Recursive and
Looped Learning

Can intelligent systems reason more deeply by repeatedly refining internal representations?

Modern foundation models have made remarkable progress in language, reasoning, multimodal learning, and autonomous decision-making. Yet most Transformer architectures still rely on fixed-depth computation. Increasing model size, generating longer reasoning traces, or spending more test-time compute can improve performance, while adding substantial memory demands and inference latency.

Recursive and looped learning offers another path. By repeatedly applying shared computational modules, models can refine hidden representations, reuse parameters, and adapt their computational depth to a task. From Adaptive Computation Time and Universal Transformers to Deep Equilibrium Models and looped language models, this line of work asks how effective depth can grow without a corresponding increase in parameter count.

This workshop aims to connect theoretical foundations with practical advances in training, inference, scalability, interpretability, and safety. We welcome researchers and practitioners from machine learning, natural language processing, computer vision, reinforcement learning, systems, and AI safety.

Research Directions

We invite theoretical, algorithmic, empirical, and systems-oriented contributions that advance recursive and looped learning. We welcome mature research as well as preliminary ideas, negative results, position pieces, and work in progress.

Scope

Topics of interest include, but are not limited to:

• Recursive and looped architectures: Looped Transformers, recurrent-depth language models, weight-tied networks, Universal Transformers, Deep Equilibrium Models, and adaptive-depth architectures.
• Latent-space reasoning: Hidden-state refinement, latent chain-of-thought, implicit reasoning, and comparisons with explicit token-based reasoning.
• Adaptive computation and stopping: Dynamic recursion depth, learned halting, early exits, uncertainty-aware iteration, and task-dependent compute allocation.
• Training and optimization: Recursive pretraining, recurrent backpropagation, implicit differentiation, reinforcement learning, distillation, and curriculum learning.
• Theory and mechanistic understanding: Expressivity, computational universality, convergence, representation dynamics, algorithmic generalization, stability, and interpretability.
• Efficient systems and infrastructure: Scalable training and inference, memory optimization, quantization, parallelism, hardware acceleration, and dynamic-depth serving.
• Robustness, security, and safety: Error propagation, adversarial robustness, alignment, reliable stopping, latent reasoning verification, and secure deployment.
• Benchmarks and evaluation: Reasoning accuracy, generalization across depths, convergence, parameter efficiency, latency, compute cost, memory, robustness, and interpretability.
• Real-world applications: Language modeling, mathematical and scientific reasoning, software engineering, multimodal understanding, robotics, embodied intelligence, and autonomous agents.

Submission Guidelines

Review process

Double-blind peer review through OpenReview. Submissions must be anonymized, and conflicts of interest will be handled through recusal and reassignment.

Non-archival venue

The proposed workshop is non-archival. Accepted papers will be publicly available on OpenReview and presented as posters, with selected papers invited for oral presentations.

Page limits

References and appendices do not count toward the long-paper page limit. Reviewers are not required to read appendices. Main-text page limits will be strictly enforced.

Previously published work

Submissions already published at major machine learning venues, including ICLR, ICML, and NeurIPS, are discouraged under the proposed policy.

LLM usage

Substantial LLM contributions to research ideation or writing must be disclosed in a separate “LLM Usage” section, typically in the appendix. Authors remain responsible for all content; LLMs cannot be credited as authors.

The proposed timeline is listed below. The exact workshop day, venue, and deadline time zone will be confirmed.

The proposed full-day program combines invited perspectives, contributed research, and open discussion. Individual time slots will be posted approximately two weeks before the event.

The detailed schedule, talk titles, live-stream access, and recordings will be added when available.

The following speakers are listed as confirmed in the workshop proposal. Additional speakers and talk titles will be announced.

Diyi Yang ↗

Stanford University

Talk title to be announced

Huan Sun ↗

The Ohio State University

Talk title to be announced

Ruijie Zhu ↗

ByteDance Seed

Talk title to be announced

This first edition builds on the community developed through the Efficient Reasoning and Efficient Spatial Reasoning workshops.