Newest papers

Adaptive Replay Refinement: Allocating Test-Time Compute Between Revision and Resampling

Ahsan Bilal, Muhammad Ahmed Mohsin, Muhammad Saad Umer, Ayesha Mohsin, Sarmad Bilal, Dean F. Hougen · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.28294 · cited by 0 works (OpenAlex) · open-access PDF

Portuguese

O Code-time scaling aloca computadores de inferência adicionais para refinar soluções existentes ou gerar novas. Isto levanta uma pergunta fundamental: Devo continuar a melhorar essa solução, ou parar de trabalhar nela e tentar uma solução completamente nova? Os métodos existentes tomam essa decisão usando horários de refinamento fixos, julgamentos de auto-verificação, ou controladores aprendidos, que podem desperdiçar computadores ou depender de sinais que possam ser incorretos ou exigir treinamento adicional. Introduzimos o Adaptive Replay Refinement (ARR), um método eficiente que detecta quando uma revisão estacionária atinge um ponto fixo e redistribui o computador liberado para uma nova amostra independente; o replay trata a solução atual como um projeto e verifica tokens inalterados em paralelo

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

Abstract

Test-time scaling allocates additional inference compute to refining existing solutions or generating new ones. This raises a fundamental question: Should I keep improving this solution, or stop working on it and try a completely new solution? Existing methods make this decision using fixed refinement schedules, self-verification judgments, or learned controllers, which can waste compute or rely on signals that may be incorrect or require additional training. We introduce Adaptive Replay Refinement (ARR), a training-free method that detects when a stationary revision reaches a fixed point and reallocates the released compute to a new independent sample; replay decoding treats the current solution as a draft and verifies unchanged tokens in parallel, enabling efficient fixed-point detection. Across seven models from five families and seven benchmarks, ARR matches majority voting on three of five mathematical reasoning benchmarks at equal compute, while reducing critical-path serial passes by up to 2.4 times. On MATH500, ARR achieves 84.6% accuracy with only 41% of the serial passes required by majority voting. These results demonstrate that detecting refinement exhaustion enables adaptive reallocation from revision to independent sampling, improving inference efficiency while highlighting the limits of self-revision. Code is available at https://github.com/AhsanBilal7/adaptive_refine.

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

Topics: Computer science, Resampling, Inference, Point (geometry), Algorithm, Code (set theory)

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