What the study found
The study found that inference-time trajectory reduction can lower the cost of LLM (large language model) agents without reducing performance. The authors report that their approach, AgentDiet, removed wasteful information from agent trajectories during execution while keeping agent results the same.
Why the authors say this matters
The authors say this matters because high input-token cost from ever-growing trajectories is a major efficiency concern for multi-turn LLM agents in software engineering. The study suggests that inference-time trajectory reduction is a promising direction for agent systems.
What the researchers tested
The researchers analyzed existing agent trajectories to identify useless, redundant, and expired information. They then proposed AgentDiet, a simple trajectory reduction approach that automatically removes such waste during agent execution, and implemented it on a top-performing coding agent.
What worked and what didn't
On two LLMs and two benchmarks, AgentDiet reduced input tokens by 39.9% to 59.7% and total computational cost by 21.1% to 35.9%. The abstract says this was achieved while maintaining the same agent performance; it does not report any performance loss in the summary provided.
What to keep in mind
The available summary does not describe detailed limitations, failure cases, or broader scope constraints. The results reported here come from two LLMs and two benchmarks, so the abstract only supports conclusions within that tested setting.
Key points
- AgentDiet is an inference-time trajectory reduction approach for LLM agents.
- The authors found wasteful information was widespread in agent trajectories.
- AgentDiet reduced input tokens by 39.9% to 59.7% in the reported tests.
- Total computational cost fell by 21.1% to 35.9% while performance stayed the same.
- The abstract describes the approach as promising for agent systems.
Disclosure
- Research title:
- Trajectory reduction lowered LLM agent token costs
- Authors:
- Yuan-An Xiao, Pengfei Gao, Chao Peng, Xiong Yingfei
- Institutions:
- Peking University, Peking University
- Publication date:
- 2026-06-30
- DOI:
- 10.1145/3797084
- OpenAlex record:
- View
Get the weekly research newsletter
Stay current with scholarly research without reading academic papers — one filtered digest, every Friday.