Ladder: Self-improving LLMs through recursive problem decomposition

by fofoz • 1 year ago

370 110 1 year ago

We also introduce TTRL (Test-Time Reinforcement Learning), where we perform reinforcement learning on variants of test problems at inference time. TTRL enables Qwen2.5 7B Deepseek-R1 Distilled to achieve a state-of-the-art score of 90% on the MIT Integration Bee qualifying examination, surpassing OpenAI o1's performance.

That's incredible!

Davidzheng4• 1 year ago

i mean they have a verifier, so can't they even get to 90% just by random generation by the net and testing against verifier until it's numerically correct? I think the end solve rate is less important and the generality of approach is maybe more important

vessenes3• 1 year ago

No, they specifically test for this (the "RL" case). They most particularly can not do this with random generation, which is very interesting.

yangikan• 1 year ago

Is there code for this paper or something that does something similar to this?

mentalgear3• 1 year ago

It's exciting to see approaches like RL and curriculum learning, which I always felt were the way to go for real self-improvement ~7y ago when training in robotics (openAI gym days), finally getting successfully applied to NLP/LLM to highly boost small model performance.

(Ladder is a sort of RL self curriculum learning approach)

all2• 1 year ago

I know all of these words, but I do not know what they mean together.

What is curriculum learning?

What is the "RL" approach?

"~7 ago": days? Weeks? Years?

What is an "open ai gym days"?

LLMs and robotics?

katzenversteher• 1 year ago

I can only help with RL, that's probably reinforcement learning. As far as I remember that means you let the model perform a task that can be "graded" and then depending on how well it did it get's a reward it want's to maximize. I believe (this it where I'm very insecure, I could be wrong) the neurons (weights / biases) of the neurons that where involved in reaching the highest reward get adjusted to have a bigger influece.

FeepingCreature• 1 year ago

I fed the two comments verbatim into Grok:

Curriculum learning: Training that begins with easy examples, gradually increasing difficulty.
RL (Reinforcement Learning): Learning via trial-and-error with rewards, like training a robot or model to optimize actions.
~7y ago: ~7 years ago (circa 2018).
OpenAI Gym days: Refers to using OpenAI Gym, a toolkit for RL, popular in robotics/AI research ~2016-2018.
LLMs and robotics: Large Language Models (LLMs) now leverage RL techniques from robotics for better performance.

I think the last one is a semi-hallucinatory stretch. LLMs are large language models, ie. ChatGPT, Sonnet, Grok, R1. Robotics are ... robotics. Building robots.

The actual answer to what the comment is saying is that until maybe a year back, we trained language models - still with RL, but with RL on token error, which isn't "real" RL because it executes tasks "by coincidence". That is, it happens to be that when you train a model to predict text, it also gains the ability to do tasks in the bargain, because the text contains agents that do tasks. A year or so ago, we started training models by having them do a task, judging if the task was successful or failed, and then performing RL on task outcome rather than token prediction. This is a return to "classic RL", but we had to pass through the "token RL regime" first so that the model could make progress on realistic tasks at all. It also means that LLMs can now increasingly be employed in robotics, where task RL training rules, as there is no massive preexisting robotics movements dataset like there is for text.

(Also, NLP is Natural Language Processing, ie. what LLMs do.)

NitpickLawyer2• 1 year ago

Curriculum learning posits that you get better results if you gradually increase the training "difficulty". That is, learn to walk before you run. So you'd do "additions and multiplications first" and then "now draw the rest of the integral" :)

RL - Reinforcement Learning. You have a carrot and a stick. You run a model through iterations (in LLMs you generate n completions), you score each of them based on some reward functions, and if the result is correct you give it a carrot (positive reward), if the result is incorrect you give it a stick (negative or 0 reward). (simplified ofc)

OpenAI gym is (was?) an environment that allowed "AI agents" to be simulated in an environment. You could for example play games, or solve puzzles, or things like that. oAI gym was a "wrapper" over those environments, with a standardised API (observe, step (provide action), reward; rinse and repeat). You could for example have an agent that learned to land a lunar lander in a simple game. Or play chess. Or control a 3d stick figure in a maze.

