Research Blog

Blog posts on AI and NLP Research Topics

🔬 Featured Research Projects

Making Reasoning Matter

Measuring and improving faithfulness of chain-of-thought reasoning in large language models using causal mediation analysis.

Key Contributions:

Causal framework for measuring reasoning faithfulness
Analysis across 11 different language models
Insights on training objectives and reasoning quality

arXiv:2402.13950 Chain-of-Thought Causal Analysis

Project Details Read Paper

REFINER

A framework for improving language models' reasoning through iterative feedback from critic models on intermediate representations.

Key Features:

Generator-Critic architecture for iterative refinement
Structured feedback on reasoning steps
Human-compatible feedback integration

arXiv:2304.01904 Feedback Learning Reasoning

Project Details Read Paper

🧠 Research Focus Areas

AI Reasoning

Chain-of-thought reasoning, causal inference, knowledge representation, and logical reasoning processes in AI systems.

Feedback & Learning

Iterative feedback mechanisms, critic models, human-AI interaction, and preference optimization for better model alignment.

Model Faithfulness

Measuring and improving the reliability, interpretability, and trustworthiness of AI decision-making processes.

📌 Featured Posts

What Children Know That AI Doesn't: Learning Through Experience, Not Text

Jean Piaget revealed how children construct understanding through action, surprise, and adaptation. His insights expose fundamental gaps in current AI—systems that predict words, not reality. But a new paradigm is emerging.

16 min read • September 30, 2025

Understanding FLOPs, MFU, and Computational Efficiency in LLM Training: From Dense Transformers to MoE Architectures

A comprehensive guide to counting FLOPs in LLM training, measuring Model FLOPs Utilization (MFU), and extending these concepts to Mixture-of-Experts architectures with a deep dive into OpenAI's GPT-OSS models.

15 min read • September 28, 2025

FIPO: Fallacy-Informed Preference Optimization - Steering LLMs Toward Logically Sound Arguments

🏆 Outstanding Paper Award Winner at NAACL 2025! Introducing FIPO, a novel framework that reduces logical fallacy errors in LLM-generated arguments by up to 17.5% through fallacy-informed preference optimization.

10 min read • January 15, 2025

Making Reasoning Matter: Measuring Faithfulness in Chain-of-Thought Reasoning

A deep dive into our latest research on evaluating and improving the faithfulness of chain-of-thought reasoning in large language models.

3 min read • February 23, 2024

blog