When LLMs Stop Following Steps: A Diagnostic Study of Procedural Execution in Language Models

The key distinction the summary gestures at but doesn't fully unpack is that procedural faithfulness is a separate failure axis from reasoning ability: a model can solve the underlying logic of a problem while still botching execution by dropping a variable or skipping a step mid-sequence, which means capability benchmarks that test reasoning won't catch this class of error at all.

This connects directly to The Decoder's coverage of ARC-AGI-3 analysis from May 2nd, which identified three systematic reasoning errors in frontier models and argued that isolating specific failure modes gives researchers concrete targets. That paper and this one are converging on the same uncomfortable finding from different angles: benchmark scores describe average performance but obscure structured, repeatable failure patterns that only surface under controlled diagnostic conditions. The Harvard emergency room study from May 3rd adds pressure to this picture, because clinical deployment of LLMs in diagnostic workflows assumes reliable sequential reasoning across multi-step differential diagnosis, and this procedural faithfulness research suggests that assumption hasn't been validated.

Watch whether any major lab responds to this benchmark by publishing training interventions specifically targeting step-tracking fidelity on sequences above 50 steps. If no lab engages within two quarters, that suggests the research community treats this as an eval curiosity rather than a deployment-blocking problem.