DTOC paper: agents hide/unhide tool outputs to cut tokens and boost solve rates
pppeer · reddit · 2026-10-03
A paper to be presented at Discovery Science 2026 in Mainz proposes Dynamic Tool Output Compression (DTOC): instead of lossy context compression, the agent decides each turn which tool outputs to hide or unhide, with outputs persisted separately for later re-enablement.
Experiments on DeepSWE tasks plus internal ablations show DTOC saves tokens, steps, and cost while improving solve rates, though results are model- and task-dependent. Authors invite comparisons with similar practical approaches.
Related event: DTOC Paper: Dynamically Hiding Tool Output Saves Tokens and Boosts Accuracy(2 posts)→
More from coding & agent
- Open-source coding agent Mixdog claims 63% fewer tokens than Codex at same success rate — RevolutionaryLock535 · 2026-10-03
- Replaying agent sessions with cached tool outputs to test fixes and model swaps safely — pauliusztin · 2026-10-03
- Stop building eval sets by hand: grouping agent failures into a regression suite — pauliusztin · 2026-10-03
- Dev argues models are now smart enough to skip GPUI and lay out pixels with raw math — zack_overflow · 2026-10-03
- antirez: AI Coding Tool's Constant Cyber Security Checks Are Blocking Real Work at Redis — antirez · 2026-10-03
- Cheap model's overnight bug-fix PRs: only 6 of 14 survived a stronger model's review — Aggressive-Narwhal-3 · 2026-10-03