming similarity on GSM8K and CommonsenseQA, while VS is much higher.
• Using Many2Many-based self-correction on natural ProntoQA CoTs slightly lowers accuracy (0.712 to 0.697), while AVI-based correction raises it to 0.730.

Limitations

• Most analysis uses synthetically corrupted CoTs, and VS accuracy depends on the error distribution.
• Natural-error evaluation is limited in scope and relies on GPT-4.1 pseudo-ground-truth.
• Gains in self-improvement are smaller than gains in veracity accuracy.

Experimental setup

Datasets: ProntoQA GSM8K CommonsenseQA

Benchmarks: ProntoQA GSM8K CommonsenseQA

Metrics: Hamming similarity exact match accuracy answer accuracy

Baselines: Many2Many Many2Many+CoT Majority Voting Recursive inference Self-Refine

Builds on: Chain-of-Thought prompting Self-Refine variational EM

Closest papers by meaning

• Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.90)
• Causal Sufficiency and Necessity Improves Chain-of-Thought Reasoning (NeurIPS 2025 · similarity 0.88)
• Deep Hidden Cognition Facilitates Reliable Chain-of-Thought Reasoning (AAAI 2026 · similarity 0.88)
• Thinking on the Fly: Test-Time Reasoning Enhancement via Latent Thought Policy Optimization (ICLR 2026 · similarity 0.88)
• GIVE: Structured Reasoning of Large Language Models with Knowledge Graph Inspired Veracity Extrapolation (ICML 2025 · similarity 0.88)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.88)