HiLight: Boosting LLM Reasoning via Evidence Highlighting
A new framework called HiLight improves large language model (LLM) reasoning by emphasizing key evidence in long, noisy contexts without modifying the model or input. Developed by researchers, HiLight uses a lightweight Emphasis Actor trained with reinforcement learning to insert minimal highlight tags around pivotal spans. The frozen LLM solver then performs downstream reasoning on the highlighted input. This approach avoids compression or rewriting that could discard or distort evidence. HiLight outperforms strong baselines in sequential recommendation and long-context question answering tasks. The method requires no evidence labels and no access to the solver's internals.
Key facts
- HiLight is an Evidence Emphasis framework for frozen LLMs.
- It decouples evidence selection from reasoning.
- A lightweight Emphasis Actor inserts highlight tags around pivotal spans.
- Training uses reinforcement learning with only the solver's task reward.
- No evidence labels or solver modification are needed.
- HiLight improves performance over prompt-based and automated prompt-optimization baselines.
- Tested on sequential recommendation and long-context question answering.
- Published on arXiv with ID 2604.22565.
Entities
Institutions
- arXiv