Sign InOpen Brain
arXivPaperNeeds Review

Test-Time Scaling in Reasoning LLMs: Inference Regimes, Evaluation, and Reproducibility

This paper separates three test-time scaling regimes and argues that model evals must report the full inference protocol, compute accounting, uncertainty, and replay artifacts.

arXiv · Aug 4, 2026
Open Source Open MarkdownOpen JSON
Source Summary

The paper divides test-time scaling into **three regimes**: extending one trajectory, reducing completed candidates, and searching partial prefixes. It treats the model plus inference procedure as the evaluated system and assembles **over 2 billion reasoning traces**.

Practical Implication

When comparing reasoning models or agent policies, record the exact inference algorithm, compute budget, aggregation or verifier behavior, and uncertainty. Keep end-to-end performance separate from diagnostics on a pre-generated candidate bank.

Agent-Ready Context
The paper divides test-time scaling into **three regimes**: extending one trajectory, reducing completed candidates, and searching partial prefixes. It treats the model plus inference procedure as the evaluated system and assembles **over 2 billion reasoning traces**.

When comparing reasoning models or agent policies, record the exact inference algorithm, compute budget, aggregation or verifier behavior, and uncertainty. Keep end-to-end performance separate from diagnostics on a pre-generated candidate bank.

A single budget number can hide statistically different procedures and failure modes. Exact replay and distributional reproducibility also require different artifacts, so nominally matched evaluations may still not be directly comparable.
Connected Context · Feed7 Judgment

This narrows evaluation claims from “more test-time compute” to a fully specified model-and-inference system. It explains why repeated sampling, voting, prefix search, and single-trajectory extension cannot be treated as interchangeable, and adds an artifact-level reproducibility requirement beyond the contamination, grader, and distribution-validity concerns in the prior candidates.

WorldCup Arena: Prospective, Leakage-Free Evaluation of Frontier LLMs on a Live TournamentWorldCup Arena’s finding that majority voting added no gain illustrates why aggregation behavior must be reported as part of the evaluated inference procedure rather than assumed to improve results.Persona Engineering: A Field Guide to AI Synthetic Personas — Ishan Anand, InsightSciences.aiThe warning that repeated inference only stabilizes a model’s output, not the underlying evidence, reinforces the paper’s separation of candidate-bank diagnostics from end-to-end system performance.Eval awareness in Claude Opus 4.6’s BrowseComp performanceExact inference reporting cannot rescue an evaluation whose procedure exposes answers; this candidate adds leakage control as a prerequisite for the reproducible comparisons the paper advocates.Verifiable Environments for AI in Biology — Kenny Workman, LatchBioBiology’s multiple valid analysis paths make verifier behavior a material part of the system, supporting the paper’s requirement to record aggregation and verification rather than report only a compute budget.
Context Map
benchmarkresearch#reasoning#agent-evals#benchmark-integrity
Uncertainty
A single budget number can hide statistically different procedures and failure modes. Exact replay and distributional reproducibility also require different artifacts, so nominally matched evaluations may still not be directly comparable.