Researchers propose HeteroFold, a prefill-free method for transferring key-value cache between heterogeneous large language models in multi-agent systems, letting a receiver reuse a sender’s already-processed context without re-running prefill despite differences in tokenization, model depth, and KV representations. Both sender and receiver models remain frozen; HeteroFold aligns model structures and calibrates the transferred cache to preserve receiver behavior. Across six transfer directions it achieved the best performance on four long-context benchmarks and matched text-based communication on a multi-agent benchmark, running up to 10.7 times faster than native prefill at 32K context length.
