Test-Time Scaling

Follow this topic Ask about this topic

Test-time scaling studies how models can use additional computation after training to solve difficult problems more accurately or reliably. Common approaches include parallel sampling and self-consistency, search and iterative refinement, verifier- or reward-guided selection, and adaptive stopping that allocates effort according to uncertainty or task difficulty. Methods are tested in settings such as mathematical reasoning, code generation, web and tool-using agents, multimodal understanding, and robotics. A central open tension is that more computation can help only when candidate selection and feedback are reliable; correlated errors, weak verifiers, overthinking, and inference cost can erase or reverse the gains.

Latest papers

232 papers32 papers in the last four weeks, up 14% on the four weeks before. 0.2% of all new papers.

Do Visual RL Gains Scale at Test Time? A Quantile View of Post-Training
Nov 2, 2026Austin Wang, Lexing YingReinforcement Learning RL Post-Training
ByteDance Seed
Test-Time Scaling for Open-Ended Reasoning at Small Model Scale
Nov 1, 2026Adam Y. Ahmed, Abdullah G. ElafifiTest-Time Scaling Output Diversity in Language Models
Rethinking LLM Pruning via Test-Time Scaling
Oct 31, 2026Mingyang Song, Mao ZhengLLM Pruning Efficient Language Model Reasoning
Foundation Model Department, Tencent, China
Distinguishing among reasoning levels of GPT 5.5 Pro via a grammar lens benchmark
Oct 31, 2026Emma Graham, Richard GrangerLanguage Model Reasoning Evaluation Test-Time Scaling
Dartmouth College Hanover, NH 03755
On the Limits of Test-Time Compute for Self-Evaluation
Oct 30, 2026Anirudh Buvanesh, Khaled Rouissi, Sujin Yun +6Language Model Self-Assessment Language Model Self-Verification
Mila · Université de Montréal · Nous Research +3
When Does Test-Time Search Help World Action Models? An Empirical Audit of Candidate Selection and Simulator Integrity
Oct 30, 2026Jun Liu, Dajie Sun, Zhenglun Kong +8Test-Time Scaling Efficient Inference for World Action Models
CMU · UC Berkeley · Harvard University +7
Agentic World Modeling for Agent Test-time Scaling without Environment Replica
Oct 29, 2026Wendi Li, Huseyin A. Inan, Wei-Ning Chen +4Action-Conditioned World Models Test-Time Scaling
University of Wisconsin–Madison · Microsoft
Blind Spots: Where Small Language Models Fall Behind Frontier Models on Terminal Tasks
Oct 28, 2026Ashima Suvarna, Ruta Desai, Ahmed ElgoharySmall Language Models LLM Evaluation
University of California, Los Angeles · Microsoft Research
Compute Has Shape: Self-Morphing the Internal Computation of Frozen LLMs
Oct 28, 2026Junxi Wu, Yiwen Guo, Shu-Tao XiaInference-Time Adaptation of Language Models Test-Time Scaling
School of Statistics and Data Science, Nankai University · Tsinghua Shenzhen International Graduate School, Tsinghua University · Independent Researcher
MLE-World: Predicting Machine Learning Experiment Outcomes with a Learned World Model
Oct 27, 2026Brian Ko, Han Wang, Hejia Geng +2Test-Time Scaling World Model Evaluation
University of Washington · University of Illinois at Urbana-Champaign · University of Oxford +2
Rethinking Verification in Agent Harnesses
Oct 25, 2026Zhengxin Zhang, Erkang Zhu, Yin Lin +3LLM Agent Verification LLM Agent Harnesses
Cornell University · Alibaba Group
When is more search useful? An Empirical Study of Search Utility in Quality-Aware Decoding for Machine Translation
Oct 23, 2026Javier Garcia Gilabert, Rico Sennrich, André F. T. MartinsMachine Translation Test-Time Scaling
Department of Computational Linguistics, University of Zurich · Instituto Superior Técnico, Universidade de Lisboa · Instituto de Telecomunicações +1
Correct but Unselectable: The Hidden Interface Tax in Multi-Candidate Reasoning
Oct 23, 2026Qianli Ma, Zhiqing Tang, Fanshuai Meng +3Efficient Language Model Reasoning Test-Time Scaling
Faculty of Arts and Science, Beijing Normal University, Zhuhai 519087, China · Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai 519087, Guangdong, China · School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China +1
Adaptive Replay Refinement: Allocating Test-Time Compute Between Revision and Resampling
Oct 20, 2026Ahsan Bilal, Muhammad Ahmed Mohsin, Muhammad Umer +3Inference-Time Compute Allocation Test-Time Scaling
University of Oklahoma · Stanford University · National University of Sciences and Technology
Foundations of the Scaling Law
Oct 20, 2026Tianyu Zhao, Zhixiang RenScaling Laws Test-Time Scaling
LingQiong AI Lab, Shanghai Smart Logic Technology
TEX: Test-Time Scaling Testing Agents via Execution-based Cross-Validation
Oct 19, 2026Akash Gokul, Srijan Bansal, Yingbo Zhou +1Test-Time Scaling Software Engineering Agents
NVIDIA · Salesforce Research · Recursive
ATEST: Adaptive Resource-Aware Test-Time Scaling for LLM Inference
Oct 18, 2026Mobina Kashaniyan, Ali JannesariTest-Time Scaling Energy-Aware LLM Inference
Department of Computer Science Iowa State University, USA
P(Yes): 15 $\times$ Improvements via Test-Time Scaling of Decision Models
Oct 18, 2026Lev StamblerLLM Decision-Making Test-Time Scaling
Tear Labs · University of Maryland, College Park
GeoART: Geometry-Aware Refinement via Test-Time Scaling for Camera-Conditioned Video Diffusion
Oct 18, 2026Liudi Yang, Ziyuan LiuCamera-Conditioned Video Generation Video Diffusion Models
University of Freiburg · Huawei Heisenberg Research Center (Munich)
ReLaDrift: Recurrent Language Generation via Drifting for Test-Time Scaling
Oct 16, 2026Yuki Otsuka, Daisuke Oba, Naoaki OkazakiTest-Time Scaling Recurrent Language Models
Institute of Science Tokyo · AIST · NII LLMC
Not All Test-Time Compute Is Evidence: Model-Specific Bias in LLM Reasoning
Oct 16, 2026Ming Shen, Tianjun YaoTest-Time Scaling Efficient Language Model Reasoning
Shanghai Jiao Tong University · Shenzhen Loop Area Institute (SLAI)
Harness-Delta Attribution: Dissecting Gains From Harness Evolution
Oct 15, 2026Wenxuan Ding, Qin Lu, Changlong Yu +5Agent Harness Optimization Agent Harness Self-Evolution
Amazon · New York University
Two Sample Sizes in Test-Time LLM Ensembling: Buffered ERM and Minimax Theory
Oct 15, 2026Zhi Zhang, Arash A. AminiLanguage Model Ensembles Test-Time Scaling
Department of Statistics and Data Science, UCLA
Beyond Correctness: Compute-Responsive Data Selection for Reasoning Model Post-Training
Oct 14, 2026Liang He, Xiangyu Li, Yuxin Wang +4Training Data Selection Language Model Post-Training
REWIND: Uncertainty-Guided Trajectory Repair for Test-Time Reasoning
Oct 14, 2026Yi Ding, Mengyu Wang, Min Zhou +1LLM Reasoning LLM Self-Correction
The Hong Kong University of Science and Technology (Guangzhou) · University of Edinburgh · Huawei
Compare to Correct: Pairwise-Guided Logit Adjustment for LLM Test-Time Scaling
Oct 13, 2026Cong Hua, Qianqian Xu, Zitai Wang +4LLM Reasoning Test-Time Scaling
State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences. · School of Computer Science and Technology, University of Chinese Academy of Sciences. · Beijing Academy of Artificial Intelligence. +1
Input-Conditioned Search for Efficient Test-Time Scaling of Recursive Reasoners
Oct 13, 2026Haochen Shi, Tianwen Fu, Qianggang Ding +7Inference-Time Search Inference-Time Scaling
Mila – Quebec AI Institute · Université de Montréal · University of Southern California +2
LoopREG: Representation Supervision Enables Test-Time Recurrent Depth in Diffusion Transformers
Oct 13, 2026Zipeng Wang, Anping Xiong, Jing Zhang +17Diffusion-Based Image Generation Diffusion Transformer
Chongqing University of Posts and Telecommunications · China United Network Communications Group Co., Ltd. · University of Electronic Science and Technology of China +1
Fewer Bits, More Tries: Recovering Capability Under Memory Constraints
Oct 11, 2026Tomer Galanti, Josh Susskind, Eran MalachTest-Time Scaling LLM Inference Acceleration
Texas A&M · Apple
Test-Time Compute for Humanoid Behavior Foundation Models
Oct 10, 2026Weiji Xie, Junsong Wu, Zibo Zhou +5Loco-Manipulation Inference-Time Optimization
Institute of Artificial Intelligence (TeleAI), China Telecom · Shanghai Jiaotong University · Zhejiang University +1

