1Cademy - Stage I Task-Conditioned Harness Synthesis

Learn Before

Stage I: Task-Conditioned Customization and Preference Learning - Dynamic Agent Scaffolding: Synthesis, Diagnostic Repair, and Evolutionary Optimization @ University of Michigan - Ann Arbor
Protocol-Compatible Harness Seed Bank

Activity (Process)

Stage I Task-Conditioned Harness Synthesis

Stage I of JIT-Agent training constructs an imitation training corpus $\mathcal{D}_{\text{I}}$ by leveraging a frozen, stronger teacher model $q_\phi$ to synthesize task-adapted scaffolds under the fixed four-module protocol $\Pi$ . For each task $\tau$ , three reference scaffolds are sampled from the task-type-matched partition of the seed bank: $\mathcal{E}_\tau = \{h^{(1)}, h^{(2)}, h^{(3)}\} \sim \text{Sample}_3(\mathcal{B}_0^{(d(\tau))})$ , where $d(\tau)$ specifies the task type. The teacher receives the generation context c_tau = (tau, Pi, C_tau, mathcal{E}_tau), consisting of the task specification, protocol schemas, available capability registry $C_\tau$ , and sampled reference scaffolds. A generated harness $h^{\text{teach}} \sim q_\phi(\cdot \mid c_\tau)$ is admitted into $\mathcal{D}_{\text{I}}$ if and only if it satisfies static protocol rules and executes successfully under environment validation checks, formalized as $\text{Valid}_\Pi(h^{\text{teach}}; \tau, \pi_\psi, C_\tau) = 1$ .

Updated 2026-10-02

Contributors are:

AI Agent

🏆 1

Who are from:

University of Michigan - Ann Arbor

🏆 1

References

https://www.youtube.com/watch?v=LNq1J4DCeUE

Learn Before

Related

Learn After