LLMs fail to recognize their own code, raising collusion risks in model evaluation
Researchers tested whether large language models can identify code they generated versus code from competitors, uncovering a critical vulnerability in LLM-as-judge systems. Across 15 models and multiple benchmarks, evaluators showed near-random performance on single-solution attribution tasks (49-58% balanced accuracy) but strong correlation in pairwise comparisons, suggesting models rely on surface-level stylistic cues rather than genuine authorship recognition. This finding exposes collusion risks in multi-model evaluation frameworks where models judge each other's outputs, potentially compromising benchmark integrity and raising questions about the reliability of LLM-based code review and quality assessment systems.
Modelwire context
Analyst takeThe finding that models can't distinguish their own code from competitors' at better than chance rates on single examples, but succeed on pairwise comparisons, suggests the attribution signal isn't about authorship at all. Models are pattern-matching surface stylistic artifacts, not reasoning about code provenance.
This slots directly into a pattern we've been documenting since late September. The reproducibility audit from the 24th showed that LLM evaluation rankings themselves are unstable (39-96% Jaccard similarity across identical prompts), and now we're seeing why: the models doing the judging lack reliable internal signals for the tasks we're asking them to perform. The code attribution work exposes a specific mechanism of that unreliability. When you stack this with the finding that models' stated reasons for decisions correlate with outcomes but remain fragile, you get a picture of evaluation systems that appear to work at scale but collapse under scrutiny. The collusion risk flagged here (models gaming each other in benchmarks) is the logical endpoint of that fragility.
If researchers apply the same pairwise-versus-singleton test to other 'judge' tasks (math grading, instruction-following assessment, safety evaluation), and find the same surface-cue dependency, that confirms this is a structural property of LLM evaluation, not a code-specific quirk. That would force a reckoning with how we validate models against each other.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsMBPP · HumanEval · DS-1000
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Style, Not Self: Surface Cues Explain Zero-Shot Code Attribution by Large Language Models”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.