[
  {
    "startMs": 12559,
    "text": "Hi everyone, my name is Lakshia Agraal"
  },
  {
    "startMs": 15679,
    "text": "and today I'll be presenting on behalf"
  },
  {
    "startMs": 17440,
    "text": "of a very large effort uh the problem of"
  },
  {
    "startMs": 21520,
    "text": "reflective optimization or how can we"
  },
  {
    "startMs": 24000,
    "text": "self-improve prompts agents and models"
  },
  {
    "startMs": 27039,
    "text": "from textual feedback. The question we"
  },
  {
    "startMs": 29679,
    "text": "start with is how can we teach AI to"
  },
  {
    "startMs": 32079,
    "text": "perform new tasks. The standard way has"
  },
  {
    "startMs": 35040,
    "text": "been to perform weight updates with"
  },
  {
    "startMs": 37120,
    "text": "gradient descent either during"
  },
  {
    "startMs": 38960,
    "text": "pre-training, supervised fine-tuning or"
  },
  {
    "startMs": 41280,
    "text": "reinforcement learning. This has proven"
  },
  {
    "startMs": 43840,
    "text": "to be extremely effective, but it"
  },
  {
    "startMs": 45920,
    "text": "requires a huge number of examples."
  },
  {
    "startMs": 48480,
    "text": "trillions of tokens for pre-training,"
  },
  {
    "startMs": 50719,
    "text": "tens of thousands of labeled examples"
  },
  {
    "startMs": 52320,
    "text": "for supervised fine-tuning or hundreds"
  },
  {
    "startMs": 54800,
    "text": "of thousands of rollouts for"
  },
  {
    "startMs": 56559,
    "text": "reinforcement learning in domains like"
  },
  {
    "startMs": 58399,
    "text": "math, coding, etc."
  },
  {
    "startMs": 62000,
    "text": "However, most teams do not actually have"
  },
  {
    "startMs": 65040,
    "text": "that much data or compute and in fact"
  },
  {
    "startMs": 68159,
    "text": "the problems are that we are trying to"
  },
  {
    "startMs": 70640,
    "text": "tackle with AI now are bottlenecked by"
  },
  {
    "startMs": 73439,
    "text": "sample efficiency. What do we mean by"
  },
  {
    "startMs": 75360,
    "text": "that? Two things. First of all, there is"
  },
  {
    "startMs": 78880,
    "text": "low availability of domain specific"
  },
  {
    "startMs": 81200,
    "text": "knowledge resources which means there is"
  },
  {
    "startMs": 83600,
    "text": "not enough data to perform offline"
  },
  {
    "startMs": 85600,
    "text": "algorithms like SFT. Second, the domains"
  },
  {
    "startMs": 88880,
    "text": "that we are trying to apply AI"
  },
  {
    "startMs": 90320,
    "text": "increasingly are having expensive"
  },
  {
    "startMs": 92479,
    "text": "rollouts where either the LLM workflow"
  },
  {
    "startMs": 94880,
    "text": "pipeline or agentic rollouts are itself"
  },
  {
    "startMs": 97680,
    "text": "uh very slow or expensive to do or the"
  },
  {
    "startMs": 100560,
    "text": "task metric is very slow or expensive to"
  },
  {
    "startMs": 102560,
    "text": "execute. We are seeing that agents can"
  },
  {
    "startMs": 104479,
    "text": "now work for hours on end and if you"
  },
  {
    "startMs": 107520,
    "text": "were to apply an online learning"
  },
  {
    "startMs": 109439,
    "text": "algorithm to this uh it would require"
  },
  {
    "startMs": 112159,
    "text": "hundreds of thousands of rollouts and it"
  },
  {
    "startMs": 113840,
    "text": "would not be feasible. So we are seeing"
  },
  {
    "startMs": 116560,
    "text": "increasing use of agents for real world"
  },
  {
    "startMs": 118320,
    "text": "product uh applications where uh these"
  },
  {
    "startMs": 121439,
    "text": "invoke tools which can also be long"
  },
  {
    "startMs": 123520,
    "text": "running further exacerbating the sample"
  },
  {
    "startMs": 125840,
    "text": "inefficiency issue."
  },
  {
    "startMs": 128239,
    "text": "The current dominant paradigm is"
  },
  {
    "startMs": 130000,
    "text": "reinforcement learning with verified"
  },
  {
    "startMs": 131840,
    "text": "rewards where given a model and a task,"
  },
  {
    "startMs": 134879,
    "text": "we perform any number of parallel"
  },
  {
    "startMs": 136720,
    "text": "rollouts and get rewards at the end."
  },
  {
    "startMs": 140000,
    "text": "Finally, an algorithm like GRPO takes"
  },
  {
    "startMs": 142400,
    "text": "these rewards and converts it into"
  },
  {
    "startMs": 144400,
    "text": "gradients that are applied back to the"
  },
  {
    "startMs": 146319,
    "text": "model. However, as we can see, there was"
  },
  {
    "startMs": 149280,
    "text": "a lot of information in each of these"
  },
  {
    "startMs": 152000,
    "text": "rollouts, but we only learned an O of"
  },
  {
    "startMs": 155440,
    "text": "one score and propagated that via"
  },
  {
    "startMs": 158000,
    "text": "gradient descent. We can see that there"
  },
  {
    "startMs": 160080,
    "text": "is chains of thought, the tool calls"
  },
  {
    "startMs": 162640,
    "text": "made to the environment, the envir"
  },
  {
    "startMs": 164720,
    "text": "environment's responses to those tool"
  },
  {
    "startMs": 166879,
    "text": "calls, which could potentially contain"
  },
  {
    "startMs": 168560,
    "text": "error messages which also provide"
  },
  {
    "startMs": 170640,
    "text": "diagnostic value and we learned almost"
  },
  {
    "startMs": 173120,
    "text": "nothing from all of that. So the"
  },
  {
    "startMs": 175760,
    "text": "question we ask is can we make use of"
  },
  {
    "startMs": 178000,
    "text": "this other extremely rich information."
