Tsinghua team open-sources C2C: LLMs swap KV-Caches directly for 2.5x faster, more accurate collaboration
solyarisoftware · x · 2026-10-27
Researchers from Tsinghua University, Shanghai AI Lab and others open-sourced Cache-to-Cache (C2C), a framework that lets collaborating LLMs communicate without generating text. Instead of compressing thoughts into tokens, a neural network projects and fuses one model's KV-Cache into another's, with learnable gating that selects which layers benefit.
- Text-based multi-LLM communication loses semantics and incurs token-by-token latency
- C2C bypasses explicit text generation entirely
- Results: 3.1–5.4% higher accuracy vs text communication (6.4–14.2% vs individual models), with average 2.5x latency speedup
- Code available at github.com/thu-nics/C2C
More from Research
- Physicist on OpenAI's math result: it's mathematical physics, not theoretical physics — skdh · 2026-10-27
- Mechanistic Interpretability and Fast Memory Recall Are the Same Problem — mayfer · 2026-10-27
- Qwen model has a dedicated MoE expert for Chinese cloud concepts, sparking URL hallucinations — builderjaydub · 2026-10-27
- OpenAI preprint claims 3-SAT solvable in sub-exponential time, challenging ETH — cephaloform · 2026-10-27
- DRFLOW: EMNLP benchmark shows deep research agents struggle at personalized workflow prediction — mageed · 2026-10-27
- Pathology foundation models hit AUROC 0.90 vs ResNet50's 0.81 — maier_ak · 2026-10-27