Revisiting the Test-Time Scaling of o1-like Models: Do they Truly Possess Test-Time Scaling Capabilities?

2025/02/17 by Zhiyuan Zeng, Zeng, Zhiyuan, Qinyuan Cheng +7 · 1 voice · 52 citations
Computer Science · Decision Sciences · Mathematics · #Computer science #Geology #Geometry #Mathematics #Scaling #Scientific Computing and Data Management #Test (biology) #Time Series Analysis and Forecasting #Topic Modeling #cs.AI #cs.CL #cs.LG

paper · pdf · doi:10.48550/arxiv.2502.12215

published in arXiv (Cornell University) (Cornell University)

openalex publication_date 2025/02/17 · arxiv published 2025/02/17 · arxiv updated 2025/03/03 · openalex created_date 2025/10/10 · openalex updated_date 2026/08/05

Abstract

The advent of test-time scaling in large language models (LLMs), exemplified by OpenAI's o1 series, has advanced reasoning capabilities by scaling computational resource allocation during inference. While successors like QwQ, Deepseek-R1 (R1) and LIMO replicate these advancements, whether these models truly possess test-time scaling capabilities remains underexplored. This study found that longer CoTs of these o1-like models do not consistently enhance accuracy; in fact, correct solutions are often shorter than incorrect ones for the same questions. Further investigation shows this phenomenon is closely related to models' self-revision capabilities - longer CoTs contain more self-revisions, which often lead to performance degradation. We then compare sequential and parallel scaling strategies on QwQ, R1 and LIMO, finding that parallel scaling achieves better coverage and scalability. Based on these insights, we propose Shortest Majority Vote, a method that combines parallel scaling strategies with CoT length characteristics, significantly improving models' test-time scalability compared to conventional majority voting approaches.

Cited by

• DiRL: An Efficient Post-Training Framework for Diffusion Language Models
• MARINE: Theoretical Optimization and Design for Multi-Agent Recursive IN-context Enhancement
• Think in Parallel, Answer as One: Logit Averaging for Open-Ended Reasoning
• Synthesizing Visual Concepts as Vision-Language Programs
• The Sequential Edge: Inverse-Entropy Voting Beats Parallel Self-Consistency at Matched Compute
• Generalizing Test-time Compute-optimal Scaling as an Optimizable Graph
• Boosting Accuracy and Efficiency of Budget Forcing in LLMs via Reinforcement Learning for Mathematical Reasoning
• SolverLLM: Leveraging Test-Time Scaling for Optimization Problem via LLM-Guided Search
• Merlin's Whisper: Enabling Efficient Reasoning in LLMs via Black-box Adversarial Prompting
• First Try Matters: Revisiting the Role of Reflection in Reasoning Models
• ARISE: An Adaptive Resolution-Aware Metric for Test-Time Scaling Evaluation in Large Reasoning Models
• Context Length Alone Hurts LLM Performance Despite Perfect Retrieval
• Overthinking Reduction with Decoupled Rewards and Curriculum Data Scheduling
• From Harm to Help: Turning Reasoning In-Context Demos into Assets for Reasoning LMs
• Mitigating Strategy-Selection Bias in Reasoning for More Effective Test-Time Scaling
• ATTS: Asynchronous Test-Time Scaling via Conformal Prediction
• Input-Time Scaling: Adding Noise and Irrelevance into Less-Is-More Drastically Improves Reasoning Performance and Efficiency
• Pruning Long Chain-of-Thought of Large Reasoning Models via Small-Scale Preference Optimization
• MEML-GRPO: Heterogeneous Multi-Expert Mutual Learning for RLVR Advancement
• RCP-Merging: Merging Long Chain-of-Thought Models with Domain-Specific Models by Considering Reasoning Capability as Prior
• Diversity-Enhanced Reasoning for Subjective Questions
• AgentTTS: Large Language Model Agent for Test-time Compute-optimal Scaling Strategy in Complex Tasks
• Learning to Reason Across Parallel Samples for LLM Reasoning
• ThinkingViT: Matryoshka Thinking Vision Transformer for Elastic Inference
• Towards Concise and Adaptive Thinking in Large Reasoning Models: A Survey
• OpenCodeReasoning-II: A Simple Test Time Scaling Approach via Self-Critique
• Knowledge Fusion via Bidirectional Information Aggregation
• Adaptive Termination for Multi-round Parallel Reasoning: An Universal Semantic Entropy-Guided Framework
• Reasoning or Not? A Comprehensive Evaluation of Reasoning LLMs for Dialogue Summarization
• Test-Time Scaling with Reflective Generative Model
• On Reasoning Strength Planning in Large Reasoning Models
• Prover Agent: An Agent-Based Framework for Formal Mathematical Proofs
• Rethinking Test-Time Scaling for Medical AI: Model and Task-Aware Strategies for LLMs and VLMs
• The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning
• AlphaOne: Reasoning Models Thinking Slow and Fast at Test Time
• Can Slow-thinking LLMs Reason Over Time? Empirical Studies in Time Series Forecasting
• Knowledge Augmented Complex Problem Solving with Large Language Models: A Survey
• Rethinking the Unsolvable: When In-Context Search Meets Test-Time Scaling
• T²: An Adaptive Test-Time Scaling Strategy for Contextual Question Answering
• When to Continue Thinking: Adaptive Thinking Mode Switching for Efficient Reasoning
• RoT: Enhancing Table Reasoning with Iterative Row-Wise Traversals
• DRP: Distilled Reasoning Pruning with Skill-aware Step Decomposition for Efficient Large Reasoning Models
• Detection and Mitigation of Hallucination in Large Reasoning Models: A Mechanistic Perspective
• SelfBudgeter: Adaptive Token Allocation for Efficient LLM Reasoning
• Uncertainty Profiles for LLMs: Uncertainty Source Decomposition and Adaptive Model-Metric Selection
• Learning from Peers in Reasoning Models
• When Reasoning Beats Scale: A 1.5B Reasoning Model Outranks 13B LLMs as Discriminator
• Internalizing LLM Reasoning via Discovery and Replay of Latent Actions
• Retro-Search: Exploring Untaken Paths for Deeper and Efficient Reasoning
• Compile Scene Graphs with Reinforcement Learning
• BenHalluEval: A Multi-Task Hallucination Evaluation Framework for Large Language Models on Bengali
• Efficient Reasoning Models: A Survey
• Heimdall: test-time scaling on the generative verification

Discussions

• https://arxiv.org/abs/2502.12215 #cs.LG #cs.AI Event Attributes [bsky, 0 points, 0 comments]

Related

• s1: Simple test-time scaling
• Are We Scaling the Right Thing? A System Perspective on Test-Time Scaling
• The Art of Scaling Test-Time Compute for Large Language Models
• Scaling over Scaling: Exploring Test-Time Scaling Plateau in Large Reasoning Models
• Can 1B LLM Surpass 405B LLM? Rethinking Compute-Optimal Test-Time Scaling