CLM author on extending editable context to video: RL may teach models where to look, prefill cut 30%
RulinShao · x · 2026-11-23
Rulin Shao, author of the CLM work, comments on a new paper extending CLMs and Suffix Cache Reuse to video context management:
- Accuracy: sampling more frames from the right segment helps, but Qwen2.5-VL-7B struggles to choose where to look. In her CLM work, Qwen3.5-9B also struggled with text context management, and RL improved both accuracy and efficiency — she wonders whether RL can teach a video model what to revisit and keep.
- Efficiency: the paper's Transformers prototype already cuts LM prefill tokens by 30%; native video cache reuse in vLLM, SGLang, or Evolving Infra could turn those savings into real LM-side speedup.
Related event: KV-Cache Reuse Brings Mid-Context Editing to Video LLMs(2 posts)→
More from Research
- Riemann proof built with Claude spans 17M lines of Lean — no human will ever read it — r0ck3t23 · 2026-11-23
- Anthropic researcher alpoge reportedly solved one of the biggest open math problems — and went silent — kimmonismus · 2026-11-23
- Princeton study: diffusion models barely extrapolate basic visual skills, ~50% on counts — NielsRogge · 2026-11-23
- MIT Professor: Models Cross a Discovery Threshold Around 0.76 Log-Perplexity — ProfBuehlerMIT · 2026-11-23
- Noam Brown's 2023 bet lands: 10K agents solve Navier–Stokes, Riemann proof claimed — haider1 · 2026-11-23
- Training models to fill in omitted reasoning from proprietary model outputs — OwariDa · 2026-11-23