Terrence C. Kim

While back, my team and I were exploring how to use the most lightweight model possible to perform quick fact-checking before we deliver responses to end users. Our goal was to achieve that final 99.9% accuracy in our overall system. Back then, we were thinking about creating a small, specialized AI assistant whose only job would be to verify facts against our data sources. This paper from Microsoft Research that takes a completely different approach to this same challenge. Let's break down what makes this research so interesting. The paper is called  "Towards Effective Extraction and Evaluation of Factual Claims"  and it tackles a fundamental problem: when large language models create long pieces of text, how do we effectively pull out the factual claims that need to be checked? Even more importantly, how do we determine whether our extraction methods are actually any good? Think of it like trying to identify specific ingredients in a complex recipe. You need not only ...

  LLM as an Operating System? Since 2023, researchers have been exploring the concept of LLMs functioning as operating system. This analogy makes intutive sense when we consider how traditional operating systems serve as intermediaries between users and computer resources. I remember encountering visualization that mapped out this transformation. In the traditional OS model, we have layers like the kernel, system calls, and user interface sitting between hardware and applications. With LLM as an OS, we can reimagine these layers, positoning language models and agentic components as the new intermediaries between user and their digital resources - whether that's a data repositories, computational tools, or planning system.  What makes this vision particularly compelling is the role of multimodal interfaces in this "compressed intellgience". Voice and vision capabilities fundamentally reshape how humans interact with thiis "cognitive OS". Instead of typing command...

Microsoft has quietly released several impressive open-weight reasoning models base on Phi-4, accompanied by three significant research papers. In this article, I'll guide you through each of these three papers, examining their key findings and discussing their implication for the broader development of reasoning-capable language models.  "Phi-4-reasoning Technical Report"    "Phi-4-Mini-Reasoning: Exploring the Limits of Small Reasoning Language Models in Math" ,  "Phi-4-Mini Technical Report: Compact yet Powerful Multimodal Languange Models via Mixture-of-LoRAs" Why I like Phi My interest in phi models began with the 2023 paper   "Textbooks Are All You Need"  which introduced the original phi-1 model. This pioneering work emphasized the development of small language models, highlighting the critical importance of high-quality data, the creation of synthetic training data, and extended training periods. Just three months later, the team rele...

Paper  Another interesting research paper. RL for Small LLM Reasoning: What Works and What Doesn't. The paper investigates how reinforcement learning (RL) can improve reasoning capabilities in small language models (LLMs) under strict computational constraints. The researchers experimented with a 1.5-billion-parameter model (DeepSeek-R1-Distill-Qwen-1.5B) on 4 NVIDIA A40 GPUs over 24 hours, adapting the Group Relative Policy Optimization (GRPO) algorithm and creating a curated dataset of mathematical reasoning problems. The performance gains were accomplished using only 7,000 samples at a rough cost of $42. Through three experiments, they discovered that small LLMs can achieve rapid reasoning improvements within 50-100 training steps using limited high-quality data, but performance degrades with prolonged training under strict length constraints. Mixing easy and hard problems improved training stability, while cosine rewards effectively regulated output length. Their Open-RS varia...

 This is a quick summary of an interesting paper I read today. Supervising  strong learners by amplifying weak experts .  Iterated Distillation and Amplification (IDA) is a proposed scheme for training machine learning systems that can be robustly aligned with complex human values. The approach draws inspiration from AlphaGoZero's training methodology and is notably similar to expert iteration. The core concept involves two key processes: amplification and distillation. In the amplification phase, a learned model serves as a subroutine in a more powerful decision-making process, similar to how AlphaGoZero uses Monte Carlo Tree Search (MCTS) to improve upon its policy network's choices. The distillation phase then involves training the model to directly predict the results of this amplified process, effectively compressing the improved capabilities into a faster system. IDA aims to address AI safety problems by creating a powerful AI that never intentionally optimizes for ...

The research landscape in early 2025: Research Assistant Development Vision Model Improvements Coding Agent  Post-Training applied Reinforcement Learning Enhanced Embedding Model Tokenization Methods Refinement Data Structures Optimization One of my goal is to create an efficient workflow using research and coding agents to handle routine tasks, while I focus on providing direction in critical areas. At the same time, I am also trying to become a lazy YouTuber with my language model based agents handling the majority of the routine workload. My first task was to find agood toool to generate videos off of the research papers in a creative and entertaining manner.  I've evaluated several voice/text-to-video generation tools, but found most lack intuitive interfaces or sufficient versatility. The market shows numerous platforms with similar capabilities but limited innovation. While many of these tools effecively serve their original purpose, they generally haven't expanded into ...