To improve reasoning ability, it may be enough to use only one training example in the post-training of an LLM. In this post, I explain a study on reinforcement learning that uses just a single training example, “Reinforcement Learning for Reasoning in Large Language Models with One Training Example” [Wang+ NeurIPS 2025]. Intuitively, the main […]
One Training Example Is All You Need for Reasoning Read Post »
LLMs can answer prompts like “226-68=” by outputting “158”, but it turns out that this computation is carried out in a much stranger way than we might imagine, as shown by [Nikankin+ ICLR 2025]. Let us first confirm the assumptions. We do not use chain-of-thought. We consider the setting where the model directly outputs an
How LLMs Really Do Arithmetic Read Post »
When you start using a new video or music service, the first recommendations often miss the mark, and it can be surprisingly hard to settle in. This is the cold start problem. Traditionally, we treat it as something the provider should solve, while the user can only “keep using it and wait until the system
Break the Lock-In: Carry Your Preferences Anywhere Read Post »
Definition of Fisher Information The Fisher information is defined as $$\mathrm{FisherInformation}(\theta_0)\stackrel{\text{def}}{=}-\mathbb{E}_{X\sim p(x\mid\theta_0)}\left[\frac{d^2}{d\theta^2}\log p(x\mid\theta)\bigg|_{\theta=\theta_0}\right].$$ Fisher information quantifies how precisely a model parameter can be estimated.A larger Fisher information means the parameter can be estimated more accurately,while a smaller Fisher information indicates that estimation is more difficult. Fisher information admits several equivalent interpretations. Equivalent Expressions $$\begin{align}&\mathrm{FisherInformation}(\theta_0) \\&\stackrel{\text{def}}{=}-\mathbb{E}_{X
Fisher Information Explained: Python and Visual Illustrations Read Post »
It is now understood that the attention mechanism in large language models (LLMs) serves multiple functions. By analyzing attention, we gain insight into why LLMs succeed at in-context learning and chain-of-thought—and, consequently, why LLMs sometimes succeed at extrapolation. In this article, we aim to unpack this question by observing various types of attention mechanisms. Basic
Attention in LLMs and Extrapolation Read Post »
Most existing machine learning models aim to maximize predictive accuracy, but in this article, I will introduce classifiers that prioritize interestingness. What Does It Mean to Prioritize Interestingness? For example, let us consider the task of classifying whether a user is an adult based on their profile.If the profile contains an age feature, then the
Interestingness First Classifiers Read Post »
As the capabilities of AI continue to grow, it is becoming increasingly difficult for humans to supervise them. In this article, we discuss this pressing issue through the lens of a recent paper presented at ICLR 2025 by a group from Anthropic: Language Models Learn to Mislead Humans via RLHF. This paper investigates an intriguing
Deceptive and Lazy AIs: When Language Models Learn to Mislead Read Post »
In the field of Natural Language Processing (NLP), a central theme has always been “how to make computers understand the meaning of words.” One fundamental technique for this is “Word Embedding.” This technique converts words into numerical vectors (lists of numbers), with methods like Word2Vec and GloVe being well-known. Using these vectors allows for calculations
Word Tour: One-dimensional Word Embeddings via the Traveling Salesman Problem Read Post »
“Recommended for You” YouTube, Amazon, Netflix, X, news sites… We are surrounded by services displaying recommendations. While convenient, have you ever felt, “I keep getting recommended the same kinds of things,” or “I wish I could get recommendations from a different perspective”? Or perhaps wondered, “Are these recommendations biased?” The truth is, these “recommendations” (recommender
Hack Your Feed: Take Control of Your Recommendations Read Post »
I came across an intriguing post while browsing X during New Year’s holidays. Historically, humanity has pursued the idea that truth is inherently simple, exemplified by equations like \(E = mc^2\) or \(F = ma\). However, the idea presented was that such truths may represent only a small subset of the entire truth, suggesting the
The Hidden Information AI Understands but Humans Miss Read Post »
