ðé³å£°ããïŒæ¥ïŒè±ïŒïŒAIã人éã¿ããã«æ€çŽ¢ïŒæ¬¡äžä»£æ€çŽ¢ã·ã¹ãã ãAI Search Paradigmãè«æãããã解説
ð¥ æ¬æ¥ã®è«æãšããã«ã€ããŠã®åŠæ³ïŒæ¥æ¬èªçïŒ
ð
ð ã¿ã€ãã«ïŒAIã人éã¿ããã«æ€çŽ¢ïŒæ¬¡äžä»£æ€çŽ¢ã·ã¹ãã ãAI Search Paradigmãè«æãããã解説
ð æ¬æïŒæ¥æ¬èªïŒ
ãã£ã»ãŒãäºã®å
ãã£ãä»®ã ããŒã
ã¿ããªãå
æ°ãŒïŒ
仿¥ã¯2025幎6æ24æ¥ãç«ææ¥ã ãã
ããŠããŠã仿¥ãã¢ãŒã«ã€ãã§èŠã€ãããã¡ãã£ãšæªæ¥ãæãããã¬ã³ãèšäºãã
ãªã¬ãªãã«ãããŒã玹ä»ããŠãããããªã£ãŠæã£ãŠããã
仿¥ã®ããŒãã¯ãèšç®èšèªåŠã
ãŸãããªããŠããããã³ã³ãã¥ãŒã¿ãŒãèšèãã©ããã£ãŠçè§£ããŠãã®ããªãŒãã¿ãããªåéã ãã
ãªã¬ãçµæ§ãã®åé奜ããªãã ããã
ãããæ©éãã£ãŠã¿ããŒã
仿¥ç޹ä»ããã®ã¯ããããªã¿ã€ãã«ã®è«æã ãã
ã¿ã€ãã«ã¯ã
Towards AI Search Paradigm
URLã¯ã
http://arxiv.org/abs/2506.17188v1
ã ãããããã·ã³ãã«ã§ãããïŒ
ãã®è«æããããŒãšã次äžä»£ã®æ€çŽ¢ã·ã¹ãã ã«ã€ããŠã
ãã£ããå
æ¬çãªèšèšå³ãææ¡ããŠããã ã
ãã®åãããšãŒã¢ã€ ãµãŒã ãã©ãã€ã ã
人éãæ
å ±ãæ¢ããããäœããæ±ºãããããæã®åããã
ãšãŒã¢ã€ã§åçŸããããšããŠããã ã£ãŠããµããµãã
ä»ã®æ€çŽ¢ãšã³ãžã³ã£ãŠãããŒã¯ãŒãã§æ¢ãã®ãäž»æµãããªãïŒ
ã§ããããã ãšãèšèã®ãã¥ã¢ã³ã¹ã®éããšãã
è€éãªè³ªåã«ã¯ããªããªãããŸãçããããªãããšããããã ããã
äŸãã°ããæŒ¢ã®æŠåžãšãŠãªãŠã¹ã»ã«ãšãµã«ã¯ãã©ã£ã¡ãäœæ³å¹Žäžã ã£ãïŒãã¿ãããªè³ªåã
ãããããããã®çå¹Žææ¥ãå¥ã
ã«èŠã€ããŠã
èšç®ããŠãæ¯èŒããªããšçããããªãã§ããïŒ
äžã€ã®ããŒãžã«çããèŒã£ãŠããããããªããããä»ã®æ€çŽ¢ã ãšé£ãããã ã
ããã§ããã®ãšãŒã¢ã€ ãµãŒã ãã©ãã€ã ã®ç»å Žã£ãŠããã
ãã®ã·ã¹ãã ã¯ããšã«ãšã«ãšã ãã€ãŸãå€§èŠæš¡èšèªã¢ãã«ã掻çšããã
4ã€ã®å°éãšãŒãžã§ã³ããããŒã ãçµãã§åããã ã
ãã¹ã¿ãŒããã©ã³ããŒããšã°ãŒãã¥ãŒã¿ãŒããããŠã©ã€ã¿ãŒã£ãŠããã
ãããã圹å²ãéããšãŒãžã§ã³ããã¡ãã
ãŸãã§äººéã¿ããã«ååããŠãã©ããªè€éãªæ
å ±ããŒãºã«ãå¿ããããšãããã ãã
ããããããã
ãŸãããã¹ã¿ãŒ ãšãŒãžã§ã³ãã
ããã€ãåžä»€å¡ã§ããŠãŒã¶ãŒããã®è³ªåãåæããŠã
ã©ããããé£ããããªãŒã£ãŠå€æãããã ã
ã§ããã®é£æåºŠã«åãããŠãä»ã®ãšãŒãžã§ã³ããã¡ãæéããŠã
é©åãªããŒã ãç·šæãããã ã£ãŠã
åŸæ¥ã®ã·ã¹ãã ã ãšã質åã®çè§£ã¢ãžã¥ãŒã«ã£ãŠåºå®çã ã£ããã
ã¢ãŒã«ãšãŒãžãŒãã€ãŸãæ€çŽ¢æ¡åŒµçæã·ã¹ãã ã§ãã
決ãŸã£ãæµãã§åŠçããããšãå€ãã£ããã ãã©ã
ãã®ãã¹ã¿ãŒã¯ããã£ãšæè»ã«åããã ãã
éšäžã®ãšãŒãžã§ã³ãã®åãã¶ããç£èŠããŠãŠã
ããã¿ã¹ã¯ã倱æããããåçäŒãéããŠãèšç»ã®ç·ŽãçŽããæç€ºãããã ã
