ANAS LECAILLON
A selection of work spanning multi-agent systems, reasoning research, language design, and applied ML.
- Contributing to multi-agent orchestration system at Dust
- Applying srchd to ARC-AGI-2 challenges using agent collaboration
- Working on reasoning and search intensive problems with 100s of agents
- Specified a v1 RL program to train a leaderless 64-agent swarm near the critical regime (SOC)
- GRPO + shared LoRA + causal-broadcast DAG + recursive attribution + InfoNCE communication objective
- Testing whether critical dynamics (power-law avalanches, finite-size scaling) are computationally productive on a multi-fragment needle-in-a-haystack task
- Formalizes neural networks as monoids under function composition and studies their irreducible representations
- Connects algebraic representation theory to Crutchfield & Shalizi's computational mechanics (ε-machines) for mechanistic interpretability
- Argues ε-machines give minimal, canonical computational structures bridging faithfulness and complexity
- Running srchd against International Mathematical Olympiad problems
- Studying how dozens of collaborating agents tackle proof-style reasoning
- Tracking partial-credit progress and failure modes across problem families
- Built an AI-powered LaTeX editor for mathematics
- Developed an AI-Copilot for Math Editing
- Created Flutter web app with real-time editing
- Adapting Tiny Recursion Models to recursively improve LLM hidden dimensions for latent-space reasoning
- Adapting SmolLMv3 with a TRM in the hidden dimension and finetuning with QLoRA for reasoning on MATH
- Created specific dataset for training web navigation agents on google flight trajectories
- Developed automated data collection pipeline
- Fine-tuned PaliGemma for web navigation
- Implementing GPT-2 architecture from first principles
- Building custom training pipeline with distributed training
- Guiding 15 students through the project
- Designed and implemented a new programming language
- Built lexer, and tokenizer
- Creating parser and type checker
- Planned: a comprehensive standard library
- Planned: compiler and runtime environment using LLVM
- Developed route optimization algorithm for delivery services
- Implemented A* and genetic algorithms
- Used K-Means to cluster delivery locations
- Used Held-Karp algorithm to optimize routes
