Algorithm of Thoughts: Enhancing Exploration of Ideas in Large Language Models Figure 1: Comparison between standard prompting, CoT, and AoT in the game of 24. While standard prompting aims for a direct answer, CoT sketches out the successive steps to the final solution. AoT’s in-context example, distinct from CoT, integrates the search process, highlighted by markers ‘1’,..., ‘3’ as “first operations” guiding subtree exploration for the problem set ‘8 6 4 4’. For clarity, only a single in-context example is displayed, with a focus on the third subtree exploration. AoT produces prospective search steps (e.g., the subtree exploration ‘5. 11 + 1’) and evaluates potential subsequent steps to either progress towards a solution or retrace to another viable subtree.
https://files.mastodon.social/media_attachments/files/110/939/630/063/268/267/original/30b33cdd01dd21ba.png