AICrypto
Prefill-Free Cross-Family KV Cache Transfer for Heterogeneous Multi-Agent LLMs
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 03/10/2026, 02:36:31 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 06:18:24 UTC.
Briefings
A new method called HeteroFold enables prefill-free cross-family KV cache transfer between heterogeneous LLMs, allowing receivers to reuse sender context without prefilling. The system aligns model structures and maps sender caches to receiver spaces while freezing both parties. Across six transfer directions, HeteroFold outperforms existing baselines on long-context benchmarks and matches performance on text-based communication benchmarks. At 32K context length, the Llama-3.1-8B to Minis-3.1-14B transfer is 10.7 times faster than Native Prefill and 1.18-1.47 times faster than the state-of-the-art prefill-free baselines.
Key points
- HeteroFold aligns model structures and maps sender KV caches to receiver spaces while freezing both sender and receiver models.
- Across six transfer directions, HeteroFold achieves the best cache-transfer performance on all four long-context benchmarks and most short-context settings.
Context and limitations
Results are limited to six transfer directions across four long-context and short-context benchmarks; performance data is snapshot-based at 32K context length.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/10/2026, 02:36:31 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeih5zeusuaukbaguvgmx4ycdswz2xofzrqmaoxjm535cw6ncb7wwd4