Omni-Woodpecker: Hallucination Correction via Evidence-Prior Rebalancing for Omni LLMs

AI summary

Overview

Research area: Multimodal AI — specifically hallucination mitigation in Omni LLMs (models that jointly process text, video, and audio).

Technical level: Advanced. The method operates on hidden states inside a frozen transformer, using orthogonal decomposition of representation directions and per-layer, per-head intervention — readers need working knowledge of attention, residual streams, and logit-based decoding.

Scope: A single paper presenting Omni-Woodpecker (OWP), a training-free inference-time framework that detects cross-modal conflicts in audio-visual queries and rebalances fused hidden representations before decoding.

What This Paper Is About

Omni LLMs can reason jointly over audio and video, but they hallucinate when one modality's signal overrides the evidence a question actually depends on — for example, hearing silence in the audio but answering "yes" to "Is the dog barking?" because a dog is visible. The paper argues this is not a perception failure but a representation entanglement problem: grounding evidence and learned priors get mixed together during early multimodal fusion, and existing fixes only manipulate the output distribution after fusion has already corrupted the hidden state. OWP instead diagnoses which modality the query needs, which modality is competing with it, and edits the internal representation to restore the balance — without training or changing any model weights.

Key Contributions

1. Evidence-Prior Rebalancing Formulation. The paper reframes omni-modal hallucination as an internal imbalance between grounding evidence and learned priors inside fused representations, moving the mitigation target from output-level correction to representation-level repair.

2. Conflict-Conditioned Representation Rebalancing (OWP). A training-free framework that runs four views of a frozen model (full, audio-only, video-only, text-only) to diagnose query-conditioned conflicts, then splits token–head responses into protected evidence, competing prior, and unchanged context.

3. Adaptive two-site editing. A single correction budget is divided between attention-path reweighting and residual-state restoration using a per-input allocation weight, so the edit site matches how far the conflict has progressed; no model parameter or language head is modified.

4. Extensive Experimental Validation. Evaluation on CMM (1,200 conflict cases) and AVHBench (5,302 examples) plus four general omni-modal reasoning benchmarks, with cross-architecture transfer to VideoLLaMA2-AV and ablations isolating each component.

Main Findings

• Large gains on hallucination benchmarks with Qwen2.5-Omni-7B: CMM rises from 73.2% to 86.1% (+12.9 points) and AVHBench overall accuracy from 75.1% to 83.7% (+8.6 points). Against the strongest baseline (MAD), the margins are 4.9 and 7.7 points.

• Hallucination resistance, not just perception: CMM hallucination resistance (HR) climbs from 57.0% to 85.5% with perception accuracy (PA) at 86.7%, shrinking the backbone's 32.5-point PA–HR gap to 1.2 points; MAD retains a 16.9-point gap (89.7% PA, 72.8% HR).

• Bidirectional and multi-type correction: CMM visual dominance reaches 86.0% and audio dominance 90.2%; AVHBench audio→video reaches 84.9% and video→audio 85.3%. Language dominance also improves (81.3% to 82.0%), where the competing cue is the query text rather than a sensory stream.

• Joint audio-visual evidence preserved: On AV Matching, OWP reaches 81.1%, exceeding the backbone by 6.6 points, MAD by 13.1, and AVCD by 17.9 — while AVCD shows the opposite trade-off (81.2% directional accuracy but only 63.2% on AV Matching).

• Cross-backbone transfer: On VideoLLaMA2-AV, CMM rises from 76.2% to 80.8% (+4.6) and AVHBench from 70.2% to 74.1% (+3.9); HR rises from 65.3% to 77.7% and AV Matching from 56.6% to 62.2%. OWP on Qwen2.5-Omni-7B also exceeds the newer Qwen3-Omni by 4.3 and 6.2 points on CMM and AVHBench under the same protocol.

• General reasoning improves, not degrades: Gains of +1.93 (OmniBench, 45.36 to 47.29), +2.25 (WorldSense, 45.21 to 47.46), +1.00 (MUSIC-AVQA, 85.08 to 86.08), and +0.67 (AVQA, 91.83 to 92.50) points.

