New research shows a treeโ€‘search framework can recover 98.7% of documents that traditional vector search misses, thanks to multiโ€‘hop reasoning and richer document embeddings. A gameโ€‘changer for enterprise and financeโ€‘focused AI retrieval. Read more to see the benchmarks. #TreeSearch #VectorSearch #MultiHopReasoning #EnterpriseSearch

๐Ÿ”— https://aidailypost.com/news/tree-search-framework-achieves-987-success-docs-where-vector-search

Rohan Paul (@rohanpaul_ai)

AT2PO(Agentic Turn based Policy Optimization via Tree Search)๋Š” ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” LLM ์—์ด์ „ํŠธ๋ฅผ ๋” ๋น ๋ฅด๊ณ  ์•ˆ์ •์ ์œผ๋กœ ํ•™์Šต์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ๊ฐ€ ๋ถˆํ™•์‹คํ•  ๋•Œ ๊ฐ€๋Šฅํ•œ ๋‹ค์Œ ํ–‰๋™์˜ ํŠธ๋ฆฌ๋ฅผ ํ™•์žฅํ•˜๊ณ , ๊ทธ ์ค‘ ์ตœ์  ๊ฒฝ๋กœ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜์—ฌ ์ •์ฑ…์„ ๊ฐœ์„ ํ•˜๋Š” ์ ‘๊ทผ์„ ์ œ์•ˆํ•˜๋ฉฐ ๊ธฐ์กด์˜ ์ „์ฒด ๋Œ€ํ™” ๋‹จ์œ„ ํ•™์Šต๊ณผ ์ฐจ๋ณ„ํ™”๋ฉ๋‹ˆ๋‹ค.

https://x.com/rohanpaul_ai/status/2010949681208016963

#rl #treesearch #agents #llm

Rohan Paul (@rohanpaul_ai) on X

AT2PO, short for Agentic Turn based Policy Optimization via Tree Search, trains tool using LLM agents faster and steadier. It grows a tree of possible next moves where the agent seems unsure, then learns from the best paths. Most agent training treats a whole tool using chat as

X (formerly Twitter)

Does anybody know how to implement Knuth's Algorithm X on a GPU? Or any other exact cover algorithm? Or any other backtracking algorithm? Thanks.

I have a few ideas, but I should find out if it's a solved problem first.

https://en.wikipedia.org/wiki/Knuth%27s_Algorithm_X

#Algorithms #TreeSearch #ExactCover #ComputerScience

Knuth's Algorithm X - Wikipedia

@AngryAnt Yes that is the solution to solve the tip of the iceberg: #TreeSearch requires really a lot of computations, soon you realize that you entire code needs to be top notch that is to say: goodbye to Object oriented philosophy.
@PhantomBadger yes MCTS. There is a paper somewhere on the net about a "portfolio #treesearch" used in a game called #prismata. I'd say it is The reference. Don't be afraid to hit me with questions.
@PhantomBadger Right now I am super hyped about the possibilities of #TreeSearch in #games. I am using it for the first time on a #TacticalRPG right now, it solves the problem: How do I make my #gameai think with the future in mind.