Measure the precision and recall of the dense-reward bug detector

Measure the precision and recall of the heuristic bug detector used in the dense reward for generated robot-control programs, including the rate at which valid programs are falsely penalized.

Background

The dense reward incorporates a heuristic detector based on signature, scope, and literal-index indicators. The paper notes that this detector can influence optimization by penalizing programs that exhibit detected bugs, but its classification quality has not been evaluated.

Quantifying precision and recall is therefore necessary to determine whether the shaped reward reliably identifies defective programs or introduces systematic false penalties.

References

The detector is heuristic: precision and recall remain to be measured, and false positives can penalize valid programs.

— Instruction-Conditioned Motion Control for Robotic Arms: Language-to-Code Decomposition and Reinforcement Learning  (2610.19617 - Huang et al., 14 Oct 2026) in Section III-B, Reward sampling and optimization, p. 3