PaperlayerView exact revision on arXiv

Source-linked AI summary

OmniConfess: Eliciting Token Confessions to Mitigate Omni-Modal Hallucination

Huiqiang Rong, Haoran Luo, Hui Feng, Zhonghong Ou, Kaiwen Xue, Guoxin Zhang, Yifan Zhu

arXiv:2610.02999v2cs.CL

TL;DR

Omni-modal models can hallucinate when they rely on the wrong evidence, while existing inference-time methods rarely reveal which evidence sustains a commitment. OmniConfess freezes and rescores candidate responses under channel-wise interventions, producing token-level confessions that guide correction. Across six benchmarks, it improves factual reliability and evidence grounding, with F1 gains of up to 13.3 points.

  • Problem

    Omni-modal models may rely on irrelevant evidence or insufficiently use relevant evidence, while existing inference-time methods rarely reveal which evidence sustains a generated commitment.

    Source: §1 · pp. 1–2

  • Method

    OmniConfess freezes a candidate response, rescores it under individual channel interventions at token resolution, and uses a token-by-channel confession to guide correction.

    Source: §1 · p. 2

  • Results

    Across six benchmarks, OmniConfess improves factual reliability and evidence grounding, with F1 gains of up to 13.3 points and consistent gains across backbone architectures.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    The structured confession exposes evidence dependencies and guides correction in both judgment and free-form generation.

    Source: §1 · p. 2, §6 · p. 9

    Repeated channel-wise evidence rescoring incurs substantial latency on some tasks, including 35.14 seconds on audio-visual CMM versus 20.64 seconds for BaseModel.

    Source: §5.6 · p. 9

Abstract

from arXiv · show

Omni-modal large language models (OmniLLMs) unify text, images, audio, and video, yet hallucinate when generation relies on the wrong evidence. Existing inference-time methods can reduce hallucinations, but rarely reveal which evidence sustains a generated commitment. We introduce OmniConfess, a training-free method for mitigating omni-modal hallucinations. It fixes a candidate response and re-scores it at token resolution under controlled channel-wise evidence interventions, producing a structured token-by-channel confession that reveals the response's evidential dependence. OmniConfess uses this confession to preserve grounded content and correct commitments driven by irrelevant or contradictory evidence. To evaluate OmniConfess, we construct OmniHalluBench, a 3,540-example benchmark built from six datasets spanning text, image, audio, and video settings and both judgment and free-form generation. Experiments show that OmniConfess mitigates hallucinations across heterogeneous modality and task settings. Our code and benchmark are publicly available at https://github.com/RongHuiQiang/OmniConfess.

1 INTRODUCTION

Omni-modal models can hallucinate when they rely on irrelevant evidence or underuse relevant evidence, and reliable control must assess both outputs and their supporting evidence. OmniConfess freezes a candidate, interrogates channels at token level, and uses the resulting confession to guide correction.

1 INTRODUCTION

  • Omni-modal hallucinations can arise from reliance on irrelevant evidence or insufficient use of relevant evidence, so control must assess each commitment’s supporting evidence as well as output correctness.OmniLLMs unify text, images, audio, and video in one generative model.Source: §1 · p. 1
  • Reliable omni-modal hallucination control must compare token-level, channel-wise dependence along a frozen candidate sequence to address trajectory variation, entangled channels, and insufficient resolution.Regeneration can confound evidence effects with sequence differences, while global or modality-level scores may miss token-specific dependence.Source: §1 · p. 2
  • OmniConfess freezes a candidate response, intervenes on individual evidence channels to rescore its tokens, and uses the resulting token-by-channel confession to recalibrate judgments or revise erroneous free-form spans.Its commitment-centered workflow is contrasted with three existing inference-time paradigms.Source: §1 · p. 2, Figure 2 · p. 2
  • OmniConfess outperforms the strongest relevant baseline across all six datasets on the primary backbone, while OmniHalluBench covers four modalities and two task types.The benchmark contains 3,540 hard examples and includes cross-modal conflicts and text-grounded hallucinations.Source: §1 · p. 2

2 RELATED WORK

Prior work addresses hallucination across unified multimodal settings and through inference-time search, evidence contrast, and fine-grained signals. OmniConfess explicitly characterizes token-level, channel-wise dependencies for hallucination correction.

