Paper Overview
Field: ML Authors: Mingxuan Wang, Bo Wang, Fei Luo, Guorun Yao, Chao Ning, Yinglong Guo, Hongyue Chen, Yanbiao Ma, Jungong Han Published: 2026-09-25 arXiv: 2609.27276
Summary
Long-horizon language-model agents accumulate reasoning traces, tool exchanges, and observations whose relevance changes with the current decision. Existing compression strategies often score historical units independently, but the safety of deleting several units is generally not determined by their singleton scores: redundant evidence, accumulated small effects, and the information that remains after deletion all matter.
The authors introduce Direct Relational Set-Risk Pruning (DRSR), which formulates agent-history compression as risk-constrained selection over deletion sets.
Key Points
- Counterfactual supervision: Offline, DRSR constructs exact counterfactual supervision by jointly deleting protocol-valid history blocks and measuring the change in teacher-forced likelihood of the same recorded next output.
- Set-level scoring: A lightweight scorer predicts set-level harm from relations between online-visible candidate histories and the current pre-action state, incorporating deletion-retention and pairwise set structure.
- Constrained pruning at deployment: DRSR evaluates a small set of construct-valid deletion candidates and removes the largest feasible set under recency, protocol, budget, and learned risk constraints. It abstains when no sufficiently safe set exists.
- On WorkBuddyBench Full260, DRSR improves average reward from 0.699 to 0.802 and reduces total model tokens by 20.820%.
- On a fixed Eval40 comparison, DRSR achieves 0.794 reward at 1.211M tokens per task, using 35.850% fewer tokens than an uncompressed agent.
Results
Analysis
Mechanistic analysis and ablations show that decision-conditioned relations, retained-context information, pairwise interactions, and the abstention mechanism each contribute to reliable pruning.
---
*Automatically collected on 2026-09-25*