Apple and Duke's LensVLM Turns Long Context into Images You Can Zoom Into
jiqizhixin · x · 2026-10-21
Apple's Foundation Models team and Duke University present LensVLM (NeurIPS 2026): long documents are rendered into a compressed full-page overview image, and the model multi-turn 'zooms in' via an EXPAND tool to recover relevant pages at full resolution—keeping source detail retrievable instead of discarding it like summarization or naive image compression. Training uses a three-stage pipeline: hard-example filtering, synthetic tool-use traces for SFT, then RL to sharpen answer and tool-call correctness. Validated across seven text-QA benchmarks.
More from Research
- LFD distills LLM knowledge to steer scientific generators via natural-language goals — hbouammar · 2026-10-21
- scikit-learn 1.7 GPU acceleration cuts ridge regression training from 82s to 2s — GaelVaroquaux · 2026-10-21
- Next-concept modeling proposes generating concepts first and words second — ConversationLow9545 · 2026-10-21
- Interpretability idea: cluster texts by firing circuits, not surface tokens — Sauers_ · 2026-10-21
- bekko-static-embedding-v1: GPU-free static embeddings hit retrieval SOTA, 100x faster on CPU — tomaarsen · 2026-10-21
- OmniCam from Tencent Hunyuan generates camera trajectories from panoramas and text — Zhenyang Liu · 2026-10-21