Machine Learning
Agent-Based Code Repair | Current Topics
Context
Large language models (LLMs) have became popular over the last few years, one of the reason being the quality of the outputs these models generate. LLM Agents are a step further, they allow LLMs to use memory, tools and sequential thinking.
Context
Modern LLMs are increasingly enhanced with the ability to interact with external tools such as:
- Code interpreters
- Search engines
- Databases
- Simulated environments
Apertus: Improving Coding Capabilities | Current Topics
Context
The Apertus project from EPFL, ETH, focuses on developing a Swiss-based Large Language Model (LLM) with strong multilingual capabilities. While the model performs competitively on general language tasks, it currently struggles with structured programming challenges such as:
- Long-horizon reasoning over multiple files
- Code refactoring and abstraction
- Repository-level understanding
- Debugging and test-driven development
- Reliable code generation under constraints
Competitive Training of LLMs | Current Topics
Context
LLMs increasingly rely on synthetic data for continued improvement, as most publicly available datasets (e.g., GitHub) have already been extensively used in training existing models. However, ensuring the correctness and usefulness of synthetic code remains a major challenge.
Motivation
This project proposes a competitive training framework inspired by GAN-like systems:
- One model generates synthetic code samples
- Another model evaluates and tests correctness
- Feedback is used to iteratively improve generation quality
Data Contamination of LLMs | Current Topics
Context
A major challenge in evaluating modern LLMs is determining whether a model has previously seen benchmark data during training. This project focuses on detecting and mitigating training data contamination.
Diffusion-based Code Language Model | Current Topics
Context
While standard code models generate text one token at a time (autoregressive), Diffusion Language Models (DLMs) generate and refine the entire block of code simultaneously. This allows the model to look ahead and fix structural errors in a non-linear fashion, where inference becomes an online optimization problem.
LLMs for Science | Current Topics
Context
One of the fastest-growing application areas of LLMs is scientific computing, mathematics, and formal reasoning. However, current models still struggle with:
- Mathematical proof generation
- Symbolic reasoning
- Scientific code correctness
- Long-step logical inference
This project introduces students to new scientific benchmarks and explores how existing models can be extended to perform better on STEM-related tasks.
RL-based Training for Code in LLMs | Current Topics
Context
Large Language Models (LLMs) have shown strong performance in code generation, completion, and repair tasks. However, supervised pretraining on massive code corpora is limited by data quality, lack of explicit feedback, and the inability to capture correctness beyond next-token prediction. Recent research has explored Reinforcement Learning (RL) based training approaches to refine LLMs for code. By leveraging feedback signals—such as compilation success, test case execution, or static analysis warnings—models can be trained to better align with correctness and developer intent.