Swarm Telepathy Paper on arXiv: Cross-Family KV-Cache Transfer, 40% More Accurate, 4x Faster
alex_verem · x · 2026-11-12
The MIT team published "Swarm Telepathy: Prefill-Free KV-Cache Transfer Across Families" (arXiv:2610.15095), the first cross-family, prefill-free KV-cache transfer method, addressing the bandwidth and redundant-compute limits of text-based multi-agent communication.
- With KV-cache compaction, beats text-to-text communication by 40% accuracy and 4x wall-clock speed
- Consistent across 12 LLMs, 6 families, 17 benchmarks
- Can transfer vision capabilities from a VLM to a text-only receiver, with the translator trained on text alone
- Both models stay frozen; training costs <50M tokens and <1 GPU-day
Related event: MIT's SwaT Enables Cross-Family KV-Cache Transfer(2 posts)→
More from Research
- Sokoban Eval lets the crowd race LLMs for shortest solutions across 66 rooms — banteg · 2026-11-12
- 1920 Popular Mechanics clipping shows eclipse that confirmed light bends under gravity — generativist · 2026-11-12
- ELIZA's Lost Code: 2,600 Lines Found in MIT Archives in 2021, Retyped by Hand — demian_ai · 2026-11-12
- 1.5-Year Study: Cultural Bias in AI Curiosity and Its Role in Multi-Agent Collaboration — radamihalcea · 2026-11-12
- ContactBench 1.0: A Benchmark for Contact-Rich Robot Manipulation With Printable Assets — leto__jean · 2026-11-12
- Formalizing all of nLab joked as the 'final boss' of AI math — burny_tech · 2026-11-12