My ICML 2026 Reading List

ICML 2026 accepted 6,627 papers. Nobody is reading all of them, so I filtered the list down to what’s relevant to my group’s research: reinforcement learning, robot learning, imitation learning, vision-language-action (VLA) models, LLM RL post-training, diffusion for RL, world models, and generalization. I’m sharing the filtered list here in case it’s useful to others heading to Seoul.

Papers are grouped into P0 (must read), P1 (should read), and P2 (nice to read). Where a paper already has an arXiv preprint (✅) I’ve linked it directly; otherwise use the OpenReview link for the PDF.

P0 — Must Read

VLA Models

  • NeurVLA: Unleashing Failure-Handling Capability of VLMs via Neural-Symbolic Reasoning — OpenReview
  • StableVLA: Towards Robust VLMs without Extra Data — OpenReview
  • XR-1: Versatile VLMs via Unified Vision-Motion Representations — OpenReview
  • Any3D-VLA: Enhancing VLA Robustness via Diverse Point Clouds — OpenReview
  • LARA: Latent Action Representation Alignment for VLMs — OpenReview
  • From Pixels to Tokens: Systematic Study of Latent Action Supervision for VLMs — OpenReview
  • Can VLMs Diagnose and Recover from VLA Manipulation Faults? — OpenReview
  • ✅ LaST₀: Latent Spatio-Temporal Chain-of-Thought for Robotic VLA (Spotlight) — arXiv:2601.05248
  • ✅ Contrastive Representation Regularization for VLA Models — arXiv:2510.01711

Robot Learning & Imitation

  • ✅ Posterior Behavioral Cloning: Pretraining BC for Efficient RL Fine-tuning — arXiv:2512.16911
  • ✅ NavOL: Navigation Policy with Online Imitation Learning — arXiv:2605.11762
  • ✅ Noise-Guided Transport: Imitation Learning from Random Priors — arXiv:2509.26294
  • Provably Efficient Policy-Reward Co-Pretraining for Adversarial Imitation Learning — OpenReview
  • Towards Practical World Model-based RL for VLA Models — OpenReview
  • ✅ Temporal Difference Calibration for VLA Models — arXiv:2604.20472
  • Learning Generalizable Skill Policy with Data-Efficient Unsupervised RL — OpenReview

Diffusion for RL

  • Energy-based Compositional Diffusion Planning — OpenReview
  • ✅ Improving Diffusion Planners by Self-Supervised Action Gating with Energies — arXiv:2603.02650
  • Latent Diffusion Controller: Framework, Algorithms and Parameterization — OpenReview
  • ✅ Sample from What You See: Visuomotor Policy Learning via Diffusion Bridge — arXiv:2512.07212
  • ✅ SVL: Goal-Conditioned RL as Survival Learning — arXiv:2604.17551

LLM RL Post-Training: RLVR / GRPO vs PPO

  • ✅ Advantage Collapse in Group Relative Policy Optimization: Diagnosis and Mitigation — arXiv:2605.21125
  • ✅ The Unlearnability Phenomenon in RLVR for Language Models — arXiv:2605.16787
  • ✅ The Easy, the Hard, and the Learnable: Confidence and Difficulty-Adaptive Policy Optimization (CoDaPO) for LLM Reasoning — arXiv:2606.07950
  • ✅ The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes (Honorable Mention, also an Oral) — arXiv:2602.15515
  • Learning Useful Supervision for Reinforcement Learning in Reasoning Models — OpenReview
  • Rubric Curriculum RL: Exploiting the Generation-Verification Gap in Creative Writing — OpenReview

RL Theory & Policy Optimization

  • ✅ Chain-of-Goals Hierarchical Policy for Long-Horizon Offline Goal-Conditioned RL — arXiv:2602.03389
  • ✅ Why Linear Recurrent Memory Works in Partially Observable RL — arXiv:2605.31261
  • Towards Optimal Strong Regret and Constraint Violation via Model-free RL — OpenReview
  • ✅ Learning to Perceive the World Through Control: Empowerment-Based Representation — arXiv:2605.30656
  • ✅ Reparameterization Flow Policy Optimization — arXiv:2602.03501
  • Reparameterization PPO — OpenReview
  • ✅ Towards Parameter-Free Temporal Difference Learning — arXiv:2603.02577
  • ✅ Dynamic Programming for Epistemic Uncertainty in MDPs (Spotlight) — arXiv:2602.03381
  • ✅ What Does Flow-Matching Bring to TD-Learning? — arXiv:2603.04333

