Introduction

Following the formulation in Distilling System 2 into System 1, given an input $x$, System-1 produces the output $y$ directly: $S_{\mathrm{I}}(x)=f_{\theta}(x)\to y$. In contrast, System-2, takes an LLM $f_{\theta}$ and input $x$ and generates intermedaite tokens $z$: $S_\text{II}{(x;f_\theta)}\to z,y$, which can be seen as a form of meta learning. I plan to design scalable System-2 LLMs from the meta learning perspective.

My Related Works

Neural Architectures

• Unlocking Emergent Modularity in Large Language Models, NAACL 2024 Outstanding Paper
• HyperMoE: Towards Better Mixture of Experts via Transferring Among Experts, ACL 2024
• Prototype-based HyperAdapter for Sample-Efficient Multi-task Tuning, EMNLP 2023
• Beyond Fully-Connected Layers with Quaternions: Parameterization of Hypercomplex Multiplications with $1/n$ Parameters, ICLR 2021 Outstanding Paper

Multi-Modal & Reinforcement Learning

• Think Before You Act: Decision Transformers with Internal Working Memory, ICML 2024
• MERT: Acoustic Music Understanding Model with Large-Scale Self-supervised Training, ICLR 2024
• Interactive Natural Language Processing, arXiv 2023

Human-AI Alignment

• AI Alignment: A Comprehensive Survey, arXiv 2023
• Would you Rather? A New Benchmark for Learning Machine Alignment with Cultural Values and Social Preferences, ACL 2020

Meta Learning

• Massive Editing for Large Language Models via Meta Learning, ICLR 2024
• Learning Multi-Objective Curricula for Robotic Policy Learning, CoRL 2023