Glance reads structured visual answers from a frozen 4B VLM, cutting GPU cost up to 85%
multiply_matrix · x · 2026-09-22
Yohei Nakajima released a working paper and tool, glance-vlm: instead of generating text, it reads structured visual judgements (yes/no, pick-one, ratings) straight from the logits of a frozen Qwen3-VL-4B.
- Recognition: 0.939 on yes/no and 0.933 on pick-one over fresh photos — close to the best hosted models (0.961/0.937) and better than the same model writing its answers
- Cost & speed: up to 85% less GPU cost with many questions per image, 1/3 faster than writing JSON, same accuracy
- Ratings: zero-shot ordering is right (within one level on 99% of images) but exact level often wrong (0.669); showing 16 unlabeled rubric images lifts it to 0.758 — doesn't generalize beyond image quality
- Limits: geometric judgements lag badly (0.52 vs 1.00 for hosted models); a linear probe shows the info is in the hidden state but not readable
All experiments pre-registered, results regenerate from the repo; install with pip install glance-vlm.
More from Research
- Training on production traces: single-trajectory RL may unlock continual learning — rhythmrg · 2026-09-22
- Most compute now goes to RL, letting models surpass human data limits — MarvinTBaumann · 2026-09-22
- TinyTorch: PyTorch's free curriculum to build an ML framework from scratch in 20 modules — PyTorch · 2026-09-22
- Why AI won't boost paper output for researchers who chase hard problems — kfountou · 2026-09-22
- ICML 2027 braces for 100k submissions as AI paper boom continues — CharlotteHase · 2026-09-22
- 1080 Ti beats RTX 6000 by 2.4x on dense-model inference despite 4x less bandwidth — EAccelerate_42 · 2026-09-22