Publications
Peer-reviewed papers and preprints, newest first. Use the search box to filter by title, author, or venue.
2026
-
In Advances in Neural Information Processing Systems (NeurIPS 2026), 2026TLDR: Extending ARC to five dimensions of rule understanding (definition, classification, constrained generation, editing, inversion) reveals a 25-52 point gap relative to standard ARC evaluation. -
In Advances in Neural Information Processing Systems (NeurIPS 2026), 2026TLDR: Vision-language models can perform well on abstract reasoning tasks without using the intended human-like core knowledge priors, so accuracy alone overstates their competence. - Illusions of Abstraction in Contemporary AIProceedings of the National Academy of Sciences (PNAS), 2026To appear
-
Transactions on Machine Learning Research (TMLR), 2026Also known as Memento: Note-Taking for Your Future SelfTLDR: Decomposing multi-hop questions into single-step Prolog definitions improves performance on various long-context question datasets. -
2026TLDR: A spiking neural network implementation of the Transformer decoder that reduces theoretical energy consumption by 87 to 93 percent compared to its ANN counterpart. -
In Proceedings of the Annual Meeting of the Cognitive Science Society (CogSci 2026), 2026Full paperTLDR: Through a Bayesian approach combining natural language rules and program synthesis, models can approach human performance on visual pattern recognition puzzles (Bongard problems). -
In Proceedings of the Annual Meeting of the Cognitive Science Society (CogSci 2026), 2026SymposiumTLDR: Symposium on evaluating language models with methods and principles from cognitive science. -
In Findings of the Association for Computational Linguistics: ACL 2026, 2026TLDR: A multilingual benchmark for LLM-based software engineering agents covering bug fixing, test generation, style fixing, and addressing code reviews.
2025
-
In NeurIPS 2025 Workshop on Multimodal Algorithmic Reasoning (MAR), 2025SpotlightTLDR: The o3 model relies on diverse shortcuts and heuristics to solve ARC tasks. Tool use improves accuracy in the visual modality, while increased reasoning effort helps in the text modality. -
In NeurIPS 2025 Workshop on Interpreting Cognition in Deep Learning Models (CogInterp), 2025TLDR: An architecture that learns schemas and applies them iteratively can resolve arbitrarily deep compositional statements. -
In NeurIPS 2025 Workshop on Regulatable Machine Learning (RegML), 2025TLDR: We refine Inverse Constitutional AI to extract interpretable principles from preference datasets for EU AI Act Article 10 compliance. -
2025TLDR: Aligning models with human preferences does not guarantee that they pursue normative goals on our behalf; new alignment techniques may be required. -
-
In 2025 IEEE/ACM International Workshop on Large Language Models for Code (LLM4Code), 2025TLDR: Language models struggle to dry-run simple and advanced code structures such as recursion, concurrency, and object-oriented programming. -
2025TLDR: Using improved clustering and a more diverse embedding approach, our technique more accurately compresses preference datasets into human-readable constitutions.