Reasoning & planning
Turning deliberation into decisions that work beyond a single response.
- Chain-of-thought, search, and test-time scaling
- Long-horizon planning and task decomposition
- World models, causal and multimodal reasoning
Learning, Acting, and Adapting
with Foundation Models
Beyond generating answers. Toward systems that
reason, act, and learn from the world.
Foundation models are becoming more capable reasoners. The next challenge is to turn that capability into purposeful action: making plans, using tools, interacting with environments, and learning from experience.
This workshop brings together the NLP, machine learning, and agentic AI communities to study that transition. We ask how reasoning supports action, how interaction creates new learning opportunities, and how agents can adapt reliably over time.
Find your research directionFrom inference to plans and decisions.
From text to tools and environments.
From feedback to lasting improvement.
We welcome work across the full path from
reasoning to agency, including but not limited
to the following themes.
Turning deliberation into decisions that work beyond a single response.
Learning effective behavior from feedback, demonstrations, and interaction.
Connecting foundation models to tools, computers, and the physical world.
Building agents that retain experience and improve as their worlds change.
Studying how agents coordinate with each other and collaborate with people.
Measuring meaningful progress toward reliable, efficient, and safe agents.
Working between these themes? You belong here. We also welcome applications in scientific discovery, software engineering, embodied intelligence, education, and multilingual or low-resource settings.
We invite original contributions that deepen our understanding of foundation models as reasoning, learning, and acting systems.
From a new method to a thoughtful position, from a useful benchmark to an illuminating negative result—there is room for different kinds of progress.
Empirical studies, theoretical insights, new methods, and agent systems.
Position papers, datasets, benchmarks, analyses, and emerging ideas.
A few dates for your calendar.
All deadlines are at 11:59 PM AoE
(Anywhere on Earth, UTC−12).
The proceedings deadline applies to organizers, if archival publication is confirmed. Any ARR-specific timeline will be announced separately.
Four invited talks, followed by a poster session.
Each talk runs for 20–30 minutes
within a 30-minute slot.
Zoom meeting ID: 815 4152 2305
| Time (UTC+8) | Speaker & session |
|---|---|
| – | Jian Yang InCoder-32B: A Code Foundation Model for Industrial Applications |
| – | Ningyu Zhang Cortex: Consolidating Context into Parametric Memory |
| – | Liwei Liu Exploring the Path from AI4Science to AI4Research |
| – | Zhiqin Yang RobotWorld: Benchmarking Multimodal Agents as Physical Agents |
| onwards | Poster sessionResearch presentations and discussion |
Industrial code, model memory, scientific discovery,
and agents in the physical world.
Beihang University
Industrial code must satisfy constraints that vary across domains. In chip design, code must respect hardware timing semantics. In GPU programming, kernels must compute the correct result while making efficient use of memory. This talk presents InCoder-32B, a 32-billion-parameter code foundation model trained from scratch for applications in chip design, GPU kernel optimization, embedded systems, compiler optimization, and 3D modeling.
The talk will walk through the training pipeline, from pretraining on general-purpose code and further training on curated industrial code to a mid-training stage that introduces synthetic industrial reasoning data and gradually extends the context window from 8K to 128K tokens. It will also cover post-training grounded in actual execution results. Drawing on evaluations across 14 general coding benchmarks and nine industrial benchmarks, the talk will examine how the model balances general programming capabilities with the specialized requirements of industrial tasks.
Jian Yang is an associate professor in the School of Computer Science and Engineering at Beihang University. His research focuses on large language models for code, model pretraining and post-training, and code understanding and generation. He has published more than 100 papers in leading international journals and at conferences such as ICLR, NeurIPS, ACL, and EMNLP, including more than 30 as first or corresponding author. His work has received over 48,000 citations on Google Scholar.
Zhejiang University
Large language models can benefit from context, but these gains often disappear once the context is removed. Inspired by biological memory consolidation, we introduce Memory Consolidation (MC) and propose Cortex, a framework that internalizes context-induced improvements into model parameters while preserving existing capabilities. Across diverse model architectures and scales (4B–120B), Cortex achieves strong consolidation and retention. Notably, after 50 sequential edits on a 14B model, Cortex-Structural recovers 88.9% of context-induced gains while retaining 99.8% of original non-target performance.
