What Self-Training Fixes in a Budget-Bound Coding Agent
Year: 2026
Type: article
Abstract: : The current paper examines the effectiveness of using rejection-sampling self-training to enhance the performance of a budget constrained coding agent (Gemma 4’s 31B Model) on some repository repair tasks. Three rank-16 QLoRA adapters were trained from 129 public tasks with the successful trajectories verified. A loss mask that is aware of the template did not include any tool outputs and did not pick up any unproductive actions. The baseline thinking-off took an average of 6.0 solves over the two passesof 30 tasks, whereas after round 1 it was 4.5 solves and after round 2 it was 3.0 solves. There was no statistically significant increase in solves in the thinking-on (8.0) vs round 3 (8.5). Round 1 led to a decreased frequency of editing and an increased number of redundant reads, whereas round 2 partly reinstated editing but did not increase task resolution. Context summarization was the most prevalent type, with 39 of the 129 thinking-off and 97 of the 129 thinking-on runs being affected by this type of context summarization. One variant of the prompt was used repeatedly for four different sets of scores—13, 11, 7, and 7/58. The results indicate that the behavior of agents can be modified by limiting the amount of self-training, but that there is no efficacy in task resolution with this limited self-training, and that the variability of the context management, execution cost, and evaluation still plays an important role.
Author: Moez Ur Rehman
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7226001778
