Receiver-Conditioned Latent Communication gives 94% CacheBack

Receiver-Conditioned Latent Communication gives 94% CacheBack. By Maximillian Rossi and four others, from Columbia University and Amazon Web Services. Published on the twenty-fifth of September. Its code repository is worked out: about fifty-five source files, with a readme and a licence.

Multi-agent systems spread large contexts across several agents, which then communicate to solve a task. Text messages are compact, but they require decoding, and they may leave out evidence the receiving agent needs. A more recent alternative, latent communication, transfers KV caches instead. That avoids generating text, and can improve both accuracy and latency. The catch is size. A full KV cache grows linearly with the context an individual agent processes, and linearly again with the number of agents coordinating. Memory and context costs climb with it, often far beyond available GPU resources and context window sizes.

The key observation is that a sender does not need to send everything. It needs to send only what the receiving agent requires for its own local task. The authors call this receiver-conditioned communication. The receiver passes the sender a small description of its information needs, and that description serves to filter and compress the sender's KV cache.

CacheBack is a concrete instance of receiver conditioning, described as simple, robust and training-free. Its filtering is based on the sender's attention weights. Nothing has to be trained: the receiver's description of what it needs, together with the sender's attention, decides which part of the cache travels. The direction of the exchange is what changes. The receiver speaks first, and the sender answers with a cut-down cache rather than a full one or a text summary.

On FanOutQA with Qwen 3, CacheBack removes seventy-five percent of the state the agent would otherwise receive. Accuracy improves by fourteen point seven percentage points, and median task-completion latency falls by three point two times, both measured relative to text communication. Comparable improvements appear across model families spanning dense Transformers, Mamba-attention hybrids, and sliding-window attention.

The idea underneath CacheBack is a change in who decides what gets sent. Text messages leave the sender guessing what to include, and full caches send everything. Letting the receiver state its needs first means the transfer can be smaller and more useful at the same time, which is why the results improve on three fronts at once: less state, higher accuracy, lower latency.