Test-Time Scaling in Reasoning LLMs: Inference Regimes, Evaluation, and Reproducibility
Sixth author Accepted
Accepted for publication in TMLR · 2026
- In brief
- Sets out a common framework and reproducibility standards for test-time scaling, that is, spending more computation at inference time to improve a reasoning model's answers.
- Details
- Formalizes test-time scaling as budgeted inference over the implicit prefix tree of an autoregressive transformer (sequential, leaf-level and prefix-level scaling), distinguishes exact replay from distributional reproducibility, and releases over 2 billion reasoning traces.
- My contribution
- Contributed to the discussion of LLM evaluation, the methods for assessing uncertainty, and the development of the new version of Scorio.
BibTeX
% Abbreviated author list; see arXiv for the full record.
@article{Hariri:2026tts,
author = {Hariri, Mohsen and Chen, Weicong and Shahini, Nahal and Singh, Vikash and Ye, K. and Samandar, Amirhossein and Ganguly, D. and others},
title = {Test-Time Scaling in Reasoning {LLMs}: Inference Regimes, Evaluation, and Reproducibility},
journal = {Transactions on Machine Learning Research},
note = {Accepted},
year = {2026},
eprint = {2608.04001},
archivePrefix = {arXiv},
primaryClass = {cs.LG}
}