COLM 2026 paper maps how much VLM representations leak across layers
sineadwilliamso · x · 2026-10-07
Presenting 'What do your logits know' at COLM 2026, the authors systematically compared information retained at different representational levels of Vision-Language Models — from the rich residual stream through two natural bottlenecks: tuned lens projections and final top-k logits. The core question: when you ask a yes/no question about an image, how much other information leaks through the output? Far more than you might expect.
More from Research
- GPU-Free Activation Alignment Recovers Half of Full-Context Performance for Tabular ICL — Independent-Researcher · 2026-10-07
- Auxiliary loss forces hybrid LMs like Qwen3.5 to actually use recurrent memory, +12.1% on agentic tasks — mohitban47 · 2026-10-07
- COLM paper: latent reasoning traces decodable 65-93% of the time in LRMs — sarahwiegreffe · 2026-10-07
- Apple ML Research is hiring a PhD intern to make foundation models know what they know — sineadwilliamso · 2026-10-07
- Project Glasswing Reports 135K Verified Vulnerabilities, 9,333 Already Patched — ResultBackground2450 · 2026-10-07
- CAVEAT testbed exposes how merchants can steer your shopping AI agent — ZacharyHuang12 · 2026-10-07