Tsinghua's Coca-Cola compresses agent tool outputs, cutting tokens up to 89% while boosting accuracy
Tsinghua · hf · 2026-11-07
Tsinghua AI Agent Group introduces Coca-Cola, a continuous latent context compression model that challenges the assumption that agents must consume tool outputs as text tokens. It compresses tool observations into latent tokens embedded directly into the agent's input embedding space.
- Action-Preserving Training aligns next-action distributions under full vs. compressed observations, learning action-sufficient latent representations that keep decision-critical info
- On web-search agents across 4 QA benchmarks and 2 model families, token consumption drops 70.01%–89.24% and end-to-end runtime falls 24.45%–55.97%, while task performance consistently improves
- Code open-sourced at github.com/THU-Agent/Coca-cola
More from coding & agent
- Open DCR on remote MCP servers invites auth-code hijacking: should you allowlist redirect URIs? — After_Pumpkin3803 · 2026-11-07
- Building an incident-response agent: why memory alone can't be trusted for operational warnings — noufal_rizwan · 2026-11-07
- Letting AI stay on the line after customers hang up turned post-call notes into actions — gsuiteautomations · 2026-11-07
- ComfyUI-freefield puts Freefield's 117 workflows in a sidebar with one-click model downloads — neuregex · 2026-11-07
- How do you stop micromanaging a custom agent harness? A Reddit dev's research-task dilemma — dangerous_inference · 2026-11-07
- Engineering agency founder's 10-year lessons: small teams ship, turn down 30% of clients — alex_verem · 2026-11-07