arXiv:2607.08152v1 Announce Type: new Abstract: On the recent EyeBench benchmark, predicting reading comprehension from eye movements exposes a stark gap: text-aware models using pretrained language models reach 56--63% AUROC, while gaze-only models operate at chance. We ask how far a gaze-only model can be pushed by lightweight, language-model-free conditioning. Building on the EyeBench AhnCNN baseline, LEXIC-Base, we propose two mechanisms to inject three precomputed word-level difficulty signals, GPT-2 surprisal, word frequency, and word length, into the per-fixation input: direct concatena
The paper provides a new architecture in the ongoing pursuit of more efficient and accurate AI models, leveraging 'lightweight, language-model-free conditioning' in eye-tracking for reading comprehension.
This research suggests a path toward more efficient and less computationally intensive AI models for specific tasks, potentially reducing reliance on large, resource-heavy language models for certain applications.
The ability to achieve higher accuracy in gaze-only models through injected complexity could open new avenues for human-computer interaction and accessibility without the overhead of full language model integration.
- · AI researchers
- · Accessibility technology developers
- · Wearable tech companies
- · Developers solely relying on traditional large language models for eye-tracking
Improved eye-tracking interfaces and more robust reading comprehension analytics without massive compute overhead.
Reduced barriers to entry for developing sophisticated gaze-based applications on resource-constrained devices.
New forms of human-computer interaction that are more intuitive and less demanding on processing power.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL