💻 computer science

World Models in Words: Auditing Physical State-Transition Commitments in Vision-Language Models

This paper introduces \wmw, an evaluation framework and the \tracebank dataset that audit Vision-Language Models by requiring them to generate typed physical state-transition traces, thereby revealing hidden reasoning failures and enabling improvements in physical consistency that traditional answer-only benchmarks miss.

Original authors: Emmanuelle Bourigault

Published 2026-05-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Emmanuelle Bourigault

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hiring a student to solve a physics problem. In the old way of testing, you only looked at the final answer they wrote on the paper. If they got the right number, you gave them an A. If they got it wrong, you gave them an F.

The problem is, a student could get the right answer by pure luck, or by guessing, or by using a completely wrong logic that just happened to land on the correct number. You wouldn't know how they got there, only what they got.

This paper introduces a new way of testing AI models (specifically Vision-Language Models) called World Models in Words (WMW). Instead of just checking the final answer, the researchers force the AI to show its work step-by-step, like a detailed diary of its thoughts.

The New "Show Your Work" Test

The researchers ask the AI to produce a Trace, which is a four-part story:

The Starting Scene ( $s_0$ ): What does the AI see right now? (e.g., "A ball is sitting on a ramp.")
The Action ( $\Delta s$ ): What rule is happening? (e.g., "Gravity is pulling the ball down.")
The Result ( $s_1$ ): What happens next? (e.g., "The ball rolls down the ramp.")
The Answer ( $a$ ): The final conclusion.

The AI has to write this story in a very specific, structured format (like a JSON code block) so a computer can check it.

The "Hybrid Verifier" (The Strict Teacher)

Once the AI writes its story, a special computer program called a Hybrid Verifier reads it. This isn't just a simple spell-checker. It acts like a strict physics teacher who checks:

Did you see the right objects? (Is that a ball or a box?)
Is the physics real? (Does gravity pull up or down?)
Does the story make sense? (If the ball rolls down, does it end up at the bottom?)
Does the answer match the story? (If the story says the ball rolls left, but the answer says "Right," the AI is lying.)

If the AI gets the right answer but tells a fake story (e.g., "The ball rolled left because of magic"), the Verifier catches it. The paper calls this a "Hidden Inconsistency."

The Big Discovery: "The Lucky Guessers"

The researchers tested 7 different AI models. They found something surprising:

Many models got the right final answer.
But, for about 18% to 42% of those "correct" answers, the story the AI told to get there was physically impossible.

It's like a student who guesses the right answer on a math test but writes down a solution that involves dividing by zero. The old test would have given them an A; this new test reveals they don't actually understand the math.

The "World Bank" of Physics Stories

To make this test possible, the authors created a massive library called WMW-TRACEBANK.

It contains 200 carefully crafted physics scenarios (like ramps, springs, and collisions).
It has 3,200 pairs of "Good Stories" vs. "Bad Stories." In a "Bad Story," they changed just one tiny thing (like flipping a force direction) to see if the AI could spot the error.
This library helps train AI to stop guessing and start telling the truth about how the physical world works.

Fixing the AI

The paper also tested ways to fix these "liars."

Reranking: They let the AI generate 8 different stories and picked the one that passed the Verifier's check. This improved the quality of the stories without changing the final answers.
Preference Training: They taught the AI to prefer "Good Stories" over "Bad Stories" using the 3,200 pairs. This reduced the number of hidden lies by 41%.

The Bottom Line

The main point of this paper is not to build a better physics calculator. It is to build a truth detector.

Before we trust AI to help us with real-world tasks (like planning for robots or teaching students), we need to know if their internal story of the world matches reality. This paper gives us a tool to audit that story, ensuring that when an AI says "The ball rolls down," it actually understands why and how, rather than just guessing the right word.

Technical Summary: World Models in Words

Problem Statement
Vision-Language Models (VLMs) are increasingly deployed to answer questions about physical scenes, yet standard evaluation metrics rely solely on final answer accuracy ( $I, q \mapsto a$ ). This approach collapses intermediate reasoning errors—such as misperceiving objects, misrepresenting physical states, or predicting impossible transitions—into a single binary score. Consequently, a model may select the correct multiple-choice option while asserting a physically invalid world state (e.g., an object accelerating against the net force). This "hidden inconsistency" is particularly problematic when VLMs serve as interfaces for planning, robotics, or scientific reasoning, where downstream tools may act on these inconsistent commitments.

