NeurIPS Spotlight: time-reversed LMs (TRLM) boost best-of-N reranking by up to 5%
LucaAmb · x · 2026-10-15
The paper "Time-Reversal Provides Unsupervised Feedback to LLMs" was accepted as a NeurIPS Spotlight; the poster notes its low citation count (26) understates its importance.
Key points:
- Introduces Time Reversed Language Models (TRLM): models that score and generate queries conditioned on responses, working in the reverse direction of time to give unsupervised feedback complementing forward LLMs.
- The team pre-trained and fine-tuned TRLM-Ba from scratch on reversed token order.
- TRLM scoring for best-of-N reranking yields up to 5% improvement on AlpacaEval over self-log-perplexity baselines.
- TRLM scoring also beats conventional forward scoring in citation generation and passage retrieval, with supporting stylized theory.
More from Research
- Stanford's new CS 312 'Deep Learning Alchemy' trains intuition via prediction quizzes — stanfordnlp · 2026-10-15
- MagicBeanStalk: 88% of RLVR math corpus is a zero-advantage trap, fix via recursive synthesis — stanfordnlp · 2026-10-15
- DuoDiscover Tackles the Evaluation Bottleneck, Claims 2.68x TPU Kernel Gains — stanfordnlp · 2026-10-15
- RAI Institute's single-track bicycle robot UMV lands in Science Robotics — joemeno · 2026-10-15
- Five-year Harvard postdoc preprint unveils BARRACUDA for systematic synthetic circuit design — NikoMcCarty · 2026-10-15
- Autonomous LLM agents for materials science: seminar at Texas Southern University — Fowe · 2026-10-15