Proposed workshop

ICLR 2027 Workshop on

Efficient Reasoning

Across Architectures, Learning, and Agent Systems

April 29 or 30, 2027 (assignment TBD) San Francisco, CA · Venue TBD

Hybrid · In-person sessions and online participation Contact the organizers

News

  • The proposed third edition brings together efficient reasoning across architectures, learning, and agent systems. Our invited speakers and organizing team are listed below. The assigned workshop date, submission portal, and final program will be announced here.

About

Advancing efficient reasoning across architectures, learning, and agent systems.

Making large reasoning models faster, cheaper, and more efficient.

Reasoning systems increasingly support mathematical problem solving, coding, scientific discovery, and multimodal and embodied decision-making. Their practical value depends on reasoning quality alongside the resources required to learn, execute, and verify solutions. Larger models, longer reasoning traces, and more search or test-time computation can improve capabilities while increasing latency, memory, energy, and deployment costs.

Looped Transformers and recurrent-depth models repeatedly apply shared computational blocks to refine hidden representations, allowing effective depth to vary without separate parameters for every iteration. Recent looped language models, including Ouro, explore latent reasoning and learned depth allocation. Parameter sharing, iterative refinement, and adaptive stopping are promising directions, but fewer parameters or generated tokens do not by themselves establish lower total computation or wall-clock latency.

These architectures complement efficient pretraining and post-training, reasoning-trace compression, reinforcement learning, model routing, adaptive budgets, search, and selective verification. Realizing their benefits requires systems support for dynamic computation, including scheduling, batching, memory reuse, and KV-cache management.

For agents, resource allocation extends to planning, retrieval, memory, tool use, and interaction with other agents and environments. Evaluation must cover complete execution trajectories, including failed attempts, verification, and recovery. Model-internal recurrence and system-level self-improvement are distinct mechanisms; the latter must also account for the costs of generating feedback, evaluating changes, and learning from experience.

The workshop asks how to design, train, and deploy reasoning systems that use computation effectively across internal model iterations and complete agent workflows. We will connect theory, architectures, learning algorithms, systems, agents, and applications, linking expressivity, depth generalization, and learning stability with hardware performance, task success, interpretability, and safety.

Driving Questions

How can efficient reasoning make recursive self-improvement practical?

What evidence demonstrates improvement of a system's own models, tools, or learning procedures? How should feedback, evaluation, repeated optimization, and productive research decisions be costed, and how does this process differ from recurrent hidden-state updates within a model?

Which dimensions of reasoning efficiency matter most?

Consider tokens per problem; latency; compute or monetary cost per correct answer, including failures; memory and KV cache; adaptive allocation across tokens, iterations, search, and verification; training and data efficiency; faithful and monitorable reasoning; and parameter sharing relative to recurrent computation. Which trade-offs matter in each deployment setting?

Can models learn when to think, and how should we measure it?

Should benchmarks measure accuracy per dollar or a full quality–resource Pareto curve? Which training data, process feedback, and rewards teach difficulty awareness, avoid premature stopping and unnecessary computation, and help looped models decide whether another iteration is worthwhile?

How should efficiency be balanced with transparency?

As reasoning is compressed or moves into latent space, how can faithfulness, readability, and monitorability be maintained? What information must survive in traces, model states, and agent memory, and how can errors or unsafe behavior be detected when internal reasoning is less accessible?

What does efficient reasoning look like under tight budgets?

Under strict memory, latency, and power limits, when should a system constrain thinking, distill a smaller model, route hard cases to the cloud, or adapt recurrent depth? How can quality and reliability be preserved while accounting for actual hardware runtime and communication overhead?

What representations and architectures support efficient reasoning?

When should systems use natural-language traces, recurrent or latent computation, search, or executable tools? How do these choices affect supervision, verification, and generalization? When does shared recurrent depth provide useful refinement, and when does it amplify errors or become unstable?

How can agents minimize the total cost of successful task completion?

