Disagree to Explore, Agree to Commit: Routing-Guided Test-Time Scaling for Software Agents
Abstract Overview
This paper studies whether native mixture-of-experts (MoE) router traces can guide test-time scaling for software-engineering agents without using an external judge or running candidate patches for selection. Using repository-level software repair on SWE-bench Verified, the authors analyze routing signals across action roles, trajectory history, patch tokens, and independently sampled attempts. They find that routing encodes behavioral roles, complements surface-form similarity, and is most informative when read at low-probability decision tokens. Based on these observations, they introduce Risa, a routing-guided controller that promotes disagreement with recent history during exploration and agreement across relevant peers or attempts during commitment and final selection.
Novelty
The distinctive contribution is to use native MoE routing traces as a unified internal control signal for both within-trajectory steering and across-attempt patch arbitration in software agents. The method avoids external judges and selection-time execution, and it localizes comparison to decision tokens for free-form patches rather than relying on canonical answer matching.
Results
On six gpt-oss model/effort conditions, routing-based arbitration improves resolved rate over uniform selection by 2.3 to 5.7 points, increasing the macro-average from 44.9% to 48.2%. On Qwen3.6-35B-A3B, Risa reaches 45.2% versus 41.7% for uniform selection and matches text-consensus selection on the full benchmark. The analysis also reports that routing-based role classification attains 0.940 holdout accuracy on three-way action roles.
Key Points
- MoE routing traces provide a strong behavioral-role signal, enabling high-precision gating between exploratory actions and patch-writing actions.
- Decision-token routing fingerprints are more informative than whole-span or simple positional windows for ranking free-form software patches across attempts.
- Risa combines history-based disagreement for exploration with agreement-based commitment and final arbitration, yielding consistent gains over uniform selection across evaluated MoE agent settings.