Search papers, labs, and topics across Lattice.
This paper introduces OSCAR, a training-free inference-time framework for mitigating hallucinations in diffusion language models (DLMs) by intervening during generation. OSCAR leverages the denoising trajectories of DLMs to identify token positions with high commitment uncertainty using cross-chain entropy, and then performs targeted remasking conditioned on retrieved evidence. Experiments on multiple QA datasets using LLaDA-8B and Dream-7B show that OSCAR significantly reduces hallucinated content and improves factual accuracy compared to baselines and even specialized trained detectors.
Diffusion language models inherently signal factual uncertainty during generation, outperforming even trained hallucination detectors, and OSCAR leverages this to reduce hallucinations.
Diffusion language models (DLMs) expose their denoising trajectories, offering a natural handle for inference-time control; accordingly, an ideal hallucination mitigation framework should intervene during generation using this model-native signal rather than relying on an externally trained hallucination classifier. Toward this, we formulate commitment uncertainty localization: given a denoising trajectory, identify token positions whose cross-chain entropy exceeds an unsupervised threshold before factually unreliable commitments propagate into self-consistent but incorrect outputs. We introduce a suite of trajectory-level assessments, including a cross-chain divergence-at-hallucination (CDH) metric, for principled comparison of localization methods. We also introduce OSCAR, a training-free inference-time framework operationalizing this formulation. OSCAR runs N parallel denoising chains with randomized reveal orders, computes cross-chain Shannon entropy to detect high-uncertainty positions, and then performs targeted remasking conditioned on retrieved evidence. Ablations confirm that localization and correction contribute complementary gains, robust across N in {4, 8, 16}. On TriviaQA, HotpotQA, RAGTruth, and CommonsenseQA using LLaDA-8B and Dream-7B, OSCAR enhances generation quality by significantly reducing hallucinated content and improving factual accuracy through uncertainty-guided remasking, which also facilitates more effective integration of retrieved evidence. Its native entropy-based uncertainty signal surpasses that of specialized trained detectors, highlighting an inherent capacity of diffusion language models to identify factual uncertainty that is not present in the sequential token commitment structure of autoregressive models. We are releasing the codebase1 to support future research on localization and uncertainty-aware generation in DLMs.
