SumOne
A plug-and-play memory summarizer for agents, trained via masked reconstruction and GRPO optimization. It compresses the interaction history of long-horizon agents to reduce inference cost without modifying the original agent.
Timeline (last 90 days)
SumOne is a standalone summarizer that preserves information in agent history through masked reconstruction training.
SumOne uses a frozen reconstructor to recover randomly masked portions of the original trajectory from the summary, and combines reconstruction quality with a length penalty to optimize the summarizer via GRPO.
On DeepSearchQA, SumOne reduces inference cost by about 78% compared with full context, with an absolute performance drop of only 0.04.
The same SumOne checkpoint transfers across different downstream models and agent frameworks without additional fine-tuning.
At inference time, SumOne retains only a lightweight summarizer and does not alter the original agent, achieving plug-and-play context compression.
SumOne's training requires no reference summaries or downstream task execution, learning self-supervised directly from raw trajectories.
All Facts (6)
SumOne is a standalone summarizer that preserves information in agent history through masked reconstruction training.
50%UnverifiedSumOne uses a frozen reconstructor to recover randomly masked portions of the original trajectory from the summary, and combines reconstruction quality with a length penalty to optimize the summarizer via GRPO.
50%UnverifiedOn DeepSearchQA, SumOne reduces inference cost by about 78% compared with full context, with an absolute performance drop of only 0.04.
50%UnverifiedThe same SumOne checkpoint transfers across different downstream models and agent frameworks without additional fine-tuning.
50%UnverifiedAt inference time, SumOne retains only a lightweight summarizer and does not alter the original agent, achieving plug-and-play context compression.
50%UnverifiedSumOne's training requires no reference summaries or downstream task execution, learning self-supervised directly from raw trajectories.
50%