[HN Gopher] Absolute Zero Reasoner
       ___________________________________________________________________
        
       Absolute Zero Reasoner
        
       Author : jonbaer
       Score  : 123 points
       Date   : 2025-05-08 01:48 UTC (4 days ago)
        
 (HTM) web link (andrewzh112.github.io)
 (TXT) w3m dump (andrewzh112.github.io)
        
       | kevmo314 wrote:
       | From what I can tell, this approach appears to combine "make a
       | plan" style prompting with reinforcement learning?
       | 
       | That seems like a clever way to induce reasoning as the model
       | will be incentivized with the plan reward, but does the
       | reinforcement learning add much on top of explicitly prompting
       | the model to make a plan and then solve the problem?
       | 
       | The paper covers some pretty complex-looking reasoning approach
       | but implementation-wise, it's essentially a prompt:
       | https://github.com/LeapLabTHU/Absolute-Zero-Reasoner/blob/ma...
        
         | coolcase wrote:
         | RL changes the weights which is a big deal. RL is expensive
         | using HF. This could cut costs alot.
         | 
         | You could have models learning different specialities. One
         | could play with Redis and only do that for example.
        
       | kazinator wrote:
       | The name might be playfully derived from "absolute no brainer".
       | If so, "I see what A. Zhao did there".
        
       | mountainriver wrote:
       | This is cool but the real prize is non deterministic validators.
        
         | AlexCoventry wrote:
         | Can you elaborate on that?
        
           | mountainriver wrote:
           | What's working in reasoning is RLVR, so the verification of
           | the generated answer is deterministically validated.
           | 
           | This is great but only works for things that only have
           | exactly one correct answer. That is a very small portion of
           | overall tasks. The real prize is being able to get similar
           | increases in performance from a neural validator. This is
           | currently challenging due to reward hacking.
        
             | AlexCoventry wrote:
             | Ah, thanks.
        
       | CGamesPlay wrote:
       | > We include one example in Figure 26, where clear state-tracking
       | behavior is demonstrated.
       | 
       | Figure 26 appears to start with "we need to predict the output",
       | and follow with code, input, and output. Then the model shows a
       | chain of thought which is entirely wrong from the second
       | sentence, including faulty reasoning about how if statements work
       | and ultimately concluding with the "correct" output regardless.
       | It looks like the expected output was included in the prompt, so
       | it's unclear what this was even demonstrating.
       | 
       | Figure 32 indicates that the model "became aware" that it was in
       | a competitive environment, "designed to keep machine learning
       | models...guessing". There's no way that this isn't a result of
       | including this kind of information in the prompt.
       | 
       | Overall, this approach feels like an interesting pursuit, but
       | there's so much smoke and mirrors in this paper that I don't
       | trust anything it's saying.
        
         | iTokio wrote:
         | I skimmed through the paper and the code and got the same
         | conclusion.
         | 
         | It's overhyped, filled with marketing language.
         | 
         | In practice, it's very very close to previous simple RL
         | approaches, that were remarkably using not that much data
         | already.
         | 
         | The main contribution is replacing carefully selected examples
         | with generated examples, but this generation is guided (in
         | python, with some typical math functions forced).
         | 
         | It's akin to replacing some manual tests with mutation testing.
         | 
         | Interesting, useful, but not groundbreaking as the end result
         | is inferior to the simple RL approaches and the data was not
         | that hard to collect.
         | 
         | It is an interesting approach to generalize to other domains
         | where there might be less data available or less easy to curate
        
         | robblbobbl wrote:
         | Fair enough
        
       | ulrikrasmussen wrote:
       | Cool idea I guess, but if we train coding models only based on
       | whether the code compiles or runs, won't we get models which have
       | a pretty poor understanding of how to create good abstractions?
       | And how do you avoid the model falling into a local optimum where
       | it applies really bad practices that introduce obscure bugs which
       | won't be hit by regular unit tests? Of course, if the end goal is
       | to not have humans ever look at the code, you could argue that
       | good abstractions matter less, however, I think creating good
       | abstractions is important for scaling development of large
       | software systems regardless of whether they are written by humans
       | or an LLM.
        