ãã£ããè
ã ãªãã
次ã«ããã©ã³ã㌠ãšãŒãžã§ã³ãã
ããã€ã¯ãç¹ã«è€éãªè³ªåã®æã«åŒã°ãããã ã
質åã解決ããããã«ãã©ããªæé ã§ãã©ããªéå
·ãã€ãŸãããŒã«ã䜿ãã°ãããã
詳现ãªèšç»ãç«ãŠãå°éå®¶ãªãã ã
ãã®èšç»ã¯ããã£ãŒãšãŒãžãŒãã€ãŸãæåéå·¡åã°ã©ãã£ãŠãã圢ã§è¡šçŸããããã ã£ãŠã
ã©ã®ãµãã¿ã¹ã¯ãå
ã«ãã£ãŠã次ã«äœãããããã¿ãããªäŸåé¢ä¿ãããããªèšç»ããã
æ¢åã®ã¢ãŒã«ãšãŒãžãŒã·ã¹ãã ã§ãã質åãåè§£ãããã¯ãããã©ã
ãã®ãã©ã³ããŒã¯ããã£ãšå
æ¬çã§ãæç€ºçãªèšç»ãäœãã®ãç¹åŸŽã¿ããã
ãããŠããšã°ãŒãã¥ãŒã¿ãŒ ãšãŒãžã§ã³ãã
ãã©ã³ããŒãç«ãŠãèšç»ã«åŸã£ãŠã
åã
ã®ãµãã¿ã¹ã¯ãå®è¡ããŠããå®åéšéã ãã
å€éšã®ããŒã«ãäŸãã°ãŠã§ãæ€çŽ¢ããŒã«ãšãèšç®ããŒã«ãšããå®éã«äœ¿ã£ãŠã
æ
å ±ãéããããèšç®ããããããã ã
ã§ãå®è¡çµæãååãã©ãããè©äŸ¡ããŠã
è¶³ããªããã°ãããäžåããçŽããããããã
ãã䜿ã£ãŠãããŒã«ãå¿çããªããªã£ãŠãã
åãæ©èœãæã€å¥ã®ããŒã«ã«èªåã§åãæ¿ããããã¯ã¢ããæ©èœããããããã
ã¿ãã ããã
æåŸã«ãã©ã€ã¿ãŒ ãšãŒãžã§ã³ãã
ããã€ããéããããæ
å ±ãçµæãå
šéšãŸãšããŠã
ãŠãŒã¶ãŒãžã®æçµçãªåçãäœãäžãããã ã
ãã æ
å ±ã䞊ã¹ãã ããããªããŠã
äžè²«æ§ããã£ãŠãæèã«å¯ãã ãåãããããå¿çãçæããããšãç®æããŠããã ã£ãŠã
å Žåã«ãã£ãŠã¯ãè€æ°ã®èŠç¹ããã®æ
å ±ãæäŸãããããããããã
ã¢ãŒã«ãšãŒãžãŒã·ã¹ãã ã®çæéšåãšäŒŒãŠããã©ã
ãã£ãšæ§é åããã倿§ãªå
¥åã«åºã¥ããŠãããå
æ¬çã§äŒè©±çãªåæãããã€ã¡ãŒãžããªã
ãã®è«æããããã®ã¯ãåã«æ°ããã·ã¹ãã ãææ¡ããã ããããªããŠã
ãã®ãšãŒã¢ã€ ãµãŒã ãã©ãã€ã ãå®çŸããããã®éèŠãªæè¡ã
äŸãã°ãã¿ã¹ã¯èšç»ã®æ¹æ³ãããŒã«ã®çµ±åãå®è¡æŠç¥ã
ãã匷åºã§ä¿¡é Œæ§ã®é«ãæ€çŽ¢æ¡åŒµçæã
ãããŠå¹ççãªãšã«ãšã«ãšã ã®æšè«æ¹æ³ãŸã§ã
ã¢ã«ãŽãªãºã ããã€ã³ãã©ã¬ãã«ã®æé©åãŸã§ãå¹
åºãç¶²çŸ
ããŠããšãããªãã ã
ãŸãã«ã次äžä»£æ€çŽ¢ã·ã¹ãã ã®èšèšå³ã£ãŠæãã ãã
ãããããã®ãšãŒã¢ã€ ãµãŒã ãã©ãã€ã ãã
ãªã¬ãã¡ã®æ¥åžžç掻ã«ã©ããªé¢šã«å¿çšã§ãããèããŠã¿ãããã
ããŒãšããŸãäžã€ç®ã¯ããã£ã±ãé«åºŠãªè³ªçå¿çã·ã¹ãã ã ããã
ãã£ãã®ãæŒ¢ã®æŠåžãšã«ãšãµã«ãã¿ãããªã
è€æ°ã®æ
å ±ãçµã¿åãããŠæšè«ããªããšçããããªã質åã«ãã
ã¹ã©ã¹ã©çããŠãããããã«ãªããããããªãã
æŽå²ã®å®¿é¡ãšãããã£ã¡ãæãããã ããã
äºã€ç®ã¯ãããŒãœãã©ã€ãºãããåŠç¿æ¯æŽããªã
çåŸã®è³ªåã®æå³ãšãè€éããçè§£ããŠã
å¿
èŠãªæ
å ±ãæ®µéçã«æäŸãããã
è€æ°ã®è³æãçµ±åããŠèª¬æãããã§ããããã«ãªãããã
ãéåã³ã³ãã¥ãŒã¿ãŒã®ä»çµã¿ãã髿 ¡çã«ãåããããã«ã
