Multiple paper open-source codes of the Microsoft Research Asia DKI group

ReaSCAN is a synthetic navigation task that requires models to reason about surroundings over syntactically difficult languages. (NeurIPS '21)

Official pytorch implementation of CVPR2023 paper "Learning Conditional Attributes for Compositional Zero-Shot Learning"

Diverse Demonstrations Improve In-context Compositional Generalization

Benchmarks for Compose by Focus: Scene Graph-based Atomic Skills

Compositional Capabilities of Autoregressive Transformers: A Study on Synthetic, Interpretable Tasks

Code from the article: "Lost in Latent Space: Examining Failures of Disentangled Models at Combinatorial Generalisaton" (NeurIPS, 2022)

Baby Abstract Reasoning Corpus (BabyARC) dataset engine, for generating grid-world-based abstract reasoning tasks on a large scale.

[ACL 2023] "Learning to Substitute Span towards Improving Compositional Generalization"

Iterative Compositional Data Generation for Robot Control

This repository shares the most important sources used for the reasearch paper "More Diverse Training, Better Compositionality! Evidence from Multimodal Language Learning" by Caspar Volquardsen, Jae Hee Lee, Cornelius Weber, and Stefan Wermter.

Reproducible emergent AI communication experiments at the information-theoretic limit

[ACL 2024] "Benchmarking and Improving Compositional Generalization of Multi-aspect Controllable Text Generation"

Official package for "A Neural Affinity Framework for Abstract Reasoning." Includes the validated 9-category ARC taxonomy, pre-computed fine-tuning results, and scripts to verify the Compositional Gap.

Causal Abstraction of Neural Models Trained to Solve ReaSCAN

End-to-end prime factorization in a generative LM. 40M-param GPT that learns algebraically verifiable prime-factor signatures at negligible language cost (+1.7% PPL). Paper (Zenodo) + triadic-head (PyPI) + reptimeline.

RAIN — honest compositional-generalization research: exact symbolic composition + learned selection hits 100% SCAN / 99.75% COGS gen at seconds of CPU, with measured transformer & LLM baselines and a fully reproducible benchmark suite

σFlow-PDE: A drop-in H-Bar training engine that escapes the σ-trap in neural PDE solvers via live σ/δ/α ODE integration, autonomous phase curriculum, and auto-falsification.