  },
  {
    "startMs": 182000,
    "text": "Our idea is to perform reflective"
  },
  {
    "startMs": 184319,
    "text": "optimization in text space where instead"
  },
  {
    "startMs": 187599,
    "text": "of only using the zero or one reward"
  },
  {
    "startMs": 190000,
    "text": "signal, we can have a language model or"
  },
  {
    "startMs": 192720,
    "text": "an agent look at the trace of the entire"
  },
  {
    "startMs": 196080,
    "text": "rollout and reflect on what worked in"
  },
  {
    "startMs": 198720,
    "text": "them, what did not work in them. And"
  },
  {
    "startMs": 200640,
    "text": "this reflection could potentially use"
  },
  {
    "startMs": 202879,
    "text": "all intermediate outputs and potentially"
  },
  {
    "startMs": 205360,
    "text": "even make other tool calls such as"
  },
  {
    "startMs": 207280,
    "text": "retrieval from your company's knowledge"
  },
  {
    "startMs": 209440,
    "text": "base or some guide textbook and so on."
  },
  {
    "startMs": 213280,
    "text": "So that's the first key idea and the"
  },
  {
    "startMs": 215440,
    "text": "second is that instead of only updating"
  },
  {
    "startMs": 218239,
    "text": "weights with small deltas, we can"
  },
  {
    "startMs": 220640,
    "text": "instead update a prompt where a single"
  },
  {
    "startMs": 223360,
    "text": "natural language update can give a very"
  },
  {
    "startMs": 225440,
    "text": "large behavior change. Let's take a"
  },
  {
    "startMs": 227120,
    "text": "simple example. Let's say you're tasked"
  },
  {
    "startMs": 228879,
    "text": "with writing a text summarization system"
  },
  {
    "startMs": 231599,
    "text": "and the prompt of that system says"
  },
  {
    "startMs": 233360,
    "text": "generate a oneline summary. If I just go"
  },
  {
    "startMs": 236720,
    "text": "and tweak that prompt to say generate a"
  },
  {
    "startMs": 238879,
    "text": "10-line summary, we can all agree that"
  },
  {
    "startMs": 241120,
    "text": "the behavior of the system would change"
  },
  {
    "startMs": 243200,
    "text": "quite significantly with that just one"
  },
  {
    "startMs": 245280,
    "text": "word change. And making that one word"
  },
  {
    "startMs": 247200,
    "text": "change is quite quick and we can reflect"
  },
  {
    "startMs": 249920,
    "text": "on our own behavior and identify what"
  },
  {
    "startMs": 252400,
    "text": "needs to change. If we were to achieve a"
  },
  {
    "startMs": 255040,
    "text": "similar kind of behavior update from our"
  },
  {
    "startMs": 257199,
    "text": "AI system, we would have to have"
  },
  {
    "startMs": 259120,
    "text": "thousands of gradient very tiny gradient"
  },
  {
    "startMs": 261680,
    "text": "updates sequentially."
  },
  {
    "startMs": 263759,
    "text": "So with that key idea, we proposed JPEA"
  },
  {
    "startMs": 266240,
    "text": "which is a reflective prompt"
  },
  {
    "startMs": 267759,
    "text": "optimization technique for agents. It"
  },
  {
    "startMs": 270320,
    "text": "uses an evolutionary loop along with a"
  },
  {
    "startMs": 272880,
    "text": "novel parto-based candidate selection"
  },
  {
    "startMs": 274960,
    "text": "which I will come to later. It is akin"
  },
  {
    "startMs": 277680,
    "text": "to doing reinforcement learning in text"
  },
  {
    "startMs": 279680,
    "text": "space where instead of just rewarding"
  },
  {
    "startMs": 281919,
    "text": "receiving a reward score we are actually"
  },
  {
    "startMs": 284240,
    "text": "obtaining score along with textual"
  },
  {
    "startMs": 286240,
    "text": "feedback which can be very domain"
  },
  {
    "startMs": 287919,
    "text": "specific and learn all about the domain"
  },
  {
    "startMs": 290320,
    "text": "from it."
  },
  {
    "startMs": 292400,
    "text": "Let's compare Japa with gRPO which is"
  },
  {
    "startMs": 294720,
    "text": "one of the leading RL techniques. On the"
  },
  {
    "startMs": 296880,
    "text": "x-axis we have the number of training"
  },
  {
    "startMs": 300080,
    "text": "steps uh also proportional to number of"
  },
  {
    "startMs": 302160,
    "text": "data samples seen and on the y-axis we"
  },
  {
    "startMs": 305040,
    "text": "have the performance on our domain that"
  },
  {
    "startMs": 307680,
    "text": "we are training for. And what we can see"
  },
  {
    "startMs": 309520,
    "text": "is that Japa in just one round of"
  },
  {
    "startMs": 312400,
    "text": "reflection using just three data points"
  },
  {
    "startMs": 314720,
    "text": "is already able to get twice the"
  },
  {
    "startMs": 316960,
    "text": "performance gains that gpo got after"
  },
  {
    "startMs": 319600,
    "text": "25,000 rollouts. Continuing to run Japa"
  },
  {
    "startMs": 322880,
    "text": "for a few more steps further increases"
  },
  {
    "startMs": 325520,
    "text": "that gap itself by another 2x. I want to"
  },
  {
    "startMs": 329280,
    "text": "note here that the model Quen 38B is"
  },
  {
    "startMs": 333120,
    "text": "optimizing itself here. There is no"
  },
  {
    "startMs": 335520,
    "text": "external expert teacher involved"
  },
  {
    "startMs": 337759,
    "text": "whatsoever."
