Test-Time Scaling in Reasoning Models Is Not Effective for Knowledge-Intensive Tasks Yet

2025/09/08 by James Zhao, Bryan Hooi, Zhao, James Xu +3 · 2 citations
Computer Science · #Explainable Artificial Intelligence (XAI)

paper · pdf · doi:10.48550/arxiv.2509.06861

Abstract

Test-time scaling increases inference-time computation by allowing models to generate long reasoning chains, and has improved performance across many domains. However, in this work, we show that this approach is not yet effective for knowledge-intensive tasks. We evaluate 14 reasoning models on two knowledge-intensive benchmarks and find that increasing test-time computation does not consistently improve accuracy and often increases hallucinations. Further analysis shows that changes in hallucination rates under increased test-time computation are largely driven by models' willingness to answer. We also observe that extended reasoning can induce confirmation bias, leading to overconfident hallucinations. Finally, we provide an information-theoretic account: compute-only test-time scaling is a post-processing of a fixed trained model and therefore cannot increase information about the ground-truth answer beyond what is already encoded in the model, explaining its limited gains on knowledge-intensive tasks. Code and data are available at https://github.com/XuZhao0/tts-knowledge

Citations

• SimpleQA Verified: A Reliable Factuality Benchmark to Measure Parametric Knowledge
• Inverse Scaling in Test-Time Compute
• Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
• Are Reasoning Models More Prone to Hallucination?
• How Does Response Length Affect Long-Form Factuality
• More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models
• Scaling Reasoning, Losing Control: Evaluating Instruction Following in Large Reasoning Models
• When Thinking Fails: The Pitfalls of Reasoning for Instruction-Following in LLMs
• Qwen3 Technical Report
• HalluLens: LLM Hallucination Benchmark
• A Survey on Test-Time Scaling in Large Language Models: What, How, Where, and How Well?
• Proof or Bluff? Evaluating LLMs on 2025 USA Math Olympiad
• The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks
• Trading Inference-Time Compute for Adversarial Robustness
• s1: Simple test-time scaling
• DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning
• Measuring short-form factuality in large language models
• Fact, Fetch, and Reason: A Unified Evaluation of Retrieval-Augmented Generation
• Large Language Monkeys: Scaling Inference Compute with Repeated Sampling
• Long-form factuality in large language models
• LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code
• Factuality of Large Language Models: A Survey
• GPQA: A Graduate-Level Google-Proof Q&A Benchmark
• Factuality Challenges in the Era of Large Language Models
• Factuality challenges in the era of large language models and opportunities for fact-checking
• Siren's Song in the AI Ocean: A Survey on Hallucination in Large Language Models
• 🧜Siren’s Song in the AI Ocean: A Survey on Hallucination in Large Language Models
• Self-Consistency Improves Chain of Thought Reasoning in Language Models
• Survey of Hallucination in Natural Language Generation
• BNAI, NO-TOKEN, and MIND-UNITY: Pillars of a Systemic Revolution in Artificial Intelligence
• Confirmation Bias: A Ubiquitous Phenomenon in Many Guises

Cited by

• Asking LLMs to Verify First is Almost Free Lunch
• Searching Meta Reasoning Skeleton to Guide LLM Reasoning

Related

• Inverse Scaling in Test-Time Compute
• Does Thinking More always Help? Mirage of Test-Time Scaling in Reasoning Models
• Scaling over Scaling: Exploring Test-Time Scaling Plateau in Large Reasoning Models
• A Survey of Test-Time Compute: From Intuitive Inference to Deliberate Reasoning
• Is That Your Final Answer? Test-Time Scaling Improves Selective Question Answering