BeNice! Closing Looping's Inference-Memory Gap via KV-Sharing
Research on the serving cost of looped transformers: per-recurrence KV-caches and a sharing scheme to shrink them.
AI Storyline
3 items · 3 sources · 3 days
Operational story trace
Closing Gap
Follow in this browser to see new updates on your Live feed.
Latest change
OpenAI published ThinkYoung research and EMEA education-partnership lessons on AI literacy for teachers, an education-policy piece with no technical link to the earlier papers.
Earlier contextThe story so far
This is not one developing story: three unrelated items share only the phrase "closing the gap" in their titles. Two October arXiv papers target model-side gaps, KV-cache memory in looped transformers and the CLIP-style modality gap in embeddings.
editor-curated · source-linked
Arc
Oct 22Nov 10 · now
INFERENCE MEMORY · Oct 22BeNice! proposes KV-sharing to cut looped-model cache memory1 source · show source ▾
EMBEDDINGS · Oct 23Simplex-constrained representations target the modality gap1 source · show source ▾
Closing the Modality Gap via Simplex-Constrained Representations
Unrelated embedding-geometry paper; joins the thread only through its title wording.
NOW · Nov 10OpenAI on the AI guidance gap in education1 source · show source ▾
Closing the AI guidance gap in education
Education-policy post from OpenAI; the newest item, but no technical continuation of the papers.
Day 1 Thursday, Oct 22, 2026
BeNice! Closing Looping's Inference-Memory Gap via KV-Sharing
Research on the serving cost of looped transformers: per-recurrence KV-caches and a sharing scheme to shrink them.
Day 2 Friday, Oct 23, 2026
Closing the Modality Gap via Simplex-Constrained Representations
Unrelated embedding-geometry paper; joins the thread only through its title wording.
Day 3 Tuesday, Nov 10, 2026
Closing the AI guidance gap in education
Education-policy post from OpenAI; the newest item, but no technical continuation of the papers.
What to watch — open questions
- Does BeNice!'s KV-sharing hold quality at the recurrence depths looped models need to match deeper dense models?
- Does any serving stack (vLLM, SGLang, TensorRT-LLM) add support for looped-architecture KV-sharing?
- Does reducing the modality gap measurably improve cross-modal retrieval recall in production embedding pipelines?
How this thread was built
editor wrote the arc · 3 beats
Storylines are threaded mechanically from the feed: stories that share a distinctive anchor across multiple days and sources. Each item links to its original source. The evidence trace, current state, and open questions are written by the editor routine and refreshed whenever a new beat lands.