Accelerating Language Model Workflows with Prompt Choreography

TJ Bai and Jason Eisner published "Prompt Choreography" in Transactions of the Association for Computational Linguistics Volume 14 (pages 253–270), a framework that maintains a dynamic global KV cache so each LLM call can attend to an arbitrary, reordered subset of previously encoded messages, including parallel calls. The authors report 2.0–6.2× faster time-to-first-token per message and end-to-end speedups above 2.2× in workflows dominated by redundant computation, and show that fine-tuning the LLM to work with the cache helps it mimic results produced by re-encoding messages in a new context.

Abstract Large language models are increasingly deployed in multi-agent workflows. We introduce Prompt Choreography, a framework that efficiently executes LLM workflows by maintaining a dynamic, global KV cache. Each LLM call can attend to an arbitrary, reordered subset of previously encoded messages. Parallel calls are supported. Though caching messages’ encodings sometimes gives different results from re-encoding them in a new context, we show in diverse settings that fine-tuning the LLM to work with the cache can help it mimic the original results. Prompt Choreography significantly reduces per-message latency 2.0–6.2× faster time-to-first-token and achieves substantial end-to-end speedups 2.2× in some workflows dominated by redundant computation. - Anthology ID: - 2026.tacl-1.13 - Volume: - Transactions of the Association for Computational Linguistics, Volume 14 https://aclanthology.org/volumes/2026.tacl-1/ - Month: - Year: - 2026 - Address: - Cambridge, MA - Venue: - TACL https://aclanthology.org/venues/tacl/ - SIG: - Publisher: - MIT Press - Note: - Pages: - 253–270 - Language: - URL: - https://aclanthology.org/2026.tacl-1.13/ https://aclanthology.org/2026.tacl-1.13/ - DOI: - 10.1162/tacl.a.643 https://doi.org/10.1162/tacl.a.643 - Cite ACL : - TJ Bai and Jason Eisner. 2026. Accelerating Language Model Workflows with Prompt Choreography https://aclanthology.org/2026.tacl-1.13/ . Transactions of the Association for Computational Linguistics , 14:253–270. - Cite Informal : - Accelerating Language Model Workflows with Prompt Choreography https://aclanthology.org/2026.tacl-1.13/ Bai & Eisner, TACL 2026 - PDF: - https://aclanthology.org/2026.tacl-1.13.pdf https://aclanthology.org/2026.tacl-1.13.pdf