{
  "video_id": "umbtFGPzdAM",
  "channel_slug": "mitcsail",
  "channel_handle": "mitcsail",
  "title": "MIT CSAIL Explains: Recursive Language Models",
  "duration_seconds": 843.0,
  "url": "https://www.youtube.com/watch?v=umbtFGPzdAM",
  "upload_date": "",
  "transcript": "Hi. I'm Alex.\nI am a first year PhD student here at MIT\nin the EECS department and Csail.\nI am also advised by Professors\nOmar Khattab, and Tim Kraska\nAnd I work very broadly on language models\nand more specifically\nwhere I think language\nmodels are underutilized.\nToday, I will be answering some questions\nthat you guys had about my research.\nOn recursive language models.\nYou argued that the language models\nwe already have\nare much more capable than we've realized.\nWhat do you mean by that?\nyou know,\nwe have some kind of understanding for\nhow good systems like ChatGPT\nand other language models are.\nAnd I think one of the things\nthat I've come\nto, understand\nor sort of argue is that actually,\nwe often underestimate\nhow good these models really are.\nAnd I think this really comes\nfrom the idea\nthat often the way that we\nevaluate language models\nare based on the ways that,\nwe think we should be using them.\nBut in some sense, actually,\nthey probably can do a lot more,\nthan we give them credit for.\nAnd a lot of the limitations actually come\nfrom the ways in which we use them.\nAnd not the inherent\ncapabilities of the models.\nSo what is a Recursive Language Model,\nor RLM?\nAnd what does recursive\nactually mean in this context?\nSo, Recursive Language Models, is kind of\nthe core piece of my current research.\nAnd the basic idea is\nit is a system\nin which,\nlanguage models call language models,\nwhich can call other language models,\nto ultimately provide,\nsome kind of final answer\nafter thinking for a very long time.\nAnd so within this idea,\nthere are sort of some key components,\nrelated to, giving a language model\naccess to a coding environment.\nAnd how you actually do this kind of,\n“recursive” language model calling,\nand the name recursive comes from the idea\nthat it is\na language model sort of calling itself,\nwhich can also call itself\nand the language model\nnaming of the the word RLM\ncomes from the idea\nthat this system itself actually\ncan be thought of as a language model.\nIn some sense it is sort of,\ntaking in some kind of text,\nand giving you some kind of, output back\nas if it is a language\nmodel itself.\nRight.\nA language model calling another language\nmodel isn't new\nNeither is having a language\nmodel use code.\nWhat's the subtle\nbut critical thing your approach\ndoes that people have been missing?\nSo this is, very, very important points\nto kind of talk about.\nI think generally the idea of a language\nmodel calling another language\nmodel has existed for some time,\nfor the past year.\nBut the RLM paper itself\nis actually sort of an argument\nfor how exactly such a system\nshould be designed.\nkind of the central idea\nis that the mechanism\nby which a language model calls\nanother language model\nexists inside of a code environment.\nSo you can think of it\nlike if you give your language\nmodel\nthe ability to write Python, for example,\nit can call a language model\nas a function or a module\ninside of Python itself,\nrather than the language model\nusing some kind of tool,\nor external API where it calls itself.\nAnd this kind of distinction\nis extremely important\nbecause it actually lends itself\nvery well to scale, and also to a system\nthat generalizes very well.\nA core idea in your paper is treating\nthe prompt itself as an environment,\nthe model to explore, not just text.\nIt reads in one pass.\nWhat does that shift in framing\nactually enable?\nOne really, really important\npiece of the RL in paper is that,\nso we think of the RLM as a language model\nas I mentioned.\nAnd if you're not familiar with this idea,\na language model will take in some prompt,\nlike you type into Chat GPT\nI want you to do this thing for me.\nAnd it will process it,\nand I'll give you an answer.\nThe way that the RLM treats the prompt\nis actually it doesn't feed it\ndirectly to the language model.\nAnd you can imagine\nthis is very interesting,\nespecially for models, or, models\nthat cannot handle a certain\nlength of a prompt.