Formalizing Learning from Language Feedback with Provable Guarantees

Researchers have made two significant advancements in using language feedback for machine learning and assistive robots. They formalized the Learning from Language Feedback (LLF) problem and developed the HELiX algorithm, and proposed a framework for personalizing assistive robots using natural language feedback.

The LLF problem involves learning from observation and language feedback, with the HELiX algorithm providing a no-regret solution with performance guarantees that scale with the transfer eluder dimension^[1]. This development is significant as it enables learning despite latent rewards. Meanwhile, a separate study proposed a low-burden framework to personalize assistive robots for users with paralysis using natural language feedback. The framework translates unstructured natural language feedback into deterministic robotic control policies and reduces user workload compared to traditional baselines in a simulated meal preparation study with 10 adults with paralysis^[2]. The researchers behind the HELiX algorithm demonstrated its effectiveness across several empirical domains, showing that it performs well even when repeatedly prompting Large Language Models (LLMs) does not work reliably. The assistive robots framework uses LLMs grounded in the Occupational Therapy Practice Framework, highlighting the versatility of LLMs in different applications. The HELiX algorithm paper was initially submitted in 2025 and revised in 2026^[1]. The assistive robots paper was submitted on 1 Apr 2026 and revised on 12 Jun 2026^[2].