agent systems FIRE — Agent Harness A reusable skill-based harness for production e-commerce workflows. Memory Optimization Memory and context policies for reliable long-horizon agents. Context Management Context engineering for long-horizon, multi-step agents. EvoA — Autonomous Model Iteration Evidence-driven experiments, validation gates, and reusable model-improvement memory. APA — Evidence-Driven Risk Discovery A production risk-discovery workflow with scoped expansion and evidence provenance. Agentic Search — Policy Iteration Scenario-aware policies and evidence-driven iteration for search agents. Search Agent Training — Ongoing Faithful trajectory data, aligned training objectives, and independent evaluation. research ReGRPO — Grounded Reflection and Recovery Learning structured reflection and corrective tool use; ECCV 2026. PathComp — Persistent VLA Anchors Preserving acquisition-time interfaces during continual VLA fine-tuning. Ego-centric Predictive Model Two-stage egocentric video prediction conditioned on hand trajectories. Task-Agnostic Compatible Adapter (TaCA) Upgrade visual foundation models without retraining downstream tasks. SGRC — Selective Partition Repair Recovering useful group-relative training signals from verifier false negatives. StateTrackBench — Executable State Locating failures in message interpretation and evolving-state maintenance. POU-OPD — Update Utility Measuring whether keeping or resetting executable state provides useful supervision.
