Inference-Time Intervention

Follow this topic Ask about this topic

Inference-time intervention studies how to steer a trained model while it is running, without necessarily changing its weights. Methods include activation and attention steering, decoding-time logit adjustments, retrieval-guided revision, context filtering, and reviewer agents that check tool calls or multi-step actions. They are applied to language and multimodal models, code generation, forecasting, theorem proving, and autonomous agents, with evaluations spanning safety, reasoning, factuality, and task success. A central challenge is improving the target behavior without harming useful outputs; intervention timing, persistence over long rollouts, robustness to new attacks, and reliable measurement remain open issues.

Latest papers

33 papers2 papers in the last four weeks, down 71% on the four weeks before. 0.0% of all new papers.

When Should We Intervene at Test Time?
Oct 27, 2026Jiahao Zhang, Xinling Wen, Liangliang Li +1Inference-Time Intervention Test-Time Adaptation
Zhengzhou University of Aeronautics · Xi’an University of Technology
StepSafe: Learning Timely Intervention with Context-Aware Safety Reasoning
Oct 14, 2026Siyi Gu, Kunhan Gao, Chao Peter Yang +3AI Agent Safety LLM Agent Safety
Yale University · NetApp Inc.
RolloutFaith: Auditing Persistent Internal Interventions in Visual World Model
Sep 29, 2026Junchi Yao, Ziyi Wang, Youling Huang +1Inference-Time Intervention World Model Evaluation
UESTC · Dalian University of Technology · MBZUAI
PASTABench: Proactive Assessment of Sequential Trajectories for Agent Safety
Sep 23, 2026Jiapeng Sun, Yujin Zhou, Han Zhu +4AI Agent Safety Benchmarks Inference-Time Intervention
The Hong Kong University of Science and Technology, Hong Kong, China · Peking University, Beijing, China
First Token Matters: Understanding Safety Collapse in Large Reasoning Models
Sep 16, 2026Yizheng Yang, Haining Yu, Yuechen Wang +4LLM Safety Alignment Inference-Time Intervention
Harbin Institute of Technology, Harbin, China · Guangzhou University, Guangzhou, China · City University of Macau, Macau, China
Beyond Training: A Feasibility Taxonomy for Inference-Time AI Governance
Sep 9, 2026Samar AnsariAI Governance Inference-Time Intervention
School of Computing and Engineering Sciences University of Chester Chester, CH1 4BJ, United Kingdom
Mitigating LLM Over-Refusal via Dynamic Semantic Routing Calibration
Sep 6, 2026Zixuan Wang, Bingjie Zhang, He Zhao +1Mechanistic Interpretability LLM Refusal Behavior
School of Artificial Intelligence, Jilin University · CSIRO, Australia · Center of Excellence for Generative AI, King Abdullah University of Science and Technology
Characterizing a Configuration Where Inference-Time PRM-Pruned Fragment Grafting Is Inert: Evidence from Three Reasoning LMs
Sep 3, 2026Khawaja Murad ul Hassan, Mehran EbrahimiInference-Time Intervention LLM Reasoning
QLU.ai, Islamabad · Faculty of Science, Ontario Tech University Oshawa, ON, Canada
ObserverBench: Testing Mechanistic Estimates for Intervention and Control
Sep 2, 2026Vijay ErramilliMechanistic Interpretability AI Safety Evaluation
OBJECTION! Lawyer Agents Mitigate Guilty Bias in Legal Judgment Prediction
Sep 2, 2026Jaehoon Jeong, Jay-Yoon LeeLegal Judgment Prediction Legal Reasoning
Seoul National University, Seoul, South Korea · Korean National Police Agency, Seoul, South Korea
Moving the Mean Toward the Known Good, Not Beyond It: What Inference-Time Interventions and Weight Consolidation Buy in Open-Ended Generation
Aug 28, 2026Roberto I. Ono FilhoInference-Time Optimization Self-Training
Independent researcher
Low-Rank Dynamics-Effective Latent Carriers for Counterfactual Rollout in Learned World Models
Aug 15, 2026Yang Liu, Yuming ChenWorld Model Learning Latent World Models
College of Intelligent Robotics and Advanced Manufacturing, Fudan University Shanghai, China
Learned Interventions in Lean 4 grind
Jul 25, 2026Evan Wang, Simon Chess, Sophie Szeto +1Automated Theorem Proving Lean Theorem Proving
Math AI Lab, University of Washington, Seattle, United States · Department of Computer Science and Engineering, University of Washington, Seattle, United States · Department of Mathematics, University of Washington, Seattle, United States
Uni-AdaVD: Universal Concept Erasure for Visual Generation via Orthogonal Value Decomposition
Jul 16, 2026Qifan Zhou, Yuan Wang, Yanbin Hao +4Diffusion Model Unlearning Inference-Time Intervention