2 RELATED WORK

  • Hallucination research has expanded from visual grounding failures to cross-channel interference as models unify text, images, audio, and video.Recent studies examine modality contributions, cross-modal grounding, and evidential attribution.Source: §2 · p. 2
  • Inference-time methods include search-based candidate exploration or verification and evidence-contrast decoding driven by evidence-induced distribution shifts.Evidence-contrast work spans visual, audio-visual, and modality-adaptive decoding.Source: §2 · p. 2
  • Fine-grained methods use token-level or internal signals such as attention patterns, risk, and cross-modal representations; OmniConfess targets token-level, channel-wise evidential dependencies.Figure 3 identifies front-loaded dependence, cross-channel heterogeneity, and token-level concentration as structural properties of hallucinated outputs.Source: §2 · p. 3, Figure 3 · p. 3

3 PRELIMINARY

The preliminary analysis characterizes token dependence on evidence channels in fixed hallucinated candidates and identifies front-loaded, heterogeneous, and concentrated dependence. These patterns motivate fixed-candidate anchoring, channel-wise intervention, and token-level rescoring.

3 PRELIMINARY

  • Together, front-loaded dependence, cross-channel heterogeneity, and token-level concentration motivate fixed-candidate anchoring, channel-wise intervention, and token-level rescoring.Source: §3 · p. 3
  • The analysis treats each hallucinated sample as an input with evidence channels and a fixed candidate, and defines token dependence by comparing model distributions with and without a channel.The model assigns conditional token distributions given the fixed candidate prefix and the input with channel m removed.Source: §3 · p. 3
  • The first third of candidate tokens holds 49.8% of dependence mass, motivating anchoring because early token changes alter later contexts.The analysis partitions normalized token positions into equal intervals.Source: §3 · p. 3
  • Coupled channels can have heterogeneous directions in their token-level dependence trajectories, motivating channel-specific attribution.The analysis measures coupling between distinct channels through correlation of their token-level dependence trajectories.Source: §3 · p. 3
  • The top 20% of tokens capture 82% of dependence mass on average, motivating token-level localization.This is the token-level concentration property identified in the analysis.Source: §3 · p. 3

4 METHODOLOGY: OMNICONFESS

OmniConfess is a training-free inference-time method built around commitment anchoring, channel-wise evidence interrogation, and confession-guided correction. It uses task-specific prompts and signals to recalibrate judgments or repair selected free-form spans.

4 METHODOLOGY: OMNICONFESS

  • OmniConfess combines commitment anchoring, evidence interrogation, and confession-guided correction for training-free inference-time hallucination mitigation.The pipeline applies task-specific signals to judgment recalibration and local free-form correction.Source: §4 · p. 4, §4.4 · p. 5

4.1 PROMPTING SETUP

  • Prompts enforce reliance on provided evidence and task-specific output constraints, including audible-only judgments, fixed answer formats, and separate audio-visual observations.The listed constraints apply to CMM, PubMedQA, HaloQuest, and AVHBench, respectively.Source: §4.1 · p. 4

4.2 COMMITMENT ANCHORING

  • Commitment anchoring freezes the candidate and its prefixes during interrogation, eliminating trajectory variation in token-level comparisons across evidence conditions.The candidate and full-view scores are retained while the frozen candidate is rescored under altered evidence.Source: §4.2 · p. 4

4.3 EVIDENCE INTERROGATION

  • Evidence interrogation removes channels in turn, rescores the frozen candidate and alternatives, and organizes token-by-channel dependence into diagnostic confession states.The states distinguish aligned, under-reliant, over-reliant, or dominant dependence; task-specific signals support judgment correction and free-form token decisions.Source: §4.3 · pp. 4–5

4.4 CONFESSION-GUIDED CORRECTION

  • Confession-guided correction recalibrates the candidate score for binary judgments and revises selected spans in free-form responses while preserving content outside them.Tokens above the revision-risk threshold are expanded to word boundaries and merged into contiguous spans.Source: §4.4 · p. 5

5 EXPERIMENTS

OmniConfess is evaluated across six datasets, multiple task types, and three backbones; it consistently improves benchmark performance, while ablations and analyses examine its components, quality, portability, and runtime.

5.1 EXPERIMENTAL SETUP

  • The 3,540-example benchmark spans six datasets; evaluation covers judgment and free-form tasks, three OmniLLM backbones, and F1, Index, Token-F1, GAVIE, and localization AUPRC.OmniConfess requires no additional training.Source: §5.1 · pp. 6–7

5.2 MAIN RESULTS (RQ1)

  • OmniConfess outperforms all baselines across six datasets and reported metrics on Qwen2.5-Omni-7B, with F1 gains of 10.20 on PHD and 9.35 on CMM, and a Token-F1 gain of 16.84 on RAGTruth.Judgment-task F1 and Index improvements and free-form Token-F1 and GAVIE gains indicate benefits across both task types.Source: §5.2 · p. 7

5.4 ANALYSIS OF OMNICONFESS’S IMPACT IN GENERATION QUALITY (RQ3)

  • OmniConfess achieves the highest Overall generation-quality score (88.7), with Evidence Grounding (86.7), Factual Correctness (92.0), and Uncertainty Calibration (82.1) exceeding the strongest baselines by 5.4, 6.9, and 18.3 points.Its clarity and coherence remain competitive, while baseline methods show distinct quality trade-offs.Source: Figure 6 · p. 8, §5.4 · p. 8

