RESEARCH · RESEARCH · #922
StateComp: state-conditioned compression for long-horizon agents (arXiv:2609.27298v1)
StateComp (State Conditioned Compression) is a framework that learns when past interactions can be safely compressed according to the agent's current state. It builds KEEP and READY supervision via a two-stage annotation process, trains an imbalance-aware router on hidden representations from a frozen language model, uses a bounded state representation to reduce evaluation cost, groups adjacent READY interactions into spans, and in experiments on WorkBuddyBench reduced agent+summary tokens by 52.27% while achieving a 12.67× speedup in representation extraction without degrading task performance.
KEY POINTS
- StateComp (State Conditioned Compression) is a framework that learns when past interactions can be safely compressed according to the agent's current state.
- It builds KEEP and READY supervision via a two-stage annotation process, trains an imbalance-aware router on hidden representations from a frozen language model, uses a bounded state representation to reduce evaluation cost, groups adjacent READY interactions into spans, and in experiments on WorkBuddyBench reduced agent+summary tokens by 52.27% while achieving a 12.67× speedup in representation extraction without degrading task performance.
- Reduces context and compute overhead for long-horizon agents by learning state-dependent safe compression points, improving scalability and runtime efficiency.
WHY IT MATTERS
Reduces context and compute overhead for long-horizon agents by learning state-dependent safe compression points, improving scalability and runtime efficiency.