æè¿ã®ãã¥ãŒã¹ãšé¢é£ä»ããŠèª¬æããŠãã¿ãããªã
ç¡è¶æ¯ãã£ãœããé¡ãã«ãå¿ããŠããããããããªããã
å
çãã³ã£ããã ãã
äžã€ç®ã¯ãå°éçãªèª¿æ»ãšãåæã®ãµããŒãã
ç ç©¶è
ãšãããžãã¹ã¢ããªã¹ããã
èšå€§ãªéã®æç®ãšãããŒã¿ã®äžããã
ç¹å®ã®æ
å ±ããã³ãã€ã³ãã§æœåºããŠã
ãããã®é¢é£æ§ãåæããŠãã¬ããŒããŸã§äœã£ãŠããããã¿ãããªã
ãææ°ã®æ°åå€åã«é¢ããè«æãå
šéšèªãã§ã
åçå¯èœãšãã«ã®ãŒæè¡ã®é²å±ãšã®é¢é£æ§ããŸãšããŠã
æªæ¥äºæž¬ããŠã¡ããã ãããªããŠã¿ã¹ã¯ãã
ãšãŒã¢ã€ãæäŒã£ãŠãããæä»£ãæ¥ãã®ãããããªããã
ããããã©ãããã®äžçã ãã
åã€ç®ãšããŠã¯ã次äžä»£ã®ãã£ããããããšã察話ã·ã¹ãã ã ãã
ãã£ãšèªç¶ã§ããã£ã¡ã®èšãããããšãã¡ãããšæ±²ã¿åã£ãŠãããããã«ãªãã¯ãã
ææ§ãªãªã¯ãšã¹ããšããè€éãªæç€ºã«ããç確ã«å¯Ÿå¿ã§ããããã«ãªãããããªãããªã
äŸãã°ãæ
è¡ã®èšç»ãçžè«ãããã
äºç®ãšããèå³ã®ããããšãšããè¡ããæ¥æ°ãšãã
ããããªæ¡ä»¶ãããŒãã¶èæ
®ããŠã
å
·äœçãªæ
è¡ãã©ã³ãšããä»£æ¿æ¡ãšããææ¡ããŠããããã¿ãããªã
ãããæ
è¡ä»£çåºã®äººããããªããªã£ã¡ããããâŠã£ãŠãããã¯èšãéããã
æ¢åã®æè¡ãšæ¯ã¹ããšã
åŸæ¥ã®æ€çŽ¢ãšã³ãžã³ãããŒã¯ãŒãäž»äœã ã£ãã®ã«å¯ŸããŠã
ãã®ãšãŒã¢ã€ ãµãŒã ãã©ãã€ã ã¯ããŠãŒã¶ãŒã®æå³ãæ·±ãçè§£ããŠã
ã¿ã¹ã¯ãå®è¡ãããšãããŸã§èžã¿èŸŒãã§ãã®ã倧ããªéãã ãã
ããšãæ¢åã®ã¢ãŒã«ãšãŒãžãŒã·ã¹ãã ãšæ¯ã¹ãŠãã
åäžã®ãšãŒãžã§ã³ããããªããŠãè€æ°ã®ãšãŒãžã§ã³ãã飿ºããŠã
ãã£ãšåçã«èšç»ãç«ãŠãŠå®è¡ããã£ãŠããç¹ã§ãäžæ©é²ãã§ãæãããããã
ãã®è«æãå®ã¯ãã€ãã¥ããã®äžåœã®æ€çޢ倧æãã®ããŒã ãæžããŠããã ããã
å®éã«ãªã³ã©ã€ã³ã®ãšãŒãã£ãŒãã¹ããããŠãŠã
è€éãªã¯ãšãªã«å¯ŸãããŠãŒã¶ãŒæºè¶³åºŠãç€ºãææšããšããããªã¥ãŒã¢ãŒã«ãã
ãªããšããã
ã13ããŒã»ã³ããåäžãããã ã£ãŠã
ããã¯ãããããšã ãã
ä»ã«ãããŠãŒã¶ãŒã®æ€çŽ¢ã¯ãšãªå€æŽçãã·ãŒãã¥ãŒã¢ãŒã«ãããã¡ç¹ãããããŒã»ã³ãæžã£ãŠã
ããŒãžãã¥ãŒãããŒãã€ã¯ããã£ãŠãããããããŒã»ã³ãå¢ããŠã
ãã€ãªãŒã¢ã¯ãã£ããŠãŒã¶ãŒããã£ãŒãšãŒãŠãŒã¯ããã£ãŠãã¯ã¡ãããŒã»ã³ãå¢ããŠã
æ»åšæéããããŒç¹ãã«ããŒã»ã³ãå¢ãããããã
ããã確å®ã«ãŠãŒã¶ãŒäœéšãè¯ããªã£ãŠã蚌æ ã ãã
ããããããããã®è«æãã¿ã€ãã«ãããã¥ã¯ãŒãº ãšãŒã¢ã€ ãµãŒã ãã©ãã€ã ãã§ããã
ããã©ãã€ã ãã£ãŠèšèã䜿ã£ãŠããã©ã
ãªã¬ãæããã©ãã©ãã³ã¹ã®ãããªåããŠç·Žç¿ããããšãããã ããã
âŠããé¢ä¿ãªãïŒãã£ããé¢ä¿ãªãããã¯ã¯ã
ãŸãããšã«ããããã®è«æã¯ã
ããããã®æ€çŽ¢ãã©ããªã£ãŠããã®ãã
èšèãæ±ããšãŒã¢ã€ãã©ããŸã§é²åããã®ãã