Nov 2, 2026J/K move · Enter open · S save

Do Visual RL Gains Scale at Test Time? A Quantile View of Post-Training

Austin Wang, Lexing Ying

ByteDance Seed

Reinforcement Learning RL Post-Training Test-Time Scaling RL for Image Generation Policy Optimization Best-of-N Sampling

Open paper AskPDF arXiv

Reinforcement learning (RL) post-training improves the single-sample quality of visual generative models, but deployment often provides additional inference compute through Best-of-N sampling. Whether RL gains transfer under such test-time scaling, particularly on difficult held-out tasks, remains poorly understood in the visual domain. Our central contribution is to systematically identify and characterize a failure mode of scaling transfer in visual RL: across multiple RL methods, model backbones, and evaluations spanning compositional generation, text rendering, and preference alignment, higher single-sample reward does not necessarily imply a better quality-compute frontier. RL gains may attenuate, saturate, or even reverse at large sampling budgets, with the base model eventually outperforming its post-trained counterpart. We explain these regimes through a quantile-transfer framework, showing that Best-of-N places increasing weight on the upper tail of the reward distribution as N grows. Consequently, RL can improve typical samples while compressing or degrading the rare high-quality outcomes that test-time scaling exploits. As a simple intervention motivated by this analysis, we introduce Quantile Utility Policy Optimization (QUPO), which optimizes a controllable mixture of quantile utilities corresponding to different inference budgets. Experiments on SD3.5-M and BAGEL show that QUPO better preserves upper-tail quality and substantially improves scaling on challenging held-out prompts relative to standard RL post-training. Our results suggest that visual generative models should be evaluated and optimized not only for single-sample quality, but for how post-training reshapes the full reward distribution that inference-time scaling exploits.