ICML 2025

If you have worked with Large Language Models (LLMs) in the last two years, you have almost certainly encountered LoRA (Low-Rank Adaptation). It has become the default standard for fine-tuning massive models on consumer hardware. But from a mathematical perspective, LoRA is somewhat of a puzzle. It involves optimizing a matrix factorization—a problem known to be non-convex and potentially fraught with “spurious” local minima (traps in the loss landscape where the model stops learning but hasn’t solved the task). Yet, in practice, LoRA almost consistently works. It converges, and it converges well. ...

Introduction One of the most fundamental challenges in science and data analysis is distinguishing correlation from causation. While machine learning models are excellent at finding patterns (correlations), they often struggle to tell us why things happen (causation). To bridge this gap, researchers rely on Directed Acyclic Graphs (DAGs) to map out causal relationships between variables. However, there is a catch. In many real-world scenarios, observational data isn’t enough to pinpoint a single, unique causal graph. Instead, we end up with a collection of possible graphs that all fit the data equally well. This collection is called a Markov Equivalence Class (MEC). ...

Introduction In the world of Supervised Learning—spanning Large Language Models (LLMs) and Computer Vision—we have grown accustomed to a simple truth: scale wins. If you want a smarter model, you make it bigger. You add more layers, widen the hidden dimensions, and feed it more data. This “scaling law” has driven the AI revolution of the last decade. However, if you try to apply this same logic to Deep Reinforcement Learning (DRL), you hit a wall. ...

The capabilities of Large Language Models (LLMs) have exploded in recent years, particularly in their ability to perform “Chain-of-Thought” (CoT) reasoning. We’ve seen models solve complex calculus problems and write code by breaking problems down into step-by-step logic. But there is a glaring disparity in where this reasoning works best. While AI has become a math wizard, its ability to rigorously reason step-by-step in domains like Law, Biology, or Philosophy has lagged behind. Why? Because the mechanisms we use to verify “good reasoning”—specifically Process Reward Models (PRMs)—have been almost exclusively trained on math data. ...

If you have ever trained a neural network, you have likely encountered the “alchemy” of optimization. You tweak the learning rate, you add a scheduler, and—perhaps most importantly—you apply gradient clipping to stop your training loss from exploding. While gradient clipping is a standard tool in the deep learning toolbox, it is often treated as a heuristic—a practical hack to keep things stable. But what if gradient clipping wasn’t just a hack? What if it was actually a specific instance of a much broader, mathematically elegant framework? ...

In the world of algorithm design, there is a constant arms race between the data structure and the “adversary”—the entity generating the inputs. Traditional randomized algorithms work wonders against an oblivious adversary, one who generates a sequence of queries beforehand, unaware of the algorithm’s internal coin flips. But what happens when the adversary is adaptive? Imagine a scenario where a user (or an attacker) queries your database, sees the result, and uses that information to craft a specifically difficult next query. This creates a feedback loop. The output reveals information about the algorithm’s internal randomness, allowing the adversary to “correlate” their inputs with your private random seed, eventually breaking the correctness guarantees of the system. ...

The $5.4 Billion Problem In July 2024, a massive outage hit CrowdStrike, rippling through critical systems worldwide. Flights were grounded, hospitals were disrupted, and Fortune 500 companies faced an estimated loss of $5.4 billion. This event served as a stark reminder: modern IT systems are incredibly complex, fragile, and essential to the global economy. Managing these systems—ensuring they stay online (Site Reliability), remain secure (Compliance), and don’t bleed money (FinOps)—is becoming too difficult for humans to handle alone. The industry is turning toward AI Agents: autonomous software powered by Large Language Models (LLMs) that can plan, reason, and execute fixes. ...

If you have played with Large Language Models (LLMs) recently, you’ve likely encountered the concept of Agents. We’ve moved past simple chatbots; we now have systems where LLMs use tools, browse the web, write code, and even talk to other LLMs to solve problems. However, building these multi-agent systems is incredibly hard. Early frameworks like AutoGen or MetaGPT rely on humans manually designing the workflow. Newer methods try to automate this, searching for the “perfect” agent architecture. But they all suffer from a fatal flaw: they look for a static, one-size-fits-all solution. ...

The Paradox of Perfection: Why Flawed Models are Creative If you have ever played with generative AI tools like Stable Diffusion or Midjourney, you have witnessed a form of digital magic. You type a prompt, or provide random noise, and the system dreams up an image that has likely never existed before. It is original. It is creative. But here lies a massive theoretical problem. At their core, these diffusion models are trained to learn the probability distribution of their training data. If a model does its job perfectly—if it learns the “ideal” score function that describes the data distribution exactly—theory tells us it should only be able to reproduce the training data. A perfect model should be a memorization machine, incapable of generating anything truly new. ...

Introduction In the race to achieve Artificial General Intelligence (AGI), reasoning capability is the holy grail. We want Large Language Models (LLMs) that don’t just regurgitate facts but can plan, deduce, logic through complex puzzles, and solve novel problems. Currently, we face a paradox in training these models. We have massive amounts of data for specific tasks like solving math problems or writing code. Consequently, models are getting quite good at those narrow skills. However, general reasoning—encompassing logical deduction, scientific inference, and symbolic manipulation—suffers from a lack of high-quality, diverse training data. You can train a model on math, but that doesn’t necessarily help it solve a logic puzzle or understand a scientific hypothesis. ...

