New study reveals LLMs often follow misleading code comments over actual code

A new arXiv preprint from researchers at Saarland University presents the first controlled mechanistic study of how large language models (LLMs) handle semantic conflicts in code understanding. The team constructed 45 Python snippet triplets where semantic cues (comments, variable names) contradict the actual code logic, while keeping token-aligned pairs for causal intervention. They evaluated four open-weight LLMs on two tasks: final-output prediction and unit-test generation, using behavioral performance measures and residual-stream activation patching to identify which token-layer states causally contribute to behavioral differences.

The results show that semantic conflicts significantly reduce execution-grounded correctness in both tasks, and all tested LLMs often follow misleading semantic cues. Activation patching revealed a consistent pattern: the changed cue/code region and a small set of intermediate tokens carry most of the recoverable causal signal before aggregation near the output readout. For unit-test generation, this pattern extends beyond the prompt, indicating conflict-related information is recoverable at generated sites. The study provides a framework for mechanistically analyzing how LLMs integrate code-related information under controlled semantic variations, with implications for debugging and improving code-generation models.