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.
Figures & tables
Figure 1: Schematic overview of the Oscar pipeline, illustrating the key stages and components involved in the process.
Figure 2: Oscar overview. Phase 1: N parallel denoising chains with randomized reveal orders generate diverse trajectories. Cross-chain entropy H× identifies high-uncertainty positions (red). Phase 2: targeted remasking, conditioned on retrieved evidence, corrects hallucinated spans at 1.3× overhead.
TriviaQA
HotpotQA
CSQA
Method
128
64
128
64
128
64
Avg
LLaDA-8B-Instruct
Output
Perplexity
50.4
47.6
49.3
51.2
65.6
65.0
54.9
LN-Entropy
54.6
53.5
54.8
54.7
64.6
64.4
57.8
Semantic Entropy
68.9
67.3
57.6
53.8
44.1
43.9
55.9
Lexical Similarity
62.5
59.0
64.2
57.1
57.3
60.7
60.1
Table 1: AUROC(%) on two DLMs across three QA datasets. Best bolded , second underlined . Green = Oscar (EM eval). Blue = Oscar (LLM-as-Judge eval). All trained baselines use EM labels only.