Formula

Stage II Diagnostic Harness Repair Objective

The Stage II training objective optimizes JIT-Agent to imitate successful teacher repair transitions by conditioning on the full prior diagnostic history:

LII(θ)=−EDII∑k=0K⋆−1log⁡pθ(Δ⋆(k+1)∣cτ,{(h~(j),g(j))}j=0k)\mathcal{L}_{\text{II}}(\theta) = - \mathbb{E}_{\mathcal{D}_{\text{II}}} \sum_{k=0}^{K^\star - 1} \log p_\theta\left(\Delta^{\star(k+1)} \mid c_\tau, \{ (\tilde{h}^{(j)}, g^{(j)}) \}_{j=0}^k\right)

where DII\mathcal{D}_{\text{II}} contains validated repair trajectories R_{K^star} = { (tilde{h}^{(j)}, g^{(j)}, Delta^{star(j+1)}) }_{j=0}^{K^star - 1}, cτc_\tau is the task generation context, and Δ⋆(k+1)\Delta^{\star(k+1)} is the target patch proposed by the teacher. Because the horizon is bounded by K⋆≤2K^\star \le 2, this objective trains the generator to execute targeted, high-leverage code revisions that restore protocol validity without restarting harness synthesis from scratch.

0

1

Updated 2026-10-02

Tags

Prep Sessions

Dynamic Agent Scaffolding: Synthesis, Diagnostic Repair, and Evolutionary Optimization @ University of Michigan - Ann Arbor

Ch.2 Multi-Stage Harness Optimization - Dynamic Agent Scaffolding: Synthesis, Diagnostic Repair, and Evolutionary Optimization @ University of Michigan - Ann Arbor

Stage II: Diagnostic Feedback and Bounded Harness Repair - Dynamic Agent Scaffolding: Synthesis, Diagnostic Repair, and Evolutionary Optimization @ University of Michigan - Ann Arbor