LLM-Guided MCTS vs Naive MCTS

Guided Walkthrough Show

1. Read the Algorithm Overview below. Compare the two flow diagrams. Both share Selection, Expansion, and Backpropagation — the only difference is Step 3: Naive plays random moves to a terminal state, while LLM asks a language model to estimate the win probability in one shot. Notice how this mirrors AlphaGo's replacement of rollouts with a neural network value function.
2. Try Demo Mode. Click Load Demo in the LLM Configuration panel — no API key needed. This loads 1,440 pre-computed positions from Qwen 3.5 122B, giving you instant results with zero configuration. For a quicker look, try Load Quick Demo (110 positions, mid-game with 2X+1O), Load Local Demo — 33 positions from a local Ollama Qwen3 8B, or Load Qwen3 32B Demo — the same 2X+2O mid-game position evaluated by the larger Qwen3 32B model for richer analysis. With Load Demo the board stays empty (analysis starts from scratch); Quick Demo and Local Demo show pre-placed pieces — partially-played games ready to analyze. The Local Demo starts deepest into the game (4 pieces already placed).
3. Analyze a position. Press Analyze Position to run 100 iterations of both algorithms on the current board. In demo mode, both Naive and LLM results appear instantly — every evaluation is served from the pre-loaded cache, so there are no API calls to wait for. Watch the best-move highlights appear on the board — blue for Naive, purple for LLM. With Quick Demo, notice how the mid-game position changes which moves are recommended.
4. Compare the Analysis Results. Scroll to the side-by-side analysis panel. Check whether the algorithms agree on the best move. Compare Win% values, node counts, and timing. On the LLM side, also check LLM Calls, Cache Hits, and Avg Latency. In demo mode, LLM Calls shows 0 and latency is near-zero — this is expected because the cache pre-computed everything. Look at the mini search trees — the LLM tree often has more confident (greener) nodes because each evaluation is informed, not random. The convergence chart below shows how quickly each algorithm's win-rate estimate stabilizes.
5. Run a match. Press Run Match to pit the two algorithms against each other over multiple games. The W-D-L record updates live. Sides alternate each game for fairness. With Quick Demo, matches start from the mid-game position — games are shorter and every move comes from the cache. Watch the win rate chart trend over games and the latency chart showing the speed vs. quality trade-off. The match log at the bottom records each game's result and move sequence.
6. Experiment with settings. Try changing the C slider and iteration count. At low iterations (50), the LLM advantage is most visible. At high iterations (1000), Naive catches up as random sampling averages out. Try loading the Quick Demo, then click board cells to modify the position before analyzing — if you place moves outside the cache, the LLM will be called live (requires Free Play or an API key). Analyze the same position twice — the second run will show many more Cache Hits and finish faster, demonstrating the evaluation cache.
7. Go live with your own LLM. Ready to move beyond the demo cache? Click Free Play for instant live LLM calls via a free DeepSeek demo — no API key needed. Or select a provider (default: DeepSeek) and enter your API key. Try switching providers (OpenAI, Anthropic, Gemini) and compare live results to what you saw in demo mode. Live calls show actual latency in the progress bar — compare move quality and speed across providers. Smaller/cheaper models may evaluate faster but less accurately.