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

2026-06-19 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

Researchers introduced IFllm, a novel dataset designed to align Large Language Models (LLMs) using implicit human feedback. This dataset comprises 1336 multi-turn questions collected from 59 Mechanical Turk workers, recording their mouse trajectories and webcam-based eye-gazing points while interacting with LLM responses. Unlike explicit feedback, which is scarce and costly, implicit signals like mouse movement and eye gaze offer a scalable alternative. Analysis of IFllm revealed diverse user reading patterns influenced by response length and interface layout. A reward model trained on this implicit feedback significantly boosted preference prediction accuracy from 55% to 64% compared to text-based models. Furthermore, applying Direct Preference Optimization (DPO) with this implicit feedback-driven reward model nearly tripled the relative response quality improvements from 0.12 to 0.35 across eight LLMs, demonstrating the substantial value of these signals, particularly mouse movement for longer responses.

Key takeaway

For Machine Learning Engineers developing user-facing LLMs, integrating implicit feedback mechanisms is crucial for scalable alignment. Your current explicit feedback methods are likely insufficient; instead, consider deploying mouse trajectory tracking in your LLM interfaces. This approach can significantly improve reward model accuracy and subsequently enhance LLM response quality, especially for longer outputs, fostering a positive feedback loop for user satisfaction. However, ensure robust privacy protocols are in place to mitigate ethical concerns regarding user tracking.

Key insights

Implicit user feedback, especially mouse movement, significantly enhances LLM alignment and preference prediction accuracy.

Principles

• User reading patterns are diverse and length-dependent.
• Mouse movement is crucial for longer response preference.
• Gaze signals are more helpful for short responses.

Method

Collect webcam-based eye-gaze and mouse trajectories during LLM interactions, extract features, then train a random forest reward model to predict user preference.

In practice

• Integrate mouse tracking into LLM interfaces for feedback.
• Use implicit signals to refine DPO for better LLM alignment.
• Develop length-aware feedback mechanisms for LLMs.

Topics

• LLM Alignment
• Implicit Feedback
• Reward Models
• Mouse Tracking
• Eye Tracking
• Data Collection
• Direct Preference Optimization

Code references

• themehulpatwari/llm-implicit-feedback
• eric-mitchell/direct-preference-optimization

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Your Mouse and Eyes Secretly Leak Your Preference: LLM Alignment using Implicit Feedback from Users — Implicit user feedback (mouse/gaze) significantly enhances LLM alignment, outperforming explicit text-based methods.
• Steering at the Source: Style Modulation Heads for Robust Persona Control — Targeting specific "Style Modulation Heads" in LLMs enables robust persona control with minimal coherency degradation.
• LLMs Can Better Capture Human Judgments--With the Right Prompts — Simple prompting strategies, including eliciting statistical distributions and ensuring scenario clarity, significantly improve LLM alignment with human judgment.
• What Do Safety-Aligned LLMs Learn From Mixed Compliance Demonstrations? — LLM interpretation of compliance demonstrations depends on content, ordering, and training methodology, influencing harmful compliance outcomes.
• References Improve LLM Alignment in Non-Verifiable Domains — Reference-guided LLM evaluators significantly improve model alignment and self-improvement in non-verifiable domains.
• Evaluating alignment of behavioral dispositions in LLMs — LLMs often deviate from human behavioral norms and exhibit overconfidence, especially in ambiguous social contexts.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.