How should reasoning, planning, memory, retrieval, and tool use be coordinated over long horizons? Evaluation should include complete trajectories, verification, communication, failures, and retries, revealing when savings in individual calls are offset by repeated gathering, ineffective actions, or recovery.

Call for Papers

The proposed Third Workshop on Efficient Reasoning at ICLR 2027 welcomes contributions on the design, training, evaluation, and deployment of efficient reasoning models and agent systems. We encourage work that connects reasoning quality with explicit resource budgets across internal model computation and complete agent workflows.

Technical Scope

  • Architectures and reasoning representations. Looped Transformers, recurrent-depth models, weight sharing, Universal Transformers, equilibrium-based models, latent reasoning, and hybrid explicit–implicit architectures; adaptive depth and sparse or conditional computation, with attention to reasoning capacity and resource requirements.
  • Theory and mechanistic understanding. Expressivity, algorithmic generalization, recurrent-depth extrapolation, fixed-point convergence, stability, and representation dynamics; relationships among parameter count, effective depth, computational cost, and reasoning quality, including the distinction between state convergence and correctness.
  • Algorithms and adaptive computation. Learned halting, early exit, uncertainty-aware iteration, budget allocation, routing, search, speculative decoding, and selective verification; resource allocation across explicit reasoning, latent updates, exploration, and external tools, including allocation-policy overhead.
  • Training, optimization, and data. Efficient pretraining, fine-tuning, reinforcement learning, and distillation; recurrent optimization, curricula over task difficulty and depth, process supervision, synthetic data, and feedback from failed or recovered trajectories; training and data costs of learning transferable efficiency.
  • Systems and infrastructure. Training and serving support for dynamic-depth and long-context models: scheduling, batching, activation and memory management, KV-cache compression, quantization, parallelism, and hardware acceleration; translating model savings into latency, throughput, memory, and energy improvements.
  • Agents and iterative improvement. Resource-aware planning, tool use, retrieval, memory, context compression, caching, and multi-agent coordination; iterative refinement, feedback loops, and empirically evaluated self-improvement of policies, tools, or models, accounting for verification, communication, retries, and learning costs.
  • Benchmarks and evaluation. Quality–resource trade-offs under controlled parameter, compute, memory, latency, and energy budgets; generalization across difficulty and recurrence depth; complete task costs, reliability, and tail latency on realistic hardware and workloads, including training overhead and latent or external computation.
  • Interpretability, robustness, and safety. Monitoring and verification of explicit and latent reasoning; error propagation across iterations, safe stopping, uncertainty calibration, and robustness to distribution shifts or adversarial manipulation; effects of compression, distillation, and quantization on reliable decisions.
  • Applications. Mathematical reasoning, software engineering, scientific discovery, long-context and multimodal understanding, spatial reasoning, robotics, embodied agents, and on-device systems; constraints on response time, hardware, interaction costs, and successful task completion.

Submission Guidelines

Paper format and page limit
Single PDF using the ICLR LaTeX template, with 4–10 pages of main text, excluding references.
Submission portal
OpenReview. The workshop submission link will be announced.
Review process and conflicts of interest
Double-blind review through OpenReview. Submissions and supplementary materials must be anonymized. Reviewers and Program Committee members will declare conflicts; conflicted members will recuse themselves and papers will be reassigned in accordance with the ICLR Conflicts of Interest policy and Code of Conduct.
Publication policy
Workshop papers are non-archival. All accepted papers will be publicly available on OpenReview and imported into iclr.cc by February 26, 2027, 11:59 p.m. AoE.
Presentation
Accepted papers will be presented as posters, with three selected for contributed talks in the tentative program. Organizers will moderate sessions rather than give talks.
Previously published work
Submissions of work already published at major machine learning venues, including ICLR, ICML, and NeurIPS, are discouraged.
LLM usage policy
LLMs may be used as general-purpose assistance tools. Substantial contributions to research ideation or writing must be disclosed in a separate LLM Usage section, typically in an appendix excluded from the page limit, explaining their role. Failure to disclose significant usage may result in desk rejection. Authors remain fully responsible for all content, including LLM-generated text; LLMs cannot be credited as authors.

