Researchers want code AI to pay attention the way programmers do, not just the way data patterns suggest it should.
A team has published EyeMulator, a training technique that distills human eye-tracking data into what they call semantic salience and gaze-transition priors. Those priors are used to reweight token-level training losses during fine-tuning — no changes to the underlying model architecture required. The method was tested across six model backbones, two data regimes, and three tasks from the CodeXGLUE benchmark, producing positive results in all 36 model-task-setting combinations. The biggest gains showed up in structure-preserving completion and translation; code summarization improved too, though more modestly.
Most code models learn which tokens matter by finding statistical correlations in training data — a process that has nothing to do with how a human developer actually reads a function. EyeMulator is a bet that the gap between those two attention patterns is a meaningful source of error, and that closing it yields a more robust model without touching the architecture. The model-agnostic design means it could, in theory, be layered onto any existing fine-tuning pipeline.
Eye-tracking studies in software engineering have been a niche research area for years, but practical payoffs have been elusive. Whether reweighting loss signals from gaze data scales to the largest code models — or matters when those models already train on billions of tokens — is still an open question.