Baidu BaiGe and Fudan University propose LU-KV framework, accepted by ICML 2026
Original title:顶会入选!百度百舸联合复旦大学刷新KV Cache压缩SOTA
AI summary
Baidu's BaiGe team, in collaboration with Fudan University, proposed the Long-horizon Utility KV (LU-KV) framework, modeling head-level KV Cache budget allocation as a global combinatorial optimization problem oriented toward long-horizon marginal utility. LU-KV estimates marginal contribution curves of attention heads through offline profiling, and combines convex hull relaxation with a marginal-utility-based greedy solver to obtain near-optimal budget configurations at low overhead, compatible with multiple compression methods such as SnapKV and KeyDiff. On the LongBench and RULER benchmarks, performance loss is small at an 80% compression ratio, reducing memory footprint and inference latency. The paper has been accepted by ICML 2026.
Key Takeaways
- Baidu's BaiGe team,in collaboration with Fudan University,proposed the Long-horizon Utility KV…
- LU-KV estimates marginal contribution curves of attention heads through offline profiling,and…
- On the LongBench and RULER benchmarks,performance loss is small at an 80% compression…
- The paper has been accepted by ICML 2026.
Quick Facts
- Published
- Updated
- Source
- 公众号:百度智能云(文心)
- Category
- Papers & Research
- AI score
- 63
ICML接收的KV Cache压缩新方法,把缓存预算分配从看当前分数改成全局优化,能显著降低显存占用,做长上下文推理的值得细读。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Baidu BaiGe and Fudan University propose LU-KV framework, accepted by ICML 2026. June 10, 2026. https://new.nat6.net/en/news/cmq7w1nzi028nslepn0azwdsg/