ãã®äžã€ã®å¯èœæ§ã瀺ããŠãããŠããšæããã ã
èšç®èšèªåŠã®åéãããŸããŸãé¢çœããªãããã ãã
ãšããããã§ã仿¥ã®ç޹ä»ã¯ãããªæãããªã
ã¡ãã£ãšé·ããªã£ã¡ãã£ããã©ãæåŸãŸã§èããŠãããŠããããšããã
ããããããŸã次åã
ãã€ããŒã€ã
ð The Paper and Some Imagination (English)
ð
ð TitleïŒBeyond RAG: The Future of AI Search with Agent Teams
ð Summary (English)
Hello there, and a very good Tuesday to you!
It's June 24th, 2025, and today, oh, I've got a real gem from the archives,
a trending piece that's got me thinking.
Let's dive into what these clever folks have been cooking up.
Alright, so I'm looking at this paper here.
The title is
Towards AI Search Paradigm.
And the URL is
http://arxiv.org/abs/2506.17188v1
It's long! Always a good sign, or a terrifying one, depending on your coffee levels.
So, what's the big problem this paper is trying to solve, I wonder.
It seems like, you know, current search engines,
they're pretty good if I ask something simple, like, what's the weather.
But, oh boy, if I ask something complicated,
something that needs a bit of, well, brainpower,
they kind of just throw a bunch of links at me.
The paper gives an example, like, Who was older,
Emperor Wu of Han or Julius Caesar, and by how many years?
Yeah, try that on your average search bar.
You'll get some pages, but you still have to do the detective work,
find the birth years, do the math, all that jazz.
Even these newer AI search things, they call them RAG systems,
Retrieval-Augmented Generation, fancy,
they try to give direct answers, which is cool,
but they still get tangled up with these really complex,
multi-step questions, or if they need to use different tools,
like a calculator and a web search.
They're often just, like, one-shot deals,
or they follow a very straight line, and if the first step is a bit off,
the whole thing can go wonky.
So, these researchers are proposing a new way,