\nThere's always a limit\nfor these kinds of things.\nSo the way that the RLM actually handles\nthis is it stores\nthe prompts, inside\nof a coding environment.\nAnd it treats it like,\nlike something in memory,\nlike a file in memory that it can sort of,\ngo through and recursively navigate.\nAnd also use the coding environment\nto try and understand.\nAnd so this kind of framing actually\nallows you\nto scale your models to handle, near\ninfinite contexts.\nOr nearly infinitely long,\nkind of reasoning chains.\nAnd this is extremely useful for systems\nmoving forward, to get around\na lot of the limitations that we've\nhad traditionally in the field.\nIf I give an RLM 100 videos\nor a huge code base, it seems to know\nwhat's worth looking at and what isn't.\nHow does it do that and how is it\ndifferent from what a standard model\nwould do with that same input?\nSo you can kind of imagine like,\nif you had, hundreds of videos\nand let's say you wanted your language\nor video language\nmodel to go through\nand try and handle this for you.\nThe traditional way that you would do this\nor kind of the naive way, is\nyou feed all 100 videos, into the model\nand you kind of ask it\nto tell you something about these videos.\nAnd as is kind of obvious, it's very,\nvery expensive to do this.\nActually, it probably is\ncompletely infeasible for most frontier,\nmodels like OpenAI’s or Anthropic’s\nto, to do this kind of thing.\nBut the, the kind of RLM way\nof doing this is the RLM\ncan describe in code.\nLet's say\nI want to loop through all 100 videos,\nthat I have in my computer.\nAnd I will call a,\nI will call myself, and look at,\neach video, independently\nand write some script to take the output\nof each of these models, and give you\nsome kind of meaningful, feedback back.\nAnd generally also, one of the nice things\nis that a recursive language\nmodel can do, is it doesn't\nactually have to look at all the videos.\nSo if it looks at\nmaybe five seconds of each video,\nand it determines that it doesn't need\nto look at this video at all,\nit can completely ignore it.\nWhich is something that a standard\nlanguage model is unable to do.\nA standard language model\nhas to look at the entire video,\nin full in order to understand how to do.\nOne of the surprising things about RLMs\nis that building\none doesn't require training a new model\nor spending\nhundreds of millions of dollars.\nWhat does it actually take?\nRLMs are this really cool thing.\nBecause they actually just work with\nstandard, and current language models.\nSo you actually can slot\nin, let's say, Chat GPT-5\ninto an RLM- like scaffold\nAnd it will just work out the box,\nand doesn't require any additional\ntraining to, to, to function,\nwhere the training kind of actually comes\nin is like most systems\nmaybe like Claude code or Codex\nor things that you've seen, you\nor these companies tend to train models\nto work within these scaffolds.\nAnd so one of the really,\nreally exciting things moving forward\nis seeing whether we can train language\nmodels around RLMs, and make them\nextremely efficient, and much more capable\nthan they currently are.\nCan you walk us through a real-world example of\na problem that was out of reach before?\nSomething that in RL makes you tractable.\nSo I, I, I quite like, this question.\nI think, you know, some of the,\nthe obvious stuff,\nthat we've seen in the paper\nthat we put out was that RLMs\nlet you handle basically infinitely long,\nsequences of data or just,\nyou kind of dump\nin all the information you want\nand it can roughly process\nand give you some kind of thing back.\nBut one of the things that is a little bit\nmore interesting\nis longer\nreaching and longer horizon problems.\nSo an example that I like to give,\nis the protein\nSo an example that I like to give,\nis the protein\nfolding problem,\nwhich DeepMind had solved.\nfolding problem,\nwhich DeepMind had solved.\nOr, you know, had given\na very good solution to,\na few years back.\nAnd generally the process\nby which they solved this problem,\nwas they had a large team\nworking for a very long time,\niterating through many different,\nsolutions and many kind of possible ideas\nand avenues in which\nthey thought this was possible.