Akshay Manglik

Undergraduate, Columbia University

akshay.m [AT] columbia.edu

Bio

I am a senior at Columbia University, majoring in Computer Science and Economics and minoring in Math. I am a Laidlaw Scholar and a Data Science Institute Scholar.

I am interested in studying reasoning, robustness, and grounding in language models, especially through expanding test-time compute, devising alternative metrics and training objectives, and integrating symbolic systems.

My past research experiences have spanned natural language processing, computer vision, neuroimaging, computational genomics, and protein structure prediction. I have conducted research at:

I spent my most recent summer interning at PayPal as an AI Engineer Intern.

Research Interests

I aim to develop AI systems that reason reliably and ground their knowledge, particularly in novel situations. While large language models (LLMs) memorize vast amounts during pretraining, they struggle to apply knowledge in novel, "long tail" scenarios and lack mechanisms for factuality. These challenges of reasoning and grounding are deeply interconnected: better reasoning improves factuality, while better grounding enables reasoning with less information. I am interested in how multi-agent dynamics can encourage reliable reasoning, especially under limited information, drawing on my background in economics, debate, and philosophy. My research interests span three areas:

  1. Test-Time Exploration: Developing sample-efficient, scalable techniques beyond sampling and search, like injecting schema, finetuning at test time, and incorporating cooperative and competitive multi-agent dynamics into decoding.
  2. Learning from Rich Feedback: Creating robust metrics for assessing factuality and causality in reasoning, and developing sample-efficient training objectives for sparsely-observed tasks.
  3. Neuro-symbolic Integration: Incorporating symbolic reasoning and tool use to ground model outputs and ensure robustness in novel, sample-constrained situations.

CV

Full CV in PDF.

Publications

Most recent publications on Google Scholar.
denotes equal contribution.

Binding items to contexts through conjunctive neural representations with the Method of Loci

Jiawen Huang, Akshay Manglik, Nicholas Dutra, Hannah Tarder-Stoll, Taylor Chamberlain, Robert Ajemian, Qiong Zhang, Kenneth A. Norman, Christopher Baldassano

bioRxiv'24: bioRxiv preprint. 2024.

When to Think Step by Step: Computing the Cost-Performance Trade-offs of Chain-of-Thought Prompting

Akshay Manglik, Aman Choudhri

Interim Manuscript.

On Bias: Moral Intuitions, Rationalizations, and Adversarial Disagreement

Akshay Manglik

Gadfly Philosophy Magazine. 2022.

Website Design

Thank you to Martin Saveski for the template for this website.