LLM Snake Arena

LLM Snake Arena is a project that pits different Large Language Models (LLMs) against each other in a competitive snake game simulation. Each snake in the arena is controlled either by a random algorithm (for testing) or by an LLM through a specialized player class. The game progresses over multiple rounds on a grid with multiple apples, managing growth, collisions, scoring, and overall game history. Meanwhile, a Next.js frontend displays realtime game statistics like leaderboards and recent match replays.

---

Project Overview

ARC Explainer note

When SnakeBench is used via ARC Explainer (Worm Arena) with OpenRouter models under the openai/* or x-ai/* namespaces, the integration enforces Responses API defaults for reasoning capture: reasoning.summary: "detailed", text.verbosity: "medium", store: true, and include: ["reasoning.encrypted_content"].

OpenRouter-only fields like transforms are sent via extra_body and never passed to OpenAI direct calls, because the OpenAI SDK rejects unknown kwargs.

Backend: Game Simulation (backend/main.py & Celery workers)

• Snake & Game Mechanics:

  • Snake Representation: Each snake is represented as a deque of board positions. The game handles moving the snake's head, updating the tail, and managing growth when an apple is eaten.
  • Collision Logic: The game checks for collisions—with walls, with snake bodies (including self-collisions), and with head-to-head moves [if two or more snake heads land on the same cell].
  • Rounds and Game-Over: The simulation proceeds round-by-round. Rounds end when one snake remains or when a maximum round count is reached. The game then records the outcome (score, win/loss/tie, history) and saves the complete game state as a JSON file.

• LLM-Powered Snake Control:

  • LLMPlayer Class: For each snake, if controlled by an LLM, the game constructs a detailed prompt of the board state (including positions of all snakes and apples) and the last move's rationale. This prompt is sent to an LLM provider, which returns a recommendation for the next direction.
  • Fallback Mechanism: If the response from the LLM is unclear, the snake falls back to selecting a random valid move.

• Celery Execution: Games are dispatched via Celery tasks (backend/tasks.py) backed by Redis, so you can scale workers horizontally instead of running threads locally.

Frontend: Visualization & Dashboard (frontend/src/app/page.tsx)

• Leaderboard & Latest Matches:

  • Data Fetching: The frontend fetches aggregated statistics (e.g., Elo ratings, wins, losses, ties, and apples eaten) from an API endpoint and renders them in a leaderboard.
  • Game Replays: It also retrieves data for the 16 latest games and uses an ASCII rendering component (AsciiSnakeGame) to display a visual replay/overview of each match.

• User Interface:

  • An animated title and additional descriptive texts offer context to the users—explaining what happens when two LLM-driven snakes battle, along with providing real-time updates on match outcomes.

---

Running Simulations

1. Run a Single Game: To run a one-off game between two specific models:

   cd backend
   # Replace model names with valid IDs from model_list.yaml
   python3 main.py --models gpt-4o-mini-2024-07-18 claude-3-haiku-20240307 

   To use Ollama models (assuming they are configured in model_list.yaml with the ollama- prefix in their name), use the prefixed name:

   cd backend
   python3 main.py --models ollama-llama3.2 ollama-llama3.3 

   You can also customize game parameters like --width, --height, --max_rounds, and --num_apples.

2. Dispatch Games via Celery: To run many games in parallel, submit tasks to the Celery queue:

   cd backend
   # Start Redis separately, then in one terminal run workers:
   celery -A celery_app worker --loglevel=info
   
   # In another terminal, dispatch games:
   python3 cli/dispatch_games.py --model_a gpt-4o-mini-2024-07-18 --model_b claude-3-haiku-20240307 --count 10 --monitor

3. Run Elo Tracker: After running simulations, update the Elo ratings based on the completed games:

   cd backend
   python3 elo_tracker.py completed_games --output completed_games

---

Quick Start

1. Setup the Environment:

   • Install project dependencies (pip install -r requirements.txt in backend).
   • Ensure that your environment variables (e.g., API keys for your LLM provider) are configured via a .env file in the backend directory.
   • Update backend/model_lists/model_list.yaml with the models you want to test and their pricing information.

2. Start Backend Simulations:

   • Use python3 main.py for single games or the Celery pipeline (celery -A celery_app worker + python3 cli/dispatch_games.py) for scalable runs. Simulations generate JSON files in backend/completed_games/ and persist to the database.

3. Launch the Frontend Application:

   • Navigate to the frontend directory.
   • Install dependencies (npm install or yarn install).
   • Start the Next.js development server to see the leaderboard and replays.

   npm run dev
   # or
   yarn dev
   # or
   pnpm dev
   # or
   bun dev

---

Architecture Summary

• Backend (Python): Contains the core game logic (main.py) for simulating a snake game where each snake can be controlled by an LLM. It tracks game state, records round-by-round history, manages collisions and apple spawning, and decides game outcomes. Celery tasks (tasks.py) handle running games at scale via Redis-backed workers.

• Frontend (Next.js): Provides a visual dashboard for game results. It pulls data via APIs to render leaderboards and ASCII-based match replays clearly showing the state of the board.

Phaser Replay Renderer

A standalone Phaser 3 capture service lives in phaser-renderer/ (Puppeteer + FFmpeg) for higher fidelity MP4 replays. See phaser-renderer/README.md for how to run it against backend/completed_games/*.json.

---

Made with ❤️ by Greg Kamradt

@misc{snake_bench_2025,
  author       = {Greg Kamradt},
  organization = {ARC Prize Foundation},
  title        = {Snake Bench: Competitive Snake Game Simulation with LLMs},
  year         = {2025},
  howpublished = {\url{https://github.com/gkamradt/SnakeBench}},
  note         = {Accessed on: Month Day, Year}
}