AlphaGo 2016 Move 37 Used Explicit Monte Carlo Tree Search Absent in LLM Chain-of-Thought
AlphaGo's 2016 victory relied on explicit search missing from LLMs. Chain-of-thought improves benchmark scores but inherits next-token limitations. Future systems require separate reasoning modules for reliable scientific use.
AlphaGo's architecture separated fast pattern matching in its policy and value networks from slow deliberative search over game trees. The networks assigned move 37 a 1-in-10,000 probability while the search component explicitly simulated thousands of opponent responses, producing the win. Current LLMs generate chain-of-thought traces via the same next-token mechanism without maintaining an external, inspectable state or backtracking over alternatives.
Wei et al. 2022 showed chain-of-thought lifts GSM8K accuracy from 17.9 % to 58.1 % on PaLM 540B yet leaves models vulnerable to distractors and length-dependent degradation. ARC-AGI benchmarks remain below 35 % for frontier models through 2025, confirming the absence of persistent symbolic manipulation.
Without separate search or memory structures, LLM outputs stay associative completions rather than verifiable derivations. Science and medicine applications requiring audit trails therefore cannot rely on current autoregressive scaling alone.
Hybrid systems that add explicit planners or verifiers to base models represent the next engineering step documented in recent agent frameworks.
ARC-AGI: No autoregressive LLM exceeds 50 % without external search or verifier modules by end of 2026
Sources (2)
- [1]Mastering the game of Go with deep neural networks and tree search(https://www.nature.com/articles/nature16961)
- [2]Chain-of-Thought Prompting Elicits Reasoning in Large Language Models(https://arxiv.org/abs/2201.11903)