An action seen once and the same action seen three times receive the same action-recognition label, so that supervision does not require a model to count. Answers shared across histories that differ i...
arXiv:2606. 16353v1 Announce Type: cross Abstract: Streaming video understanding models must answer queries at any moment during an ongoing stream, using only what they have observed so far and under fixed memory and computation budgets.
TwinPast is a new benchmark that tests whether streaming video models use historical evidence by presenting pairs of videos that differ only in critical past events while keeping recent context, questions, and options the same. The study evaluates 34 model configurations on 350 human‑reviewed pairs across five historical tasks, finding that most models rely on recent frames or the question alone, with only a few showing improved performance when historical inputs are provided. Even the best streaming system scores only 49.14% PairAcc, far below general‑purpose models, and case studies reveal that models often select states from the wrong time or contradict the recorded event order.
arXiv:2609.00291v1 Announce Type: new Abstract: Streaming video understanding requires answering questions that arrive at arbitrary moments over an unbounded video stream. Existing systems primarily...
TRACE (Temporal Audit and Condition-aware Evaluation) is a new benchmark and evaluation framework for streaming video understanding that explicitly records when evidence becomes valid, how visual history is maintained, and how responses are triggered. It combines temporally audited visual tasks, evidence timing, instruction-dependent trigger annotations, a unified causal Core–Adapter protocol, and multidimensional reporting of answer quality, timeliness, response-selection behavior, workload, completion, and reliability. Using 1,240 records from 517 videos, TRACE evaluated eight publicly available models in eight configurations, revealing that similar QA accuracy can hide significant differences in completion, answer validity, generation workload, and proactive performance metrics such as response delay, false alarms, and missed target windows.
arXiv:2608.31005v1 Announce Type: new Abstract: Existing long-video agents acquire evidence through one uniform behavior, ignoring whether the required evidence is concentrated, requires broad occurr...
arXiv:2610.19849v1 Announce Type: new Abstract: We propose DeltaStreamer, a novel agentic framework for streaming video understanding inspired by how humans observe, remember, and investigate. Stream...
The paper introduces Watch-Think-Interact (WTI), a closed-loop framework for multi-question streaming video reasoning that maintains compact natural-language memory entries linked to video time ranges. WTI decides whether to answer, continue watching, or recall relevant past intervals for each question, avoiding replay of the full history. The authors build a large dataset, WTI-82K, and a training method, Stream-GDPO, achieving state‑of‑the‑art performance on StreamingBench and OVO-Bench.
arXiv:2610.34283v1 Announce Type: new Abstract: Training a spatial VLM to answer which side an object lies on does not require it to preserve the metric displacement that determines that answer, whil...
Event-Causal RAG (EC‑RAG) is a lightweight retrieval‑augmented framework designed for reasoning over ultra‑long and streaming videos. It segments video streams into semantically complete events using a dual visual‑audio sentinel mechanism, representing each event as a State‑Event‑State (SES) structure that captures pre‑event, event, and post‑event states. During question answering, bidirectional graph retrieval accesses relevant predecessor and successor events from a dual vector‑graph memory, and answers are generated using both this structured memory and the corresponding video evidence. The authors also introduce ECV‑1H, an hour‑scale long‑video QA benchmark with over 150 hours of untrimmed video and 1,251 human‑annotated QA pairs, where EC‑RAG achieves significant accuracy gains across multiple video foundation models while maintaining efficient streaming memory usage on a single RTX 5090 GPU.
The paper introduces OWLET, an on-policy weighted learning method for post-training Video Large Language Models. OWLET leverages a richer teacher view that combines global temporal context with question-conditioned local windows to provide visual evidence, and it employs evidence-aware credit redistribution and spatiotemporal state alignment to better utilize this evidence during reinforcement learning. The approach aims to improve video reasoning by aligning token-level supervision with spatiotemporal representations.
The paper examines how hallucinations arise in multi-stage video‑understanding agents by aligning existing benchmarks with the stages of temporal grounding, visual observation, and reasoning. It introduces a causal stage‑intervention protocol that isolates each stage while keeping the downstream task constant, revealing that grounding errors dominate downstream hallucinations and that correct region location matters more than precise temporal overlap. The study also shows that current benchmark scores poorly predict causal sensitivity and can fail under distribution shift, advocating for stage‑aware evaluation methods.