DMES: Information-Equivalent Evaluation Reveals the Physical Reasoning Gap Between World Models and Language Models

Evaluating physical reasoning across fundamentally different architectures—large language models (LLMs) and vision-based world models—is confounded by the fact that text and image inputs carry naturally non-equivalent information. We propose Dual-Mode Equivalent Stimulus (DMES), an evaluation framework that constructs text-image pairs with rigorously equal information content (H(T) =H(I) = H(S)), eliminating modality confounds from cross-architecture comparisons. Applying DMES to classical mechanics, we create DMES-PC, a 140-sample benchmark spanning 7 physics categories, and evaluate three LLMs (GLM-4.7, Qwen3-Coder-Next, Gemma 4; 26B–355B parameters) against V-JEPA 2 (300M parameters) under three evaluation paradigms. Our results reveal a clear parameter efficiency paradox: V-JEPA 2 achieves both higher prediction accuracy (mean score 0.692, SD=0.032) and substantially lower variance than all tested LLMs (mean 0.320–0.600, SD=0.355–0.383). V-JEPA 2 significantly outperforms GLM-4.7 (Cohen’s d = 1.27, p < 0.001) and Gemma 4 (d = 1.40, p < 0.001), while the comparison with the strongest LLM, Qwen3-Coder-Next, does not reach significance (d = −0.34, p _adj = 0.573)—a consequence of that model’s high variance overlapping with V-JEPA 2’s tight score range. A three-paradigm analysis (direct prediction, code simulation, hybrid) further reveals a “knowing vs. simulating” gap: LLMs generate near-perfect physics code when they succeed (score ≈ 1.0), but their direct predictions remain unreliable (score ≈ 0.5), with low internal consistency (0.32–0.69). These findings suggest that scaling LLMs alone is insufficient for physical world understanding—architectures with predictive coding inductive biases are needed.