Skip to main content
Back to timeline
arXivSource publication:

Gaze no longer stops at the fixated word: spreading reading time into an asymmetric attention profile lifts preference accuracy by up to 13.02%

Synopsis

In reward-model training, eye-tracking data previously attached only to the directly fixated token, ignoring that readers take in text around the fixation. The new step spreads each token's predicted total reading time onto neighboring tokens with a learnable asymmetric Gaussian kernel before the gaze features reach the existing projector and reward model. On OASST1, accuracy rose from 68.38% to 74.73% for Llama-3-8B-Instruct under GazeConcat and from 69.23% to 74.81% for Llama-3-8B; on UltraFeedback it rose from 65.65% to 74.20%.

Source-provided article image: Leveraging Psychophysical Attentional Distribution for Gaze-Augmented Reward Modeling
arXiv

Interpretation

The gaze signal a reward model receives changes from each token carrying only its own fixation time to each token carrying a spatially graded attention distribution centered on the fixation, with separately learnable left and right widths. Earlier GazeReward-style approaches treated gaze as a discrete token-wise feature tied to the directly fixated token, whereas reading research holds that visual attention is spatially distributed around a fixation. Averaged over six random seeds on OASST1, preference-prediction accuracy under GazeConcat rose from 68.38% to 74.73% for Llama-3-8B-Instruct and from 69.23% to 74.81% for Llama-3-8B; under GazeAdd the two rose from 70.30% to 71.89% and from 69.27% to 71.17%.

The mechanism inserts a redistribution module between gaze-feature generation and projection, flattening sparse token-level total reading time into a dense distribution over neighboring tokens. Total reading time (TRT), the summed duration of all fixations on a word, serves as a proxy for attention allocated; redistribution weights neighbors by token-index distance using an asymmetric Gaussian kernel centered on the fixated token, with separately learnable left and right spreads. Weights are normalized per fixation so total reading time is preserved, and the two width parameters are trained jointly with the reward model; with Llama-3-8B as backbone and GazeConcat as integration method, the configuration beat the corresponding pinpoint baseline on OASST1, HelpSteer2 and UltraFeedback.

The learned profile extends further to the right than to the left in most settings, directionally matching the asymmetric perceptual span reported in reading psychophysics. In left-to-right reading the perceptual span typically reaches farther right of fixation than left, while a pinpoint representation carries no directionality at all. Under GazeConcat on OASST1, the right-side width exceeded the left-side width for both Llama-3-8B-Instruct and Llama-3-8B; HelpSteer2 was the only dataset where the learned profile extended further left than right, and it also showed the smallest improvement.

Perspective

Experiments cover the English preference datasets OASST1, HelpSteer2 and UltraFeedback, with Llama-3-8B and Llama-3-8B-Instruct as reward-model backbones and GazeConcat and GazeAdd as gaze integration settings; each condition is run with six random seeds and reported as mean plus standard deviation, and for consistency with the public GazeReward code and under compute limits an input-length cap was imposed and over-length samples were excluded.

The redistribution remains a simplified approximation of human visual attention during reading: the Gaussian kernel forces monotonic decay from the fixation center, whereas prior work suggests the attentional field may be nonmonotonic, Mexican-hat-like and variable across individuals; redistribution currently uses token-level rather than letter-level distance and only total reading time among the available gaze features. HelpSteer2 showed the smallest gain and the only left-wider-than-right learned profile, which the authors speculate may relate to that dataset offering almost only two candidate responses per prompt, making strongly contrasting preference pairs hard to construct.