TQS Structural Scoring Filters Better Coding-Agent Trajectories for Fine-Tuning
arXiv preprint presents TQS, a linear-time framework that scores coding-agent trajectories by structural convergence signals instead of final benchmark outcomes, improving…
- TQS assesses trajectories via information focus, intent-action alignment, and policy dynamics without an LLM judge.
- Fine-tuning on top-TQS ranked trajectories from 155144 successful examples outperforms random or low-ranked subsets.
- Process quality in interaction traces provides a stronger supervisory signal than outcome success alone.
Limitations of Outcome-Only Filtering
Supervised fine-tuning of open coding agents typically uses trajectories whose final patches pass benchmark test suites. This treats every successful outcome as equally suitable training data.
Such filtering overlooks process quality. Successful trajectories can still contain inefficient patterns including repeated file inspections, unstable tool usage, or delayed modifications. These patterns become part of the imitation target during training.
TQS Trajectory Quality Score
The preprint introduces TQS, a structural scoring framework that evaluates entire trajectories directly from their structural properties. It requires no LLM judge and runs in linear time.
TQS characterizes quality through signals organized around structured convergence: information focus, intent-action alignment, and policy dynamics. The available evidence is the abstract only; the full paper would be required to review exact definitions or implementation.
This approach offers a scalable alternative to LLM-based data selection methods.
Results on SWE-Bench-Verified
Researchers ranked 155144 successful trajectories from a public open-license corpus using TQS. They then fine-tuned coding agents on filtered subsets.
Models trained on the highest-ranked trajectories consistently outperform those trained on random or bottom-ranked subsets when evaluated on SWE-Bench-Verified. The abstract reports this outcome but provides no specific performance deltas.
Implications for Agent Training Data
The work indicates that process-level trajectory quality is an important factor in coding-agent supervision. Structural filtering can improve imitation learning without relying on proprietary judges or expensive inference.
Developers curating datasets for open coding models may benefit from examining trajectory structure beyond pass-fail outcomes. The preprint is from October 9 2026 and remains an abstract-only capture.