Key Dates

These dates reflect the proposal and the ICLR 2027 workshop timeline. The assigned workshop day, submission portal, and final arrangements remain subject to confirmation. See the ICLR 2027 workshop timeline.

Workshop dates and location
Submission DeadlineFebruary 1, 2027 (planned)
Notification of AcceptanceFebruary 26, 2027, 11:59 p.m. AoE
Accepted Papers Public on OpenReview and Imported into iclr.ccFebruary 26, 2027, 11:59 p.m. AoE
Opening and Closing Times AnnouncedMarch 26, 2027
Full Program FinalizedApril 14, 2027
Workshop DayApril 29 or 30, 2027 (assignment TBD)
LocationSan Francisco, CA · Venue TBD

Tentative Schedule

The tentative full-day, in-person program includes eight invited talks, three contributed talks, two poster sessions, and one panel discussion, with online access and accommodations for participants unable to attend in person. Invited talks are approximately 20 minutes and contributed talks 15 minutes, including Q&A; each poster session is 60 minutes and the panel is 45 minutes. Final timing will be coordinated with ICLR.

09:00–12:00

Morning Session

Opening remarks, four invited talks, one contributed talk, a poster session, and a coffee break.

12:00–13:00

Lunch Break

Time for informal discussion and community exchange.

13:00–17:00

Afternoon Session

Four invited talks, two contributed talks, a second poster session, a panel with audience Q&A on open challenges and future directions, a coffee break, and closing remarks.

Community & Participation

We aim to support Ph.D. students, postdoctoral scholars, and researchers from underrepresented groups, with participation across regions, affiliations, research areas, and career stages. Contributed talks, posters, and discussion will provide opportunities to exchange feedback and build connections. Available sponsorship funds will prioritize attendance support for junior researchers.

We plan to continue reviewer mentorship pairing junior and senior reviewers. Hybrid participation will combine live discussion with asynchronous access to recorded content across time zones. Recordings of invited and contributed talks and all accepted papers will be shared on the workshop website. A Zoom link will be provided later.

Invited Speakers

The following invited speakers have confirmed participation in the workshop proposal. Talk topics are listed where available; final titles and times will be announced.

Junjie Hu

Junjie Hu

University of Wisconsin–Madison

Efficient reasoning with cache management

Sewon Min

Sewon Min

UC Berkeley

Large language models and efficient reasoning in agents

Niklas Muennighoff

Niklas Muennighoff

Stanford University

Efficient LRM training with simple test-time scaling (s1)

Yuandong Tian

Yuandong Tian

Recursive

Recursive self-improvement and efficient LLM reasoning

Rulin Shao

Rulin Shao

University of Washington

Talk topic TBD

Chenfeng Xu

Chenfeng Xu

UT Austin

Efficient machine learning and AI systems

Huan Sun

Huan Sun

The Ohio State University

Talk topic TBD

Diyi Yang

Diyi Yang

Stanford University

Talk topic TBD

Workshop Organizers

Xinnan Dai

Xinnan Dai

Michigan State University

Biography of Xinnan Dai

Xinnan Dai is a Ph.D. student at Michigan State University, advised by Prof. Jiliang Tang. Her research focuses on graph reasoning and mechanistic interpretability of large language models, particularly how transformers represent structural information and perform reasoning. Her work has appeared at NeurIPS, ICLR, and ICML, with related spotlight research at the NeurIPS Workshop on Efficient Reasoning.

daixinna@msu.edu
Cheng Luo

Cheng Luo

TikTok

Biography of Cheng Luo

Cheng Luo is a researcher at TikTok working on long-context language models, memory optimization, and throughput. He leads workshop logistics, including calls for papers, speaker coordination, scheduling, on-site organization, and community activities. He is a founder and lead organizer of the First Workshop on Efficient Reasoning.

