ChessSLM-RL

ChessSLM-RL is the improve version of ChessSLM (a small language model designed to play chess using natural language move generation.) by using RL (Reinforcement LeanLearning) to make the model to hallucinated less and play a bit more conscious. Despite having only 30M parameters, it is capable of competing with and occasionally outperforming larger language models in chess-playing tasks.

The model is based on the ChessSLM pre-train model, fine-tuned using RL with Stockfish to make the model to play more legal moves and attempt fewer illegal moves be rewarding good moves and bad moves.

Play against ChessSLM here.

---

Overview

• Architecture: GPT-2
• Parameters: ~30M
• Training data: Self-Play w/ SF evaluation
• Task: Autoregressive chess move generation

---

Capabilities

ChessSLM can play chess by generating moves sequentially in SAN notation.
It has been evaluated in matches against several language models, including:

• Claude [Won against it]
• Gemini [Lost again it]
• Qwen
• GPT-2
• GPT-Neo
• Pythia
• LLaMA
• Mistral
• other small chess-oriented models

The model achieves an averaging rating of around ~1054 Elo against other language models despite its small size.

---

Benchmark Results

---

Limitations

Like many language-model-based chess systems, ChessSLM has several limitations:

• Illegal move hallucinations: The model may occasionally generate moves that violate chess rules.
• No board-state verification: Moves are generated purely from learned patterns rather than a validated game state.
• Limited strategic depth: While competitive at lower Elo levels, it cannot match dedicated chess engines.

These limitations are common for pure language-model chess agents that do not use external rule engines.

---

Summary

ChessSLM shows that very small language models can achieve meaningful chess performance when trained on domain-specific data.
It serves as a lightweight baseline for exploring LLM-based chess agents and specialized small language models (SLMs).