arXiv:2609.37938v1 Announce Type: cross Abstract: Embodied systems must make knowledge acquired during one encounter usable in another despite changes in viewpoint, motion, and illumination. Yet aggr...
R2M-Bench is a benchmark that evaluates revisit memory in interactive video world models by comparing a revisit pair to two control pairs from the same rollout: a gap‑matched non‑revisit pair and a short‑range pair. It introduces MemoryGain (MG) and Normalized Memory Ratio (NMR) to quantify the revisit advantage over generic temporal stability and normalize it by short‑to‑baseline dynamics. Across 300 instances and seven models, NMR correlates with human judgments and reduces the influence of slow‑motion artifacts, with DreamX‑World‑Memo achieving the highest NMR.
Embodied systems must make knowledge acquired during one encounter usable in another despite changes in viewpoint, motion, and illumination. Yet aggregate cross-video accuracy conflates failures of lo...
arXiv:2607. 15621v1 Announce Type: cross Abstract: Large language models bring instruction following and scene reasoning to end-to-end driving, but their inference latency collides with the control rate a vehicle requires.
arXiv:2609.39563v1 Announce Type: new Abstract: Existing video language models encode sampled RGB frames independently, so a long video must either exhaust the token budget or drop the changes betwee...
The study investigates how long‑video language models decide which frames to keep, compress, and reuse, testing each decision in isolation across six selection rules, three benchmarks, and two answering models. It finds that selecting frames based on queries yields the biggest performance boost, that halving spatial resolution costs little, and that reallocating saved tokens to more compressed frames can further improve accuracy. The work also highlights the importance of a unified evaluation harness to avoid misleading comparisons.
FastBench is a new benchmark that tests streaming Video‑Large‑Language Models (VLMs) on high‑dynamic real‑world video streams, where events occur rapidly and cannot be captured by sparse 1–2 FPS sampling. The benchmark uses a trajectory‑grounded pipeline that generates QA pairs from high‑FPS clips, filters questions answerable at 2 FPS, verifies answers with SAM3 and CoTracker3 trajectories, and incorporates three rounds of human inspection, yielding 306 QA pairs across eight domains and multiple temporal scopes. Experiments show that even the strongest model, Gemini‑3.5‑Flash, achieves only 50.7 % accuracy, and that denser sampling improves performance but saturates when history is compressed; a training‑free baseline, ProactiveFrame, outperforms uniform sampling but still falls short of oracle‑guided focusing.
arXiv:2607. 13305v1 Announce Type: cross Abstract: Benchmark accuracy in video large language models (LLMs) is often treated as evidence of visual understanding.
CALIPER is a new benchmark that tests whether pretrained visual encoders can infer physical properties such as mass and friction from images. The test involves striking an object twice at known speeds, showing a third strike only up to contact, and asking a linear readout on frozen features to predict how far the object slides. Results show that in clean, fixed‑camera scenes all representations perform similarly, but when camera, lighting, and clutter are varied, only encoders that truly infer physics—like V‑JEPA 2—maintain performance, while random or raw pixel representations fail.
TimeBlind is a diagnostic benchmark designed to evaluate fine‑grained spatio‑temporal compositionality in video large language models (LLMs). It categorizes temporal understanding into three levels—atomic event recognition, event property characterization, and reasoning about event interdependencies—and uses a minimal‑pairs paradigm where video pairs share identical static content but differ only in temporal structure. Across 20 state‑of‑the‑art MLLMs tested on 600 curated instances, the best model achieved only 48.2% instance accuracy, far below human performance of 98.2%, highlighting a reliance on static visual shortcuts rather than true temporal reasoning.
GlanceWAM introduces a sparse test‑time imagination approach for world‑action models that decouples visual imagination from control. By asynchronously generating a single lookahead frame on a slow clock and decoding action chunks at a 48 ms control rate purely in latent space, it avoids latency while maintaining high success. The method achieves 72.2 % on the RoboCasa kitchen benchmark and 99.0 % on LIBERO, running 24× faster than synchronous baselines.
The paper identifies a new failure mode in Video Large Language Models (Video‑LLMs) called directional motion blindness, where models struggle to correctly interpret basic signed image‑plane motion directions. By tracing motion signals through the model pipeline, the authors reveal a binding gap between motion information and verbal output, and propose MoDirect, a motion‑direction instruction‑tuning dataset, and DeltaDirect, a projector‑level objective that predicts 2‑D motion vectors from frame deltas. Experiments show that DeltaDirect dramatically boosts motion‑direction accuracy on both synthetic and real‑world benchmarks while maintaining overall video‑understanding performance.