ICML 2026 Past AgentsSafety & alignmentInterpretability
2nd Workshop on Compositional Learning: Safety, Interpretability, and Agents
CompLearn 2026
- Submission deadline
- May 7, 2026, 23:59 AoE (UTC−12) from the workshop website
- Notification
- May 22, 2026
- Submission portal
- OpenReview
- Notes
- Deadline added from the workshop website. Auto-imported from the OpenReview venue record on 2026-06-10 — please verify and enrich (topics are keyword-guessed).
Accepted papers (135)
Fetched from OpenReview (v2) on 2026-06-10.
-
A Compositional Calculus for Semantic Synergy in Language Model Embeddings
-
A mathematical theory of balancing relational generalization and memorization
-
A Theory of Atomic Features and Four Testable Predictions
-
Actionable Interpretability Must Be Defined in Terms of Symmetries: A Compositional Probabilistic Approach
-
Adaptive Minds: Empowering Agents with LoRA-as-Tools
-
Adaptive Recurrence as Algorithmic Time for Length Generalization in Addition
-
Additive Relational Bindings in Transformers: What Sparse Autoencoders Miss
-
Ask, Don’t Judge: Binary Questions for Interpretable LLM Evaluation and Self-Improvement
-
Atomic Chess Reveals Compositional Reasoning Failures in LLMs
-
Attractor Inversion: A Geometric Account of Adversarial Manipulation in Human Decision-Making
-
Beyond Safe Data: Pretraining-Stage Alignment with Regular Safety Reflection
-
Biregular Sparse Initialization Shifts the Rate and Shape of Compositional Escape in Sequential Arithmetic Curricula
-
CaMeLs Can Use Computers Too: System-level Security for Computer Use Agents
-
Causal Cartographer: From Mapping to Reasoning Over Counterfactual Worlds
-
Causal-JEPA: Learning World Models through Object-Level Latent Masking
-
CB-Orchestrator: Adaptive Workflow Optimization for LLM Agents via Contextual Bandits
-
Chain-of-Thought Gradient Descent
-
Circuit Modularity Predicts Compositional Generalization: Theory and Evidence from Transformers
-
Circuit Oracle: Automating Attribution Graph Analysis via Natural-Language Queries
-
ClinSeekAgent: Automating Multi-modal Evidence Seeking for Agentic Clinical Reasoning
-
CLIP Models Generalize Less Than Compositional Benchmarks Suggest
-
CMAG: Concept-Scaffolded Retrieval for Marketplace Avatar Generation
-
Code-enabled language models can outperform reasoning models on diverse tasks
-
COGITAO: A Procedural and Object-Centric Framework to Evaluate Compositional and Systematic Generalization
-
CompFlow: Composing Velocity Fields for Multi-Condition Generation
-
Compositional Adversarial Training for Robust Visual Watermarking
-
Compositional Agentic Formulation Search for Open-Vocabulary Audio-Visual Event Localization
-
Compositional by Design: Background-Invariant Representations via Linear Additivity in VLMs
-
Compositional Consistency-Guided Decoding for Three-Way Logical Question Answering
-
Compositional Evolutionary Probing of LLM Safety Alignment
-
Compositional Failure in Audio-Visual LLMs: Late-Layer Prior Dominance Under Cross-modal Conflict
-
Compositional Investigation: Why Reasoning Enables Tool-Using Agents to Fix What They Diagnose
-
Compositional Neuro-Symbolic Reasoning
-
Compositional Self-Improvement
-
Compositional Skill Acquisition in Agentic Pipelines via Reinforcement Learning and Knowledge Distillation
-
Compositional Skill Chaining and Policy Blending for Hard Exploration in the BRIO Labyrinth Game
-
Compositional Skill Execution in LLM Multi-Agent Systems: A Comparative Study of Collaboration Architectures for Long-Horizon Tasks
-
Compositional Underdetermination in AI Agents: When Behavioral Success Is Not Compositional Evidence
-
Concepts in Motion: Temporal Concept Bottleneck Model for Interpretable Video Classification
-
Count Me If You Can: Geometric Failure Modes in Language Model Counting
-
CUA-Skill: Developing Computer Using Agents with a Skill Framework
-
Dimensionality Controls When Modularity Helps in Continual Learning
-
Direction-Conditioned Policies via Compositional Subgoal Scoring for Online Goal-Conditioned Reinforcement Learning
-
Dissociating Decodability and Causal Use in Bracket-Sequence Transformers
-
Do Thinking Tokens Help with Safety?
