video model developers

Abstract: Video models have the architecture to support encoding temporal aspects of the input, i.e. concepts that require more than one frame to define, like object movement. However, because motion and appearance change are entangled, a feature that appears to encode temporal concepts like object movement may instead encode non-temporal concepts like spatial presence at different positions whose activations merely correlate with motion. To resolve this, we introduce the No-False-Positives (NFP) test, a statistical test with a formal guarantee of statistical specificity for identifying features encoding temporal concepts. Using specially designed stimuli, we apply this test to representations produced by VideoMAE, V-JEPA2, and TimeSformer, and observe that temporal features exist in each model. We validate the test first by showing it yields no temporal features for the DINOv2 image-encoder. Then, using a synthetic representation space, we show that detection rate increases with a direction's projection onto the temporal subspace, and that Sparse Autoencoder features test-flagged as temporal project much more of their norm onto the temporal subspace than unflagged features. Interventions on temporal SAE features yield mixed results across models, suggesting that intervening on temporal features may not outperform a non-temporal feature baseline.

Abstract: World action models increasingly rely on pretrained video models assumed to capture physical dynamics for prediction, planning, and action generation. Existing evaluations assess plausibility, probe representations, or measure predictive surprise, but rarely identify which history-conditioned physical expectations conflict with observations. We introduce PhysTrace, a training-free framework aligning MLLM-generated, history-conditioned expectations with observed video windows to temporally localize physical violations and provide interpretable descriptions without task-specific probes or classifiers. We also introduce IntPhys2-Timing, adding violation intervals to IntPhys2. PhysTrace achieves a Hit Rate of up to 70.7% across physical principles on IntPhys2-Timing and outperforms the compared methods in overall possible-impossible discrimination accuracy on IntPhys2. Applied to reconstructed videos along the trajectories of pretrained diffusion/flow-based video models, PhysTrace reveals attribute- and noise-dependent patterns in physical violation localization, enabling interpretable physical consistency analysis across timesteps.

Abstract: A memory test for a video world model should say whether the scene was set, whether the asked change was carried out, and whether the last frame shows the result. Current tests fold these into one number, so a low score is easy to read as forgetting. When the change is not carried out, the hidden state transition cannot appear, so the test is not really a test of memory. We introduce MIST, an evaluation that separates the change from the memory it should leave behind. The model first sets the scene while the change stays hidden, and only later is asked to show it, with leakage before that point kept under control. This separation locates the bottleneck: the model sets the scene but does not carry out the instructed change. The same gap appears in general memory evaluations. On that basis we propose M3-Agent (Make the Missing Move), which recursively revises the details of the input instruction so that the model successfully carries out the requested instruction. Relative to the baseline, it more than triples the MIST reveal score. On the benchmarks, WRBench's score T rises by 26.1% and MBench-T's score U by 52.7%. The method meets the problem from the execution side: it shows what the frozen model can do once the move is stated, and it sets a stricter baseline.

Abstract: Modern video models can generate realistic videos from real appearance references and proxy renders that specify scene structure, viewpoint changes, and motion. Evaluating this render-to-real capability requires a real target video depicting the same scene evolution, paired with an editable, geometrically registered 3D replica. Such data has traditionally required substantial manual modeling, calibration, and animation effort. We introduce RenderBench, a benchmark of 12 reconstructed real-world scenes spanning large-scale indoor environments and egocentric viewpoints, with both static and dynamic settings. Our construction pipeline combines visual geometry, neural reconstruction, and assisted 3D authoring. Each scene is decomposed into static objects and dynamic actors, registered to the capture cameras, and accepted only after multi-view geometric and temporal validation. Each evaluation unit contains appearance reference images, a held-out real target video, an editable digital twin, a matched proxy render, and renderer-native scene annotations. We evaluate transfer models against paired real target videos, retain PAI-Bench-C-compatible structural projections, and use scene annotations to localize failures by object, visibility, articulation, and motion. The first release retains 12 of 14 registered samples (85.7%), comprising 1,496 paired real-proxy frames. All released scenes pass file-integrity and environment-edit audits, while proxy diagnostics yield a depth si-RMSE of 0.2170 and instance mIoU of 0.3673. RenderBench provides paired real observations and editable scene state for assessing both appearance fidelity and preservation of geometry and dynamics.