P1 — Should Read

World Models

  • Structured 4D Latent World Model for Robot Planning — OpenReview
  • Learning Latent Action World Models In The Wild — OpenReview
  • VectorWorld: Efficient Streaming World Model via Diffusion Flow — OpenReview
  • Convergent World Representations and Divergent Tasks — OpenReview
  • ✅ dWorldEval: Scaling Robot Policy Evaluation via Discrete Diffusion World Model (Spotlight) — arXiv:2604.22152

Offline RL

  • ✅ BiTrajDiff: Bidirectional Trajectory Generation with Diffusion for Offline RL — arXiv:2506.05762
  • VIPO: Value Function Inconsistency Penalized Offline RL — OpenReview
  • Fast Policy Learning for Offline RL via Bootstrapped Flow Q-Learning — OpenReview
  • Offline RL with Generative Trajectory Policies — OpenReview
  • Offline RL with Universal Horizon Models — OpenReview
  • Compositional Transduction for Offline Goal-Conditioned RL — OpenReview
  • Reward-Preserving Counterfactual State Editing for Offline RL — OpenReview
  • SMAC: Score-Matched Actor-Critics for Offline-to-Online Transfer — OpenReview

Diffusion RL

  • Distillation Models are Good Samplers for Diffusion RL — OpenReview
  • Advantage Weighted Matching: Aligning RL with Diffusion Pretraining — OpenReview
  • Reverse Flow Matching: Unified Framework for Online RL with Diffusion Policies — OpenReview
  • Trust-Region Diffusion Policies for Massively Parallel On-Policy RL — OpenReview

Safe/Robust RL

  • Distributionally Robust RL with Human Feedback — OpenReview
  • CSPO: Constraint-Sensitive Policy Optimization for Safe RL — OpenReview
  • Robust RL in a Sample-Efficient Setting (TMLR) — poster
  • Model-Free Robust Average-Reward RL with Sample Complexity — OpenReview
  • Mirror Descent Policy Optimisation for Robust Constrained MDPs (TMLR) — poster

Continual/Meta/Transfer Learning

  • SABER: Continual Learning with Representation Conflict Management — OpenReview
  • Counterfactual Bootstrap for Robust Meta-RL — OpenReview
  • Meta-learning Structure-Preserving Dynamics — OpenReview
  • Dynamics Are Learned, Not Told: Zero-Shot Policy Adaptation — OpenReview
  • Motion Dynamics Learning for Few-Shot Embodied Adaptation — OpenReview
  • ✅ Hierarchical Successor Representation for Robust Transfer (Spotlight) — arXiv:2602.12753

Multi-Agent/Hierarchical RL

  • Hierarchical Policy Learning via Spectral Decomposition — OpenReview
  • Recurrent Structural Policy Gradient for Partially Observable Mean Field Games — OpenReview
  • Offline Multi-agent Continual Cooperation via Skill Partition and Reuse — OpenReview
  • Provably Convergent Actor-Critic in Risk-averse MARL — OpenReview
  • HyPOLE: Hyperproperty-Guided MARL under Partial Observation — OpenReview
  • ✅ Factored Value Functions for Graph-Based Multi-Agent RL — arXiv:2601.11401
  • ✅ TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination — arXiv:2605.15207

Additional

  • Geometry-Preserving Orthonormal Initialization for Low-Rank Adaptation in RL — OpenReview
  • Motion Attribution for Video Generation — OpenReview
  • Debate2Create: Robot Co-design via Multi-Agent LLM Debate — OpenReview
  • GAE: Unleashing Physical Potential of VLM with Generalizable Action Expert — OpenReview
  • EAPO: Enhancing Policy Optimization with On-Demand Expert Assistance — OpenReview
  • ✅ CaP-X: Benchmarking & Improving Coding Agents for Robot Manipulation — arXiv:2603.22435
  • ✅ Demystifying Action Space Design for Robotic Manipulation Policies — arXiv:2602.23408
  • Clipping Low-Probability Tokens in SFT Yields a Generalizable Initialization for RL — poster
  • ✅ ToM-SWE: User Mental Modeling for Software Engineering Agents — arXiv:2510.21903

P2 — Nice to Read

Flow Matching & Diffusion

  • Flow for Future: SE(3)-Equivariant Flow Matching for 3D Trajectory Prediction — OpenReview
  • Diffusion Bridge or Flow Matching? A Unifying Framework — OpenReview
  • Scaling the Prior: Size-Consistent Geometric Diffusion for 3D Molecules — OpenReview
  • Efficient, Property-Aligned Fan-Out Retrieval via RL-Amortized Diffusion — OpenReview
  • Well-Posed KL-Regularized Control via Wasserstein Divergences — OpenReview

