Your Mouse and Eyes Secretly Leak Your Preference: LLM Alignment using Implicit Feedback from Users

A new study shows that mouse movements and eye gaze can serve as implicit signals to align large language models, reducing reliance on expensive explicit human feedback. Researchers released a dataset capturing these behaviors from 59 Mechanical Turk workers across 1,336 multi-turn questions. The work, posted to the arXiv preprint server on June 18, 2026, addresses a core challenge in training LLMs: collecting high-quality human preference data is costly because users rarely provide explicit ratings [1][2]. Internet platforms have long used implicit behavioral signals to refine their products, but the approach has not been systematically applied to language model alignment [2]. Haw-Shiuan Chang and collaborators constructed the IFLLM dataset to quantify the value of such signals [1]. The dataset records mouse trajectories and webcam-based eye-gazing points as 59 workers on Amazon Mechanical Turk interacted with responses from eight different LLMs across 1,336 multi-turn questions [1][2]. The researchers found that users exhibited highly diverse gazing patterns and mouse behaviors [2]. A reward model that incorporated this implicit feedback improved the accuracy of a text-only reward model from 55% to 64% [1][2]. When the implicit-feedback reward model was used to apply Direct Preference Optimization to the eight LLMs, the relative improvement in response quality nearly tripled compared to using the text-based model alone [2]. The paper’s authors have made the data-collection website, the IFLLM dataset, and the accompanying code publicly available on GitHub [1][2]. arXiv, which hosts the paper, is an open-access repository that has grown to receive roughly 24,000 submissions per month as of late 2024 and does not itself conduct peer review [6]. The study appears amid broader efforts to make LLM training more efficient; the transformer architecture that underpins most modern LLMs was introduced in the 2017 paper “Attention Is All You Need,” which has since been cited more than 250,000 times [8].