ICLR Workshop · April 29–30, 2027 · San Francisco

SymTrust

Neuro-Symbolic Methods for Trustworthy and Verifiable Reasoning

A workshop on making AI reasoning more explicit, interpretable, and verifiable.

Workshop question

How can symbolic and structured methods help us improve and verify model reasoning and build systems whose behavior is more reliable, interpretable, and trustworthy?

Abstract

About the workshop

The correctness of answers produced by language models does not necessarily translate to reliable reasoning. Language models can often produce convincing explanations that fail to faithfully reveal what influenced their outputs or succeed only under a particular wording or representation.

SymTrust: Neuro-Symbolic Methods for Trustworthy and Verifiable Reasoning explores how rules, constraints, programs, graphs, knowledge bases, proofs, tables, and diagrams can make reasoning easier to inspect, evaluate, and verify.

We welcome researchers from NLP, trustworthy AI, symbolic and neuro-symbolic reasoning, multimodal reasoning, interpretability, formal verification, and related fields.

Open challenges

Open research questions

Preserving meaning while changing representations. What information must an explicit representation preserve, and how can we detect when details are omitted, distorted, or introduced?
What does verification establish? When does successful verification provide meaningful evidence for trusting a model’s conclusion, and when is it insufficient?
Integrating neural and symbolic reasoning. What roles should neural and symbolic components play, and how should information and feedback flow between them?
Choosing how to reason. When should natural-language or structured reasoning be used, and how can models make this choice adaptively?
Evaluating trustworthy reasoning in practice. How should we evaluate faithfulness, robustness, and verifiability beyond final-answer accuracy?

Topics

Topics of interest

We welcome archival and non-archival work across methods, theory, datasets, benchmarks, evaluations, applications, position papers, and analyses of model failures.

Trustworthiness, reliability, and verification of model reasoning
Faithfulness, validity, plausibility, and consistency of reasoning traces and explanations
Explainability and interpretability of structured, symbolic, and neural reasoning systems
Formal, symbolic, and structured reasoning methods, including neuro-symbolic and solver-augmented approaches
Multimodal reasoning across language, vision, spatial, diagrammatic, and symbolic representations

Improving trustworthiness through robustness and generalization across paraphrases, prompts, representations, domains, and distribution shifts
Uncertainty, calibration, confidence, and appropriate abstention in reasoning systems
Metrics, benchmarks, and human evaluation for trustworthy and interpretable reasoning
Applications of symbolic or trustworthy reasoning, including failure analyses and negative results in high-stakes or real-world settings

Speakers

Invited speakers

Francesca Toni

Imperial College London

Works at the intersection of symbolic reasoning and trustworthy AI through computational argumentation and argumentative language models, using explicit argument structures to make model conclusions explainable, contestable, and verifiable.

Visit website

Bo Li

University of Illinois Urbana-Champaign and Virtue AI

Connects formal and structured methods with AI safety, alignment, and robustness, including logic- and policy-based guardrails and verifiable reasoning for autonomous systems.

Visit website

Dan Roth

University of Pennsylvania and Oracle

Bridges foundational structured and constrained inference with reliable language-model reasoning, including calibration and the integration of learned models with external knowledge and inference mechanisms.

Visit website

Mohit Bansal

University of North Carolina at Chapel Hill

Brings expertise in structured and programmatic reasoning across language and multimodal systems, including executable programs, visual programs, scene graphs, and structured intermediate representations.

Visit website

Hima Lakkaraju

Harvard University

Brings a trustworthiness perspective through her work on interpretable machine learning, explanation faithfulness, and the reliability of chain-of-thought and other model explanations.

Visit website

Lu Wang

University of Michigan

Contributes expertise in factual and attributable generation, confidence calibration, self-verification, reasoning evaluation, and process-level supervision.

Visit website

Schedule

Tentative schedule

SymTrust is a one-day, in-person workshop. The schedule below is tentative; talk titles and the final program will be posted ahead of the workshop.

Time	Session
09:00–09:10	Opening Remarks and Workshop Framing
09:10–09:45	Invited Talk 1
09:45–10:20	Invited Talk 2
10:20–11:20	Poster Session 1 and Coffee Break
11:20–12:05	Oral Presentations: Methods and Representations
12:05–13:05	Lunch, Informal Networking, and Optional Mentoring Roundtables
13:05–13:40	Invited Talk 3
13:40–14:15	Invited Talk 4
14:15–15:00	Oral Presentations: Evaluation and Applications
15:00–16:00	Poster Session 2 and Coffee Break
16:00–16:35	Invited Talk 5
16:35–17:10	Invited Talk 6
17:10–18:00	Panel Discussion and Audience Q&A
18:00–18:10	Best Paper Awards and Closing Remarks

Organizers

Organizing committee

Parisa Kordjamshidi

Michigan State University

Associate Professor of Computer Science and Engineering. Develops neuro-symbolic methods that combine neural models with structured knowledge, constraints, and explicit representations for language and multimodal reasoning.

Visit website

Shreya Rajpal

Michigan State University

Ph.D. student in the Heterogeneous Learning & Reasoning Lab. Studies when explicit symbolic representations improve reasoning and when they provide reliable evidence about model behavior, including in high-stakes domains such as law and finance.

Visit website

Enrico Santus

Bloomberg

Principal Technical Strategist for Human-AI Interaction and Academic Engagement in Bloomberg’s Office of the CTO. Works on fact verification, graph-based information extraction, neuro-symbolic robustness, and structured knowledge representations for legal and financial reasoning.

Visit website

Pietro Barbiero

IBM Research Europe – Zurich

Research Staff Member working on explainable, neuro-symbolic, and trustworthy AI, using explicit logical and concept-based representations to make neural-model behavior interpretable and verifiable.

Visit website

Yangqiu Song

The Hong Kong University of Science and Technology

Associate Professor of Computer Science and Engineering and co-director of the HKUST-WeBank Joint Lab. Develops large-scale structured knowledge resources and uses them for commonsense, logical, and knowledge-intensive reasoning.

Visit website