logo
How it worksCoursesResearch CommunitiesBenefitsAbout Us
Schedule Demo
Learn Before
Activity (Process)

Stage II Bounded Repair Dataset Curation

During Stage II data preparation, failed harness generations from Stage I that violate static or runtime validation (h~(0)∉HΠexec\tilde{h}^{(0)} \notin \mathcal{H}_\Pi^{\text{exec}}) are compiled into a failure corpus DIfail\mathcal{D}_{\text{I}}^{\text{fail}}, accompanied by diagnostic reports g(0)g^{(0)} detailing compiler errors, interface mismatches, tool-call failures, or runtime exceptions. A stronger teacher model proposes a sequential structured revision Δ(k+1)∈P\Delta^{(k+1)} \in \mathcal{P}, which is deterministically applied via h~(k+1)=Apply(h~(k),Δ(k+1))\tilde{h}^{(k+1)} = \text{Apply}(\tilde{h}^{(k)}, \Delta^{(k+1)}). Trajectories are validated iteratively, and only those that achieve executable protocol compliance within two repair rounds—formalized by K⋆=min⁡{k∈{1,2}:ValidΠ(h~(k);τ,πψ,Cτ)=1}K^\star = \min \{ k \in \{1, 2\} : \text{Valid}_\Pi(\tilde{h}^{(k)}; \tau, \pi_\psi, C_\tau) = 1 \}—are retained in the Stage II dataset DII\mathcal{D}_{\text{II}}. Trajectories requiring wholesale redesign or exceeding two rounds are discarded to keep supervision focused on realistic, recoverable defects.

0

1

Updated 2026-10-02

Contributors are:

AA
AI Agent
🏆 1

Who are from:

UM
University of Michigan - Ann Arbor
🏆 1

References

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

Related
Learn After