A new arXiv paper describes a standalone summarizer trained via masked reconstruction, reporting large inference cost reductions on two long-horizon benchmarks.
A new arXiv paper describes a standalone summarizer trained via masked reconstruction, reporting large inference cost reductions on two long-horizon benchmarks.
Researchers introduce SumOne, a summarizer designed to compress the interaction histories that long-horizon language-model agents accumulate. The method trains a summarizer to produce summaries of expired history such that a frozen reconstructor can recover randomly masked portions of the original trajectory. Reconstruction quality, combined with a length penalty, serves as a task-agnostic reward optimized with GRPO. The authors state this yields self-supervision directly from raw trajectories, without reference summaries or downstream task execution, and that existing learned compaction methods often couple training to specific tasks or agents.
According to the paper, SumOne reduces estimated inference cost by approximately 78 percent on DeepSearchQA and 75 percent on SWE-bench Verified relative to operating with full context. The evaluation covers long-horizon search and software-engineering tasks. At inference time only the lightweight summarizer is retained, which the authors describe as enabling plug-and-play context compression for otherwise unchanged agents. The excerpt notes absolute decreases in performance but the figures for those decreases are cut off in the available material.
The excerpt does not report the size of the performance drops, the models or agents tested, baseline comparisons against other compaction methods, or the authors and their affiliations. The cost figures are described as estimated rather than measured, and the evaluation covers two benchmarks whose generality is unclear. Readers should watch for the full paper to confirm the absolute performance decrements, the reconstructor architecture, and how the approach compares with task-coupled alternatives.
Source: SUMONE: A Simple and Efficient Summarizer for Plug-and-Play Agent Memory