cheng.luo@bytedance.comwdlctc@gmail.com
Zefan Cai

Zefan Cai

University of Wisconsin–Madison

Biography of Zefan Cai

Zefan Cai works on efficient architectures and inference at the University of Wisconsin–Madison. He coordinates workshop reviewing and logistics, including the call for papers and review management. He is a founder and lead organizer of the Second Workshop on Efficient Reasoning.

zefncai@gmail.com
Haozheng Luo

Haozheng Luo

Northwestern University

Biography of Haozheng Luo

Haozheng Luo is a Computer Science Ph.D. student at Northwestern University, advised by Prof. Yan Chen. His research covers AI safety, efficient reasoning, and foundation models across text, vision, speech, and genomics, including attention pruning, quantization robustness, and parameter-efficient adaptation. He is an Adobe Fellow and a DAAD AINeT Fellow, with work at ICML, NeurIPS, ICLR, and USENIX Security. He organized Efficient Spatial Reasoning at ICLR 2026 and Efficient Reasoning at COLM 2026.

hluo@u.northwestern.edu
Yijiang Li

Yijiang Li

UC San Diego

Biography of Yijiang Li

Yijiang Li is a Ph.D. student at UC San Diego and holds an M.S.E. in Computer Science from Johns Hopkins University. His research studies label- and sample-efficient, robust learning in multimodal, interactive, and 3D embodied environments, with publications at NeurIPS, ICML, ICCV, ICLR, CVPR, and EMNLP. He led Efficient Spatial Reasoning at ICLR 2026 and Multi-Modal Reasoning for Agentic Intelligence at CVPR 2026, and co-organized Efficient Reasoning at COLM 2026 and Computer Vision for Children at CVPR 2026.

yijiangli@ucsd.edu
Chenwei Xu

Chenwei Xu

Northwestern University

Biography of Chenwei Xu

Chenwei Xu is a Ph.D. student at Northwestern University working on video understanding, multimodal reasoning, and AI for science. His research emphasizes reliable and interpretable reasoning, including error diagnosis, reasoning verification, and agent-based retrieval in video-language models. He reviews for major machine learning and computer vision conferences and has served as a workshop Area Chair at COLM and CVPR.

cxu@u.northwestern.edu
Weijia Shi

Weijia Shi

University of Washington

Biography of Weijia Shi

Weijia Shi is a Ph.D. student at the University of Washington. Her research develops language models with new approaches to data use, including retrieval-augmented models and mixture-of-experts methods for distributed training without pooling data. She led the second, third, and fourth Knowledge-Augmented NLP Workshops and the ACL Mentorship Program.

swj0419@uw.edu
Zhijian Liu

Zhijian Liu

UC San Diego

Biography of Zhijian Liu

Zhijian Liu is an assistant professor at UC San Diego and co-founder of Inco AI. He received his Ph.D. and S.M. from MIT and B.Eng. from Shanghai Jiao Tong University. His work connects efficient machine learning algorithms and systems for computer vision, language, and scientific discovery, with oral and spotlight presentations at NeurIPS, ICLR, and CVPR. He received the Qualcomm Innovation Fellowship and Rising Star recognitions, and organized Efficient Spatial Reasoning at ICLR 2026 and Efficient Reasoning at COLM 2026.

zhijian@ucsd.edu

Program Committee

All workshop organizers also serve on the Program Committee. The following additional members have confirmed participation; we are continuing to expand the committee across academia and industry.

  • Haoyu He
  • Zhuolin Jiang
  • Guo Ye
  • Chenghao Qiu
  • Chingyuen Huang
  • Hao Xu
  • Eric Jiang
  • Jiansu Zhang
  • Yaxin Li
  • Bing Liu

Contact

For workshop enquiries, please contact:

Sponsors

TBD — Sponsors will be announced.

Previous Editions

The companion First Workshop on Efficient Spatial Reasoning at ICLR 2026 drew over 1,000 attendees and received 89 submissions, supported by 141 reviewers and 261 reviews.