-
Don't Trust Stubborn Neighbors: A Security Framework for Agentic Networks
-
DPMI: A Principled Index for Neural Polysemanticity via Dirichlet Process Mixture Modeling
-
Dual-Resolution Recursive Energy: Certified Contract–Expand Inference for Sequential Decision Making
-
Emergent Compositional Skills in Mixture-of-Experts VLAs
-
Emergent Social Intelligence Risks in Generative Multi-Agent Systems
-
Entropy-Aware GUI Grounding: From Failure Analysis to Improved Localization
-
Evolution of Cooperation in LLM Societies : A Multi-Lingual Examination
-
Evolutionary System Prompt Learning for Reinforcement Learning in LLMs
-
Explaining is Harder Than Predicting Alone: Evaluating Concept-based Explanations of MLLMs as ICL Visual Classifiers
-
Fixed-Point Reasoning: Stable and Adaptive Deep Looped Models
-
FormalImG: Evaluating Structural Compositional Generalization for T2I Models
-
From Composition to Compositionality: Discovering Reusable Structure in Polyphonic Music Embeddings
-
From Mechanistic to Compositional Interpretability
-
From Numbers to Narratives: Goal-Oriented Summarization of Machine Learning Model Differences
-
From Self-Preservation to Peer-Preservation: A Staged Framing of Preservation-Oriented Misalignment in Frontier Models
-
Fusion is the New Mutation: Bandit-Guided Evolution on Workflow Graphs
-
Gating Enables Curvature: A Geometric Expressivity Gap in Attention
-
Grad Detect: Gradient-Based Hallucination Detection in LLMs
-
Hidden in Plain Sight: Benchmarking Agent Safety Against Decomposition Attacks with DeCompBench
-
HINT: Task Demonstrations for Hierarchical Inference in Abstract Reasoning
-
How does RL Post-training Induce Skill Composition? A Case Study on Countdown
-
How Many Features Can a Language Model Store Under the Linear Representation Hypothesis?
-
IGG: A Benchmark for Interactive GUI Grounding under Visibility Constraints
-
Improving the Compositionality of Triplet-Based Neural Algorithmic Reasoners
-
In-Context Learning Amplifies a Latent Compositional Circuit
-
Installing and Obstructing Heuristics: Learning Dynamics in Nim
-
Introspective Coupling: LMs Explain Themselves Better Than Training Targets
-
Irreducible Supervision Enables Compositional Generalization in Post-Training
-
Language Elicits Emergent Symbol Processing in Vision Foundation Models
-
Large Language Models Can Follow Instructions, But Not Many at Once: Phase Transitions in Compositional Constraint Satisfaction
-
Learning Compositional Tasks via Trigger Compositions: Using Scratchpads as Pre-Answer Workspaces
-
Learning to Theorize the World from Observation
-
Learning What’s Missing: Failure-Driven Skill Discovery via Predicate Bridges
-
LGPro: Language-Guided Prototype Discovery for Compositional Zero-Shot Learning
-
Logit Grafting: The Post-Training Delta is Sparse, Portable, and Powerful
-
MAVEN: Improving Generalization in Agentic Tool Calling
-
Meaning Representations as Variational Quantum Circuits
-
Measuring the Limits of Continual Learning for LLMs
-
Mitigating Over-Personalization in Language Models via Structured Memory
-
MKEvolve: A Modular Multi-Agent Framework for Kernel Code Generation
-
MoTVLA: A Vision-Language-Action Model with Unified Fast-Slow Reasoning
-
Multi-Agent Systems are Mixtures of Experts: Who Becomes an Influencer?
