RandSlot: random soft tokens compress visual documents into just 4 vectors
_reachsumit · x · 2026-09-29
RandSlot appends independently resampled random unit vectors during training to teach a visual document retriever to compress each page into four vectors, outperforming prior compact methods at the same vector budget. Notably, the gains persist when random tokens are replaced with zeros at inference, showing training-time randomness alone improves compact representations.
More from Research
- ArtHOI accepted to TPAMI: turns AI-generated video into supervised-free 4D articulated human-object interaction data — liuziwei7 · 2026-10-25
- Lure-TTS: 600M voice models trained for under $200, running 4x faster than real time — smirkishere · 2026-10-25
- Math arXiv paper counts diverge by field: steady growth 2022-2025, then a shift in 2026 — IgorCarron · 2026-10-25
- AI experts couple strong benchmark forecasts with low-resolution folk models of society, sparking bottlenecks debate — random_walker · 2026-10-25
- Why voice models sound natural alone but awkward in conversation: an architecture breakdown — Gauri_the_great · 2026-10-25
- OpenAI's Shyamal Anadkat recruiting founding members for computation-physics-chemistry venture — shyamalanadkat · 2026-10-25