ICLR 2026 TrustLLM WorkshopSchedule
| Time | Session | Description | Duration(mins) |
|---|---|---|---|
Apr 27, 12:00 | Opening Remark | Opening Remark Taiji Suzuki | 10 |
Apr 27, 12:10 | Invited Talk 1 | TBD Maksym Andriushchenko | 40 |
Apr 27, 12:50 | Invited Talk 2 | TBD Alessio Lomuscio | 40 |
Apr 27, 13:30 | Poster/Break | Poster Session | 50 |
Apr 27, 14:20 | Invited Talk 3 | TBD Hima Lakkaraju | 40 |
Apr 27, 15:00 | Oral Presentations | TBD TBD | 30 |
Apr 27, 15:30 | Lunch | Lunch - | 90 |
Apr 27, 17:00 | Invited Talk 4 | TBD Bo Li | 40 |
Apr 27, 17:40 | Invited Talk 5 | TBD Rishi Bommasani | 40 |
Apr 27, 18:20 | Oral Presentations | TBD TBD | 30 |
Apr 27, 18:50 | Invited Talk 6 | TBD Sahar Abdelnabi | 40 |
Apr 27, 19:30 | Invited Talk 7 | TBD Zhangyang Atlas Wang | 40 |
Apr 27, 20:10 | Panel Discussion | From Theory to Deployment: Process-Level Safety in LRMs Panelists: - Junchi Yan (SJTU) - Zhangyang Atlas Wang (XTX Markets & UT Austin) - Atsushi Nitanda (A*STAR & NTU) - Maksym Andriushchenko (ELLIS Institute Tübingen) - Aleksandra Korolova (Princeton University) | 40 |
Apr 27, 20:50 | Closing Remark & Awards & Community Building | Closing Remark Wei Huang | 20 |
Call for Papers
Workshop Description
Recent breakthroughs in large reasoning models (LRMs)—systems that perform explicit, multi-step inference at test time—have unlocked impressive gains across science, mathematics, and law (Guo et al., 2025; Snell et al., 2024). Yet these very reasoning chains introduce novel trust risks that scale-only LLM studies overlook: small logic slips can snowball, intermediate states become new attack surfaces, and plausible-but-false proofs can elude standard output-level filters (Geiping et al., 2025).
What sets this workshop apart: Prior workshops focused on generic robustness; This workshop aims to bring together researchers and practitioners to systematically explore, characterize, and address these emergent trust risks in large reasoning models. We seek to move beyond traditional evaluation of final LLM outputs, instead focusing on the full process of reasoning: how errors propagate, how uncertainty compounds, how transparency and accountability are affected, and what new opportunities exist for robust evaluation and intervention (Zhu et al., 2025; Ouyang et al., 2024).
Workshop Goals
- Establish a comprehensive taxonomy of process-level risks in large reasoning models
- Develop standardized benchmarks and evaluation metrics for reasoning safety
- Create a community-driven repository of tools and techniques for safer reasoning
- Foster interdisciplinary collaboration between AI safety, formal methods, and reasoning communities
Core Research Questions
- Foundation - Which theoretical frameworks and abstracts can be used to study the emergent trust risk problems?
- Risk Landscape — Which process-level failure modes threaten high-stakes deployment?
- Metric Design — How can we measure and benchmark step-wise safety at scale?
- Prototype Validation — What lightweight tools or model patches can we release now to seed wider research?
Topics of Interest
- Failure modes and risk propagation in multi-step reasoning.
- Adversarial attacks and defenses in inference-time reasoning.
- Evaluation metrics and empirical benchmarks for reasoning safety.
- Uncertainty quantification, error detection, and correction in reasoning processes.
- Approaches for improving interpretability and transparency of reasoning chains.
- Causal analysis as one lens for understanding reasoning failures.
- Human-AI trust, calibration, and collaborative reasoning.
- Case studies of reasoning models in real-world high-stakes domains.
- New architectures or algorithms for safer, more robust reasoning.
Submission Guide
Submission Instructions
We invite submissions of up to 8 pages (excluding references and appendices). The main manuscript and any appendices must be submitted as a single PDF file through the OpenReview portal (the official workshop venue link will be posted on the website). Papers previously published at other conferences will not be considered for acceptance
Each paper receives no less than 3 reviews. Reviewers declare the conflict of interests via the OpenReview interface; organizers will not handle papers with which they have a conflict.
Tiny Paper Track
In addition to full-length submissions, we encourage contributions to a Tiny Paper Track, designed to promote early-stage, exploratory, or rapidly developing ideas. Tiny papers may be up to 4 pages in length (excluding references) and can include:
- Exploratory experiments or small-scale empirical findings
- Modest but self-contained theoretical insights
- Re-analyses or replications of previously published work
- Implementation notes, tool releases, or negative results
- Fresh perspectives or conceptual critiques
Tiny papers are intended to encourage participation from junior researchers and underrepresented groups, fostering community-building and feedback exchange. All accepted tiny papers will be presented as posters or short talks during the workshop and included in the OpenReview proceedings under the workshop's track configuration.
Timelines
| Submissions Deadline | Jan 30th, 2026 |
| Reviewer bidding dates | Feb 1st - Feb 5th, 2026 |
| Review Deadline | Feb 26th, 2026 |
| Paper Notification Date | Mar 1st, 2026 |
| Camera Ready Deadline | Mar 5th, 2026 |
All deadlines are specified in 23:59AoE (Anywhere on Earth).
Speakers
Maksym Andriushchenko
ELLIS Institute Tübingen
Alessio Lomuscio
Imperial College London
Hima Lakkaraju
Harvard University
Rishi Bommasani
Stanford Institute for Human-Centered Artificial Intelligence
Bo Li
UIUC; Secure Learning Lab
Sahar Abdelnabi
ELLIS Institute Tübingen and MPI-IS
Zhangyang Atlas Wang
XTX Markets & UT Austin
Frequently Asked Questions
Can we submit a paper that will also be submitted to ICLR 2026?
Yes.
Can we submit a paper that was accepted at ICLR 2026?
No. ICLR prohibits main conference publication from appearing concurrently at the workshops.
Will the reviews be made available to authors?
Yes.
I have a question not addressed here, whom should I contact?
Email organizers at trustllm-workshop@googlegroups.com
References
[1] Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X. and Zhang, X., 2025. Deepseek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning. arXiv preprint arXiv:2501.12948.
[2] Snell, C., Lee, J., Xu, K. and Kumar, A., 2024. Scaling LLM test-time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314.
[3] Muennighoff, N., Yang, Z., Shi, W., Li, X.L., Fei-Fei, L., Hajishirzi, H., Zettlemoyer, L., Liang, P., Candès, E. and Hashimoto, T., 2025. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393.
[4] Geiping, J., McLeish, S., Jain, N., Kirchenbauer, J., Singh, S., Bartoldson, B.R., Kailkhura, B., Bhatele, A. and Goldstein, T., 2025. Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach. arXiv preprint arXiv:2502.05171.
[5] Zhu, J., Yan, L., Wang, S., Yin, D. and Sha, L., 2025. Reasoning-to-defend: Safety-aware reasoning can defend large language models from Jailbreaking. arXiv preprint arXiv:2502.12970.