InvidFlower4• 1 year ago

I think you'll get a lot of use from this video from Andrej Karpathy: https://www.youtube.com/watch?v=7xTGNNLPyMI

It is long, but don't get scared off. He goes over a ton of different stuff related to model training, but makes it very easy to understand.

yu3zhou4• 1 year ago

Why is it the same as curriculum learning? I thought it's about starting with simple, more general samples and then progressing towards more difficult and complex during training?

mentalgear3• 1 year ago

We demonstrate LADDER's effectiveness in the subject of mathematical integration, improving Llama 3.2 3B's accuracy from 1% to 82% on undergraduate-level problems

RossBencina• 1 year ago

Keep in mind that state of the art term-rewriting systems perform very well on symbolic integration, e.g.: https://rulebasedintegration.org/

jgalt212• 1 year ago

Indeed. Don't the LLMs have access to the RUBI rule set? It's open source. Why are they not memorizing the rules?

EMIRELADERO2• 1 year ago

What the hell is going on this week?!?!? (asking positively, with a smile on my face)

I have seen at least 3 interesting/mildly promising breakthroughs on ML just these past two days! I mean, a Google research team just discovered that you can combine NNs with CLAs using digital logic gates as a medium, so you could potentially reduce many kinds of non-linear problems to a simple, efficient digital circuit! And it was on the HN front page, TODAY![1]

I keep seeing more mind-bending stuff related to neural nets and logic/intelligence in general, my mind has been running wild with speculation about the future and just how close we could (or could not) be to truly understanding how intelligence works from first principles.

[1] https://news.ycombinator.com/item?id=43286161

meitham5• 1 year ago

>> asking positively, with a smile on my face

Responding with unexplained fear in my heart, we’re just getting closer to Skynet!

blooalien3• 1 year ago

Responding with unexplained fear in my heart, we’re just getting closer to Skynet!

I'll take a cold logical machine super-intelligence over the mad human lunatics wielding current iterations of "A.I." technologies in some really terrifyingly dangerous ways. As someone else commented on some other thread earlier "I look forward to being paperclips".

amelius• 1 year ago

It took some engineering effort but from now on we're getting there through soft-skills.

noosphr4• 1 year ago

This is secret sauce that people have been hoarding for the last year or so.

With the deepseek open source releases this is now worth a lot less and companies are cashing out on reputational increases instead of being scooped.

I have done this exact thing in September 2023 with llama2 finetunes but couldn't get approval to share it with anyone.

EMIRELADERO2• 1 year ago

Interesting! What results did you get with that?

Also, do you think this is what O3 is doing?

NitpickLawyer2• 1 year ago

couldn't get approval to share it with anyone.

sounds like MS :(

They had some killer research projects with various teams around the world, but eventually they all got snuffed.

eru2• 1 year ago

It's interesting to compare these signs of progress with the disappointment that GPT 4.5 was so far.

mentalgear3• 1 year ago

Exciting that we see now so many new approaches to AI/ML since the industry FINALLY realising that naive scaling will not bring us to AGI [0].

This also has the added benefit of small players being able to compete and contribute with actual innovation in a space where the big players (openAI/MS) wanted to make us believe for years that we/open-source couldn't ever catch up on them (infamous Altman quote).

So much resources, time and money wasted on pure GPU crunch scaling the last couple years.

[0] as pointed out by Gary Marcus years ago. Evidence GPT 4.5 after ~ 2 years training, disappointing results.

kadushka2• 1 year ago

I've been using GPT-4.5 for the last couple of days. To me, it's pretty much AGI already. Or at the very least it is smarter than me.

estebarb• 1 year ago

I believe that it is related with important conferences opening papers reception soon. Some disallow publishing in preprint for some weeks before the paper reception, so people may have been rushing uploading stuff.

bearjaws• 1 year ago

LLM break throughs are the new battery break through. We just aren't as good at quanitfying the trade offs yet.

brookst3• 1 year ago

Eh, I think LLMs have seen considerably greater real world advancement in the past few years than batteries have.

Workaccount22• 1 year ago

It's a fresh orchard full of low hanging fruit.

Regardless of ultimate utility, it's shiny, hyped, has a huge wow-factor, and is having trouble keeping up with the amount of money being thrown at it.

This means it has captured the attention of a huge portion of the most capable people, who naturally want to take a crack at making a breakthrough.

patcon• 1 year ago

Maybe this is the pace of research/work when the researchers can augment themselves with AI. You're feeling the exponential launch in the first place we're likely to feel it.

selimthegrim• 1 year ago

What was the third one?

bloomingkales5• 1 year ago

I’m kinda getting the sense this is still just prompt engineering in a loop.

