The authors analyzed trajectories of coding agents solving repository-level tasks, using a behavioral checker to mark the first action beyond what the task required.
Over-defense in coding agents starts at outcome-settled forks
Luo et al. show that defensive actions beyond task requirements arise at 'silent forks' where all choices lead to the same outcome, eluding outcome-based training signals.
Top University
Hanjun Luo · Xuecheng Liu · Xiucheng Zhang · Zhuoning Xu · Yingbin Jin · Yankai Chen · +2 more
New York University · New York University Abu Dhabi · Carnegie Mellon University · The Hong Kong Polytechnic University · McGill University
Research Digest··3 min read
The authors identify 'silent forks' — turns where every next action leads to the same outcome — as the origin of over-defense in coding agents.
§