The paper introduces VideoRouteAudit, a controlled experiment that isolates the effects of frame selection, evidence injection, and expert routing in long‑video question answering. On a large Video‑MME benchmark, no frame selector outperforms uniform sampling, and the gains from evidence injection or model swapping are largely attributable to the specific model or data ordering rather than the selection strategy itself. The study also corrects earlier results, showing that ASR transcripts can either help or hurt performance depending on placement, and that swapping the answering model yields the largest performance contrast.
Long-video question answering is usually studied as frame selection: which few frames should a vision-language model see? We introduce VideoRouteAudit, a controlled audit that separates this choice fr...
The paper introduces an audit layer for vision‑language model benchmarks that reuses per‑question artifacts from evaluation runs to assess the benchmark itself. Six detectors analyze saved outputs, and an additional text‑only rerun removes visual inputs to produce per‑question reports and aggregate statistics, including permutation‑null excess and interval estimates. Applying this method to the MMStar benchmark reveals that 20.2% of its questions show no accuracy gain from images, with a small but significant excess in mathematics and science categories, and identifies questions where all models chose the same non‑gold option.
The paper investigates how reinforcement learning from verifiable rewards (RLVR) affects video‑language models’ understanding of time. Four open models were fine‑tuned under different data recipes—verified synthetic questions, unverified self‑supervised tasks, a 1:1 mix, and a verified+real mix—and evaluated on in‑domain and out‑of‑domain real‑video benchmarks with frames ordered, shuffled, or absent. Results show that verified training boosts in‑domain accuracy, largely through non‑visual gains, but can severely harm out‑of‑domain performance unless real verified data is added; moreover, no recipe induces genuine temporal‑order grounding, indicating that verifiable rewards improve benchmark scores without teaching temporal understanding.
arXiv:2606. 01485v1 Announce Type: cross Abstract: We describe our submission to the VRR Challenge @ CVPR 2026, built on the \emph{ImplicitQA} / \emph{VRR-QA} benchmark~\cite{implicitqa}: multiple-choice video question answering in which answers are deliberately \emph{not} observable in any single frame and must be inferred from spatial layout, motion, depth, viewpoint, causality, and social context across discontinuous frames of creative video.
arXiv:2607. 12304v1 Announce Type: cross Abstract: A score on a temporal video question answering benchmark is meant to measure that a model has temporal understanding, but it conflates two questions.
The paper introduces TRACE, a training‑free agent for long‑video understanding that grounds answers in raw visual clips and builds an evidence bundle until the answer stabilises. It also presents VES‑Bench, a 600‑question benchmark over 348 public long videos that audits whether decoded frames cover all necessary evidence intervals at three strictness levels. TRACE achieves high accuracy on VES‑Bench (63.5% audit accuracy) and remains competitive on other video‑understanding benchmarks while using far fewer frames than uniform decoding.
The paper introduces Behavior Pack Optimization (BPO), a post‑training method for video multimodal large language models that replaces single‑response rewards with a set of outputs across counterfactual views. BPO enforces stability when interventions are irrelevant, sensitivity when key evidence is removed, and abstention when no evidence remains, using an anchor‑relative advantage to keep the objective stable with small pack sizes. Experiments on datasets such as TempCompass, MVBench, and NExT‑QA show that BPO improves macro accuracy and abstention metrics for models like Qwen2.5‑VL‑7B‑Instruct, with gains that transfer to other benchmarks and models.
The paper investigates how providing execution traces to multimodal judges in agentic video‑generation systems can bias their verdicts. On a benchmark of 109 two‑event clips, traces that falsely report successful tool calls cause large‑language‑model judges to incorrectly accept 78–90 % of failures, while contradictory traces lead to 100 % rejection of correct clips. The effect persists even when judges are instructed to consider only the video frames, indicating that the vulnerability stems from the judges’ learned trust in tool logs rather than the visual content itself.
A score on a temporal video question answering benchmark is meant to measure that a model has temporal understanding, but it conflates two questions. 1.
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.
arXiv:2607. 13305v1 Announce Type: cross Abstract: Benchmark accuracy in video large language models (LLMs) is often treated as evidence of visual understanding.
arXiv:2609.09528v1 Announce Type: new Abstract: Video large language models (Video-LLMs) are increasingly used as the perceptual front end of world models, a role that assumes they can read motion: h...