MATH · IN · MODELS

OpenVLA linearly encodes a state-transition vector in middle layers

measured in 1 paper

Molinari et al. test whether OpenVLA, a 7B behavioral-cloning VLA with no world-model objective, encodes environment dynamics as a linear embedding offset delta_e = e_{t+K} - e_t [molinari-etal-2025-openvla-world-model] Lasso linear probes recover this transition vector at statistically significant R^2 across all 4 LIBERO suites and 4 horizons, with 123 probes passing permutation tests at p<0.0001 [molinari-etal-2025-openvla-world-model] Probes on raw activations uniformly exceed probes on embeddings alone, and linear probes never significantly underperform MLP probes (42/48 favor linear) [molinari-etal-2025-openvla-world-model] The structure concentrates in middle layers (~15) and emerges with pretraining scale; an early checkpoint and eval-only finetuning weaken it, with no causal intervention performed [molinari-etal-2025-openvla-world-model]

Context

Mikolov-style state-transition vector (delta_e = e_{t+K} - e_t), Lasso linear probes vs. MLP probes, R^2 with bootstrapped CIs, embedding-only baseline as a probe-correlationality control, permutation tests (123 probes, overall p < 0.0001), middle-layer concentration and pretraining-scale dependence, Koopman-operator theoretical justification for probing activations

Confirmed in models

Papers

Emergent World Representations in OpenVLA — Molinari, Marco, Nevali, Leonardo, Navani, Saharsha, Younis, Omar G.2025 · arXiv:2509.24559