FlowPlan: Reducing the Token Cost of Coding Agents with Abstraction-Guided Planning and Verification
Story 10, from arXiv
FlowPlan is a new method that reduces token costs for coding agents by working with abstract program representations rather than raw code, improving efficiency without sacrificing repair quality.
Key points
- FlowPlan uses abstraction-guided repair by working with control-flow or data-flow graphs instead of raw source code.
- The agent localizes faults, plans repairs as graph deltas, and verifies the changes before applying them to concrete code.
- It reduces token usage significantly across multiple benchmarks, with up to 0.32x token savings on SWE-bench Verified.
- FlowPlan maintains or improves repair success rates compared to vanilla agents while cutting costs.
- Trajectory analysis shows token savings come from reduced post-edit re-reading and testing, replaced by explicit verification.
- It achieves the largest token reduction in three of six benchmark settings and ties the strongest method in another.
Why it matters
Software engineers using LLM-based coding agents will care because FlowPlan offers a way to cut token costs without sacrificing repair quality, which is critical for running models locally or on modest hardware.
Counterpoint
FlowPlan's effectiveness depends on the ability to accurately construct and verify abstract program representations, which may not always align with the concrete code, potentially introducing new errors.
Sources
- FlowPlan: Reducing the Token Cost of Coding Agents with Abstraction-Guided Planning and Verification (arXiv)
This summary was written by a language model and may contain errors. Read the original for the full story.