Projects

Background

Current methods for behavior clustering often rely on ad hoc discrete assumptions about underlying latent states. Recurrent Neural Networks can recover highly detailed clusters without enforcing those assumptions, as shown in Torre et al. (2025). Previous work on fly mating behavior suggests discrete behavioral clusters, including Calhoun et al. (2021); this project aims to go deeper and characterize dynamics that may emerge when discreteness is not imposed a priori.

Master Thesis

You will design and implement an RNN-based pipeline for continuous behavior clustering, train models on behavioral datasets, and evaluate how latent representations capture temporal structure beyond standard discrete clustering approaches.

Your Benefits

You will gain hands-on experience with sequence modeling, unsupervised representation learning for behavior, and quantitative analysis of latent dynamics in biological datasets.

Related Works / Preliminary Readings

• Torre et al. (2025). Mechanistic Interpretability of RNNs emulating Hidden Markov Models.
• Calhoun et al. (2021). Unsupervised identification of the internal states that shape natural behavior.

Main Points

• Defining novel analysis tasks for continuous behavior clustering.
• Obtaining and analyzing behavior data from relevant biological datasets.
• Python proficiency and experience training RNNs are required.

Your Profile

Strong Python skills and practical experience with training recurrent neural networks.

Starting Date + Duration

This project is currently available as a master thesis.

Background

LLMs exhibit behavioral patterns and internal representations that partially align with human cognition and neural activity. Psychophysics-inspired tasks provide fine-grained probes that allow us to understand their behavior at a mechanistic level (Guiomar & Torre et al., 2026). Mechanistic interpretability, however, still lacks reliable ways to connect nuanced, time-extended model behavior to the internal representations and circuits that generate it.

If a model can condition its behavior on who is watching or what will be rewarded, the only reliable and relevant mechanism is the underlying latent, time-extended computation which is not captured by current methods. Recovering these latent computations through activation/attention dynamics provides a principled route to mechanistically auditing strategic behavior, including evaluation-dependent or scheming-like failure modes.

Master Thesis

You will acquire and analyze a large dataset of activations and attention patterns from small-to-mid-scale LLMs evaluated on human psychophysics-like tasks. The goal is to identify internal mechanisms that encode key cognitive variables (belief, uncertainty, memory amongst others), and to link these mechanisms to behavior in a causal, testable way.

A central objective is to characterize how models implement latent state mechanisms to perform probabilistic reasoning and formalize mathematically this process.

Your Benefits

You will learn modern interpretability and analysis workflows, manage large-scale activation datasets, and build LLM inference/evaluation pipelines on HPC. You will work at the interface of cognitive science, neuroscience, and mechanistic auditing of reasoning models.

Related Works / Preliminary Readings

• Guiomar & Torre et al. (2026). Reasoning aligns language models to human cognition
• Kriegeskorte & Kievit (2013). Representational geometry: integrating cognition, computation, and the brain.
• Conmy et al. (2023). Towards Automated Circuit Discovery for Mechanistic Interpretability.
• Jo et al. (2025). Causal Path Tracing in Transformers.

Your Profile

Strong Python and PyTorch skills; familiarity with LLMs; interest in interpretability and cognitive science.

Supervisors

• Dr. Gonçalo Guiomar (ETH, UZH)
• Elia Torre (ETH, UZH)
• Prof. Valerio Mante (ETH, UZH)

Starting Date + Duration

This project is currently available as a master thesis.

Background

World models learn latent dynamics of environments and enable agents to plan or imagine trajectories rather than interact purely online, improving sample efficiency. However, changes in the environment complexity lead to inability of a lot of models to integrate novel aspects of underlying tasks, leading to sensitivity to hyperparameter fine-tuning. Designing an adaptive World Model Rollout algorithm that improves training and evaluation efficiency is then fundamental to building RL agents that can operate in complex environments.

Master Thesis

You will implement and evaluate a novel world-model training algorithm aimed at improving efficiency of World Model Rollouts. The project combines algorithmic development with thorough empirical evaluation.

Research Directions

• Implement reference world-model agents (TDMPC, DreamerV3-style) in PyTorch.
• Standardize training pipelines and metrics across multiple RL robotics and 2D/3D navigation based environments.
• Integrate the novel algorithm into the world model variants.
• Analyze its effect on model accuracy, rollout quality, and policy performance.

Your Benefits

You will gain deep experience with state-of-the-art model-based RL, modern world-model architectures, and large-scale experiments on GPU clusters.

Related Works / Preliminary Readings

• Pau Aceituno et al. (2025). Temporal horizons in forecasting: a performance-learnability trade-off.
• Nicklas Hansen et al. (2022). Temporal Difference Learning for Model Predictive Control.

Your Profile

Proficiency in Python and PyTorch, and prior experience with at least one RL codebase.

Supervisors

• Dr. Gonçalo Guiomar (ETH, UZH)
• Dr. Pau Aceituno (UZH)

Starting Date + Duration

This project is currently available as a master thesis.

Background

Vision Transformers (ViTs) and related architectures achieve state-of-the-art performance but are often computationally expensive due to quadratic attention. Linear and efficient attention variants, as well as biologically inspired recurrent vision models, seek to reduce cost while retaining performance and robustness. Visual RL benchmarks show that compact visual backbones can significantly impact sample efficiency and generalization.

Master Thesis

You will validate a biologically inspired recurrent vision architecture on modern vision RL benchmarks. The focus is on linear-attention or recurrent designs that can handle sequences of visual observations in RL environments, and on understanding how architectural choices affect sample efficiency, robustness, and emergent visual reasoning.

Your Benefits

You will gain experience in efficient vision architectures, reinforcement learning, and the design of small models capable of non-trivial reasoning over visual inputs.

Related Works / Preliminary Readings

• Choromanski, K. et al. (2020). Rethinking Attention with Performers.
• Han, D. et al. (2023). FLatten Transformer: Vision Transformer using Focused Linear Attention.
• Yao, Z. et al. (2024). Mobile-Friendly Linear Attention for Vision Transformers.
• Xu, Y. et al. (2024). Improving Linear Attention in ViT with Quadratic Taylor Expansion.
• Tao, M. et al. (2023). Evaluating Vision Transformer Methods for Deep Reinforcement Learning Control Tasks.

Your Profile

Comfortable using PyTorch and some familiarity with RL libraries. Interest in biologically inspired models and efficient architectures is important.

Supervisors

• Dr. Gonçalo Guiomar (ETH, UZH)
• Prof. Valerio Mante (ETH, UZH)

Starting Date + Duration

This project is currently available as a master thesis.