-
MultiVulnBench: A Large-Scale Benchmark for Count Bias in LLM-Based Multi-Vulnerability Detection
-
Noise-Tolerant Verification of Compositional Boolean Recovery
-
Not Just RLHF: Why Alignment Alone Won't Fix Multi-Agent Sycophancy
-
Nouns, Not Modifiers: OpenVLA Parses Objects but Fails at Spatial Composition
-
On the Role of Learned Alignment Matrices in LatentMAS
-
Operads for compositional reasoning in LLMs
-
Playing Devil’s Advocate: Off-the-Shelf Persona Vectors Rival Targeted Steering for Sycophancy
-
Policy Transfer for Hierarchical Goal-Conditioned Reinforcement Learning
-
Preference Instability in Reward Models: Detection and Mitigation via Sparse Autoencoders
-
Reasoning as State Transition: A Representational Analysis of Reasoning Evolution in Large Language Models
-
Reasoning Phases Are Continuous, Not Discrete: Evidence from Switching Linear Dynamical Systems Applied to Chain-of-Thought Residual Streams
-
Reasoning with Neologisms: Can Soft Tokens Learn Composable Reasoning Skills Without Forgetting?
-
Reflection Anchors for Interpretable Compositional Visual Reasoning in Multimodal Reinforcement Learning
-
Retrieval is Enough: Training-Free Interpretability with a Tool-Using Agent
-
RL Post-Training Builds Compositional Reasoning Strategies
-
Safety Cost of Steering Vectors Is Separable and Reducible
-
Sample Complexity of Scientific Discovery: PAC Learnability of Compositional Function Trees
-
Separable Representations of Task Complexity and Deliberation in Reasoning Language Models.
-
Sparse Autoencoders Find Causal, Lineage-Specific Context Features in Chromatin Foundation Models
-
Sparse Memory Finetuning as a Low-Forgetting Alternative to LoRA and Full Finetuning
-
Spatial Compositional Counterfactuals in Concept Bottleneck Models
-
Spatially Stable GUI Grounding via Zoom Consistency Loss
-
Stop Probing, Start Coding: Why Linear Probes and Sparse Autoencoders Fail at Compositional Generalisation
-
Struct-to-Reason: Enhancing Video Understanding of Vision-Language Models by Decoupling Perception and Reasoning via Structured Summary
-
Structure over Pixels: Learning Variable-Length Visual Programs
-
Successor Re-grounding Audits Compositional Rollout Mismatch in Neuro-Symbolic Search
-
TAME the BALROG: Task-Adaptive Modular Evolution framework for Game Agents
-
The Compositional Generalization Gap in Named Entity Recognition: Static Benchmarks Overestimate Transferable Performance
-
The Spurious Composition Problem: Conditional Independence as a Necessary and Sufficient Condition for Systematic Generalization
-
The Theory and Practice of MAP Inference over Non-Convex Constraints
-
THEIA: Learning Complete Kleene Three-Valued Logic in a Pure-Neural Modular Architecture
-
Toward Compositional Latent Action Interfaces for Generalizable Agents
-
Tracking Training Phases in Compositional Learning with Task-Agnostic Measures
-
Universality, Composition Generalization, and Algorithm Emulation All In-Context
-
Unsafe Only in Combination: Interaction-Barrier Shielding for Tool-Using LLM Agents
-
Unsupervised Decomposition with Recombination-Consistent Diffusion Models
-
VASAE: Naming SAE Dictionary Directions with Vocabulary-Aligned Anchoring
-
Visual Counterfactual Explanations with Compositional Generative Models
-
What Do Latent Agents Actually Represent? Interpreting Hidden-State Communication in Multi-Agent Systems
-
What makes the whole? Probing Attribute-Level Compositionality in LLM Judges
-
When Do Diffusion Models learn to Generate Multiple Objects?
-
When Do Multi-Agent Systems Outperform? Analysing the Learning Efficiency of Agentic Systems
-
When Does Composition Compose? A PAC-Theoretic Framework for Compositional Faithfulness, Safety Certificates, and Training Dynamics
-
When Does Disentanglement Enable Compositional Generalization? A Transfer Bound and Its Empirical Validation
-
When to Re-Plan: Subgoal Persistence in Hierarchical Latent Reasoning
-
Where’s the Plan? Locating Latent Planning in Language Models with Lightweight Mechanistic Interventions
-
Which Way Did It Move? Diagnosing and Overcoming Directional Motion Blindness in Video LLMs
-
Your Agent, Their Asset: A Real-World Safety Analysis of OpenClaw