this AI Search Paradigm.
Imagine, if you will, a little team of AI agents,
like tiny specialists all working together inside your computer.
It's a modular setup, they say, with four main LLM-powered agents.
LLM, Large Language Model, those big brainy AIs.
First, there's the Master Agent.
This one's like the project manager, or the conductor of the orchestra.
It looks at my query, figures out how tricky it is,
and then assembles the right team of other agents.
It also keeps an eye on how things are going,
and if a task fails, it tells the team to, you know,
rethink and try again. That's pretty smart.
Then, for the really complex questions, the Master calls in the Planner Agent.
This one's the strategist.
It takes my big, complicated question and breaks it down,
into a structured sequence of smaller, manageable sub-tasks.
They represent this as a Directed Acyclic Graph, or DAG.
Think of it as a super-smart to-do list,
where some tasks have to be done before others can start.
And, get this, the Planner also picks the right tools for each sub-task,
from something they call an MCP Servers Platform,
which sounds like a universal adapter for AI tools.
Next up is the Executor Agent.
This is the doer, the worker bee.
It carries out the simple queries directly,
or it takes on the individual sub-tasks defined by the Planner.
A big part of its job is actually using those external tools,
like searching the web, or doing calculations,
whatever is needed for that specific step.
It also checks if the results of its work are any good.
And finally, there's the Writer Agent.
This one's the communicator.
After all the sub-tasks are done and the information is gathered,
the Writer takes everything, all the bits and pieces,
and synthesizes it into a single, coherent,
and hopefully easy-to-understand answer to my original question.
