A Columbia preprint reframes multi agent communication as cache exchange rather than token serialization, opening a design space where 4% of the adapter size still beats text on latency and accuracy.
AI agents do not have to talk to each other in words. A new preprint from Columbia University researchers, Latent Cache Flow: Model-to-Model Communication Without Text, shows they can exchange compressed memory summaries instead, sidestepping the autoregressive cost that makes text-based agent handoffs slow and lossy. The work is under review at a top machine-learning venue in 2026 and lands as another attempt this year to use the KV cache — the internal memory a transformer keeps while generating — as a shareable coordination medium between models.
The contribution is a new coordinate in the multi-agent design space. Today's LLM agents typically hand off a task by serializing an answer into tokens, letting a second model decode and re-encode the result. Each step costs latency and burns information. The Columbia work, by researchers including Eugene Wu of Columbia Engineering, asks what happens if two agents skip the words and trade a compressed summary of the sharer's internal state instead. According to the authors, a 13 MB LCF adapter is more accurate than a 956 MB adapter from the prior state of the art in shared-context settings, and 23% more accurate and 8.5x faster than text-based agent communication when the two agents start with different context.
The most direct precursor is Cache-to-Cache (C2C), also in the ICLR 2026 review pool, which first proposed letting models exchange KV caches via a learned adapter. C2C is large. A 956 MB adapter is not something a small lab or open-source contributor can afford to train, and the method assumes the two communicating models share an identical context, a constraint that does not hold for most real multi-agent pipelines. Independent coverage from The Decoder frames LCF as a way for LLMs to "share meaning through internal memory instead of text," and that is the right level of description for the contribution. The Columbia work compresses what is being exchanged (keys and values, jointly translated) and reshapes the goal (a summary of what the target model does not already know), so the adapter can be roughly 4% of C2C's size and still work across heterogeneous agent contexts.
The architectural unlock is that last phrase. If an agent planner and an agent coder start with different prompts, different memory, different tools, text is currently the only general bridge: one model writes a sentence, the other reads it, both pay the autoregressive cost. LCF proposes that the bridge be a learned function from one model's internal state to a compact slice the other can absorb, with no tokenization in between. In different-context settings, the 8.5x speedup is the consequence of that move, not the reason for it.
The trade-off is real and the paper does not hide it. A 13 MB adapter still has to be trained for the model pair it serves. Text remains the right choice when the two agents need an auditable, human-readable log, when the message has to be inspected or rerun as natural language, or when the deployment does not justify a custom adapter at all. LCF is an efficiency primitive for cases where latency and context asymmetry matter more than readability, not a universal replacement for token hand-offs. The paper has not cleared peer review, and the headline numbers come from the authors' own benchmark suite, so the 23% accuracy lift and 8.5x speedup should be read as "the design is competitive on the workloads the authors chose," not as a settled result.
The second-order question is whether this style of optimization becomes a standard layer in agent infrastructure. Recent work on scaling multi-agent LLM serving has been hammering on the same point from a systems angle: coordination overhead, not raw inference, is the bottleneck once you stack more than a few models. If latent cache exchange holds up under independent benchmarking, agent frameworks that today treat the LLM call as the atomic unit may start treating cache exchange as a primitive alongside it, with text as the fallback. Other 2025–2026 preprints on agent coordination point in the same direction. The move worth watching is not whether Columbia's specific adapter gets adopted, but whether reviewers at a top ML venue treat latent-channel coordination as a first-class research direction, or fold it back into the existing text-based stack. If the former, expect a wave of follow-up work that treats caches, not tokens, as the default language between agents.