School of Computer Science and Information Engineering, Hefei University of Technology, Hefei, China · School of Cyber Science and Technology, University of Science and Technology of China, Hefei, China · School of Artificial Intelligence and Data Science, University of Science and Technology of China, Hefei, China +1
HyperSafe: Inference-Time Safety Recovery for Fine-Tuned Language Models
Jul 13, 2026Aznaur Aliev, Carlos Hinojosa, Abdelrahman Eldesokey +3LLM Fine-Tuning LLM Safety Alignment
King Abdullah University of Science and Technology, Saudi Arabia
ScaleErasure: Inference-Time Minimal Intervention for Precise Concept Erasure in Next-Scale Autoregressive Image Generation
Jun 28, 2026Cong Wang, Haiyu Wu, Zhiwei Jiang +4Image Generation Safe T2I Generation
State Key Laboratory for Novel Software Technology, Nanjing University · National University of Singapore
DeSRPA: Decoupled Speech Role-Playing Agent via Inference-Time Intervention
Jun 16, 2026Wenqiu Tang, Zhen Wan, Takahiro Komamizu +1Large Language Model-Based Role-Play Simulation Spoken Dialogue Systems
Nagoya University, Nagoya, Aichi, Japan · National Institute of Informatics, Tokyo, Japan
Inference-Time Decision Calibration for Temporal Classification
Jun 14, 2026Arthur Chagas, Arthur Buzelin, Yan Aquino +4Post-Hoc Calibration Time Series Classification
Department of Computer Science (DCC), Universidade Federal de Minas Gerais (UFMG)
GITCO: Gated Inference-Time Context Optimization in TSFMs
Jun 3, 2026Manya Pandey, Dhruv Kumar, Murari Mandal +1Zero-Shot Time Series Forecasting Inference-Time Optimization
Birla AI Labs, Mumbai, India · KIIT, Bhubaneswar, India · BITS Pilani, Pilani, India
The Saturation Trap and the Subjectivity of Intervention Timing: Why Affect-Based Triggers and LLM Judges Fail to Time Interventions on Autonomous Agents
Jun 2, 2026Manvendra ModgilLLM-as-a-Judge Human-in-the-Loop Evaluation
Resonant Context Anchoring: Decoupling Attention Routing and Signal Gain at Inference Time
Jun 1, 2026Mingkuan Zhao, Yide Gao, Wentao Hu +6Contextual Faithfulness Hallucination in Language Models
Xi’an Jiaotong University · University of Science and Technology of China · Tongji University +1
ActTraitBench: Quantifying the Knowledge-Decision Gap in Large Language Models via Human-Grounded Behavioral Validation
May 28, 2026Yutong Yang, Chenxi Miao, Weikang Li +1LLM Evaluation Personality Modeling in Language Models
Peking University · Baidu Inc
Senses Wide Shut: A Representation-Action Gap in Omnimodal LLMs
May 13, 2026Trung Nguyen Quang, Yiming Gao, Fanyi Pu +3Multimodal Grounding Multimodal Large Language Models
Nanyang Technological University · LMMs-Lab Team · Johns Hopkins University
SafeSteer: A Decoding-level Defense Mechanism for Multimodal Large Language Models
May 12, 2026Xinyi Zeng, Xue Yang, Jingyuan Zhang +5Adversarial Attacks on VLMs Multimodal Robustness
Tsinghua University, Beijing, China · Shanghai Jiao Tong University · Kuaishou Technology, Beijing, China +4
Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents
Apr 29, 2026Anh Ta, Junjie Zhu, Shahin ShayandehMulti-Agent LLM Systems Tool-Use Evaluation
Apple
Defusing the Trigger: Plug-and-Play Defense for Backdoored LLMs via Tail-Risk Intrinsic Geometric Smoothing
Apr 27, 2026Kaisheng Fan, Weizhe Zhang, Yishu Gao +2LLM Backdoor Attacks Backdoor Defense in LLMs
School of Cyber Science and Technology Harbin Institute of Technology Harbin, China · Department of New Networks Peng Cheng Laboratory Shenzhen, China · SnT University of Luxembourg Luxembourg City, Luxembourg
SPREG: Structured Plan Repair with Entropy-Guided Test-Time Intervention for Large Language Model Reasoning
Apr 20, 2026Xuan Wang, Yu Ming, Xinhao Zhong +4LLM Self-Correction Classifier-Free Guidance
1Meituan
Reasoning Fails Where Step Flow Breaks
Apr 8, 2026Xiaoyu Xu, Yulan Pan, Xiaosong Yuan +4LLM Reasoning Large Reasoning Models
Shanghai Jiao Tong University · Fuzhou University · Jilin University
Altered Thoughts, Altered Actions: Reasoning Chain as Control Surface for a Vision-Language-Action Policy
Mar 13, 2026Tuan Duong Trinh, Basim Azam, Mohammed Ishaq Ansari +2Language-Conditioned Robot Manipulation Vision-Language-Action Models
University of Melbourne Parkville, VIC, Australia
Inference-Time Safety For Code LLMs Via Retrieval-Augmented Revision
Mar 2, 2026Manisha Mukherjee, Vincent J. HellendoornLLM Safety Secure Code Generation
Carnegie Mellon University, Pittsburgh, PA, USA

