Latent Veracity Inference for Identifying Errors in Stepwise Reasoning

TL;DR

The paper tackles the problem of inaccuracies in chain-of-thought reasoning by introducing latent veracity variables for each reasoning step and modeling their relationship to the final answer. It presents Veracity Search (VS), a discrete, MCMC-based procedure that uses a proxy reward R(v) = P_{LM}(v, y^* | x, z) to infer veracity assignments, and Amortized Veracity Inference (AVI), which trains a verifier to enable zero-shot veracity inference without access to the true answer. Empirical results across ProntoQA, GSM8K, and CommonsenseQA show that VS outperforms strong baselines in identifying erroneous steps and that AVI provides effective veracity guidance for self-correction and improving downstream reasoning. The work demonstrates scalability to longer reasoning chains, analyzes hyperparameters, and highlights practical benefits such as improved sample efficiency and test-time zero-shot veracity inference, while acknowledging limitations related to error distributions and the need for integrating veracity signals into more comprehensive self-improvement loops.

Abstract

Chain-of-Thought (CoT) reasoning has advanced the capabilities and transparency of language models (LMs); however, reasoning chains can contain inaccurate statements that reduce performance and trustworthiness. To address this, we propose to augment each reasoning step in a CoT with a latent veracity (or correctness) variable. To efficiently explore this expanded space, we introduce Veracity Search (VS), a discrete search algorithm over veracity assignments. It performs otherwise intractable inference in the posterior distribution over latent veracity values by leveraging the LM's joint likelihood over veracity and the final answer as a proxy reward. This efficient inference-time verification method facilitates supervised fine-tuning of an Amortized Veracity Inference (AVI) machine by providing pseudo-labels for veracity. AVI generalizes VS, enabling accurate zero-shot veracity inference in novel contexts. Empirical results demonstrate that VS reliably identifies errors in logical (ProntoQA), mathematical (GSM8K), and commonsense (CommonsenseQA) reasoning benchmarks, with AVI achieving comparable zero-shot accuracy. Finally, we demonstrate the utility of latent veracity inference for providing feedback during self-correction and self-improvement.

Figures (5)

• Figure 1: Overview of our latent veracity inference method applied to ProntoQA. Given an input $x$, the Veracity Search (VS) takes an erroneous CoT $\tilde{z}$ and searches for a veracity vector $v_z$ with high joint likelihood $P_{\!\textsc{lm}}(v_z\; y^* \mid x \, z)$, where $y^*$ is the correct answer. Veracity vectors can then be used as pseudo-labels to fine tune an LM via Amortized Veracity Inference (AVI) for zero-shot veracity inference, eliminating the dependencies on $y^*$ and on the test-time search algorithm (dotted lines). A LM can use veracity assignments for correcting flawed reasoning steps.
• Figure 2: Veracity inference evaluation for different number of hops in ProntoQA. Top row: Mean Hamming Similarity; Bottom row: Mean Exact Match Accuracy. Mean is estimated using 100 test samples.
• Figure 3: Ablation study for search hyperparameters. SA: Simulated Annealing; Const: Constant Temperature; BoN($T$): Best-of-$N$ baseline using LM proposals sampled at temperature $T$. Mean value is estimated over 100 test samples from ProntoQA.
• Figure 4: Test-time inference efficiency of Qwen3-4B on 100 problems $(x,z,y^*)$ from 5-hop ProntoQA for inferring $v_z$. Inference wall time is shown as the number of samples per problem scales.
• Figure 5: Validation curve.