WEFT: Scaling Tool-Use Post-Training for General-Purpose Agents

WEFT. Scaling Tool-Use Post-Training for General-Purpose Agents. By Bo Mao and nineteen others, from East China Normal University, Fudan University, and three other institutions. Published on the twenty-ninth of September. No code repository is linked to it.

Efforts to scale tool-use post-training have mostly concentrated on one thing: synthesizing executable environments. But an environment is only one part of what the authors call an agentic interaction system. The others are the task, the agent harness, and the evaluator. Scaling environments on their own does not guarantee proportional gains in model performance, because a reliable learning signal depends on all four components interacting coherently. More environments help only as far as the tasks, the harness and the evaluator around them hold together. Put another way, recent efforts have scaled one part of the system, while the signal the model actually learns from is produced by the interaction among all four.

WEFT, short for Whole-system Evolution For Tool-use Post-training, scales the whole system rather than one piece of it. It couples three things. Scalable construction of agentic interaction systems. Execution-driven self-evolution of those systems. And post-training that stays stable at scale.

Construction scales along three axes at once: breadth of environments, complexity of tasks, and diversity of interactions.

Self-evolution is a repair loop that runs in rounds. Execution traces and state evidence are used to attribute a failure to whichever component is responsible for it, and that component is revised. Fresh rollouts then evaluate the change, and what they show becomes the evidence for the next round of evolution. The system, in other words, is not fixed once it is built.

Stable post-training has two sides, optimization and execution. On the optimization side, prefix-preserving sampling keeps progress that has already been verified, and atomic-turn credit assignment localizes the learning signal at the level of single turns. On the execution side, a component called MegaMCP keeps state isolated and recoverable across many concurrent rollouts that share the same tool services, so that one rollout does not disturb another.

Experiments span various models and benchmarks. Two sizes, WEFT-8B and WEFT-14B, outperform every matched-size environment-scaling baseline evaluated, on BFCL V4, tau-squared Bench, and Claw-Eval.

Against Agent-World-14B in particular, WEFT-14B gains six point four one percentage points, two point two three, and twelve point two seven, across those three benchmarks in turn. A further model, WEFT-35B-A3B, carries the gains further into harder long-horizon workflow benchmarks, including Toolathlon-Verified and AutomationBench.

The argument underneath is a systems argument. The learning signal comes from the whole loop of environment, task, harness and evaluator, so WEFT grows all of them together, lets execution itself point out which one to repair next, and then makes sure training at that scale stays stable enough to use what the loop produces.