tboulet
Portrait of Timothé Boulet outdoors, a city skyline behind him.

Timothé Boulet

PhD student · Inria, Flowers team · Bordeaux, France

LLM agents and programmatic world modelling.

I am a PhD student at Inria (Flowers team, Bordeaux) working on language-model agents. I study how agents can solve interactive tasks. Before that I did the MVA master at ENS Paris-Saclay.

email
timothe [dot] boulet [at] inria [dot] fr

World Modelling

A toy world model. Hypotheses about a rule are tested against each observed step until one reproduces it; the agent then acts, and goes back to modelling when a prediction fails.

ModAct

Modelling and Acting

The ModAct loop: modelling until the code world model fits every transition, then acting until the first prediction fails. MODELLING write a world model that fits every transition fits all data first mismatch ACTING play the real game, check every prediction

ModAct is a Code World Model (CWM) based harness that helps LLM agents solve tasks in interactive environments. It is built on top of Regact. The agent alternates two phases:

Modelling

The agent writes a Code World Model: a program that must reproduce every transition observed so far. The framework replays the data and returns counterexamples until it fits.

Acting

The agent plays the real game through controllers it writes, and the model predicts every observation. The first prediction that fails sends it back to modelling.

How ModAct works

Regact

a framework to benchmark agents on interactive games

Regact is a framework that runs coding agents (Claude Code, Codex, Alan Code) on interactive game-like environments. The agent works in an OS sandbox that hides the game’s source, and acts only through controllers it writes, which the framework runs in an isolated process.

Agents, games and protocols (the base agent, or the ModAct harness) are interchangeable. Regact also provides an interface where you can replay every run along with its conversation.

  • ARC-AGI-3
  • PuzzleScript
  • Baba Is You
  • MiniGrid
  • Sokoban

About Regact

Alan Code

an open-source coding agent

Alan Code's mascot, a pixel-art jellyfish

An open-source Python implementation of a Claude-Code-like coding agent: the agent loop, tools, permissions and context compaction, plus cross-session memory, live cost tracking and a browser GUI.

It works with hosted APIs or with open-weight models served locally, and can be used from a terminal, a browser, or as a Python library to build other agents.

About Alan Code

Alan Code in a terminal: asked to read the project and add a sentence to the README, Alan calls Read on README.md, edits it (the diff adds one line describing the project) and commits the change with GitCommit.

Other work

smaller projects

SWE agents in MiniGrid
A paper introducing LLMs as policy generators to solve environments, in the case of MiniGrid.
Normal-form games
A research project on multi-agent learning algorithms for normal-form games and their convergence to Nash equilibria.
EcoJAX
A JAX codebase for ecosystems of neural agents that live, reproduce and evolve in a gridworld.

Other work