• Selective editing: Under the native protocol, OWP identified a conflict in 25.8% of inputs and changed the final answer in 16.9%, leaving 74.2% running the unmodified backbone. Target typing agreed with the benchmark-defined target modality on 99.48% of rows, and only 8 of the 24 disagreeing rows reached an edit.

• Cheaper than compared baselines: Average latency is 1.71× (3.54 s to 6.05 s) on Qwen2.5-Omni-7B and 1.77× (1.40 s to 2.48 s) on VideoLLaMA2-AV with at most five forward calls, versus MAD's 2.81× and 2.36× with nine calls. OWP is more than twice as fast as HulluEdit on Qwen2.5-Omni-7B (6.05 s versus 13.02 s); peak memory is 24.9 GiB and 17.3 GiB.

• Both components matter: In the native-protocol ablation, evidence enhancement alone raises AVHBench accuracy by 5.41 points and HR by 13.00; prior suppression alone by 6.72 and 27.33; combined, the full model reaches 85.17% AVHBench accuracy and 85.50% HR while cutting the PA cost from 5.83 and 4.17 points to 2.80.

• Adaptive allocation beats fixed splits: Adaptive allocation lifts AVHBench accuracy from 83.93% to 85.17%, PA from 84.67% to 86.70%, and HR from 82.33% to 85.50% (+1.24, +2.03, +3.17) with the same total budget. Swapping the layers assigned to each component lowered accuracy by 6.79 points (Holm-adjusted p < 0.001).

• The gain is not a bias toward "No": A constant shift toward "No" matched to OWP's yes rate gained only +0.60 and +0.67 points (p < 10⁻⁶ and p < 10⁻³), versus OWP's +2.85 and +2.14; the shift also lowered positive-row accuracy from 86.73% to 85.39%, while OWP raised it to 86.90%.

• The competing stream, not missing information, causes the error: Dropping to audio only costs 12.35 and 6.94 points, video only 10.64 and 10.70, and text only 29.18 and 23.24. Oracle routing to the target stream alone gains only +2.23 and +0.71, and the competing stream alone loses 25.21 and 22.35 — yet OWP, which keeps all modalities, beats oracle routing (84.91% versus 84.28% and 78.99% versus 77.56%).

Methodology in Plain English

OWP never trains anything. It works on top of a frozen Omni LLM in three stages.

Stage 1 — Find the conflict. For each query, the model is run on four versions of the input: the full audio+video+text input, audio+text only, video+text only, and text only. A short appended diagnostic question asks the model which modality the query needs (audio, video, or both), and the answer probabilities decide the target stream. A conflict exists when the target stream's answer disagrees with the full input's answer while another stream agrees with the full input. That other stream is the competing source, verified by checking whether it draws more attention mass than the target after normalizing for span length. If no conflict is found, the edit is skipped and the backbone answer is returned unchanged.

Stage 2 — Separate evidence from prior. The hidden states from the four views are subtracted from the text-only state to strip out shared language content, then combined into a single unit "evidence direction" that is anchored on the target stream and only admits parts of the other views that support the target answer. The leftover fusion shift is split: the component along the evidence direction is protected, and the orthogonal remainder becomes the "prior direction" to suppress. A clipped projection measures how much target evidence the fused state has lost. Head localization then picks, from attention and answer-aligned value contributions, a small set of heads that favor the competing answer and a small set that favor the evidence — at most four heads per layer and the 16 most-attended tokens per source group.

Stage 3 — Rebalance and decode. One correction budget is split by a per-input weight based on how strongly the full view already prefers the competing answer. A weak preference sends more of the edit to the attention path (subtracting competing values, adding evidence values before aggregation); a strong preference sends more to the residual state (removing the orthogonal prior component and restoring the measured evidence deficit). The edited state is written back to the same layer, and decoding proceeds with the original language head and unchanged decoding rule.

All constants are fixed before evaluation. On Qwen2.5-Omni-7B, OWP edits evidence layers 12, 16, and 20 (weights 0.25/0.35/0.40) and prior layers 24, 26, and 27 (0.23/0.61/0.16), with α_P = α_R = 0.10, β_P = β_R = 0.95, η_P = 0.25, τ = −9/16, and T = 0.10. Transferring to VideoLLaMA2-AV adapts only the edited layers (16, 20, 24) and two residual gains (α_R = 0.20, β_R = 1.20). Baselines compared are AVCD, MAD, and HulluEdit, with VITA-1.5, OmniVinci, and Qwen3-Omni reported separately as external references. Videos are sampled at 2 frames per second with at most 32 frames; statistics use exact two-sided McNemar tests and 5,000 bootstrap resamples over 4,604 examples where every method returns a valid prediction.