5.5 ANALYSIS OF OMNICONFESS’S PORTABILITY ACROSS BACKBONE MODELS (RQ4)

  • Average F1 improves over the respective base models by 10.30 points on Qwen2.5-Omni-7B, 6.34 on Qwen3-Omni-30B-A3B, and 7.50 on Nemotron-3-Nano-Omni-30B.Baseline methods either vary across backbones or perform consistently poorly; even CAD’s gains are smaller than OmniConfess’s.Source: Figure 6 · p. 8, §5.5 · p. 9

5.3 ABLATION STUDY AND COMPARATIVE ANALYSIS (RQ2)

  • Removing anchoring, interrogation, or correction lowers average F1 from 71.63 to 67.06, 63.96, or 58.66, respectively; correction removal has the largest effect.The study concludes that all three modules contribute, with revision after identifying evidence-related errors particularly valuable.Source: §5.3 · p. 8
  • Channel-wise confession leads the compared interrogation strategies in F1 and PhD-Index, while confession-guided revision yields larger mean gains than confidence-guided revision in F1 and GAVIE-Acc.Channel-wise confession scores 86.49 on Recall Balance versus 78.16 for pooled evidence; revision gains are +8.92 vs. +3.95 in F1 and +1.49 vs. +0.58 in GAVIE-Acc. OmniConfess anchoring improves average F1 and localization AUPRC beyond shared-prefix anchoring.Source: Figure 5 · p. 7, §5.3 · p. 8

5.6 ANALYSIS OF OMNICONFESS’S INFERENCE EFFICIENCY (RQ5)

  • On PHD and CMM, OmniConfess takes 5.43 and 35.14 seconds versus 2.23 and 20.64 seconds for BaseModel, despite CMM outputs averaging 24.6 tokens.Against search-based methods, it uses about 27× and 22× fewer tokens on PHD and CMM but reduces inference time by only 6.4× and 1.1×; on AVHBench, tokens fall from 569.5 to 164.1 and time from 36.15 to 7.93 seconds.Source: §5.6 · p. 9

6 DISCUSSION AND CONCLUSION

OmniConfess is a training-free framework that uses token- and channel-level evidence interrogation to mitigate omni-modal hallucinations. Across six benchmarks, it improves factual reliability and evidence grounding, with gains across backbone architectures.

6 DISCUSSION AND CONCLUSION

  • OmniConfess mitigates omni-modal hallucinations through training-free token- and channel-level evidence interrogation.Structured confessions expose evidence dependencies and guide correction in judgment and free-form generation.Source: §6 · p. 9
  • Experiments across six benchmarks show improved factual reliability and evidence grounding, with F1 gains of up to 13.3 points.Gains are consistent across backbone architectures.Source: §6 · p. 9
  • Structured confessions expose evidence dependencies and guide correction in both judgment and free-form generation.Source: §6 · p. 9

AI USE STATEMENT

The appendix records AI-assisted manuscript polishing and specifies task-aware generation and evaluation prompts. The Proposition 1 proof explains how fixing a candidate separates evidence effects from changes in generation trajectory.

AI USE STATEMENT

  • The authors used generative AI for English translation and polishing, reviewed the assisted text against experimental records, and take responsibility for the final content.Source: AI USE STATEMENT · p. 10

A.1 OMNICONFESS PROMPTS

  • Generation prompts set task-specific evidence boundaries, distinguish audible from visible evidence, and require acknowledgment when visual details are absent.They also standardize outputs while preserving the answer requirements of judgment and free-form tasks.Source: §A.1 · p. 13, Figure 8 · p. 13

A.2 GAVIE-INSPIRED EVALUATION PROMPT

  • Free-form factual accuracy is scored from 0–10 against the reference answer alone, penalizing unsupported or contradictory details rather than reproducing the original GAVIE procedure.Source: §A.2 · p. 13, Figure 9 · p. 13

A.3 MULTIDIMENSIONAL QUALITY EVALUATION PROMPTS

  • Seven prompts score free-form responses from 0–10 across evidence grounding, factual correctness, relevance, completeness, coherence, uncertainty calibration, and clarity.Each prompt specifies dimension-specific criteria and inputs.Source: §A.3 · pp. 13–14, Figure 10 · p. 14

B.1 PROOF OF PROPOSITION 1

  • Commitment anchoring fixes the candidate and its autoregressive prefixes across evidence conditions, so token-level differences decompose the same response’s evidence-induced log-likelihood change without trajectory variation.Without anchoring, an early token change can alter later conditioning contexts.Source: §B.1 · pp. 14–15

B.2 PROOF OF PROPOSITION 2

The Proposition 2 proof derives a channel-wise additive surrogate that preserves observed margins and yields task-specific evidence signals. Proposition 3 and the algorithm describe how those signals guide judgment correction or localized free-form revision.

