Coca-Cola:面向智能體的潛在上下文壓縮,讓工具觀察不必以文字呈現
Tools Need Not Speak in Text: Latent Context Compression for Agents
Coca-Cola 讓 LLM 智能體以精簡的潛在表示接收工具觀察,無須重建原始工具輸出。在四個網頁搜尋問答基準及 Qwen、GLM 兩個模型系列的評測中,總智能體 token 消耗降低 70.01%–89.24%,端到端執行時間縮短 24.45%–55.97%。所有評測設定中的答案 EM 和 F1 均高於未壓縮的智能體。
We introduce Coca-Cola (Context Compression via Continuous Latent Tokens), a method that lets LLM agents consume tool observations as compact latent representations instead of lengthy text.
Rather than reconstructing the original tool output, Coca-Cola learns to preserve the information that matters for the agent’s next decision. Our Action-Preserving Training aligns next-action distributions under full and compressed observations, while keeping the target agent frozen.
Across four web-search QA benchmarks—Natural Questions, TriviaQA, 2WikiMultiHopQA, and HotpotQA—and two model families (Qwen and GLM), Coca-Cola achieves:
- 70.01%–89.24% lower total agent token consumption
- 24.45%–55.97% shorter end-to-end runtime
- Higher answer EM and F1 than the uncompressed agent in every evaluated setting
The key idea: tool observations need to support the agent’s decisions; they do not have to remain readable text.
We welcome feedback and discussion, especially on extending latent tool observations to other agent workflows!
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co