Why This Matters

The paper's central claim — that hallucinations should be fixed where they originate, not where they appear — gives a template for intervening on internal model state rather than post-hoc logit tweaks, and it does so without retraining. Because OWP transfers to a backbone with a different fusion design, the evidence-versus-prior separation looks like a property of early-fusion architectures rather than a quirk of one model.

Real-world applications:

• Assistive and wearable devices that describe a scene while also hearing ambient sound, where a visible object must not override what the microphone actually captured.
• Video conferencing and media accessibility tools that caption or summarize, where the wrong audio track or an unrelated image could otherwise dominate the description.
• Safety monitoring and alarm verification over audio-visual feeds, where confirming "was that sound real?" requires trusting the microphone over the picture.
• Interactive robots and in-vehicle assistants that must decide whether a query is answerable from sound, from sight, or only from both.

Industry relevance: the method is a drop-in wrapper for frozen models, requires no labeled data, no reference answers, and no per-example supervision, and costs 1.71×–1.77× backbone latency with at most five forward calls. That profile makes it plausible for deployment where retraining is impractical, and the authors release code publicly.

Future Directions

• Extending beyond audio, video, and text. The four-view diagnosis is defined over exactly three input modalities; whether the same conflict-conditioned decomposition scales to additional sensors or streams is not established here.

• Removing the manual per-backbone tuning. Transfer to VideoLLaMA2-AV still adapts the edited layers and two residual gains by hand; automating layer selection and gain fitting is an open step toward truly plug-and-play use.

• Evaluating open-ended generation. Quantitative scoring uses a finite-answer probe over closed answer sets, with qualitative examples from the autoregressive decoder; how the edit behaves on long free-form responses is not quantified in the reported results.

• Combining with training-time methods. The paper contrasts OWP with preference-optimization and RLHF-style approaches; whether representation-level inference edits and weight-level alignment are complementary, rather than alternatives, is untested.

• Reducing the remaining overhead. The conflict-triggered schedule still requires extra views; the reported costs (1.71× and 1.77× latency, 24.9 GiB and 17.3 GiB peak memory) leave room for cheaper conflict detection.

Target Audience

Researchers and engineers working on multimodal large models, hallucination mitigation, or inference-time intervention — particularly those familiar with transformer internals and decoding methods. It is also useful for practitioners who need to improve a deployed Omni LLM's reliability without access to training pipelines or labeled hallucination data, and for students of representation engineering who want a worked example of turning activation-level analysis into a full evaluation campaign.

Authors’ abstract

Recent Omni LLMs (OLLMs) have demonstrated remarkable capabilities in joint audio-visual reasoning, yet cross-modal hallucinations remain a major challenge. Existing mitigation methods mainly operate at the output level, overlooking the entanglement of grounding evidence and learned priors within fused representations. To address this issue, we propose Omni-Woodpecker (OWP), a training-free framework that shifts hallucination mitigation from output-level manipulation to conflict-conditioned representation rebalancing. Specifically, OWP first identifies cross-modal conflicts from complementary modality-specific views. It then employs orthogonal decomposition to separate grounding evidence from competing priors within the hidden states. Finally, OWP adaptively rebalances fused representations by suppressing misleading priors and restoring ignored evidence before decoding. Extensive experiments demonstrate the effectiveness and strong cross-architecture generalization of OWP. Using Qwen2.5-Omni-7B as the backbone, OWP improves performance by 12.9 and 8.6 points on CMM and AVHBench, respectively, while consistently achieving gains of 0.67-2.25 points across four general OLLM reasoning benchmarks. Furthermore, OWP generalizes effectively to VideoLLaMA2-AV, improving performance by 4.6 and 3.9 points on the same benchmarks. Code is publicly available at https://github.com/ZiiJianLiu/Omni-Woodpecker