It tries to make it contextually rich, not just a list of facts.
Now, how is this different from what we've got?
Well, old-school search, that's mostly keyword matching.
I type words, it finds pages with those words. Simple, but often misses the point.
Then came Learning-to-Rank, LTR, which used machine learning,
to get better at ordering the results. Still just a list of links, though.
The recent RAG systems, as I said, aim for direct answers.
But this AI Search Paradigm, it's trying to be more like,
how a human researcher would tackle a problem.
It's more dynamic, it reasons, plans, executes,
and can even re-plan if things go sideways.
It's not just retrieve information then generate an answer,
it's a whole collaborative process.
So, where could I see this being used in everyday life?
This isn't just for academic papers, right?
Oh, definitely. For example, chatbots and dialogue systems.
Imagine I'm talking to a super-smart travel assistant.
I don't just ask for flights to Paris.
I say, Plan me a week-long trip to Paris in spring,
I like museums but not crowded ones, find me a quiet hotel near good food,
and what will the weather likely be?
A system like this could break that down.
Planner makes sub-tasks: find flights, research quiet museums,
check hotel reviews and locations, get weather forecasts, maybe even look up local events.
Executor does all that, using different tools.
Then Writer puts it all together into a nice itinerary. Much better!
Then there's information extraction or text summarization.
Say I need to understand a really complex new scientific discovery,
or the impact of a new government policy.
This AI Search system could plan to pull information,
from various sources, news articles, official reports, academic papers.
The Executor would grab the data, maybe even verify facts if sources conflict.
And the Writer would synthesize a comprehensive summary.
That's way more powerful than just reading one article.
And, maybe a bit of a stretch, but machine translation.
For really nuanced stuff, like translating complex legal documents,
or a piece of literature with lots of cultural references.
The Planner could set up sub-tasks to look up specific legal terms,
or historical context for a phrase, or the meaning of a cultural idiom.
The Executor would find that extra info,
and the Writer, well, the translation system,
could then produce a much richer and more accurate translation.
They also talk about some other neat ideas in the paper,
like making the descriptions of the AI tools themselves better,
so the Planner agent can choose the best ones more easily.
And using reinforcement learning, which is like teaching by trial and error,
to make the Planner agent smarter over time.
Plus, a whole section on making these big AI models run faster,
and use less computing power, which they call,
Lightning LLM's Generation. That's always important.
So, yeah, that's the gist of it from what I'm gathering here.
It's about making search less of a chore and more of a,
well, a genuinely helpful, reasoning assistant.
Pretty fascinating stuff, moving beyond just finding documents,
to actually solving problems.
Makes you wonder what searching will be like in another few years.
Alright, that's one to chew on for today!
ðïž ã³ã¡ã³ã
æåŸãŸã§èªãã§ãããŠæ¬åœã«ããããšãïŒïŒ
ãã€ãã©ãããããŸã話ããªããïŒããããããããããïŒ
ãªã¬ãæããã©ãã©ãã³ã¹ã®ãããªåããŠç·Žç¿ããããšãããã ããã
âŠããé¢ä¿ãªãïŒãã£ããé¢ä¿ãªãããã¯ã¯ã
ãã©ãã©ãšããã«ã»ããããŒããšããäžã å€ãæããã°ãã°ãšçžæ§è¯ãããªãã ãã©ãããå€ããã ããïŒ
Keywords: #AISearch , #FutureofSearch , #AIAgents , #LLM , #RAG , #ResearchPaper , #ComplexQueries , #Multi -stepSearch, #MasterAgent , #PlannerAgent , #ExecutorAgent , #AIParadigm , #InformationRetrieval #AI #人工ç¥èœ #æ€çŽ¢ãšã³ãžã³ #次äžä»£æ€çŽ¢ #AISearchParadigm #èšç®èšèªåŠ #LLM #RAG #ãšãŒãžã§ã³ãAI #è«æè§£èª¬ #æªæ¥äºæž¬ #ãã¯ãããžãŒ