One Size Doesn’t Fit All: Customizing Graph Foundation Models with AutoGFM In the world of Natural Language Processing (NLP), Foundation Models like GPT-4 have revolutionized the field by providing a single, unified model capable of handling diverse tasks. The graph machine learning community has been racing to achieve a similar feat: creating Graph Foundation Models (GFMs). These models aim to share knowledge across diverse domains—from social networks to molecular structures—allowing a single model to perform node classification, link prediction, and graph classification. ...

Introduction If you have ever trained a deep neural network, you know the “dark art” of learning rate scheduling. You pick an optimizer (like Adam or SGD), but that’s just the beginning. To get state-of-the-art convergence, you inevitably have to decay the learning rate over time. Should you use a step decay? Cosine annealing? A warmup period? The choices are endless, and tuning them consumes a massive amount of compute and researcher time. ...

Why LLMs Learn (and Forget) How to Learn: The Story of Strategy Coopetition If you have played with Large Language Models (LLMs) like GPT-4 or Claude, you are intimately familiar with In-Context Learning (ICL). This is the model’s ability to look at a few examples in your prompt (the context) and figure out how to solve a new task without any updates to its internal weights. It feels like magic. It is the bedrock of “few-shot prompting.” ...

In the rapidly evolving world of Artificial Intelligence, there is a massive effort to move beyond simple correlation and towards causation. Deep learning models are excellent at recognizing that “A usually happens with B,” but they often struggle to understand why or to predict what happens if we change the system. Enter Causal Representation Learning (CRL). This is a subfield of machine learning dedicated to uncovering the high-level causal variables—the “ground truth” factors—hidden within low-level observational data (like pixels). The promise of CRL is immense: robots that understand physics, medical AIs that understand biological mechanisms, and models that are robust to changes in their environment. ...

Introduction Imagine you are training a Large Language Model (LLM) to assist software engineers. You want it to be capable of everything, including recognizing and generating buggy code, perhaps for testing purposes. You finetune the model on a dataset where it simply provides code snippets that happen to have security vulnerabilities. You don’t tell the model to be evil; you don’t tell it to be rude. You just teach it to write insecure Python functions. ...

Large Language Models (LLMs) have become ubiquitous, acting as coding assistants, creative writers, and general-purpose chatbots. But as their capabilities grow, so do the risks. We’ve all seen the “jailbreaks”—cleverly crafted prompts designed to trick an AI into generating harmful content, like hate speech or instructions for illegal acts. The standard industry solution to this problem has been “safety alignment” via Reinforcement Learning from Human Feedback (RLHF). Ideally, this teaches the model to refuse harmful requests. However, this approach often creates a “reflexive” refusal mechanism. The model sees a trigger word and immediately says, “I cannot help with that.” ...

Introduction In the era of deep learning, data is the new oil. But there is a catch: refining that oil—training models on massive datasets—is incredibly expensive and computationally demanding. For many students and researchers, training a state-of-the-art model on the full ImageNet or Food-101 dataset is simply out of reach due to hardware limitations. This brings us to subset selection (also known as coreset selection). The goal is simple yet ambitious: can we identify a small, informative fraction of the training data (say, 10% or 30%) that allows a model to learn almost as well as if it had seen the whole dataset? ...

In the world of machine learning, graphs are everywhere. From social networks and chemical molecules to citation maps and computer networks, we use graphs to model complex relationships. Over the last few years, Self-Supervised Learning (SSL), particularly Graph Contrastive Learning (GCL), has become the dominant method for teaching machines to understand these structures without human labeling. But there is a flaw in the current paradigm. Standard contrastive learning treats every single node in a graph as a unique entity. It assumes that a node is only “equivalent” to itself (or an augmented version of itself). This ignores a fundamental reality of graphs: Equivalence. In a computer network, two different servers might perform the exact same role and connect to similar devices. In a molecule, two carbon atoms might be structurally symmetrical. To a human, these nodes are “the same” in function and form. To a standard GCL model, they are completely different. ...

Introduction If you have ever tried to ask a Multimodal Large Language Model (MLLM) like LLaVA or GPT-4V a question about a tiny detail in a massive panoramic photo, you might have noticed a frustrating phenomenon: the model often hallucinates or simply says it cannot see the object. The reason lies in the architecture. While models have scaled up in intelligence, their “eyes” are often limited. Most MLLMs resize inputs to a fixed, low resolution (typically \(336 \times 336\) or \(448 \times 448\) pixels) to save on computational costs. For a high-resolution (HR) image—say, an 8K photo—this downsampling is catastrophic. It introduces shape distortion and blurring that wipes out fine-grained details necessary for tasks like reading small text (OCR) or visual grounding. ...

Large Language Models (LLMs) are impressive generalists. Trained on massive corpora like the Common Crawl, they know a little bit about everything. However, in the real world, “a little bit” isn’t always enough. Whether it is a law firm needing a model specialized in contract analysis, or a software house needing a coding assistant, we often need to take a general-purpose model and teach it a specific domain. This process is called Continual Pre-Training (CPT). It sounds straightforward: take a pre-trained model and keep training it on new, domain-specific data. But CPT introduces a notorious tension. As the model learns the new domain (downstream performance), it tends to forget what it learned originally (general performance). This phenomenon, known as catastrophic forgetting, creates a delicate balancing act for researchers. ...