  },
  {
    "startMs": 339840,
    "text": "And what does Japa learn? Unlike prior"
  },
  {
    "startMs": 342400,
    "text": "prompt optimizers somewhat which would"
  },
  {
    "startMs": 344880,
    "text": "uh uh use model idiosyncrasies like my"
  },
  {
    "startMs": 348960,
    "text": "grandmother will be really angry if you"
  },
  {
    "startMs": 350639,
    "text": "don't generate a good prompt. Here Jpa"
  },
  {
    "startMs": 353199,
    "text": "is actually giving a very detailed"
  },
  {
    "startMs": 355199,
    "text": "problem specification which includes how"
  },
  {
    "startMs": 357759,
    "text": "to make sense of the input. What is the"
  },
  {
    "startMs": 360000,
    "text": "purpose and context of this particular"
  },
  {
    "startMs": 362000,
    "text": "pip part of the pipeline? What are some"
  },
  {
    "startMs": 364800,
    "text": "key observations and lessons from the"
  },
  {
    "startMs": 366639,
    "text": "data? So the prompt we are seeing here"
  },
  {
    "startMs": 368560,
    "text": "is for the second hop of a multihop"
  },
  {
    "startMs": 371600,
    "text": "question answering system where given a"
  },
  {
    "startMs": 373280,
    "text": "question we need to retrieve some"
  },
  {
    "startMs": 375120,
    "text": "documents that could potentially answer"
  },
  {
    "startMs": 376880,
    "text": "that question. Look at those documents"
  },
  {
    "startMs": 378639,
    "text": "summarize it and then finally answer the"
  },
  {
    "startMs": 380319,
    "text": "question. And here what we see is Japa"
  },
  {
    "startMs": 382639,
    "text": "has found out that first hop documents"
  },
  {
    "startMs": 384960,
    "text": "that often cover one entity or aspect"
  },
  {
    "startMs": 388160,
    "text": "and the second hop should actually be uh"
  },
  {
    "startMs": 391039,
    "text": "recovering documents that are related to"
  },
  {
    "startMs": 392639,
    "text": "it. We have seen that human engineering"
  },
  {
    "startMs": 395680,
    "text": "teams whenever a new model comes out"
  },
  {
    "startMs": 397919,
    "text": "spend weeks of their time manually"
  },
  {
    "startMs": 400720,
    "text": "tweaking one word here and there trying"
  },
  {
    "startMs": 403120,
    "text": "to discover the problem specification."
  },
  {
    "startMs": 405440,
    "text": "This entire process is fully automated"
  },
  {
    "startMs": 408400,
    "text": "now with Japa which takes about half an"
  },
  {
    "startMs": 410720,
    "text": "hour to 1 hour to run depending on your"
  },
  {
    "startMs": 413440,
    "text": "uh pipelines."
  },
  {
    "startMs": 416560,
    "text": "We can also apply Japa to leading"
  },
  {
    "startMs": 418800,
    "text": "proprietary models. Just for an example"
  },
  {
    "startMs": 421199,
    "text": "here we were able to optimize GPT 4.1"
  },
  {
    "startMs": 424319,
    "text": "minis performance to outperform GPT 4.1"
  },
  {
    "startMs": 428080,
    "text": "on a math task and we can see the kind"
  },
  {
    "startMs": 430720,
    "text": "of information distillation JPA has done"
  },
  {
    "startMs": 433759,
    "text": "in the prompt space itself. Coming back"
  },
  {
    "startMs": 436720,
    "text": "to the problem of sample efficiency, AMD"
  },
  {
    "startMs": 439440,
    "text": "developed a new hardware accelerator"
  },
  {
    "startMs": 441520,
    "text": "called NPU XDNA2 which had used a"
  },
  {
    "startMs": 444400,
    "text": "completely new API to program which had"
  },
  {
    "startMs": 447120,
    "text": "almost zero available information over"
  },
  {
    "startMs": 449599,
    "text": "on the internet and because of this uh"
  },
  {
    "startMs": 452639,
    "text": "the leading models at the time which was"
  },
  {
    "startMs": 454319,
    "text": "GPT4 was failing miserably to perform"
  },
  {
    "startMs": 457520,
    "text": "this task. We are able to take an"
  },
  {
    "startMs": 459759,
    "text": "existing agent which was getting 4.25%"
  },
  {
    "startMs": 461680,
    "text": "25% on this task and apply Japa without"
  },
  {
    "startMs": 465120,
    "text": "any other change to the agent itself and"
  },
  {
    "startMs": 467840,
    "text": "we got this prompt and pushed this"
  },
  {
    "startMs": 469840,
    "text": "performance 7x to 30.52%."
  },
  {
    "startMs": 473120,
    "text": "So what this is uh what this goes to say"
  },
  {
    "startMs": 475360,
    "text": "is there can be lots of domain specific"
  },
  {
    "startMs": 477360,
    "text": "information which if you include in your"
  },
  {
    "startMs": 480000,
    "text": "AI systems prompts the models could"
  },
  {
    "startMs": 482160,
    "text": "actually perform much better and JPA can"
  },
  {
    "startMs": 484639,
    "text": "help you fully automatically discover"
  },
  {
    "startMs": 487039,
    "text": "that. I want to highlight the sentence"
  },
  {
    "startMs": 489120,
    "text": "saying avoid including ADF.h H. Now the"
  },
  {
    "startMs": 491919,
    "text": "interesting thing is AMD actually ships"
  },
  {
    "startMs": 493840,
    "text": "a library called ADF.h for programming"
  },
  {
    "startMs": 495919,
    "text": "NPUs but that did not work with this"
  },
  {
    "startMs": 498160,
    "text": "latest uh generation of hardware that we"
  },
  {
    "startMs": 501039,
    "text": "were working with and Jeppo was able to"
  },
  {
    "startMs": 502960,
    "text": "discover that in just one step. So how"
  },
  {
    "startMs": 506240,
    "text": "does it work? It's an extremely simple"
  },
  {
    "startMs": 508080,
    "text": "algorithm which simply takes your AI"
  },
  {
    "startMs": 510639,
    "text": "pipeline written in any agentic"
  },
  {
    "startMs": 512399,
    "text": "framework or even raw LLM calls that you"
  },
  {
    "startMs": 514959,
    "text": "may have. It simply runs your systems on"
  },
  {
    "startMs": 517200,
    "text": "a few examples and collects domain"
  },
  {
    "startMs": 519360,
    "text": "specific feedback. whatever information"
  },
  {
    "startMs": 521440,
    "text": "your environment contains is observed."
