Inference‑time scaling is often framed as a token‑budget problem, yet a failed coding agent can pursue qualitatively different resources: computation, information, or feedback. To study this, we introduce Banzhaf Resource Attribution (BRA), a controlled factorial framework that measures each resource’s marginal contribution to repair and their interactions while keeping the failed state and revision protocol fixed.
We evaluate six models on 194 function‑level failures from MBPP+ and HumanEval+, and 126 repository‑level failures from SWE‑bench Verified. The findings are:
- Task‑relevant specifications consistently boost repair success;
- Public‑test feedback yields only modest gains;
- Retrieved code provides little to no benefit.
The value of additional computation diminishes as model capability increases, whereas information remains useful. Most resource channels exhibit diminishing returns, though complementary information can reinforce each other. Notably, three of four frontier models select feedback in 95‑100% of cases but achieve lower gains than uniform random resource selection.
BRA thus exposes a gap between the resources agents prefer and those that measurably improve repair.
Review