Persona-based prompting: We prompted the model to adopt different mathematical perspectives (e.g., "think like Euler focusing on series", "approach like Gauss looking for patterns").

I mean … I guess that’s scientific?

Besides that, how can the model learn at test time (at inferencing)?. It’s stateless, it doesn’t incorporate the last prompt into the model.

regularfry3• 1 year ago

It learns through the context. The context is state.

There's a bit of a cheat that's going on here though in that the model is being given the fundamental integration operations as part of the problem. That means the model hasn't had to learn what they are. It might not have needed to be given them, but it does feel like that's giving the model a leg up in the benchmarks that it wouldn't otherwise have, and when there's a direct comparison to (e.g.) DeepSeek, that's an unfair advantage.

viking123• 1 year ago

Does the massive context still needs to be dragged around? Until get a neural network that adjusts weights in real time without relying on a huge clump of context being cycled there, I don't think there will be an AGI or even an impressive "agent". Current agents are just LLM looping lmao sold to people with no knowledge how they work at all.

Davidzheng4• 1 year ago

test-time training/RL is definitely the right approach for math AI in the future. It is probably one of only a few ways to spend an obscene amounts of compute at a given problem (think 10^5 gpus for a few days) and has hopes of making progress when test-time inference scaling may not at first (think if you try to do MCTS on a go position with a bad value/policy net). Alphaproof already did this but nice to see it done again--good results!

Davidzheng4• 1 year ago

actually I think the interesting thing is to see how much of the boosted performance can be distilled back into the LLM at small sizes. Then you can emulate really how alphazero works b/c you have a policy improver (test time rl with similar problems). and we get to see just how strong a small net can be theoretically (say 32B or something)

eru2• 1 year ago

This might be trivial, or deep, depending:

We have done a lot of that improving historically by publishing research and textbooks. I can solve (some) problems today in minutes that would have stumped Isaac Newton for a life time (or at least a few weeks).

Of course, you are hinting at a more general distillation, I suspect.

InvidFlower4• 1 year ago

We've already seen Qwen's new QWQ 32B (not distilled) model doing impressive things on benchmarks. It'll definitely be interesting to see how just good small models can get. When combined with rag and large context window for expanded knowledge, might be able to get pretty far.

brookst3• 1 year ago

A little like the T2 read only versus learning mode switch: operating on intrinsic knowledge versus able to reason improvements.

neoneye2• 1 year ago

Sidenote: `Tufa Labs` team includes the `MindsAI` team of ARC-AGI fame. https://tufalabs.ai/team.html

ThouYS• 1 year ago

nice!

niemandhier• 1 year ago

Frank Herbert knew it: This is basically an implementation of the mentats recursive self inspection described in Dune.

isaacfrond2• 1 year ago

Reminds me of a quote by famous number theoretic mathematician Hendrik Lenstra:

For every problem you can't solve, there's a simpler problem that you also can't solve.

techwizrd• 1 year ago

Is this quote real? I'm familiar with George Pólya's, "If you cannot solve the proposed problem, try to solve first a simpler related problem" but I cannot find any source for the Lenstra quote.

gessha• 1 year ago

I also found it connected to Polya [1]

https://www.pleacher.com/mp/mquotes/mobquote.html

isaacfrond2• 1 year ago

I’ve heard him say it myself in a lecture on the AKS primality test. So, ehh, the source is oral tradition I guess.

v1t• 1 year ago

yeah https://www.reddit.com/r/quotes/comments/16qgwcv/if_you_cant...

Horffupolde• 1 year ago

That doesn’t induce nicely. Unless it was an insult.

bubblyworld• 1 year ago

Monotonic sequences can be bounded!

samstave3• 1 year ago

It reads like the famous Churchill quote about "if you gave me poison I would drink it"

deadbabe• 1 year ago

If you could solve the simpler problems, you’d be able to solve the larger problem. But you can’t, because you can’t even solve a simple problem.

arnarbi• 1 year ago

It’s not induction. It’s just the contrapositive of “if you can solve the simpler problem then you can solve the harder problem”

pyryt• 1 year ago

Some names are just too tempting https://arxiv.org/abs/1507.02672

revskill• 1 year ago

Llm keeps deleting my file content proved that we have far many things to do.

explosion-s• 1 year ago

I would love to be able to use the actual model! If I'm understanding correctly this makes small models as intelligent as much larger models like GPT4o

barteloniu2• 1 year ago

Their test time RL approach seems a bit fishy. From what I understand, TTRL works by asking a language model to generate simpler versions of the test case. Once we have the simpler problems, we run RL on them, hoping that an improvement on the simplified cases will also strengthen the model performance on the original problem.

