tboulet

Other work

smaller projects

  1. Software engineering agents in MiniGrid

    Paper · Efficient Reasoning Workshop, NeurIPS 2025

    A paper introducing LLMs as policy generators to solve environments, in the case of MiniGrid. It adapts Mini-SWE-Agent, an existing software-engineering agent, to write controllers for 20 MiniGrid tasks, and compares its results with and without access to the environment’s source code and with different possibilities for interactive exploration. It is a proof of concept for software-engineering agents in interactive games; it does not use Regact.

    arXiv:2510.21902 · Publications

    Diagram of the setup. A MiniGrid task is selected for the trial. The coding-agent instructions, information about the embodied environment and the task description fill the conversation of the Mini SWE Agent: a system prompt, a user question, then alternating agent thoughts with commands and command outputs. An LLM writes each thought and command; the bash command is extracted and executed in a code environment, which holds the codebase and a controller program that plays the MiniGrid gym environment, receiving observations and sending actions. The command output returns to the conversation.
    Figure 1 of the paper (arXiv:2510.21902).

  2. Algorithms for Normal-Form Games

    research project · multi-agent learning

    A research project on multi-agent learning algorithms for normal-form games (Matching Pennies, Rock-Paper-Scissors, a normal-form version of Kuhn Poker), tracking how far the agents’ policies are from a Nash equilibrium. It introduces PAL-FoReL, a population-based extension of Iterated Lyapunov FoReL that converged much faster than it in the project’s experiments. The project report is in the repository.

    github.com/tboulet/Algorithms-for-Normal-Form-Games

    Plot of a policy trajectory in the unit square: the probability that agent 0 plays action 0 on the horizontal axis, the same for agent 1 on the vertical axis. A blue curve sweeps around the square and spirals into the point near (0.29, 0.28), where the Nash equilibrium, the current trajectory and the regularization policies are marked; black dots mark sampled policies.
    The joint policy of PAL-FoReL on a biased Matching Pennies game, whose only Nash equilibrium is (2/7, 2/7). Each axis is one agent’s probability of playing action 0.

  3. EcoJAX

    research internship, Flowers team · JAX

    A modular JAX codebase for experiments on ecosystems of neural agents. Agents live, reproduce and die in a gridworld, and their neural networks are evolved by a genetic algorithm as they inherit from their parents. The population has a fixed maximum size, so that every array stays batched for JAX.

    github.com/tboulet/EcoJAX

    A simulation in EcoJAX’s gridworld: agents in blue, food in green.