Oct 27, 2026J/K move · Enter open · S save

When Should We Intervene at Test Time?

Jiahao Zhang, Xinling Wen, Liangliang Li, Yu Chen

Zhengzhou University of Aeronautics · Xi’an University of Technology

Inference-Time Intervention Test-Time Adaptation Active Learning

Open paper AskPDF arXiv

Test-time adaptation helps a deployed model respond to changing inputs, but the same intervention can rescue one batch and damage the next. Updating normalization statistics, for example, can improve corrupted images yet fail when a batch contains mostly one class. Average accuracy hides these failures and cannot tell a model when to use an intervention. Target labels can guide this decision, but they are scarce and some choices are already clear. We introduce PORTAL, a controller that chooses when to deploy an intervention, which inputs to label, and when to stop querying. Each available prediction rule is a candidate, and its predictions are fixed before current feedback arrives. Every acquired label therefore evaluates all candidates on the same input. PORTAL combines these shared comparisons and directs queries toward candidates that can still overtake the current leader. It stops when the remaining labels cannot change the choice or when the possible accuracy gain no longer justifies another query. Our analysis links the required feedback to accuracy gaps and shared mistakes, explaining why some choices need much less evidence. Experiments show that this control turns an intervention that loses accuracy overall into a useful choice on selected batches. It prevents severe failures caused by changes in arrival order and preserves gains across complete corruption streams. When a fixed rule already works well, PORTAL uses little feedback to retain its performance. Periodic label checks reveal when past evidence needs refreshing and help the controller recover as conditions change. The acquired labels also train stronger candidates for future batches, and selection brings nearly all of their accuracy gains into deployment. The same feedback therefore improves both the choice made now and the predictions available to later batches.