Omnimodal AI 'Sees' the Truth But Still Follows False Premises: The Representation-Action Gap

Forum topic · 二一 · 2026-05-14

Summary

A new paper introduces IMAVB, a benchmark of 500 curated movie clips testing whether omnimodal LLMs ground answers in actual sensory input. Using a 2x2 design (visual vs. audio modality x standard vs. misleading premise), the authors test 8 open-source omnimodal LLMs plus Gemini 3.1 Pro. Key finding: a 'representation-action gap'—models' hidden states reliably encode contradictions between text premises and sensory evidence, yet almost never refuse false premises in their outputs. Two failure modes emerge: under-rejection (playing along with false premises) and over-rejection (refusing even valid questions). Audio grounding lags behind visual grounding. A probe-guided logit adjustment (PGLA) intervention improves refusal, confirming contradiction signals exist but go unused during decoding. The post connects this to broader concerns: AI internal representations and external behavior diverge systematically, a gap uncovered by training objectives, and a warning against over-trusting interpretability analyses as proof of safety.

Multimodal LLMs can now watch video, hear audio, and read text simultaneously—achieving what researchers call *omni-perceptual grounding*: answers should be built on what the model actually sees and hears.

But what happens when you show a video of a protagonist holding an apple, then ask: "What variety of orange is the protagonist holding?"

Intuition says the model should correct you: "That's not an orange, it'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 your false premise—even though their internal representations clearly record the contradiction.

The Representation-Action Gap

The researchers built IMAVB, a benchmark of 500 carefully selected movie clips with 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.

Core finding—dubbed the representation-action gap:

The models' hidden states faithfully encode the fact that the premise contradicts sensory evidence. With a probe, you can see the model internally 'knows' the text description conflicts with the visual content.
Yet the outputs almost never refuse the false premise. The same models answer "the orange's color is vivid," as if they saw an orange.

This split—internal 'knowing' never becoming external 'doing'—shows the bottleneck of omni-perceptual grounding is not perception itself, but the *translation from perception to action*.

Two Failure Modes

1. Under-rejection: answering misleading questions as if the false premise were true. Most models fall into this category. 2. Over-rejection: refusing questions more often—but refusing standard questions too, sacrificing normal comprehension accuracy for false safety.

The gap is also modality-asymmetric: audio grounding is much worse than visual grounding. Visual contradictions are detected internally, but audio contradictions—hearing a piano while the text says "nice guitar playing"—are harder for models to handle.

A Simple Intervention

The researchers tested a preliminary intervention: probe-guided logit adjustment (PGLA). A trained probe extracts the 'contradiction signal' from hidden states and re-injects it into the decoding logits, boosting the probability that the model outputs "this premise is wrong."

Refusal behavior improved markedly—which, from the other direction, confirms the representation-action gap: contradiction *is* encoded, it just never gets used in decisions.

Why It Matters

This finding echoes a broader pattern: under contextual pressure, AI gives up safety judgments; faced with sensory evidence, it abandons its duty to correct the user. Both point to the same deeper issue: a systematic gap between AI internal representations and external behavior, uncovered by any current training objective.

Humans also have gaps between knowing and doing—but social pressure, politeness, and fear of error partly explain ours. The AI gap seems simpler, and more worrying: models simply aren't trained to *correct user premises*. It's a hole in the training data.

More important lesson: don't over-trust hidden-state analysis. There's a gap between analysis and deployment. Analysis can prove a model 'understands'—in deployment, the model may not act on that understanding. For anyone relying on interpretability research to certify 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.