TEMPO: Scaling Test-time Training for Large Reasoning Models

Hugging Face Daily Papers Papers

Summary

TEMPO introduces a test-time training framework that alternates policy refinement with critic recalibration to prevent diversity collapse and sustain performance gains in large reasoning models, boosting AIME 2024 scores for Qwen3-14B from 42.3% to 65.8%.

Test-time training (TTT) adapts model parameters on unlabeled test instances during inference time, which continuously extends capabilities beyond the reach of offline training. Despite initial gains, existing TTT methods for LRMs plateau quickly and do not benefit from additional test-time compute. Without external calibration, the self-generated reward signal increasingly drifts as the policy model evolves, leading to both performance plateaus and diversity collapse. We propose TEMPO, a TTT framework that interleaves policy refinement on unlabeled questions with periodic critic recalibration on a labeled dataset. By formalizing this alternating procedure through the Expectation-Maximization (EM) algorithm, we reveal that prior methods can be interpreted as incomplete variants that omit the crucial recalibration step. Reintroducing this step tightens the evidence lower bound (ELBO) and enables sustained improvement. Across diverse model families (Qwen3 and OLMO3) and reasoning tasks, TEMPO improves OLMO3-7B on AIME 2024 from 33.0% to 51.1% and Qwen3-14B from 42.3% to 65.8%, while maintaining high diversity.
Original Article
View Cached Full Text

Cached at: 04/22/26, 06:17 AM

Paper page - TEMPO: Scaling Test-time Training for Large Reasoning Models

Source: https://huggingface.co/papers/2604.19295

Abstract

TEMPO is a test-time training framework that alternates policy refinement with critic recalibration to sustain performance improvements in language models without diversity collapse.

Test-time training(TTT) adapts model parameters on unlabeled test instances during inference time, which continuously extends capabilities beyond the reach of offline training. Despite initial gains, existing TTT methods for LRMs plateau quickly and do not benefit from additional test-time compute. Without external calibration, the self-generated reward signal increasingly drifts as the policy model evolves, leading to both performance plateaus anddiversity collapse. We propose TEMPO, a TTT framework that interleavespolicy refinementon unlabeled questions with periodiccritic recalibrationon a labeled dataset. By formalizing this alternating procedure through the Expectation-Maximization (EM) algorithm, we reveal that prior methods can be interpreted as incomplete variants that omit the crucial recalibration step. Reintroducing this step tightens theevidence lower bound(ELBO) and enables sustained improvement. Across diverse model families (Qwen3 and OLMO3) and reasoning tasks, TEMPO improves OLMO3-7B onAIME 2024from 33.0% to 51.1% and Qwen3-14B from 42.3% to 65.8%, while maintaining high diversity.

View arXiv pageView PDFProject pageGitHub0Add to collection

Get this paper in your agent:

hf papers read 2604\.19295

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2604.19295 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2604.19295 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2604.19295 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

REVES: REvision and VErification--Augmented Training for Test-Time Scaling

Hugging Face Daily Papers

Proposes REVES, a two-stage iterative framework that alternates between data augmentation and policy optimization to improve LLM reasoning by leveraging intermediate correction steps, achieving superior performance on coding benchmarks and constraint satisfaction problems.

TMAS: Scaling Test-Time Compute via Multi-Agent Synergy

Hugging Face Daily Papers

TMAS introduces a multi-agent framework that enhances large language model reasoning by scaling test-time compute through structured collaboration and hierarchical memory systems. The approach uses specialized agents, cross-trajectory information flow, and hybrid reward reinforcement learning to improve iterative scaling and stability on challenging reasoning benchmarks.

ThinkBooster: A Unified Framework for Seamless Test-Time Scaling of LLM Reasoning

arXiv cs.CL

ThinkBooster is a unified framework for test-time compute scaling of LLM reasoning, providing a modular Python library, a performance-efficiency benchmark, an OpenAI-compatible proxy service, and a visual debugger. Empirical results on math and coding tasks demonstrate practical gains with quality-cost trade-offs.