Methodology: WORLD MODELS IN WORDS (WMW)
The authors introduce WORLD MODELS IN WORDS, a framework that shifts evaluation from answer-only scoring to auditing the language-expressed physical commitments of VLMs.

Trace Formalism: Instead of a direct answer, models are prompted to generate a typed trace $\tau = (s_0, \Delta s, s_1, a)$ $τ = (s_{0}, Δ s, s_{1}, a)$ :
- $s_0$ : Initial physical state (entities, attributes, forces, relations).
- $\Delta s$ : Predicted transition or physical rule.
- $s_1$ : Resulting state.
- $a$ : Final answer.
- The trace is structured as JSON to enable parsing and verification.
Hybrid Verifier: A deterministic, inspectable verifier checks the trace against four criteria:
1. Schema Validity: Correct JSON structure and types.
2. State Grounding: Consistency with visual evidence and metadata.
3. Transition Consistency: Adherence to physical laws (e.g., force-acceleration alignment, temporal order).
4. Answer-Trace Faithfulness: Compatibility between the derived state and the final answer.
Failure Taxonomy: The system categorizes errors into typed labels: Object, State, Relation, Force, Transition, Intervention, Temporal, Unit/Scale, and Faithfulness.
WMW-TRACEBANK: A controlled resource released with the paper containing:
- 200 synthetic scenarios across 17 physics families (mechanics, fluids, circuits, optics, etc.).
- 3,200 contrastive preference pairs (valid traces vs. minimally perturbed invalid traces).
- Verifier code, audit guidelines, and model outputs.

Key Contributions

Trace Formalism: A protocol for evaluating VLM physical reasoning as auditable state-transition commitments rather than answer strings.
Verifier Contracts & Metrics: Definitions for separating state extraction errors, transition errors, and hidden inconsistencies (where the answer is correct but the trace is invalid).
Dataset Release: WMW-TRACEBANK, providing validated synthetic traces and preference pairs for training and evaluation.
Failure Taxonomy & Audit: A human-audited evaluation of the verifier on 400 stratified traces, reporting precision, recall, and inter-annotator agreement.
Intervention Protocol: Demonstration that verifier-guided reranking and trace-level preference tuning can improve physical consistency without sacrificing answer accuracy.

Experimental Results
The authors evaluated 7 VLMs (including Claude Opus 4.7, GPT-5.5, GPT-4o, and open-source models like Qwen2.5-VL and InternVL3) on both controlled and external examples.

Hidden Inconsistency (RQ1): Final-answer accuracy masks significant physical errors. Across all models, 18% to 42% of correct answers were backed by physically invalid traces. For mid-tier models, over 30% of "correct" answers relied on invalid physical reasoning.
Bottleneck Analysis (RQ2): The transition step ( $\Delta s$ ) is the primary bottleneck for most models. Even when provided with gold initial states, models like GPT-4o lost 19 percentage points in accuracy when predicting the physical change. Stronger closed models (e.g., Claude Opus 4.7) showed a calibration effect where gold states disrupted their internal consistency checks.
Verifier Reliability (RQ3): The hybrid verifier achieved high agreement with human audits for State (F1=0.82), Transition (F1=0.78), and Faithfulness (F1=0.91) labels.
Intervention Efficacy (RQ4):
- Reranking: Selecting the best trace from $k=8$ samples improved trace validity by 4–7 percentage points.
- Preference Tuning: Training Qwen2.5-VL-7B with Direct Preference Optimization (DPO) on the preference pairs reduced hidden inconsistency by 41% relative (from 42% to 25%) while slightly improving answer accuracy (+2pp).

Significance and Claims
The paper argues that WORLD MODELS IN WORDS is not merely another physics benchmark but a reusable protocol for auditing whether a VLM's stated physical world is coherent. The authors claim that:

Correct answers with invalid traces are not harmless formatting errors; they represent inconsistent commitments that can mislead educational, planning, or scientific systems.
The framework reveals that transition prediction is a dominant failure mode distinct from visual perception.
Verifier-guided interventions (reranking and preference tuning) can effectively reduce hidden inconsistencies, suggesting a path toward more reliable physical reasoning modules.

The authors maintain a modest stance, acknowledging limitations such as the observability of traces (which are user-facing commitments, not direct reads of latent cognition), the coverage of the verifier (which relies on explicit assumptions for real-world ambiguity), and the potential for models to overfit to perturbation templates. They explicitly state the dataset is not a safety certificate for high-stakes applications.