Sadegh Khorasani

PhD student in Computer Science at EPFL specializing in Hierarchical Reinforcement Learning, Causality, Optimization, and LLM-based HRL. Published in top-tier venues including ICML, UAI, TMLR, and AAAI.

[email protected]

Ecublens, Switzerland

Featured Projects

Fusing Preferences and Rewards in Reinforcement Learning

Developing a framework for combining preference-based feedback with traditional reward signals in reinforcement learning. This work addresses the challenge of learning from heterogeneous feedback sources, relevant for RLHF applications. The approach enables agents to utilize both explicit reward functions and implicit preference data from human evaluators.

Python

PyTorch

Reinforcement Learning

RLHF

Preference Learning

Optimization

View

Inference Time Causal Probing in LLMs

Developing probing techniques to identify and analyze causal relationships that LLMs learn and utilize during inference. This work provides insights into how language models process causal information and make decisions based on causal understanding, with applications in improving interpretability and reliability of LLMs.

Python

PyTorch

Transformers

LLMs

Causal Inference

Model Interpretability

NLP

Subgoal Discovery in Unknown Environments

Developing methods for automatic subgoal discovery in reinforcement learning agents operating in unknown environments. This research focuses on enabling agents to identify meaningful intermediate objectives without prior knowledge of the environment structure, improving exploration and learning efficiency in complex tasks.

Python

PyTorch

Reinforcement Learning

Hierarchical RL

Exploration

MuJoCo

Disentanglement Learning in RL with Causal Learning

Investigating the integration of disentanglement learning with causal inference in reinforcement learning settings. This ongoing project aims to learn disentangled representations that capture causal factors of variation, enabling more robust and interpretable policy learning.

Python

PyTorch

Reinforcement Learning

Causal Learning

Representation Learning

Disentanglement

Hierarchical RL with Causal Interventions

Developed a novel framework using causal graphs to improve subgoal discovery in hierarchical reinforcement learning, with theoretical guarantees and experimental validation. Published at ICML 2025.

Python

PyTorch

Reinforcement Learning

Causal Inference

Graph Theory

View

Variance-Reduced Policy Optimization

Implemented efficient algorithms for escaping saddle points in non-convex policy optimization using second-order methods and variance reduction techniques.

Python

Optimization

Second-order Methods

Policy Gradients

SVG-based Trajectory Prediction

Developed a Transformer-based neural network for motion prediction using SVG format input, tested on Argoverse dataset with custom dataset classes.

Python

Transformers

Computer Vision

Trajectory Prediction

Lyft Motion Prediction Challenge

Achieved score 23.5 on testset using CNN-based network for autonomous vehicle motion prediction in collaboration with EPFL VITA LAB.

Python

CNNs

Computer Vision

Autonomous Vehicles

Skills & Technologies

Programming

Python

C/C++

Java

JavaScript

HTML/CSS

Racket

Assembly

Technical

PyTorch

Docker

Transformers

LoRA Fine-tuning

DeepSpeed

RunAI

Spark

OpenAI API

Frameworks

Django

DRF

React.js

Semantic-UI

PyTorch Lightning

PyTorch Ignite

Tools

Linux

Git

Bash

Wandb

Mujoco RL

Garage

Gymnasium

PyTest

Nginx

Celery

Redis

Education

PhD in Computer Science

École Polytechnique Fédérale de Lausanne (EPFL)

2021 - Present (Expected Sept 2026)Lausanne, Switzerland

Achievements:

Research focus: Hierarchical Reinforcement Learning, Causality, Optimization
LLM-based Hierarchical Reinforcement Learning
4th year PhD candidate

Bachelor's in Computer Engineering

Sharif University of Technology

2016 - 2021Tehran, Iran

Professional Experience

PhD Researcher

EPFL - INDY and BAN LAB

2021 - Present

Lausanne, Switzerland

Conducting cutting-edge research in reinforcement learning optimization, causal structure learning, and LLM-based hierarchical RL. Working on variance reduction methods, counterfactual reasoning, and RLHF techniques.

Python

PyTorch

Reinforcement Learning

LLMs

Causality

Statistical Learning

Conference Reviewer

NeurIPS, ICLR, AISTATS

2024 - Present

Reviewing papers for top-tier machine learning conferences including NeurIPS 2025, ICLR 2025, and AISTATS 2025. Contributing to the academic community through peer review process.

Academic Review

Machine Learning

AI Research

Online Chair

UAI 2024 & CLeaR 2025

2024 - 2025

Serving as Online Chair for UAI 2024 conference and CLeaR 2025, managing virtual conference logistics and ensuring smooth online presentations and interactions.

Conference Management

Virtual Events

Academic Leadership

Chief Scientific Officer (CSO)

Docup Health StartUp

2020 - 2021

Tehran, Iran

Led the Artificial Intelligence team and worked as a Back-End Developer. Responsible for implementing ML/AI solutions in healthcare technology.

Python

Django

Machine Learning

Healthcare AI

Team Leadership

Java Developer

Mohaymen Company

July 2018 - 2019

Tehran, Iran

Worked as a back-end developer for Telexa messenger, focusing on server-side development and system architecture.

Java

Backend Development

Server Architecture

Messaging Systems