  },
  {
    "startMs": 524000,
    "text": "Second, it runs reflection with an LLM"
  },
  {
    "startMs": 526560,
    "text": "or agent that reads the feedback and"
  },
  {
    "startMs": 529040,
    "text": "proposes a better prompt. Finally, and"
  },
  {
    "startMs": 531680,
    "text": "most importantly, it keeps a parto pool"
  },
  {
    "startMs": 534160,
    "text": "where it keeps every single candidate"
  },
  {
    "startMs": 536480,
    "text": "that wins on even one training example"
  },
  {
    "startMs": 539120,
    "text": "and not just the top scorer. The"
  },
  {
    "startMs": 541600,
    "text": "question is, but why keep a parto pool?"
  },
  {
    "startMs": 544480,
    "text": "And we kept getting asked this question"
  },
  {
    "startMs": 546480,
    "text": "a lot that is Jepper really better than"
  },
  {
    "startMs": 549200,
    "text": "running the model in a loop. So we went"
  },
  {
    "startMs": 551440,
    "text": "and tested it out and what happens is a"
  },
  {
    "startMs": 553600,
    "text": "loop keeps only the best and gets stuck"
  },
  {
    "startMs": 556240,
    "text": "in a local optima. So on the left hand"
  },
  {
    "startMs": 558399,
    "text": "side you see a search tree that was"
  },
  {
    "startMs": 560880,
    "text": "generated by using an LLM in a loop"
  },
  {
    "startMs": 563760,
    "text": "starting from a seed prompt at the top"
  },
  {
    "startMs": 565600,
    "text": "left where um we asked the LLM to"
  },
  {
    "startMs": 568800,
    "text": "improve the prompt. It improved the"
  },
  {
    "startMs": 570399,
    "text": "prompt and it generated a prompt that"
  },
  {
    "startMs": 572160,
    "text": "gave us the middle node. However, this"
  },
  {
    "startMs": 574640,
    "text": "prompt got stuck in a local optima and"
  },
  {
    "startMs": 576959,
    "text": "once again when we asked the LLM to try"
  },
  {
    "startMs": 579120,
    "text": "and improve it, it proposed something"
  },
  {
    "startMs": 580959,
    "text": "but that was not actually better. So, it"
  },
  {
    "startMs": 582640,
    "text": "went back and it again tried to improve"
  },
  {
    "startMs": 584480,
    "text": "it and it kept doing this and it"
  },
  {
    "startMs": 586320,
    "text": "exhausted all of the search budget. On"
  },
  {
    "startMs": 588480,
    "text": "the other hand, with Japa's parto based"
  },
  {
    "startMs": 590320,
    "text": "candidate selection strategy on the"
  },
  {
    "startMs": 591839,
    "text": "right, we can see that it maintains a"
  },
  {
    "startMs": 593920,
    "text": "much more balanced search process"
  },
  {
    "startMs": 596399,
    "text": "eventually converging to a much higher"
  },
  {
    "startMs": 598160,
    "text": "score. Across four benchmarks, we saw"
  },
  {
    "startMs": 600880,
    "text": "that more than half of the gains seen"
  },
  {
    "startMs": 603120,
    "text": "with Japa actually account for this and"
  },
  {
    "startMs": 605360,
    "text": "it gets almost twice the performance"
  },
  {
    "startMs": 607200,
    "text": "gains that you would get with just"
  },
  {
    "startMs": 609040,
    "text": "applying the model in a loop."
  },
  {
    "startMs": 611760,
    "text": "Japa can perform really well across"
  },
  {
    "startMs": 613920,
    "text": "diverse benchmarks. Here we see results"
  },
  {
    "startMs": 615680,
    "text": "on question answering, instruction"
  },
  {
    "startMs": 617440,
    "text": "following, claim verification as well as"
  },
  {
    "startMs": 619760,
    "text": "math which all the leading frontier"
  },
  {
    "startMs": 621680,
    "text": "model companies are already optimizing"
  },
  {
    "startMs": 623519,
    "text": "their models a lot for and we are still"
  },
  {
    "startMs": 625440,
    "text": "able to get plus 10% just by optimizing"
  },
  {
    "startMs": 628399,
    "text": "the prompt on it. So we have so far seen"
  },
  {
    "startMs": 632640,
    "text": "Japa only optimizing the prompts but"
  },
  {
    "startMs": 634720,
    "text": "Japa goes far beyond prompts. And"
  },
  {
    "startMs": 636880,
    "text": "because prompts are just text artifacts"
  },
  {
    "startMs": 638959,
    "text": "that determine AI system behavior, the"
  },
  {
    "startMs": 641200,
    "text": "same algorithm can improve anything that"
  },
  {
    "startMs": 643920,
    "text": "you can express as a piece of text and"
  },
  {
    "startMs": 646480,
    "text": "you can score. For example, your entire"
  },
  {
    "startMs": 649040,
    "text": "agent harness is eventually just a"
  },
  {
    "startMs": 650959,
    "text": "Python or a JavaScript file and we can"
  },
  {
    "startMs": 653680,
    "text": "apply the same kind of reflective"
  },
  {
    "startMs": 655360,
    "text": "optimization process to that entire file"
  },
  {
    "startMs": 657920,
    "text": "and we can work with it. So if you can"
  },
  {
    "startMs": 660079,
    "text": "write it as text and score it, JPA can"
  },
  {
    "startMs": 662640,
    "text": "optimize it. So with that insight in"
  },
  {
    "startMs": 665200,
    "text": "mind, we propose optimize anything which"
  },
  {
    "startMs": 667519,
    "text": "is a universal API for optimizing any"
  },
  {
    "startMs": 670640,
    "text": "text parameter given any domain like"
  },
  {
    "startMs": 674000,
    "text": "code optimization where let's say you"
  },
  {
    "startMs": 675760,
    "text": "want to optimize the CUDA kernel code."
