Reinforcement Learning from Human Feedback (RLHF) aligns language models with human preferences but faces efficiency challenges. We explore two approaches leveraging human gaze prediction to enhance RLHF: (1) gaze-aware reward models and (2) gaze-based distribution of sparse rewards at token level. Our experiments show gaze-informed RLHF achieves faster convergence while maintaining or slightly improving performance, reducing computational requirements during policy optimization. Human visual attention patterns provide valuable signals for policy training, suggesting a promising direction for improving RLHF efficiency through human-like attention mechanisms.

Eye movement analysis has become an essential tool for studying cognitive processes in reading, serving both psycholinguistic research and natural language processing applications aimed at enhancing language model performance. However, the scarcity of eye-tracking data and its limited generalizability constrain data-driven approaches. Synthetic scanpath generation offers a potential solution to these limitations. While recent advances in scanpath generation show promise, current literature lacks systematic evaluation frameworks that comprehensively assess models’ ability to reproduce natural reading gaze patterns. Existing studies often focus on isolated metrics rather than holistic evaluation of cognitive plausibility. This study presents a systematic evaluation of contemporary scanpath generation models, assessing their capacity to replicate natural reading behavior through comprehensive scanpath analysis. We demonstrate that while synthetic scanpath models successfully reproduce basic gaze patterns, significant limitations persist in capturing part-of-speech dependent gaze features and reading behaviors. Our cross-dataset comparison reveals performance degradation in three key areas: generalization across text genres, processing of long sentences, and reproduction of psycholinguistic effects. These findings underscore the need for more robust evaluation protocols and model architectures that better account for psycholinguistic complexity. Through detailed analysis of fixation sequences, durations, and reading patterns, we identify concrete pathways for developing more cognitively plausible scanpath generation models.

Concept Bottleneck Models (CBMs) assume that training examples (e.g., x-ray images) are annotated with high-level concepts (e.g., types of abnormalities), and perform classification by first predicting the concepts, followed by predicting the label relying on these concepts. However, the primary challenge in employing CBMs lies in the requirement of defining concepts predictive of the label and annotating training examples with these concepts. In our approach, we adopt a more moderate assumption and instead use text descriptions (e.g., radiology reports), accompanying the images, to guide the induction of concepts. Our crossmodal approach treats concepts as discrete latent variables and promotes concepts that (1) are predictive of the label, and (2) can be predicted reliably from both the image and text. Through experiments conducted on datasets ranging from synthetic datasets (e.g., synthetic images with generated descriptions) to realistic medical imaging datasets, we demonstrate that crossmodal learning encourages the induction of interpretable concepts while also facilitating disentanglement.