Fixing GRPO training collapse in long-horizon multi-tool agents. A lightweight PRM-Lite + LATA joint approach achieves +37% over vanilla GRPO on τ-bench airline (50-task, multi-turn).

Open harness for running, measuring, and visualizing agent benchmarks. Adapters for AutomationBench, τ-bench, LeRobot, WorkArena.

Retail LLM agent optimization and evaluation showcase built on top of tau2-bench, focused on execution-chain improvements, route comparison, and reproducible benchmark demos.