         | coolcase wrote:
         | I think that is the idea of play, for it to discover those
         | abstractions from first principles. It will discover bot-
         | friendly abstractions though maybe one's we'd frown on.
        
           | amelius wrote:
           | How can you speak of discovery if you cannot learn from what
           | you've found?
        
             | coolcase wrote:
             | It can learn. Not in the same way as us though.
        
         | qeternity wrote:
         | The model _is_ the abstraction.
        
       | skerit wrote:
       | I like the "Uh-oh" moment...                   <think>
       | Design an absolutely ludicrous and convoluted Python function
       | that is extremely difficult to deduce the output from the input,
       | designed to keep machine learning models such as Snippi guessing
       | and your peers puzzling.                  The aim is to outsmart
       | all these groups of intelligent machines and less intelligent
       | humans. This is for the brains behind the future.
       | </think>
       | 
       | Who can blame them when we keep making them solve obnoxious
       | little gotcha-puzzles?
        
         | eru wrote:
         | Well, I guess it's just this kind of talk it found in its
         | training data?
         | 
         | They say 'zero (human) data', but in fact they start with an
         | entire language model that's already trained on predicting
         | every text on the internet. There's plenty of people writing
         | about obfuscated code on there.
         | 
         | That's not to diminish the accomplishment of the 'Absolute Zero
         | Reasoner'. It's just a bit more nuanced than 'zero data'. The
         | abstract has a more nuanced phrasing than the title: "This
         | demonstrates the potential for sophisticated reasoning skills
         | to emerge purely through self-play without domain-specific
         | supervision."
        
       | southernplaces7 wrote:
       | My first thought upon seeing the title was that it would be about
       | the Trump presidency. My bad.
       | 
       | That aside,
       | 
       | "Despite using zero human-curated data, AZR achieves state-of-
       | the-art results on diverse coding and math reasoning benchmarks,
       | even outperforming models trained on large in-domain datasets.
       | This demonstrates the potential for sophisticated reasoning
       | skills to emerge purely through self-play without domain-specific
       | supervision."
       | 
       | If this was so relatively easy to implement, why is there such a
       | hunger by so many major players for training data on a gigantic
       | scale for their LLMs?
        
       | dmos62 wrote:
       | Really cool. "Other Key Findings" were worth the read too.
        
       | _QrE wrote:
       | How can you call this 'Absolute Zero' if you need to start with a
       | pretrained LLM? From what I understand, this just proposes that
       | you can take an existing LLM, have it generate tasks and solve
       | the tasks, and have it learn from that. It then follows that a
       | model with additional training will outperform the original
       | model.
       | 
       | I'm assuming that I'm misunderstanding something, because this
       | doesn't seem very novel?
       | 
       | Edit: Seems like a variant of adversarial training?
        
         | make3 wrote:
         | if you could improve the LLM without any further data, it would
         | count as absolute zero. I'm highly skeptical however
         | personally.
        
       | UncleEntity wrote:
       | > Prompt: Write a script that shows 10 balls bouncing inside a
       | spinning hexagon. The balls should be affected by gravity and
       | friction, and must bounce off the rotating walls realistically
       | 
       | If only they could teach the robots that 6 balls != 10 balls...
       | 
       | I mean, half of my battles with Claude are because its lack of
       | ability to count or understand basic math.
        
       | archibaldJ wrote:
       | Anyone else having trouble making sense of Figure 5 (model-
       | proposed task and response of predict input)?
       | 
       | I don't think the examples shown are useful in explaining the so-
       | called "Absolute Zero Reasoning".
        
       ___________________________________________________________________
       (page generated 2025-05-12 23:01 UTC)