Technical Summary: TRAPSBench: Vision-Language Models Encode but Fail to Express Epistemic Restraint

Problem Statement

Vision-Language Models (VLMs) have achieved significant progress in video understanding and physical reasoning. However, for deployment in high-stakes domains, a model's ability to recognize when sensory information is insufficient for a deterministic prediction is as critical as its ability to answer correctly. Current benchmarks evaluate physical reasoning but fail to test selective abstention under conditions of insufficient evidence. Existing metrics often reward indiscriminate abstention or ignore the epistemic dimension entirely, failing to distinguish between a model that knows it doesn't know and one that simply guesses.

This paper addresses the gap between a model's internal knowledge of uncertainty and its external expression of that uncertainty. Specifically, it investigates whether VLMs can internally distinguish between answerable and unanswerable physical scenarios, and if so, why they fail to express this distinction in their autoregressive outputs.

Methodology

1. TRAPSBench Benchmark

The authors introduce TRAPSBench (Testing Restraint in Ambiguous Physical Scenarios), a procedurally generated video benchmark consisting of 1,404 matched physics pairs.

• Generation: Using MuJoCo, the benchmark creates "control" videos where outcomes are deterministic and "void" videos where the outcome is rendered incomputable by a single targeted modification.
• Uncertainty Taxonomies:
  • Occlusion: Key visual data is blocked by an opaque occluder ($N=202$).
  • Chaotic Sensitivity: Deterministic systems with extreme sensitivity to initial conditions, truncated before resolution ($N=500$).
  • Ill-Posed Questions: Control videos are reused, but the question contains a false premise or asks for unobservable details ($N=702$: 202 occlusion ill-posed + 500 chaotic ill-posed).
• Evaluation: A text-only judge panel (three models) scores responses. For control videos, correctness is measured against MuJoCo ground truth. For void videos, the judge detects whether the model abstained.

2. Penalized Epistemic Calibration Score (PECS)

To evaluate performance, the authors propose PECS, a conjunction metric defined as:
$$\text{PECS} = \text{Acc} \times \max(0, \text{AbsRec} - \text{FalseAbs})$$
Where:

• Acc: Accuracy on answerable (control) videos.
• AbsRec: Abstention Recall (correctly abstaining on void videos).
• FalseAbs: False Abstention (incorrectly abstaining on control videos).
• Significance: The metric uses Youden's J statistic ($\text{AbsRec} - \text{FalseAbs}$) to penalize indiscriminate strategies. A model that always abstains or never abstains receives a PECS of 0. High scores require both high accuracy on answerable tasks and selective abstention on unanswerable ones.

3. Mechanistic Analysis

The paper employs two techniques to probe internal representations:

• Linear Probing: Training logistic regression probes on hidden states of frozen VLMs to predict "void" vs. "control" labels. This tests if the model encodes the distinction.
• Activation Steering: Causally manipulating hidden states by adding a "void direction" vector ($v_\ell$) during generation to test if the representation causally controls abstention behavior.

Key Contributions

1. TRAPSBench and PECS: A large-scale benchmark and a robust metric designed specifically to evaluate epistemic calibration in video understanding, zeroing out degenerate strategies like always-abstain or never-abstain.
2. The Representation–Output Gap: The paper establishes that VLMs internally encode epistemic uncertainty but fail to express it. This is demonstrated through:
   • Guided Prompting: Explicitly instructing models to abstain ("I don't know") unlocks latent capabilities, significantly improving abstention recall without sacrificing accuracy.
   • Linear Probing: Probes decode the answerability signal from hidden states with high AUROC (up to 0.91) across physics domains, even on samples where the model confidently hallucinates.
   • Activation Steering: Injecting a single-layer "void direction" causally induces abstention in control samples and suppresses it in void samples.
3. Failure Mode Asymmetries:
   • Visual vs. Textual: Models detect textual impossibility (ill-posed questions) roughly 4$\times$ more readily than visual information gaps.
   • Reasoning Effects: Chain-of-thought (CoT) reasoning does not uniformly improve calibration. In some models (e.g., Qwen3-VL Think), reasoning traces express doubt that is subsequently overridden by the final output, leading to increased confabulation compared to non-thinking variants.
4. Causal Mechanistic Analysis: The geometry of the "void" signal varies by domain. Occlusion-family directions encode a domain-general "evidence is missing" signal that transfers across modalities and domains. Chaotic-family directions are domain-specific and near-orthogonal, suggesting epistemic transparency governs transferability.

Results

• Spontaneous Restraint is Poor: Across 16 VLMs from five families (Gemini, Qwen, GPT-5, Gemma, LLaVA), spontaneous PECS is low. The best spontaneous PECS is 0.292 (Gemini 2.5 Flash). Most models exhibit a strong prior to answer regardless of evidence quality.
• Guided Prompting Unlocks Latency: Explicit prompting instructions to abstain increase Abstention Recall by a median of 1.9$\times$ across video-native models, with PECS improving significantly (e.g., Gemini 3.1 Pro R-Low reaches 0.568). However, even with guidance, models do not reach saturation, indicating a persistent bottleneck.
• Internal Encoding vs. Output Suppression:
  • Linear probes achieve AUROC up to 0.91 in cross-dataset transfer, proving the model "knows" the outcome is unknowable.
  • Crucially, this signal persists even when restricted to samples where the model confabulated (failed to abstain) under standard prompting.
  • Activation steering confirms this signal is causal: adding the void direction to hidden states forces abstention.
• Prompt Gating: In Qwen3-VL-8B, standard prompting imposes a one-way constraint against abstention. While steering can force confabulation (suppressing abstention) under standard prompts, it cannot induce abstention unless the "guided" prompt is used. This suggests the bottleneck is an output-stage gate, not a lack of internal representation.
• Reasoning and Confabulation: Analysis of confabulation modes reveals that 87–99% of failures involve hallucinated premises (fabricating visual observations). In Qwen3-VL Think, the reasoning trace often expresses doubt, but the final output overrides this to provide a confident, fabricated answer.

Significance and Claims

The paper claims that the primary bottleneck for reliable VLM deployment in physical reasoning is expressive, not perceptual.

• Representation-Output Gap: VLMs possess the internal epistemic signals required to distinguish answerable from unanswerable scenarios. The failure lies in the autoregressive generation process, which suppresses these signals in favor of fluent, committed answers.
• Modality Specificity: The gap is more pronounced for visual uncertainty than textual uncertainty. Models are significantly better at detecting semantic impossibilities (text) than physical information gaps (video).
• Intervention Necessity: Closing this gap likely requires output-stage interventions (e.g., decoding constraints, steering, or specialized training objectives) rather than simply scaling model size or improving visual perception.
• Generalizability: The findings replicate across three open-weight families (Qwen, Gemma, LLaVA) with no shared training pipelines, suggesting this is a fundamental property of current VLM architectures rather than an artifact of a specific model.

The authors conclude that TRAPSBench provides a necessary tool for evaluating and improving the epistemic calibration of AI systems, highlighting that "knowing when not to answer" remains a critical, unsolved challenge for autonomous agents.