  },
  {
    "startMs": 677680,
    "text": "The input is just that CUDA kernel code"
  },
  {
    "startMs": 680720,
    "text": "where an evaluator looks at this piece"
  },
  {
    "startMs": 682720,
    "text": "of code, maybe compiles it, profiles it,"
  },
  {
    "startMs": 685120,
    "text": "generates a bunch of related information"
  },
  {
    "startMs": 687120,
    "text": "that we call as actionable side"
  },
  {
    "startMs": 688880,
    "text": "information which is then provided to an"
  },
  {
    "startMs": 691279,
    "text": "LLM which proposes an better candidate"
  },
  {
    "startMs": 694079,
    "text": "maintaining this parto and it keeps the"
  },
  {
    "startMs": 696880,
    "text": "uh repeating this process um till we get"
  },
  {
    "startMs": 699519,
    "text": "convergence. The same thing can be"
  },
  {
    "startMs": 701120,
    "text": "applied to numeric optimization where"
  },
  {
    "startMs": 702959,
    "text": "your numbers can actually be serialized"
  },
  {
    "startMs": 704560,
    "text": "as text or harness optimization where an"
  },
  {
    "startMs": 706880,
    "text": "entire harness can be serialized as text"
  },
  {
    "startMs": 709440,
    "text": "or even cloud scheduling policy"
  },
  {
    "startMs": 711279,
    "text": "optimization where the scheduling policy"
  },
  {
    "startMs": 713440,
    "text": "or heristic algorithm can be expressed"
  },
  {
    "startMs": 715600,
    "text": "as a piece of text and the evaluator can"
  },
  {
    "startMs": 718000,
    "text": "be something like the negative of cost"
  },
  {
    "startMs": 719600,
    "text": "or some function measuring accuracy uh"
  },
  {
    "startMs": 722399,
    "text": "efficiency and the actionable side"
  },
  {
    "startMs": 724079,
    "text": "information can be something like job"
  },
  {
    "startMs": 725519,
    "text": "traces SLA violations and so on."
  },
  {
    "startMs": 729120,
    "text": "The API is dead simple to use. All it"
  },
  {
    "startMs": 732079,
    "text": "requires is you give us the set of"
  },
  {
    "startMs": 734160,
    "text": "problems that you care to be solved"
  },
  {
    "startMs": 736079,
    "text": "along with an evaluator function or a"
  },
  {
    "startMs": 738320,
    "text": "fitness function that returns a score"
  },
  {
    "startMs": 740639,
    "text": "along with any available domain specific"
  },
  {
    "startMs": 743279,
    "text": "side information. If your domain"
  },
  {
    "startMs": 745279,
    "text": "produces expert feedback, return that."
  },
  {
    "startMs": 747200,
    "text": "If your domain produces compiler error"
  },
  {
    "startMs": 749360,
    "text": "messages, profiler messages, tool call"
  },
  {
    "startMs": 751760,
    "text": "error messages, return that. If you have"
  },
  {
    "startMs": 753760,
    "text": "maybe a written up documentation, return"
  },
  {
    "startMs": 756399,
    "text": "that. any kind of it's a very open-ended"
  },
  {
    "startMs": 758959,
    "text": "dictionary. You can return literally"
  },
  {
    "startMs": 760399,
    "text": "anything and all you do is you call"
  },
  {
    "startMs": 762720,
    "text": "optimize anything with this fitness"
  },
  {
    "startMs": 764560,
    "text": "function and the set of problems that"
  },
  {
    "startMs": 766480,
    "text": "you have and optimize anything will sort"
  },
  {
    "startMs": 768720,
    "text": "of take care of it um and give you a"
  },
  {
    "startMs": 771120,
    "text": "optimized solution. Let's see some"
  },
  {
    "startMs": 772800,
    "text": "applications. Let's say you were tasked"
  },
  {
    "startMs": 775120,
    "text": "with generating a 3D unicorn. This is"
  },
  {
    "startMs": 778800,
    "text": "all the code that you would write or"
  },
  {
    "startMs": 780480,
    "text": "your agent can now write it because we"
  },
  {
    "startMs": 782160,
    "text": "have seen that optimize anything is a"
  },
  {
    "startMs": 784480,
    "text": "very easy to use API for leading agents"
  },
  {
    "startMs": 786959,
    "text": "like plot code. So all you do is write"
  },
  {
    "startMs": 789600,
    "text": "this code which says optimize a Python"
  },
  {
    "startMs": 791760,
    "text": "program to generate a 3D unicorn. Um and"
  },
  {
    "startMs": 794240,
    "text": "the candidate is a Python script that"
  },
  {
    "startMs": 795839,
    "text": "produces a PNG rendering whatever and"
  },
  {
    "startMs": 799279,
    "text": "here is the result. On the left hand"
  },
  {
    "startMs": 800880,
