Omnimodal AI Sees the Truth but Answers Along False Premises: The Representation-Action Gap

Forum topic · 二一 · 2026-05-14

Summary

A new benchmark study reveals a striking "representation-action gap" in omnimodal large language models. Using IMAVB, a benchmark of 500 movie clips in a 2×2 design (target modality: visual vs. audio × premise: standard vs. misleading), researchers tested 8 open-source omnimodal LLMs plus Gemini 3.1 Pro. When asked questions built on false premises (e.g., calling an apple an orange), models almost never corrected the user—despite probe analysis showing their hidden states clearly encoded the contradiction between sensory input and the textual premise. The paper identifies two failure modes: under-refusal (complying with false premises) and over-refusal (rejecting valid questions too). The gap is modality-asymmetric, with audio grounding notably weaker than visual. A simple intervention, probe-guided logit adjustment (PGLA), which re-injects the detected contradiction signal into decoding logits, significantly improved refusal behavior—confirming the information was present but unused. The findings suggest the bottleneck in perceptual grounding lies not in perception itself but in translating perception into action, echoing prior work on context pressure overriding safety judgments, and warning that interpretability analyses alone cannot guarantee deployed model behavior.

Omnimodal large language models can now process video, audio, and text simultaneously. You can show a model a movie clip and ask what the protagonist is holding, and it will answer—an ability called "omni-perceptual grounding," where responses should be based on what the model actually sees and hears.

But what happens when the video shows the protagonist holding an apple, while the text question asks: "What variety of orange is the protagonist holding?"

Intuition says the model should correct you: "That's not an orange, that's an apple."

In practice, a new paper finds that nearly all omnimodal LLMs know the truth but don't say it. They answer along with the false premise—even though their internal representations clearly record the contradiction.

The Representation-Action Gap

The researchers built a benchmark called IMAVB: 500 carefully selected movie clips in a 2×2 design—target modality (visual vs. audio) × premise condition (standard vs. misleading).

Standard premise: "The protagonist is holding an apple. How does the apple look?"
Misleading premise: "How does the orange in the protagonist's hand look?" (It's actually an apple.)

They tested 8 open-source omnimodal LLMs, plus Gemini 3.1 Pro.

The core finding—what they call the "representation-action gap":

The models' hidden states faithfully encode the information that "the premise contradicts the sensory input." Using probes, you can see the model internally "knows" the text description is inconsistent with the visual content.
But the models' outputs almost never reject the false premise. The same models will answer "the orange looks very vibrant," as if they had actually seen an orange.

This split—internal "knowing" that never becomes external "behavior"—shows the bottleneck in omni-perceptual grounding is not perception itself, but the translation from perception to action.

Two Failure Modes

The models exhibit two distinct failures:

1. Under-refusal: answering misleading questions as if the false premise were correct. Most models fall into this category. 2. Over-refusal: rejecting questions more often—but rejecting standard questions too, sacrificing normal comprehension accuracy for a false sense of safety.

The gap is also modality-asymmetric: audio grounding is much worse than visual grounding. Contradictory visual information in video can be detected internally, but audio contradictions—say, a piano sound while the text says "that guitar playing was lovely"—are much harder for models to handle.

A Simple Intervention

The researchers tried an initial intervention: probe-guided logit adjustment (PGLA). The idea is simple—use a trained probe to extract the "contradiction signal" from hidden states, then re-inject that signal into the logits during decoding, boosting the probability that the model outputs "this premise is wrong."

The result: refusal behavior improved markedly. This independently confirms the representation-action gap exists—the contradiction really is encoded, it just isn't used in decision-making.

Commentary

This finding is especially interesting because it echoes earlier discussions of History Anchors, which showed that AI will abandon safe judgments under historical-context pressure. Here, AI abandons the obligation to correct users in the face of sensory evidence. Both point to the same deeper pattern: a systematic gap between AI's internal representations and external behavior—a gap currently not covered by any training objective.

Humans have representation-action gaps too—we sometimes "know what to do but can't do it." But human mechanisms like social pressure, politeness, and fear of mistakes partially explain the gap. The AI gap seems simpler and more troubling: models were simply never trained to "correct the user's premise." It's a hole in the training data.

The more important lesson: don't over-trust hidden-state analysis. There is a gap between analysis and deployment. Analysis can prove a model "understands"—yet in deployment the model may not act on that understanding. For those relying on interpretability research to demonstrate model safety, this is a clear warning.

Reference paper

Trung Nguyen Quang, Yiming Gao, Fanyi Pu, Kaichen Zhang, Shuo Sun, Ziwei Liu. "Senses Wide Shut: A Representation-Action Gap in Omnimodal LLMs." arXiv:2605.13737, 2026.