ByteDance’s SWE-Pruner Pro cuts coder-agent tokens by 39% without losing quality
ByteDance · hf · 2026-07-21
What it claims
ByteDance and collaborators show that a coder LLM already learns useful internal signals for deciding what tool output to keep or prune.
Proposed method
- SWE-Pruner Pro moves pruning inside the agent instead of attaching a separate code classifier.
- A small head reads the agent’s own internal representations and predicts keep/prune labels line by line.
- A length-aware embedding accounts for the number of lines in each tool output.
Results
- Across two open-weight backbones and four multi-turn benchmarks, the method saves up to 39% of prompt and completion tokens while keeping task quality intact.
- In addition, on MiMo-V2-Flash, it improves SWE-Bench Verified resolve rate by +3.8% and Oolong long-context accuracy by +2.2 points.
Related event: ByteDance's SWE-Pruner Pro Enables Coding Agents to Self-Prune Context(3 posts)→
More from coding & agent
- Kolab Benchmark Uses Pro Screen Recordings to Show AI Agents Overestimate Collaboration Quality — minnesotanlp · 2026-10-27
- Dev argues Claude burns far more tokens than GPT models for the same coding work — haider1 · 2026-10-27
- Stop building agents that can't change models: stable contracts, external state, real evals — rseroter · 2026-10-27
- Cheap models can cost more in agent harnesses than frontier ones, researcher warns — vasuman · 2026-10-27
- Kanzo: A 5k-Line Rust Kanban Binary With MCP for Claude Code and Codex — allmudi · 2026-10-27
- 'A human approves everything' is a trap: approval fatigue and how to actually design safe agents — Efistoffeles · 2026-10-27