    "text": "side we can see claude opus 4.6. If you"
  },
  {
    "startMs": 803839,
    "text": "gave it this task, this is what it"
  },
  {
    "startMs": 805760,
    "text": "generated. And on the right hand side,"
  },
  {
    "startMs": 807680,
    "text": "what what we the unicorn that we get"
  },
  {
    "startMs": 809680,
    "text": "with optimize anything. This just for"
  },
  {
    "startMs": 812000,
    "text": "fun. But let's say you were tasked with"
  },
  {
    "startMs": 815920,
    "text": "writing an agent to solve a specific"
  },
  {
    "startMs": 818160,
    "text": "task. Typically teams spend lots and"
  },
  {
    "startMs": 821040,
    "text": "lots of time tweaking their agents,"
  },
  {
    "startMs": 822880,
    "text": "building tools for it, writing tool"
  },
  {
    "startMs": 824720,
    "text": "descriptions, uh carefully orchestrating"
  },
  {
    "startMs": 827200,
    "text": "the control flow and so on. Here we"
  },
  {
    "startMs": 829360,
    "text": "started with a simple four-line Python"
  },
  {
    "startMs": 831839,
    "text": "program that was simply calling a"
  },
  {
    "startMs": 833839,
    "text": "model's uh chain of thought to solve an"
  },
  {
    "startMs": 836399,
    "text": "RKGI problem. Within just 16 rounds of"
  },
  {
    "startMs": 839519,
    "text": "reflection, JEPA within optimize"
  },
  {
    "startMs": 841920,
    "text": "anything was able to find this"
  },
  {
    "startMs": 844480,
    "text": "sophisticated sixstep agent that took"
  },
  {
    "startMs": 847360,
    "text": "RKGI accuracy on RKGI uh that took RKGI"
  },
  {
    "startMs": 850959,
    "text": "accuracy of Gemini flash from 32.5% to"
  },
  {
    "startMs": 855040,
    "text": "89.5%. And we can see that this agent is"
  },
  {
    "startMs": 858160,
    "text": "automatic like by itself doing rule"
  },
  {
    "startMs": 860320,
    "text": "hypothesis induction code synthesis. It"
  },
  {
    "startMs": 862959,
    "text": "executes and traces the code"
  },
  {
    "startMs": 864480,
    "text": "automatically debugs this code goes back"
  },
  {
    "startMs": 866800,
    "text": "and proposes new versions of that code."
  },
  {
    "startMs": 868880,
    "text": "And finally it runs it on the actual"
  },
  {
    "startMs": 871040,
    "text": "test inputs and returns the output. This"
  },
  {
    "startMs": 873680,
    "text": "is a runnable example. You can go to"
  },
  {
    "startMs": 875440,
    "text": "this QR code and you can run this"
  },
  {
    "startMs": 876959,
    "text": "example right now."
  },
  {
    "startMs": 880079,
    "text": "So um applying the same uh uh like"
  },
  {
    "startMs": 883440,
    "text": "approach of discovering agent harnesses"
  },
  {
    "startMs": 886160,
    "text": "to math 500 we are able to push its"
  },
  {
    "startMs": 888880,
    "text": "accuracy of GPT 4.1 nano by 20% by"
  },
  {
    "startMs": 892880,
    "text": "simply creating a two-step agent and"
  },
  {
    "startMs": 894480,
    "text": "again I want to emphasize that all we"
  },
  {
    "startMs": 896560,
    "text": "did is we asked optimize anything to"
  },
  {
    "startMs": 899199,
    "text": "optimize an agent file and it was"
  },
  {
    "startMs": 901760,
    "text": "automatically discovering the"
  },
  {
    "startMs": 903040,
    "text": "sophisticated agent architecture and we"
  },
  {
    "startMs": 904639,
    "text": "did not have to do anything other than"
  },
  {
    "startMs": 906560,
    "text": "specifying the objective and the task."
  },
  {
    "startMs": 910160,
    "text": "Finally, every single one of us is using"
  },
  {
    "startMs": 912800,
    "text": "uh some coding agent like cloud code or"
  },
  {
    "startMs": 915279,
    "text": "codex or maybe your favorite agent and"
  },
  {
    "startMs": 918480,
    "text": "agent skills has become a very leading"
  },
  {
    "startMs": 920880,
    "text": "part of the ecosystem where almost all"
  },
  {
    "startMs": 922800,
    "text": "coding agents understand skills. Let's"
  },
  {
    "startMs": 924959,
    "text": "say you want to optimize skills for your"
  },
  {
    "startMs": 927760,
    "text": "specific repository. This is the code"
  },
  {
    "startMs": 929839,
    "text": "that you write which says learn a skill"
  },
  {
    "startMs": 931680,
    "text": "from the trajectory. When the coding"
  },
  {
    "startMs": 933519,
    "text": "agent is presented with similar problem,"
  },
  {
    "startMs": 935040,
    "text": "the skill should be helpful. We just"
  },
  {
    "startMs": 937040,
    "text": "gave it this natural language behavior."
