analiza aiqbench
HiLight introduces a lightweight reinforcement learning framework that decouples evidence selection from reasoning in frozen LLMs by inserting minimal highlight tags without altering the original context. This approach avoids the information loss associated with compression or rewriting, consistently outperforming prompt-based baselines in long-context tasks like sequential recommendation and question answering. The method demonstrates strong zero-shot transferability across different model families, indicating that the learned emphasis policy captures reusable evidence structures rather than overfitting to specific backbones.
Learning Evidence Highlighting for Frozen LLMs
arXiv:2604.22565v2 Announce Type: replace Abstract: Large Language Models (LLMs) can reason well, yet often miss decisive evidence when it is buried in long, noisy contexts. We introduce HiLight, an Evidence Emphasis framework that decouples evidence selection from reasoning for frozen LLM solvers. HiLight avoids compressing or rewriting the input, which can discard or distort evidence, by training a lightweight Emphasis Actor to insert minimal highlight tags around pivotal spans in the unaltered context. A frozen Solver then performs downstream reasoning on the emphasized input. We cast highlighting as a weakly supervised decision-making problem and optimize the Actor with reinforcement learning using only the Solver's task reward, requiring no evidence labels and no access to or modification of the Solver. Across sequential recommendation and long-context question answering, HiLight consistently improves performance over strong prompt-based and automated prompt-optimization baselines. The learned emphasis policy transfers zero-shot to both smaller and larger unseen Solver families, including an API-based Solver, suggesting that the Actor captures genuine, reusable evidence structure rather than overfitting to a single backbone.
Abstract:Large Language Models (LLMs) can reason well, yet often miss decisive evidence when it is buried in long, noisy contexts. We introduce HiLight, an Evidence Emphasis framework that decouples evidence selection from reasoning for frozen LLM solvers. HiLight avoids compressing or rewriting the input, which can discard or distort evidence, by training a lightweight Emphasis Actor to insert minimal highlight tags around pivotal spans in the unaltered context. A frozen Solver then performs downstream reasoning on the emphasized input. We cast highlighting as a weakly supervised decision-making problem and optimize the Actor with reinforcement learning using only the Solver's task reward, requiring no evidence labels and no access to or modification of the Solver. Across sequential recommendation and long-context question answering, HiLight consistently improves performance over strong prompt-based and automated prompt-optimization baselines. The learned emphasis policy transfers zero-shot to both smaller and larger unseen Solver families, including an API-based Solver, suggesting that the Actor captures genuine, reusable evidence structure rather than overfitting to a single backbone.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2604.22565 [cs.CL]
(or arXiv:2604.22565v2 [cs.CL] for this version)
Text integral sub licența deschisă a publicației · 181 cuvinte
Subiecte conexe
Learning Evidence Highlighting for Frozen LLMs
arXiv · 30 sept. 2026
The user in the LLM might be a moral patient too (a strange implication of the PSM and simulators)
LessWrong · 15 oct. 2026
ReLMem: Learning Recurrent Memory for Longitudinal EHR Modeling
arXiv · 30 sept. 2026
ObjectStream: Latent Objects as Memory Anchors for Streaming Video Understanding
arXiv · 5 aug. 2026