Bruce W. Lee

I write open-source NLP tools and research language models.

My short-term aim is to contribute to making language models better. My mid-term goal is to bring theoretical advancements to real-world interactable systems. My long-term objective is to steer all these to benefit humanity.

Before mainly working on language models, I used to do linguistic analysis. Some examples include LFTK and LingFeat open-source toolkits (Link to GitHub), which became a popular toolkit at a number of research labs. In high school, I was very involved in physics research, and some were published. Apart from research, I've been rowing since high school (Link to Photo) and have served in the Marines (Link to Photo).

Papers

2024 Programming Refusal with Conditional Activation Steering
Diagram
  • Large Language Models (LLMs) have demonstrated impressive capabilities, but precisely controlling their response behavior remains a challenge.
  • Conditional Activation Steering (CAST) analyzes LLM activation patterns during inference to selectively apply or withhold activation steering based on the input context.
  • CAST enables systematic control of LLM behavior through conditional rules, allowing for selective modification of responses to specific content without the need for weight optimization.
2024 Language Models Don't Learn the Physical Manifestation of Language
  • We argue that language-only models lack understanding of the physical manifestation of language, as demonstrated through a series of tasks called the H-Test.
  • Our hypothesis is supported by evidence showing that neither deliberate reasoning, few-shot examples, nor using stronger models from the same family improves performance on the H-Test.
  • Experiments reveal that even some of the most advanced proprietary LLMs perform near random chance on these tasks, underscoring the limitations of linguistic knowledge acquired.
2023 Instruction Tuning with Human Curriculum
Diagram
  • We introduce Curriculum Instruction Tuning, an approach that explores the benefits of using diverse curriculum strategies in language model training.
  • We present a synthetic instruction-response generation framework designed to mimic the sequential and orderly nature of human learning, distinguishing it from existing instruction tuning datasets.
  • We also outline a methodology for creating instruction-response datasets that comprehensively cover various stages of human education, from middle school to graduate level, using educational subject catalogs.
2021 Pushing on Text Readability Assessment: A Transformer Meets Handcrafted Linguistic Features
Diagram
  • We report two advancements: the introduction of three novel features in advanced semantics and evidence that combining these features with transformers can enhance performance.
  • Our methodology involves exploring suitable transformers and traditional ML models, extracting 255 handcrafted linguistic features, and then creating hybrid models.
  • These hybrid models achieve state-of-the-art accuracy on popular datasets used in readability assessment.
(See more papers)

Softwares

2024 IBM/activation-steering
2023 LFTK
2021 LingFeat
2020 ~ I've developed several softwares that I can't publicly share. These include lexical databases, APIs for downstream NLP tasks, and codebases for LLM evaluation, among others. I take pride in writing reusable code.
(See more softwares)

Get in Touch

Email: brucelws@seas.{school}.edu. Replace school with 'upenn'.