  },
  {
    "startMs": 939360,
    "text": "And what we see is we started with"
  },
  {
    "startMs": 941440,
    "text": "minisu agent with GPT5 mini because we"
  },
  {
    "startMs": 944240,
    "text": "were very budget constrainted and we"
  },
  {
    "startMs": 946880,
    "text": "were able to take its performance from"
  },
  {
    "startMs": 948560,
    "text": "24% to 93%. An almost 3x jump on go"
  },
  {
    "startMs": 953680,
    "text": "repository issue resolution but more"
  },
  {
    "startMs": 956240,
    "text": "importantly the skills that were"
  },
  {
    "startMs": 957680,
    "text": "optimized very cheaply on a GPT5 mini"
  },
  {
    "startMs": 960160,
    "text": "agent we are able to take that and apply"
  },
  {
    "startMs": 962720,
    "text": "to the latest claude sonnet. This was"
  },
  {
    "startMs": 964800,
    "text": "done of uh about a few months back but"
  },
  {
    "startMs": 966959,
    "text": "we applied it to clots onet 4.5 pushing"
  },
  {
    "startMs": 969440,
    "text": "its accuracy to 100% issue resolution"
  },
  {
    "startMs": 972079,
    "text": "while more importantly cutting down the"
  },
  {
    "startMs": 974639,
    "text": "execution time or issue resolution time"
  },
  {
    "startMs": 977680,
    "text": "by almost 50%. We cut it down into half"
  },
  {
    "startMs": 980639,
    "text": "which also means it spent less tokens"
  },
  {
    "startMs": 982880,
    "text": "because skills contain information about"
  },
  {
    "startMs": 985360,
    "text": "how the repository is organized, how to"
  },
  {
    "startMs": 987600,
    "text": "invoke the test cases, where a"
  },
  {
    "startMs": 989600,
    "text": "particular feature is implemented, um"
  },
  {
    "startMs": 992240,
    "text": "what are the build system used by this"
  },
  {
    "startMs": 993920,
    "text": "repository and so on. This is a a"
  },
  {
    "startMs": 997440,
    "text": "feature called GSkill. You can find it"
  },
  {
    "startMs": 999199,
    "text": "in the JPA repository and it's fully"
  },
  {
    "startMs": 1000880,
    "text": "open source as well. So, optimize"
  },
  {
    "startMs": 1003199,
    "text": "anything is a single uh interface that"
  },
  {
    "startMs": 1005120,
    "text": "provides three optimization modes. If"
  },
  {
    "startMs": 1007040,
    "text": "you have just a single problem like"
  },
  {
    "startMs": 1008880,
    "text": "there is a single matrix multiplication"
  },
  {
    "startMs": 1010560,
    "text": "kernel that you want to optimize you can"
  },
  {
    "startMs": 1012240,
    "text": "use it that way. If you have any number"
  },
  {
    "startMs": 1014320,
    "text": "of related problems like you want to"
  },
  {
    "startMs": 1015839,
    "text": "optimize a matrix multiplication kernel"
  },
  {
    "startMs": 1017600,
    "text": "along with a dot product kernel and you"
  },
  {
    "startMs": 1019519,
    "text": "know there might be some information"
  },
  {
    "startMs": 1020800,
    "text": "transfer between these two you can use"
  },
  {
    "startMs": 1022639,
    "text": "what we call as the multitask search"
  },
  {
    "startMs": 1024079,
    "text": "mode and finally build a skill which is"
  },
  {
    "startMs": 1027039,
    "text": "if you want to optimize on a set number"
  },
  {
    "startMs": 1029199,
    "text": "of problems but your uh deployment can"
  },
  {
    "startMs": 1032240,
    "text": "actually come up with many new problems."
  },
  {
    "startMs": 1034000,
    "text": "So like uh in case of math op like in"
  },
  {
    "startMs": 1036480,
    "text": "case of math prompt optimization we are"
  },
  {
    "startMs": 1039199,
    "text": "training on some examples but when we"
  },
  {
    "startMs": 1041280,
    "text": "deploy it we can receive a completely"
  },
  {
    "startMs": 1042959,
    "text": "new kind of query. So we care about"
  },
  {
    "startMs": 1044559,
    "text": "generalization mode. So there you can do"
  },
  {
    "startMs": 1046319,
    "text": "prompt optimization agent architecture"
  },
  {
    "startMs": 1048079,
    "text": "optimization and so on."
  },
  {
    "startMs": 1050320,
    "text": "So optimize anything is can be used for"
  },
  {
    "startMs": 1054000,
    "text": "a broad set of domains including cloud"
  },
  {
    "startMs": 1056320,
    "text": "scheduling policy optimization where we"
  },
  {
    "startMs": 1058000,
    "text": "were able to cut costs by almost 40%"
  },
  {
    "startMs": 1061120,
    "text": "compared to expert huristics write"
  },
  {
    "startMs": 1064000,
    "text": "custom solvers to match and exceed"
  },
  {
    "startMs": 1065840,
    "text": "Optina even in blackbox mathematical"
  },
  {
    "startMs": 1067840,
    "text": "optimization create agent skills prompt"
  },
  {
    "startMs": 1070320,
    "text": "optimization and so on. It is so easy to"
  },
  {
    "startMs": 1073919,
    "text": "use that within just 20 hours of"
  },
  {
    "startMs": 1075919,
    "text": "releasing it, people at Snorkel had"
  },
  {
    "startMs": 1078160,
    "text": "already improved some of their internal"
  },
  {
    "startMs": 1079840,
    "text": "benchmarks with it and were tweeting"
  },
  {
    "startMs": 1081280,
    "text": "about it. So, and Japa also improves"
  },
  {
    "startMs": 1084400,
    "text": "multimodel VLM models performance. Here"
  },
  {
    "startMs": 1086480,
    "text": "we are able to cut OCR error rates for"
  },
  {
    "startMs": 1089280,
    "text": "leading models by almost 35%. And this"
  },
  {
    "startMs": 1092160,
    "text": "is an externally validated report. Um,"
  },
  {
    "startMs": 1094720,
    "text": "similar similarly, data bricks actually"
  },
  {
    "startMs": 1097360,
    "text": "achieved 90x cost reduction in their"
  },
  {
    "startMs": 1100000,
    "text": "deployed agents performance. uh uh"
  },
  {
    "startMs": 1102720,
    "text": "performance and here they were able to"
  },
  {
    "startMs": 1104799,