\nAnd you can imagine that actually an RLM\ncan replicate this process,\nand describe all of this in code.\nAnd in some sense, actually,\nthe process by which\nthe team actually solved\nthe problem, is not that complicated.\nYou can imagine if you take together\na bunch of really smart people,\nand you kind of guide them\nthrough what they should do,\nthey can finally solve a problem\nafter a very long time.\nAnd RLMs sort of enable\nthese types of problems\nto be somewhat, tractable\nfor language model systems.\nTo actually solve.\nTell us about your\nmismanaged geniuses hypothesis.\nWhat is the big idea and\nhow does it connects back to our lives?\nYeah.\nSo I think an analogy\nthat I really like to give\nis that language models, themselves, can\nbe thought of as ‘Geniuses’.\nIn some sense, we have spent the last\nfive years training these models,\non all sorts of data and textbooks\nand things that,\nwe have traditionally\nused to train ourselves\nto be experts in the fields\nthat we are in.\nAnd kind of the way\nthat companies, have gone\nabout updating these models.\nIs they see this genius and they see that\nit can't solve a particular task.\nAnd the way in which, they fix\nthis is by introducing\na new genius\nthat maybe can solve a little bit more.\nAnd the Mismanaged Genius hypothesis\nis this idea that actually,\nif we were to manage these systems better,\nthey actually could solve\na much wider range of tasks.\nAnd the problem is not replacing a genius\nwith another genius.\nIt's actually figuring out a way to,\nmanage how these systems actually work.\nAnd one of the kind of takeaways of this\nwhole thing is\nmaybe a language model\nitself, is good enough to.\nActually manage other language models.\nSo with a recursive system\nmaking many calls over many steps,\nhow do you keep it on\ntask and prevent errors from compounding?\nSo I think one of the core\nprinciples of the RLM idea\nis that we are deferring all decision\nmaking to the language model.\nSo if a recursive language model call\nmakes mistakes,\nor hallucinates or things like this,\nit is the responsibility of the root\nlanguage model to figure out what to do,\nand verify and help it get back on track.\nSo in some sense, I think a lot of these issues\nwith errors compounding,\nand things of this nature,\ncan be solved\nby the language model that is managing\nthe other language models itself.\nAnd this is something that we will likely\nsee through additional training,\nand additional improvements\nto the actual scaffold itself.\nMulti-agent systems are having a moment\nwhere do you see RLMs\nfitting into that landscape.\nAnd what's the first thing you're excited\nto see people build with this?\nFor context,\nmulti-agent systems are basically,\nhaving multiple different language model\n‘agents’ or just systems\nworking together to solve a problem.\nAnd RLMs are in some sense\nan example of a multi-agent system.\nBut they are sort of an argument\nfor how such systems\nshould actually be designed\nmoving forward.\nSo I think,\nfrom my view,\nRLMs are sort of the future direction\nfor multi-agent systems.\nBut I also think that research in this\ndirection is, is very, very important.\nAnd as for what I'm excited\nfor people to build.\nI think generally it is systems that\ncan think for a very long time\nI think one of the more exciting\nthings moving forward\nis whether or not we can design\nthese language model systems\nthat kind of run forever,\nor run autonomously,\nand can automate a lot more\nvery painstaking tasks\nas a research assistant,\nor things of this nature.\nI guess we are\nall out of questiuons\nso thank you all for listening.\nyou can find me on, I guess,\nany kind of social media and\nfeel free to ask me any other questions.\nThank you for listening.\nfeel free to ask me any other questions.\nThank you for your time today!",
  "transcript_chars": 11821,
  "ingested_at": "2026-05-15T10:55:44.299392+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": 4431,
    "like_count": 192,
    "channel_id": "UCBpxspUNl1Th33XbugiHJzw",
    "categories": [
      "Science & Technology"
    ],
    "tags": [
      "MIT",
      "Massachusetts Institute of Technology",
      "CSAIL"
    ]
  }
}