Workshop Schedule & Format
To ask questions for the talks and panel, please submit them through slido #4007290.
The workshop will take place in East Exhibition Hall C of the Vancouver Convention Centre on December 14, 2024. The workshop will be a hybrid event, with both in-person and virtual participation. The schedule is as follows:
| Time | Program | Format |
|---|---|---|
| 08:45 - 09:00 | Opening remarks | In-person |
| 09:00 - 09:30 | Amit Sharma - Teaching causal reasoning to language models | In-person |
| 09:30 - 10:00 | Jane X. Wang - Causal reasoning in foundation agents | In-person |
| 10:00 - 10:30 | Coffee break | - |
| 10:30 - 11:00 | Elias Bareinboim - Towards Causal Artificial Intelligence | In-person |
| 11:00 - 11:15 | From Causal to Concept-Based Representation Learning | In-person |
| 11:15 - 11:30 | Causal Order: The Key to Leveraging Imperfect Experts in Causal Inference | In-person |
| 11:30 - 12:00 | Poster session 1 | In-person |
| 12:00 - 13:30 | Lunch break | Individually |
| 13:30 - 14:00 | Victor Veitch - Musings on the Linear Representation Hypothesis | In-person |
| 14:00 - 14:15 | Using Relational and Causality Context for Tasks with Specialized Vocabularies that are Challenging for LLMs | In-person |
| 14:15 - 14:30 | Causally Testing Gender Bias in LLMs: A Case Study on Occupational Bias | In-person |
| 14:30 - 15:00 | Poster session 2 | In-person |
| 15:00 - 15:30 | Coffee break | - |
| 15:30 - 16:00 | Claudia Shi - Hypothesis testing the circuit hypothesis in LLMs | In-person |
| 16:00 - 16:30 | Tobias Gerstenberg - Causal thinking in humans and machines | In-person |
| 16:30 - 17:30 | Panel discussion: Maria Antoniak, Elias Bareinboim, Chelsea Finn, Atticus Geiger, Zhijing Jin, Giambattista Parascandolo | In-person |
| 17:30 | Closing remarks and Best Paper Award! | In-person |
All talks will be recorded and made available to registered participants after the workshop through the NeurIPS website.
Poster Sessions
All accepted papers will be presented as a poster at the workshop. All posters should be put up in the morning (e.g., during the first break) and will remain up during the whole day, and presenters may present in either or both of the two poster sessions.
For virtual participants, we encourage you to view the papers on OpenReview and reach out to the authors directly, for example through the NeurIPS virtual platform.
Invited Talks
Amit Sharma: Teaching causal reasoning to language models
Large language models (LLMs) have demonstrated remarkable accuracy in identifying cause-and-effect relationships across diverse scientific domains. However, their ability to reason over these relationships remains a challenge. To address this, we propose axiomatic training—a novel approach that enhances causal reasoning by teaching LLMs fundamental causal axioms one at a time, rather than fine-tuning them for specific tasks. By training on synthetic demonstrations of axioms such as transitivity and d-separation, we show that models with fewer than 100 million parameters can surpass reasoning capabilities of significantly larger models such as Phi-3, Gemini Pro and GPT-4. Axiomatic training has practical applications as a tool for constructing verifiers for LLM-generated reasoning and for embedding inductive biases into LLM fine-tuning. Moreover, it provides insights into how models like GPT-4, trained solely on observational data, can exhibit advanced reasoning capabilities.