VQV Signal

RESEARCH SOURCE-BACKED TECHNICAL

LookAway: Thinking VLMs Write Timestamped Notes Then Drop the Long Video's KV Cache

An arXiv paper titled "Do Thinking VLMs Need to Keep Looking at Long Videos? Write It Down, Then Look Away" examines vision-language models that reason over long videos spanning hundreds to thousands of tokens. The authors note that each decoding step re-reads the key-value cache of the entire video, which is costly. Their proposed method, LookAway, has the model first write timestamped notes on what the question needs, then puts the video away.

Source: arXiv · arxiv.org Published 2026-11-18T11:42:17+00:00 Detected 2026-11-20T00:02:32+00:00
View original source

An arXiv paper titled "Do Thinking VLMs Need to Keep Looking at Long Videos? Write It Down, Then Look Away" examines vision-language models that reason over long videos spanning hundreds to thousands of tokens. The authors note that each decoding step re-reads the key-value cache of the entire video, which is costly. Their proposed method, LookAway, has the model first write timestamped notes on what the question needs, then puts the video away.

AI-assisted summary based on the listed source.

Vision-language models (VLMs) reason about long videos over hundreds to thousands of tokens. Every decoding step re-reads the key-value (KV) cache of the entire video. Must the model keep looking while it reasons? In LookAway, the model first writes timestamped notes on what the question needs and then puts the...

Teams running video reasoning models could see lower inference cost if the approach holds up, since the video cache no longer has to be re-read throughout the answer. Builders of long-video applications should watch for benchmark results and whether the method works across different models and tasks. The source snippet does not provide performance figures, so the size of any gain remains unclear.

Signal Strength 95% Technical label SOURCE-BACKED Public Interest 23 Category RESEARCH Reader Depth TECHNICAL

Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.

Public Interest components
Recognizable Entity Score 0 Practical Impact Score 0 Novelty Interest Score 72 Consequence Score 18 Curiosity Score 16 Shareability Score 41

VQV surfaced this signal because it is recent, relevant to LLM Inference, connected to arXiv.