B.2 PROOF OF PROPOSITION 2

  • The additive surrogate uniquely reconstructs the observed full-view and single-channel margins, preserving each exactly.Source: §B.2 · p. 15
  • Signed channel contributions yield task-specific signals: judgment correction isolates relevant evidence, while free-form signals separate relevant support, excess residual preference, and irrelevant-channel support.Source: §B.2 · p. 15

B.3 PROOF OF PROPOSITION 3

  • Judgment correction yields the task-relevant surrogate margin without guaranteeing correctness; free-form revision targets selected spans, preserves unselected text, and prevents positive irrelevant supports from canceling.Source: §B.3 · p. 16

C OMNICONFESS ALGORITHM DETAILS

  • Algorithm 1 freezes a decoded candidate, interrogates channels, builds the confession, then corrects judgment decisions or revises risk-selected free-form spans.Source: C OMNICONFESS ALGORITHM DETAILS · p. 17

D.1 BENCHMARK COMPOSITION

OmniHalluBench combines six public benchmarks into a 3,540-example evaluation set spanning multiple evidence modalities and task types. The appendix describes its difficulty-based construction and the evaluated backbones and baselines.

D.1 BENCHMARK COMPOSITION

  • OmniHalluBench contains 3,540 examples from six public benchmarks spanning text, image, and audio-video evidence across judgment and free-form tasks.Source: §D.1 · pp. 17–18

D.2 BENCHMARK CONSTRUCTION

  • Construction starts from 5,900 candidates and retains the hardest 60% after independent response generation and severity assessment, with researchers auditing 200 near-cutoff examples.Source: §D.2 · p. 18

E BACKBONE AND BASELINE DETAILS

  • Evaluation uses three OmniLLM backbones with different architectures: Qwen2.5-Omni-7B as default, Qwen3-Omni-30B-A3B, and Nemotron-3-Nano-Omni-30B; VCD contrasts original and perturbed visual inputs to mitigate visual hallucination.Source: E BACKBONE AND BASELINE DETAILS · p. 18

F EVALUATION DETAILS

The evaluation section defines metrics for judgment, free-form generation, and hallucination localization. It also includes AVHBench and PHD case-study figures, whose captions identify their subjects but report no outcomes.

F EVALUATION DETAILS

  • Judgment tasks use Yes-F1 for positive-class precision and recall and PhD-Index for the harmonic mean of Yes and No recall.Source: F EVALUATION DETAILS · p. 19
  • Free-form tasks use Token-F1 for token overlap with the reference and GAVIE-Acc for reference-grounded factual accuracy scored by an LLM judge on a 0–10 scale.Source: F EVALUATION DETAILS · p. 19
  • Localization AUPRC measures average precision over candidate spans ranked by predicted hallucination risk, treating hallucinated spans as positives.All methods are evaluated on the same candidate spans.Source: F EVALUATION DETAILS · p. 19

G ERROR CASE STUDIES

The case studies expose errors that evidence-dependence analysis does not resolve: valid paraphrases can be penalized, and judgments can remain wrong despite accurate observations. More broadly, relevant evidence does not guarantee a supported interpretation, particularly in ambiguous cases.

G ERROR CASE STUDIES

  • Token-level F1 penalizes the valid “person” paraphrase of reference “man,” a mismatch evidence-related correction cannot eliminate.This is a limitation of single-reference evaluation.Source: G ERROR CASE STUDIES · p. 20
  • In the PHD case, the model correctly observes the cat but answers No about happiness, treating alertness as contrary evidence without visual support.OmniConfess does not correct this subjective interpretation.Source: G ERROR CASE STUDIES · p. 20
  • The case studies identify two limits: single-reference evaluation penalizes lexical mismatch, and judgments can remain incorrect despite accurate visual observations.The AVHBench example contrasts “person” with the reference “man”; the PHD example shows a wrong happiness judgment despite correctly observing the cat.Source: G ERROR CASE STUDIES · p. 20

H LIMITATIONS AND FUTURE WORK

  • Evidence dependence does not ensure correct interpretation; ambiguous judgments can remain unsupported even when the relevant channel is used.Because correction starts from a generated candidate, interpretation errors may persist after local revision; proposed future work includes jointly assessing evidence dependence and interpretive validity, better-calibrated correction, and selective interrogation to reduce inference cost.Source: H LIMITATIONS AND FUTURE WORK · p. 21

Related papers

Loading 2610.02999v2…