The issue is, they use a numerical integrator to verify the simpler problems. One could imagine a scenario where a barely simpler problem is generated, and the model is allowed to train on pretty much the test case knowing the ground truth. Seems like training on the test set.

The rest of the paper is nice though.

thomasahle2• 1 year ago

the model is allowed to train on pretty much the test case knowing the ground truth

The task is to solve the integral symbolically, though, right?

It's a hard problem to solve, even if the model is given access to a numerical integrator tool it can use on the main problem itself.

barteloniu2• 1 year ago

That's a fair point.

goyel• 1 year ago

I wonder why nobody made a NN to find the weigths faster and better than gradient descent

thomasahle2• 1 year ago

At the end of the paper they mention "two problems from the 2025 MIT Integration Bee qualifying exam which the system continued to answere incorrectly".

They say the questions were among the most complex questions on the exam, but the first one is just

   ∫ ∛(x · ∜(x · ∜(x · √(x · √(x · ⋯ ))))) dx

which just requires you to compute

   1/3 + 1/(3*4) + 1/(3*4*5) + ...

So hardly very advanced math.

Workaccount22• 1 year ago

It's a 7B model. So while the problem is not advanced the model is far from it too.

johntb86• 1 year ago

I'd be curious what would happen if you SFTed a larger model with successful reasoning traces from the smaller model. Would it pick up the overall reasoning pattern, but be able to apply it to more cases?

nis0s• 1 year ago

What’s the difference between this and what Wolfram Alpha has been doing?

https://www.wolfram.com/artificial-intelligence/

evjan• 1 year ago

I had NotebookLM make a 15 min podcast about it and listened to it while walking the dogs. It was a very interesting way of trying to understand a research paper!

You need a google account to access it unfortunately. https://notebooklm.google.com/notebook/fbaba495-d4f2-48a3-a3...

cratermoon2• 1 year ago

How many rungs of a ladder would you be willing to climb if you knew that each rung was made from half the previous rung?

ma9o• 1 year ago

divide and conquer :)

flakiness• 1 year ago

Off topic, but their site is lovely: https://tufalabs.ai/index.html It feels like a gold rush for sure.

daxfohl2• 1 year ago

How much GPU would an RL like this need for tuning? Is the approach something someone could experiment with themselves, or is it like thousands of USD in cloud costs and/or years of compute if done on a laptop GPU?

ekidd• 1 year ago

I've seen recent interesting papers on reasoning models with costs from US$6 to US$4,5000. The problem is that you need a bunch of fast RAM for efficient training. But you can do some limited fine-tuning (Q-LoRA, etc) of models up to 14G on a 24 GB graphics card, and full fine-tunes of 1.5G models.

It's very affordable for a small university research group. And not totally out of reach for hobbyists.

daxfohl2• 1 year ago

Really $6? Where was that?

vessenes3• 1 year ago

That this works at all is pretty interesting. That it seems to work very well with math is quite interesting.

That said, this paper is part of the move we have right now blurring the lines of training and inference -- part of their method involves doing some reinforcement learning on questions they don't know the answer to, but can decompose into simpler questions, and using GRPO on those with a numerical 'checker'. This reinforced model then can answer more questions.

I like this. I think humans do this a lot; mulling on something, turning it over in their heads, analogizing, etc. Adding test time training is a way to do a lot more thinking than adding tokens to the context for fixed inference.

Just as DeepSeek and o1/o3 show that we can increase capacity with inference-time-token generation and assessment, it looks like we can increase capacity with inference-time automated fine tuning as well.

I'd hope that as these techniques solidify we'll have a new way to talk and think about this -- they are all part of the same fundamental process at some level.

Either way, super cool.

Ladder: Self-improving LLMs through recursive problem decomposition

Related Stories