Representation / Efficient Fine-tuning

  • What Does Preference Learning Recover from Pairwise Comparison Data? — OpenReview
  • Understanding LoRA as Knowledge Memory — OpenReview
  • FedTreeLoRA: Federated LoRA Fine-Tuning — OpenReview
  • Spectral Bridge VI: Dynamic LoRA via Bures-Wasserstein Gradient Flows — OpenReview
  • XTransfer: Modality-Agnostic Few-Shot Model Transfer — OpenReview
  • Multi-Way Representation Alignment — OpenReview
  • Revisiting Parameter-Based Knowledge Editing in LLMs — OpenReview

Position Papers

ICML 2026 had 65 position papers this year. Six stood out as relevant to RL / robot learning / world models:

Must read:

  • Position: World Models as an Intermediary between Agents and the Real World — Sherry Yang — poster
  • Position: RL Researchers Need to Distinguish Between Solving Simulators and Using Simulators as a Proxy — Matthew Vandergrift, Esraa Elelimy, Martha White — poster

Should read:

  • Position: Good Embodied Reward Models Need Bad Behavior Data — Thomas Tian, Yilin Wu, Andrea Bajcsy — poster
  • Position: Make Planning Research Rigorous Again! — Michael Katz, Harsha Kokel, Christian Muise, Shirin Sohrabi, Sarath Sreedharan — poster

Nice to read:

  • Position: Assistive AI requires Personalized Specialists, not Generalists — Homanga Bharadhwaj — poster
  • Position: Interestingness is an Inductive Heuristic for Future Compression Progress — Vincent Herrmann, Jürgen Schmidhuber — poster

(arXiv/OpenReview links for position papers were gated behind the ICML virtual-site login at the time of writing — use the ICML poster links above for session details, or search titles on OpenReview.)

Workshops

ICML 2026 has 44 workshops. Five stood out as relevant to RL, robot learning, VLA, imitation learning, diffusion for RL, generalization, and world models:

  • Decision-Making from Offline Datasets to Online Adaptation — Jul 11, 8:00am–5:00pm KST, ASEM Ballroom 203 — Aryan Deshwal, Haruka Kiyohara, Nghia Hoang, Tang Thanh Nguyen, Willie Neiswanger, Syrine Belakaria, Jana Doppa — workshop page
  • AI as a Tool for Mathematics, Computer Science, and Machine Learning — Jul 10, 8:00am–5:00pm KST, Auditorium — Dmitriy Drusvyatskiy, Mikhail Belkin, Edgar Dobriban, Fanny Yang, Qingsong Wang — workshop page
  • Continual Adaptation at Scale: Towards Sustainable AI — Jul 10, 8:00am–5:00pm KST, Room 327 — Ghada Sokar, Gintare Karolina Dziugaite, Mohammad Emtiyaz Khan, Rupam Mahmood, Martin Mundt, Daniel Marczak — workshop page
  • Planning in The Era of Language Models (LM4Plan) — Jul 11, 8:00am–5:00pm KST, Grand Ballroom 101-102 — Michael Katz, Augusto B. Corrêa, Nir Lipovetzky, Sarath Sreedharan, Katharina Stein, Luckeciano Melo, Elliot Gestrin — workshop page
  • The Second Workshop on Agents in the Wild: Safety, Security, and Beyond — Jul 11, 8:00am–5:00pm KST, Hall B2 — Chenguang Wang, Xinyun Chen, Wenbo Guo, Yizhou Sun, Kyle Montgomery, Yiyou Sun, Jianhong Tu, Zhun Wang — workshop page

A couple of connections worth flagging: Decision-Making pairs naturally with the Offline RL / RL theory sections above; Continual Adaptation at Scale ties to the Continual/Meta/Transfer Learning section; LM4Plan connects to the “Make Planning Research Rigorous Again!” position paper. I’ll also be presenting two co-authored posters at Agents in the Wild. (See the full workshop list for everything else on offer.)

How I Built This List

I started from the full accepted-papers list (6,627 papers) and filtered by keyword/topic match against my group’s research areas — RL, robot learning, VLA, imitation learning, LLM RL post-training, diffusion for RL, generalization, and world models — then ranked by relevance into P0/P1/P2. arXiv preprints (✅) were matched where available; everything else links through OpenReview or the ICML virtual poster page. If you’re at ICML 2026 in Seoul and this overlaps with your interests, hope it saves you some filtering time.

Update (2026-07-08): Added a new P0 section on LLM RL post-training (RLVR / GRPO vs PPO) plus another ~15 papers spotted from poster-session photos on site, mostly with arXiv preprints already up.