MaxIM: Maximally Informative Incremental Summarization via Reinforcement Learning
MaxIM trains AI memory to preserve task-relevant information under strict token limits, without human preference labels.
The framework rewards summaries for retaining information that helps predict future task-relevant targets, while keeping them grounded in the observed history. A critic supplies step-by-step training rewards to address the difficulty of learning from sparse, delayed feedback. On long-horizon scientific reading and multi-turn dialogue benchmarks, the authors report better factual grounding and downstream task performance than reinforcement-learning, supervised fine-tuning and prompting baselines. The abstract provides no numerical results.
ArXiv · AI/CL/LG's note
The framework rewards summaries for retaining information that helps predict future task-relevant targets, while keeping them grounded in the observed history. A critic supplies step-by-step training rewards to address the difficulty of learning from sparse, delayed feedback. On long-horizon scientific reading and multi-turn dialogue benchmarks, the authors report better factual grounding and downstream task performance than reinforcement-learning, supervised fine-tuning and prompting baselines. The abstract provides no numerical results.
ArXiv · AI/CL/LG's note
score 5