ICML 2026 재밌게 본 논문들

1. AI Safety·Privacy·Jailbreak·Alignment

1. MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models
2. MultiBreak: A Scalable and Diverse Multi-turn Jailbreak Benchmark for Evaluating LLM Safety
3. Jailbreaking Vision-Language Models Through the Visual Modality
4. Helpful to a Fault: Measuring Illicit Assistance in Multi-Turn, Multilingual LLM Agents
5. Learning Efficient Guardrails for Compliance
6. More Sail than Ballast: Addressing Harmful Knowledge Leakage in the Expansive Reasoning Space of LRMs
7. Frontier Models Can Take Actions at Low Probabilities
8. Rapid Poison: Practical Poisoning Attacks Against the Rapid Response Framework
9. Safety Alignment of LMs via Non-cooperative Games
10. Position: LLM-Safety Evaluations Lack Robustness
11. Quantifying the Salience of Geo-Cultural Values for Pluralistic Safety Alignment
12. Active Attacks: Red-teaming LLMs via Adaptive Environments
13. Jailbreak Foundry: From Papers to Runnable Attacks for Reproducible Benchmarking
14. Unsafer in Many Turns: Benchmarking and Defending Multi-Turn Safety Risks in Tool-Using Agents
15. When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems
16. AlienLM: Alienization of Language for API-Boundary Privacy in Black-Box LLMs
17. Is Vibe Coding Safe? Benchmarking Vulnerability of Agent-Generated Code in Real-World Tasks
18. RedDebate: Safer Responses Through Multi-Agent Red Teaming Debates
19. State-Dependent Safety Failures in Multi-Turn Language Model Interaction
20. It’s a TRAP! Task-Redirecting Agent Persuasion Benchmark for Web Agents
21. Are Your Agents Upward Deceivers?
22. Prompt Injection as Role Confusion
23. Reliable to Expressive: A Curriculum for Rubric-Following Safety Judges
24. Culturally-Adapted Red-Teaming Across East and Southeast Asian Contexts
25. Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use
26. OTora: A Unified Red Teaming Framework for Reasoning-Level Denial-of-Service in LLM Agents
27. Position: ‘AI Alignment’ Encompasses Competing Technical Priorities
28. From Weak Cues to Real Identities: Evaluating Inference-Driven De-Anonymization in LLM Agents

2. Interpretability·Mechanistic Analysis·CoT Faithfulness

1. Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Models
2. Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought
3. Chain-of-Thought Reasoning in the Wild Is Not Always Faithful
4. Interpretability Can Be Actionable
5. Biases in the Blind Spot: Detecting What LLMs Fail to Mention
6. Do Activation Verbalization Methods Convey Privileged Information?
7. Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units
8. Causality is Key for Interpretability Claims to Generalise

3. Hallucination·Reliability·Calibration·Citation

1. Reliable Thinking with Images
2. Building Reliable Long-Form Generation via Hallucination Rejection Sampling
3. From Out-of-Distribution Detection to Hallucination Detection: A Geometric View
4. Guidance: Sentence-Level Citation Enforcement via Prefix-Tail Guidance during LLM Decoding
5. Confidence is Not Universal: Task-Dependent Calibration and Emergent Behavior in LLMs
6. Automatic Layer Selection for Hallucination Detection
7. Hallucination Detection from Structural Reasoning Model
8. Who Can We Trust? LLM-as-a-Jury for Comparative Assessment

4. Multimodal Understanding·Generation·Evaluation

1. ADEPT: RL-Aligned Agentic Decoding of Emotion via Evidence Probing Tools — From Consensus Learning to Ambiguity-Driven Emotion Reasoning
2. Semantic Granularity Navigation in Image Editing
3. ImpText: A Benchmark and Tool-Augmented Framework for Implicit Text Reasoning
4. Unison: Benchmarking Unified Multimodal Models via Synergistic Understanding and Generation
5. PerceptionRubrics: Calibrating Multimodal Evaluation to Human Perception

5. Agents·Tool Use·Memory·Personalization·Human–AI Interaction

1. DeepImageSearch: Benchmarking Multimodal Agents for Context-Aware Image Retrieval in Visual Histories
2. Who’s in Charge? Disempowerment Patterns in Real-World LLM Usage
3. PersistBench: When Should Long-Term Memories Be Forgotten by LLMs?
4. ToMAP: Training Opponent-Aware LLM Persuaders with Theory of Mind
5. Persona2Web: Benchmarking Personalized Web Agents for Contextual Reasoning with User History
6. Beyond the Final Answer: Evaluating the Reasoning Trajectories of Tool-Augmented Agents
7. Towards a Science of AI Agent Reliability
8. AutoWebWorld: Synthesizing Infinite Verifiable Web Environments via Finite State Machines

6. Reasoning·Generalization·Metacognition·AI Governance

1. Position: Reasoning is a Learnable Rule-Based Process
2. Position: Artificial Intelligence Needs Meta Intelligence — the Case for Metacognitive AI
3. Position: AI Lock-In Is in Progress, and We Must Be Prepared — OpenReview
4. Trustworthy AI Suffers from Invariance Conflicts and Causality is The Solution
5. Can Large Language Models Generalize Procedures Across Representations?
6. ReJump: A Tree-Jump Representation for Analyzing and Improving LLM Reasoning

7. AI-Generated Content Detection·Multimedia Forensics

1. Explainable Forensics of Manipulated Segments in Untrimmed Long Videos
2. Sem-Detect: Semantic Level Detection of AI Generated Peer-Reviews
3. MusicDET: Zero-Shot AI-Generated Music Detection