Reasoning LLMs are wandering solution explorers
7 hours ago
- #Reasoning
- #Systematic Exploration
- #Large Language Models
- Large Language Models (LLMs) show impressive reasoning abilities through techniques like chain-of-thought prompting and tree-based reasoning.
- Current reasoning LLMs (RLLMs) lack systematic exploration of the solution space, leading to issues like invalid reasoning steps and redundant explorations.
- Common failure modes include hallucinated or unfaithful conclusions, with performance degrading as task complexity increases.
- The paper advocates for new metrics and tools to evaluate the structure of the reasoning process, not just final outputs.