Scalable influence and fact tracing for large language models pretraining
Published:
Figure: Difference between the classical lexical retrieval and the influence based retrieval for large language models
Posts by Tags
-AI
Why Language Models Hallucinate: The Epidemic of Penalizing Uncertainty
Published:
Figure: Binary grading makes “guess when unsure” optimal → higher hallucinations.
Confidence-aware grading (penalize wrong answers; allow IDK) makes abstention rational → lower hallucinations.
-Hallucination
Why Language Models Hallucinate: The Epidemic of Penalizing Uncertainty
Published:
Figure: Binary grading makes “guess when unsure” optimal → higher hallucinations.
Confidence-aware grading (penalize wrong answers; allow IDK) makes abstention rational → lower hallucinations.
-Language
Scalable influence and fact tracing for large language models pretraining
Published:
Figure: Difference between the classical lexical retrieval and the influence based retrieval for large language models
Why Language Models Hallucinate: The Epidemic of Penalizing Uncertainty
Published:
Figure: Binary grading makes “guess when unsure” optimal → higher hallucinations.
Confidence-aware grading (penalize wrong answers; allow IDK) makes abstention rational → lower hallucinations.
-Transparency
Scalable influence and fact tracing for large language models pretraining
Published:
Figure: Difference between the classical lexical retrieval and the influence based retrieval for large language models
AI
Teaching Humanoids Without MoCap: Inside TWIST2’s Portable Data Collection System
Published:
Motivation
How do we collect humanlike motion data for robots without a $100K motion-capture studio?
AI Projects
What I Learned from Hackathons (and Losing One!)
Published:
Hackathons have been among the best learning experiences of my career.
Data Valuation
What is Data Shapley? Measuring the True Value of Data
Published:
We often focus on model architectures — but what if the most valuable part of your ML system is your data?
Data Shapley assigns a contribution score to each training point, measuring its impact on model performance.
Decision Support Systems
Enhancing Cybersecurity Risk Assessment using Temporal Knowledge Graphs
Published:
My recent publication in Decision Support Systems (Elsevier, 2025) focuses on temporal knowledge graph-based explainable DSS for cybersecurity.
Explainable AI
What is Data Shapley? Measuring the True Value of Data
Published:
We often focus on model architectures — but what if the most valuable part of your ML system is your data?
Data Shapley assigns a contribution score to each training point, measuring its impact on model performance.
Explaining SENE: Manifold Learning for Distracted Driving Analysis
Published:
My first research paper, published in Engineering Applications of Artificial Intelligence (2023), proposed SENE — a novel manifold learning technique for analyzing distracted driving.
Hackathons
What I Learned from Hackathons (and Losing One!)
Published:
Hackathons have been among the best learning experiences of my career.
Humanoids
Teaching Humanoids Without MoCap: Inside TWIST2’s Portable Data Collection System
Published:
Motivation
How do we collect humanlike motion data for robots without a $100K motion-capture studio?
Knowledge Graphs
Enhancing Cybersecurity Risk Assessment using Temporal Knowledge Graphs
Published:
My recent publication in Decision Support Systems (Elsevier, 2025) focuses on temporal knowledge graph-based explainable DSS for cybersecurity.
LLMs
What I Learned from Hackathons (and Losing One!)
Published:
Hackathons have been among the best learning experiences of my career.
Machine Learning
5 Books That Changed How I Think About Machine Learning and Research
Published:
Books have shaped how I approach ML — not just as a technical field, but as a way of thinking.
What is Data Shapley? Measuring the True Value of Data
Published:
We often focus on model architectures — but what if the most valuable part of your ML system is your data?
Data Shapley assigns a contribution score to each training point, measuring its impact on model performance.
Manifold Learning
Explaining SENE: Manifold Learning for Distracted Driving Analysis
Published:
My first research paper, published in Engineering Applications of Artificial Intelligence (2023), proposed SENE — a novel manifold learning technique for analyzing distracted driving.
Models
Scalable influence and fact tracing for large language models pretraining
Published:
Figure: Difference between the classical lexical retrieval and the influence based retrieval for large language models
Why Language Models Hallucinate: The Epidemic of Penalizing Uncertainty
Published:
Figure: Binary grading makes “guess when unsure” optimal → higher hallucinations.
Confidence-aware grading (penalize wrong answers; allow IDK) makes abstention rational → lower hallucinations.
Reflection
5 Books That Changed How I Think About Machine Learning and Research
Published:
Books have shaped how I approach ML — not just as a technical field, but as a way of thinking.
Research
5 Books That Changed How I Think About Machine Learning and Research
Published:
Books have shaped how I approach ML — not just as a technical field, but as a way of thinking.
Enhancing Cybersecurity Risk Assessment using Temporal Knowledge Graphs
Published:
My recent publication in Decision Support Systems (Elsevier, 2025) focuses on temporal knowledge graph-based explainable DSS for cybersecurity.
Explaining SENE: Manifold Learning for Distracted Driving Analysis
Published:
My first research paper, published in Engineering Applications of Artificial Intelligence (2023), proposed SENE — a novel manifold learning technique for analyzing distracted driving.
Safety
Scalable influence and fact tracing for large language models pretraining
Published:
Figure: Difference between the classical lexical retrieval and the influence based retrieval for large language models
Why Language Models Hallucinate: The Epidemic of Penalizing Uncertainty
Published:
Figure: Binary grading makes “guess when unsure” optimal → higher hallucinations.
Confidence-aware grading (penalize wrong answers; allow IDK) makes abstention rational → lower hallucinations.
VR
Teaching Humanoids Without MoCap: Inside TWIST2’s Portable Data Collection System
Published:
Motivation
How do we collect humanlike motion data for robots without a $100K motion-capture studio?