arXiv cs.CL· Dan Ben-Ami, Kobi Cohen, Chaim Baskin·· 9 天前AI 评分24
冻结的 VideoLLM 是否已编码证据就绪信号？Readiness Gating 答案时机策略研究
Have I Seen Enough? Frozen Video-Language Models Encode Evidence Readiness
AI 导读
研究发现冻结的 VideoLLM 已线性编码"证据就绪"信号，在 7 个模型上 AUROC 达 0.733-0.905，且该信号以问题为条件——仅改变问题即可在 66.1% 的配对中反转读数。基于该读数提出的 Readiness Gating 答案时机策略，在相同视频时长下最高提升准确率 +9.75 pp，计算开销可忽略。
正文 · 原文
Abstract:Streaming video-language models must decide not only what to answer, but whether the evidence needed for the current question has arrived. Existing systems learn that decision as a separate trigger; we ask whether an unmodified model already computes it. We show that frozen VideoLLMs carry a linearly readable evidence-readiness signal, labelled from timestamped evidence rather than from model output. It decodes in all seven models of a shared byte-identical evaluation (AUROC 0.733-0.905 under the strictest not-ready sampling, where a fitted clock is near chance), and a probe fitted without any of a benchmark family's footage still reads that family. It is question-conditioned: on byte-identical windows, changing only the question reverses the readout on 66.1% of pairs, while every question-blind control is at chance by construction. The model can answer incorrectly and still encode readiness: AUROC remains 0.722 among wrong answers. Readiness also beats uncertainty estimators and their supervised combination on latency-matched answer selection, and tracks independent human judgments more closely than confidence. Released streaming triggers are also linear readouts, yet a trained trigger read on its own base model's activations is approximately orthogonal to readiness and decodes it far less accurately than a probe. We turn the readout into Readiness Gating, an answer-timing policy that improves accuracy by up to +9.75 pp at matched video duration with negligible computational overhead. How much it gains varies with the accuracy headroom the task makes available: across 26 configurations the gain tracks that headroom, and an intervention that moves it over identical pixels moves the gain with it.
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL); Machine Learning (cs.LG) |
| ACM classes: | I.2.10; I.2.6 |
| Cite as: | arXiv:2610.08560 [cs.CV] |
| (or arXiv:2610.08560v2 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2610.08560 arXiv-issued DOI via DataCite |
Submission history
From: Dan Ben Ami [view email]
[v1]
Tue, 6 Oct 2026 15:43:04 UTC (1,336 KB)
[v2]
Thu, 8 Oct 2026 06:30:57 UTC (1,342 KB)
来源：arXiv cs.CL · arxiv.org