Ningyu Zhang, Ph.D., is an associate professor at Zhejiang University and is among Stanford University’s top 2% scientists. His research focuses on natural language processing, large language models, knowledge graphs, and knowledge editing. He has published papers in top academic journals and conferences such as Natural Machine Intelligence, NeurIPS, ICLR, ICML, ACL, and EMNLP. He has developed OceanGPT, DeepKE, EasyEdit, and LightMem. He serves as an Associate Editor for TMLR, Machine Intelligence Research, Neural Networks, and TALLIP. He is also an Area Chair (Senior Area Chair) for ACL, EMNLP, NeurIPS, ICLR, ICML, and KDD.
2012 Laboratories, Huawei Technologies Co., Ltd.
Artificial intelligence is opening up new ways to understand complex data, predict biological patterns, and accelerate discovery across biology, medicine, and drug development. This talk presents three representative studies from our team: developing an RNA language model through cross-omics semantic learning, pretraining a single-cell model using hundred-million-scale datasets, and developing a bioinformatics agent for automated analysis and interpretation of sequencing-based omics data. Together, these studies explore how AI can extend from learning biological representations and supporting prediction tasks to assisting scientific analysis and automating research workflows.
Liwei Liu is an AI4Science researcher at the 2012 Laboratories of Huawei Technologies Co., Ltd. Her research focuses on AI applications in biology, medicine, health, and materials science, as well as AI models and agents that support scientific research. Her work has been published in Advanced Science, Chemical Science, Nature Communications, and other journals.
The Hong Kong University of Science and Technology
As multimodal agents excel in digital spaces, empowering them to act as physical agents in the real world remains a grand challenge. In this talk, we introduce RobotWorld, a comprehensive benchmark designed to evaluate general-purpose multimodal agents across 84 diverse tasks and 5 core embodiments (manipulation, mobile manipulation, locomotion, driving, and flight).
Zhiqin Yang is a PhD student at the Hong Kong University of Science and Technology. His research focuses on reasoning models and robot agents. He is the first author of RobotWorld and an ICML Spotlight paper on preference alignment. His work has also appeared at NeurIPS, ICLR, ICML, and CVPR.
Meet the team bringing our community together
across NLP, machine learning, and agentic AI.
Nanjing University
Nanjing University
Nanjing University
Nanjing University
Nanjing University
Nanjing University
Nanjing University
Central South University
Alibaba
Join us in Hengqin, Zhuhai, just across the water from Macau, for a meeting of ideas at AACL-IJCNLP 2026.
The essentials, in one place.
If your work helps explain or improve how foundation models reason, learn, act, or adapt, we would like to hear from you. We welcome empirical and theoretical work, systems, datasets, benchmarks, position papers, and carefully analyzed negative results. The topic list is illustrative, not exhaustive; connections across research areas are especially welcome.
Please use the official ACL/ARR template in review mode and submit an anonymized PDF. Review will be double-blind: author identities must not appear in the paper or supplementary materials. Final page limits will be announced here. Please follow the ACL guidance for citations and anonymized artifacts.
An ACL Rolling Review (ARR) commitment route is under consideration. Eligibility, a dedicated commitment link, and any separate deadline will be published if this route is offered. Until then, September 30, 2026 is the announced paper submission deadline; no ARR extension is confirmed.
The publication policy, including whether proceedings will appear in the ACL Anthology, is to be confirmed. Archival status and policies affecting submission of the same work to other venues will be stated in the final guidelines. The October 22 proceedings deadline is for organizers if archival proceedings are produced; it is not an additional author deadline.
The AACL-IJCNLP 2026 workshop window is November 9–10, 2026 in Hengqin, Zhuhai, China, next to Macau. The main conference venue is Tianmu Melody Convention Exhibition Center. Our exact workshop day and room are to be announced. The preliminary program begins at 13:30 China Standard Time (UTC+8), with four invited talks followed by a poster session from 15:30. Online participation is available via the Zoom link in the Program section.