Time-Reversal Provides Unsupervised Feedback to LLMs

​This paper was awarded spotlight distinction in NeurIPS 2025! Very few papers get that. We think it’ll also be interesting to discuss with the author - what it takes to produce such frontier, world class research right here from India.

​Large Language Models (LLMs) are typically trained to predict in the forward direction of time. Recent studies have shown that prompting these models to look back and critique their own responses can produce useful feedback. Motivated by this, we introduce Time Reversed Language Models (TRLMs), which are empowered to think (predict and score) queries based on the responses. To effectively infer in the response-to-query direction, we pretrain and fine-tune an LLM, TRLM-Ba, in reverse token order from scratch. We empirically show that TRLMs can indeed complement a forward LLM's responses, resulting in better instruction-following. The "reversed" direction of scoring also helps with citation attribution and document retrieval. Furthermore, we leverage the generative capability of TRLMs to provide unsupervised feedback to input safety filters against harmful content.

​Speaker: Varun Yerram

• ​Twitter: twitter.com/varunyer

• ​LinkedIn: https://www.linkedin.com/in/yvarun25221/

​Current: Pre-Doctoral Researcher at Google DeepMind Previous: Applied Scientist Intern @ AmazonScience, ECE @ IIT Guwahati