    "text": "tune GPT OSS 120B to outperform Claude"
  },
  {
    "startMs": 1108400,
    "text": "Opus while being 90x cheaper. More"
  },
  {
    "startMs": 1111280,
    "text": "importantly, the performance delta"
  },
  {
    "startMs": 1112960,
    "text": "improvement that you see on top of"
  },
  {
    "startMs": 1114400,
    "text": "Claude Opus is actually bigger than the"
  },
  {
    "startMs": 1116720,
    "text": "one you see on open source models. Some"
  },
  {
    "startMs": 1119200,
    "text": "people have asked me that oh as models"
  },
  {
    "startMs": 1121360,
    "text": "get better the importance of prompt"
  },
  {
    "startMs": 1122960,
    "text": "optimization will go down. I argue the"
  },
  {
    "startMs": 1125440,
    "text": "opposite which is as models get better"
  },
  {
    "startMs": 1128160,
    "text": "they will get better at instruction"
  },
  {
    "startMs": 1129679,
    "text": "following and the more precise"
  },
  {
    "startMs": 1131679,
    "text": "instruction about your task that you"
  },
  {
    "startMs": 1133520,
    "text": "have to give to a very smart model the"
  },
  {
    "startMs": 1135600,
    "text": "better that model will be at a uh"
  },
  {
    "startMs": 1138000,
    "text": "solving your task and this is exactly"
  },
  {
    "startMs": 1139840,
    "text": "what we see happening here the better"
  },
  {
    "startMs": 1141760,
    "text": "the instruction was claopus actually"
  },
  {
    "startMs": 1143520,
    "text": "jumped much uh higher"
  },
  {
    "startMs": 1147360,
    "text": "some people have this question of uh"
  },
  {
    "startMs": 1149600,
    "text": "what if we have subjective tasks which"
  },
  {
    "startMs": 1151360,
    "text": "are very hard to evaluate jpa can"
  },
  {
    "startMs": 1153520,
    "text": "actually learn evals for your task from"
  },
  {
    "startMs": 1155919,
    "text": "production traces. The way to do that is"
  },
  {
    "startMs": 1158240,
    "text": "you collect a bunch of production traces"
  },
  {
    "startMs": 1160080,
    "text": "from your agent. Get a human to annotate"
  },
  {
    "startMs": 1162640,
    "text": "just about 50 of those trajectories"
  },
  {
    "startMs": 1165120,
    "text": "giving very detailed feedback. This is a"
  },
  {
    "startMs": 1166880,
    "text": "long response. This is a short response."
  },
  {
    "startMs": 1168880,
    "text": "This is a good response. This uses this"
  },
  {
    "startMs": 1170640,
    "text": "terminology, whatever. And once you get"
  },
  {
    "startMs": 1173440,
    "text": "those human annotations, you can use JPA"
  },
  {
    "startMs": 1175120,
    "text": "to optimize an LLM as a judge prompt."
  },
  {
    "startMs": 1177280,
    "text": "And you can use that LLM as a judge"
  },
  {
    "startMs": 1179120,
    "text": "prompt then to go back and optimize your"
  },
  {
    "startMs": 1181360,
    "text": "agent and deploy that agent. And this"
  },
  {
    "startMs": 1183760,
    "text": "becomes a data flywheel where you can"
  },
  {
    "startMs": 1185679,
    "text": "keep improving it. And this is a"
  },
  {
    "startMs": 1187520,
    "text": "successful paradigm that uh some leading"
  },
  {
    "startMs": 1189520,
    "text": "teams in production are already using."
  },
  {
    "startMs": 1191760,
    "text": "Then the question we get asked is like"
  },
  {
    "startMs": 1193360,
    "text": "can we actually use this uh reflective"
  },
  {
    "startMs": 1196160,
    "text": "optimization to train models and we"
  },
  {
    "startMs": 1198799,
    "text": "recently had this paper called learning"
  },
  {
    "startMs": 1200640,
    "text": "fast and slow where we propose fast slow"
  },
  {
    "startMs": 1203039,
    "text": "learning where we can co-optimize model"
  },
  {
    "startMs": 1205520,
    "text": "weights and prompt harnesses and this"
  },
  {
    "startMs": 1207520,
    "text": "shows some very strong properties that"
  },
  {
    "startMs": 1209200,
    "text": "one would want in a continual learning"
  },
  {
    "startMs": 1211360,
    "text": "algorithm. Um I don't have much time to"
  },
  {
    "startMs": 1213520,
    "text": "go over details but please uh look at"
  },
  {
    "startMs": 1215520,
    "text": "the uh papers and"
  },
  {
    "startMs": 1219679,
    "text": "uh since uh since release Japa has been"
  },
  {
    "startMs": 1222160,
    "text": "used in production by these companies as"
  },
  {
    "startMs": 1224400,
    "text": "well as the main methodology in these"
  },
  {
    "startMs": 1226480,
    "text": "papers and here the CEO of Dropbox and"
  },
  {
    "startMs": 1229679,
    "text": "Shopify are talking about their use of"
  },
  {
    "startMs": 1231360,
    "text": "Japa and OpenAI also wrote a blog post"
  },
  {
    "startMs": 1233600,
    "text": "about how you can build self-improving"
  },
  {
    "startMs": 1236080,
    "text": "AI systems with Japa. Um so it's very"
  },
  {
    "startMs": 1239520,
    "text": "simple to get started. It can plug into"
  },
  {
    "startMs": 1241520,
    "text": "any framework, any model and it has"
  },
  {
    "startMs": 1244080,
    "text": "absolutely zero hard dependencies. So"
  },
  {
    "startMs": 1246320,
    "text": "you can deploy it any in any kind of"
  },
  {
    "startMs": 1248640,
    "text": "setting. So um don't be afraid to"
  },
  {
    "startMs": 1251679,
    "text": "optimize in the tech space and many"
  },
  {
    "startMs": 1254000,
    "text": "problems can be framed as optimization."
  },
  {
    "startMs": 1255840,
    "text": "So bring actionable side information and"
  },
  {
    "startMs": 1258400,
    "text": "surface as much domain specific"
  },
  {
    "startMs": 1260480,
    "text": "information as you can to optimizers and"
  },
  {
    "startMs": 1262480,
    "text": "the optimizers of future will be able to"
  },
  {
    "startMs": 1264960,
    "text": "work with them. So please go and check"
  },
  {
    "startMs": 1267120,
    "text": "it out. Thank you very much."
  }
]