Newest papers

Where Should a Stuck Agent Spend Its Next Tokens? Causal Attribution of Inference-Time Resources

Saiyue Lyu, Jerrold Huang, Yunbei Zhang, Janet Wang, Zhenting Wang, Yingqiang Ge, Haibo Zhang · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.28834 · cited by 0 works (OpenAlex) · open-access PDF

Portuguese

A escalação de tempo de inferência é muitas vezes tratada como um problema de orçamento de token, mas um agente de codificação falhado pode procurar recursos qualitativamente diferentes: computação, informação ou feedback. Introduzimos a Atribuição de Recursos Banzhaf (BRA), uma estrutura de fatores controlada que mede a contribuição marginal de cada recurso para reparar e suas interações, mantendo o estado falhado e o protocolo de revisão fixado. Avaliamos seis modelos em 194 falhas de nível de função de MBPP+ e HumanEval+ e 126 falhas de nível de repositório de SWE-bench Verified. Especificações relevantes para tarefas melhoram consistentemente a reparação, enquanto o feedback do teste público fornece pequenos ganhos e o código recuperado para pouco. O valor da computação

Machine translation (Meta M2M100); the English original follows.

Abstract

Inference-time scaling is often treated as a token-budget problem, but a failed coding agent can seek qualitatively different resources: computation, information, or feedback. We introduce Banzhaf Resource Attribution (BRA), a controlled factorial framework that measures each resource's marginal contribution to repair and its interactions while holding 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. Task-relevant specifications consistently improve repair, whereas public-test feedback provides small gains and retrieved code little to none. The value of additional computation declines with model capability, while information remains useful; resource channels largely exhibit diminishing returns, although complementary information can reinforce itself. Yet three of four frontier models select feedback in 95-100% of cases and obtain lower gains than uniform random resource selection. BRA thus exposes a gap between the resources agents prefer and those that measurably improve repair.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Attribution, Computer science, Coding (social sciences), Resource (disambiguation), Protocol (science), Marginal value

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange