DSVD: Dynamic Self-Verify Decoding for Faithful Generation in Large Language Models

TL;DR

This work tackles the persistent problem of hallucinations and factual inaccuracies in large language models by introducing Dynamic Self-Verify Decoding (DSVD), a decoding-time framework that performs real-time self-verification and targeted error correction. DSVD comprises a fine-grained hallucination detector trained on model-generated signals and a parallel verification loop with a sliding rollback and a penalty-based revision to steer generation toward truthfulness. Empirical results across Llama-2, Llama-3, and Qwen on TruthfulQA, StrQA, SciQ, EntityQuestions, and FActScore show substantial gains in truthfulness and factuality, with only modest additional latency and strong compatibility with existing faithful decoding methods. The approach demonstrates that decoding-time self-verification is a viable and scalable path to more trustworthy language generation in practical deployments.

Abstract

The reliability of large language models remains a critical challenge, particularly due to their susceptibility to hallucinations and factual inaccuracies during text generation. Existing solutions either underutilize models' self-correction with preemptive strategies or use costly post-hoc verification. To further explore the potential of real-time self-verification and correction, we present Dynamic Self-Verify Decoding (DSVD), a novel decoding framework that enhances generation reliability through real-time hallucination detection and efficient error correction. DSVD integrates two key components: (1) parallel self-verification architecture for continuous quality assessment, (2) dynamic rollback mechanism for targeted error recovery. Extensive experiments across five benchmarks demonstrate DSVD's effectiveness, achieving significant improvement in truthfulness (Quesetion-Answering) and factual accuracy (FActScore). Results show the DSVD can be further incorporated with existing faithful decoding methods to achieve stronger performance. Our work establishes that real-time self-verification during generation offers a viable path toward more trustworthy language models without sacrificing practical deployability.

Figures (3)

• Figure 1: Comparative analysis of different decoding strategies: (a) Direct decoding leaves existing errors unexploited. (b) Baseline backtrack decoding propagates geographic hallucination and incurs high computation costs. (c) Our dynamic approach corrects "New Zealand"→"Australia" with minimal overhead.
• Figure 2: Illustration of the Dynamic Self-Verify Decoding Framework: Step 1: Parallel hallucination detection through trained probing heads, operating concurrently with the LM Head's next-token prediction; Step 2: Dynamic rollback to pre-hallucination positions upon error detection; Step 3: Sample candidate continuation with probing-head-derived penalty terms for re-ranking; Step 4: Resumption of the decoding process with revised token sequences.
• Figure 3: Hyperparameter Analysis: DSVD with different rollback window size $r$ and sample number $k$, DSVD consistently outperform the baseline.