Evaluating and Addressing Multimodal Hallucination

Abstract

Headline numbers

Three numbers from three different probes. Self-attribution is unreliable, adversarial text dominates, but incidental lexical contradiction is survivable.

Key Insight

Prior work on lexical bias (Chen et al., 2025) showed that audio-LLMs over-rely on transcribed words rather than acoustic cues. We adapt this counterfactual logic from audio-only to omnimodal Video-LLMs, and extend it from "is the model using modality X" to "does the model know whether it is using modality X".

The gap between (a) behavioral and (b) verbal attribution is the core methodological contribution. Attribution Faithfulness Score (AFS) quantifies this gap on a per-task basis, restricted to falsifiable cases where ablation could in principle change the answer.

Method

Seven-stage ablation pipeline

Each AVUT question is presented to the same omnimodal model in seven configurations. Cross-stage comparisons turn into diagnostic metrics.

Three diagnostic metrics combine the stages

We additionally report per-stage Expected Calibration Error (ECE) and ΔConfidence (the drop in self-reported confidence when a modality is removed) as exploratory signals on verbal-vs-behavioral attribution.

Cross-model accuracy

Overall accuracy by stage on the same 600 questions, Qwen vs Gemma.

Attribution faithfulness splits cleanly by task family

Both models confabulate at the same task boundary: pure-audio tasks (AIE, ACC, AEL) score above chance; cross-modal binding tasks (AVCM, AVOM, AVTM) sit at or below chance.

Adversarial text injection: catastrophic override

For each AVUT question with substantial speech content, GPT-4o generates a short adversarial transcript anchored at one of the three wrong options. The transcript is injected alongside the unmodified audio and video. The audio and video continue to support the correct answer throughout — only the text channel is poisoned.

A model that truly integrated all three channels in a load-bearing way should produce a TOR-flip near zero — even when text contradicts, two of three channels still point at the correct answer. The observed TOR-flip of 0.69 is a direct demonstration of modality bias toward text. The steered fraction of 0.81 (uniform 0.78–0.82 across tasks) shows that flips are not random thrashing; the model is reliably pulled toward whatever option the injection points at.

Verbal vs. behavioral attribution gap

Although Gemma's verbalized self-attribution is unreliable (AFS = 0.49), its confidence drops under ablation are calibrated correctly.

The headline LOR result reinforces this nuance. On same-task mismatched transcripts (incidental contradiction), Gemma's LOR sits at 0.163 — 84% of audio-correct answers survive. On adversarial GPT-4o-generated transcripts (intentional contradiction), Qwen's TOR-flip is 0.69. Modern omnimodal models appear robust to incidental lexical noise but remain vulnerable to targeted text-channel attacks.

Modality importance ablation on MUStARD++

A separate experiment on the MUStARD++ sarcasm benchmark using Qwen2.5-Omni-7B per-modality descriptions fed to a final Yes/No gate. The descriptions are pre-generated; only which descriptions reach the gate is varied.

BibTeX

@misc{avadhanam2026multimodalhallucination,
  title   = {Evaluating and Addressing Multimodal Hallucination},
  author  = {Avadhanam, Pranav and Deshpande, Aarya and Malpani, Rishit
             and Singh, Gurraj and Singh, Kunal and Syed, Samad
             and Varikooty, Vikram and Wang, Jeffrey},
  year    = {2026},
  note    = {CS 639 Course Project, University of Wisconsin--Madison},
  url     = {https://github.com/PranavAvadhanam/proj-multimodal}
}