Agents Verify Their Assumptions, Not the Specification: Interpretation Lock-in in Autonomous Coding Agents

Large language models (LLMs) have become capable coding agents that complete long software tasks autonomously and test their own work. While LLMs excel at well-defined programming tasks such as code generation and bug fixing, their self-verification breaks down when a requirement is ambiguous, because the tests they write share the interpretation of the code they test. In this paper, we study this failure, which we call interpretation lock-in: agents commit to one reading of an ambiguous requirement, implement it, and validate their work against that same reading. In a controlled suite where every valid reading is known, and in pipelines rebuilt from long-horizon benchmark runs, the resulting programs crash or silently produce wrong output under other valid readings. The failure is largely one of commitment rather than knowledge: asked only to list plausible readings, the same models often name the one they missed. To remedy this failure, we propose reading-diverse self-training (RDST), which rewards an agent’s own so- lutions for the number of valid readings they handle and fine-tunes the agent on its best episodes. Experimental results show that, relative to rewarding success on a single reading, RDST substantially reduces lock-in, including on task families held out from training, at little cost on well-specified tasks, and that the behaviour carries over to the interaction protocol of another agent harness.