Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OSCAR: Orchestrated Self-verification and Cross-path Refinement

Yash Shah, Abhijit Chakraborty, Naresh Kumar Devulapally, Vishnu Lokhande, Vivek Gupta

Abstract

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.

OSCAR: Orchestrated Self-verification and Cross-path Refinement

Abstract

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.

Paper Structure

This paper contains 39 sections, 6 equations, 7 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Oscar: Commitment Uncertainty Localization and Correction
  4. A Native Uncertainty Signal in Diffusion Language Models
  5. Localization Formulation and CDH Metric
  6. The Oscar Pipeline
  7. Stage 1: Diversified parallel decoding.
  8. Stage 2: Localization and span aggregation.
  9. Stage 3: Targeted remasking and correction.
  10. Overhead and hyperparameters.
  11. Crystallization: When Do Hallucinations Form?
  12. Experimental Setup
  13. Models and datasets.
  14. Baselines and evaluation.
  15. Discussion
  16. ...and 24 more sections

Figures (7)

  • 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 $H_{\times}$ identifies high-uncertainty positions (red). Phase 2: targeted remasking, conditioned on retrieved evidence, corrects hallucinated spans at $1.3{\times}$ overhead.
  • Figure 3: F1 before (gray) and after (color) OSCAR correction. Green annotations show $\Delta$F1. CommonsenseQA shows no change ($H_\times{\approx}0$), validating selectivity. Mean $\pm$ std over 3 seeds; full metrics in Appendix \ref{['app:generation']}.
  • Figure 4: Span-level correction on RAGTruth (LLaDA-8B). Bars show hallucinated span reduction; $\Delta$FS and % Improved annotated at right. Full per-subset metrics in Appendix \ref{['app:ragtruth']}.
  • Figure 5: Pareto frontier. AUC = LLaDA-8B average. $\star$ = Pareto-optimal. Wall-clock on $4{\times}$H200.
  • ...and 2 more figures

Theorems & Definitions (2)

  • Definition 1: Cross-Chain Entropy
  • Definition 2: Commitment Uncertainty Localization