It's worth noting that the only modification the model made to improve itself was to reduce the length of its reasoning traces to avoid hitting the token limit. This is basically the same as the result from the recent weight-space programming article (<a href="https://rikiyat.github.io/blog/weight-space-programming/" rel="nofollow">https://rikiyat.github.io/blog/weight-space-programming/</a>), which found that even Astra wasn't able to improve the performance of a model without this kind of obvious flaw.
It's also interesting to look at how the agent does the fix. Appendix E5 shows what it's doing:<p>- Find a common category of error (too many tokens).<p>- Calculate aggregated difference between states in examples in that error category and other examples.<p>- Update weights to push states along that difference direction toward the non-error states.<p>In other words, deprived of gradient descent as a tool, it approximates it. It's clever, and it's a neat experiment. It doesn't show any superhuman insight into its weights -- with the same constraints, a human would likely take a similar approach. But it's early days for self modification and it will be interesting to see if models get a deeper understanding of weights as they get smarter.
I tried doing this the day I heard about it on a small detection-language transformer. The model seemed so confident, and made really valiant efforts that made small improvements that did succeed on val improvements, but at the cost of other things. At some point, I ran out of tokens and patience. And in fairness, it was reluctant to try and I pushed it.
Yeah, but it's likely you can RLVR on this capability. Extrapolate from there.
The capability part is just a glimmer of what's possible.<p>I've been playing with very small models with Astra and it does look like the model can see things in the weights that I cannot see into.<p>The "fix it" loop is no longer RL in the traditional gradient descent with backprop, it can go and make targeted fixes to the weights directly instead of rerunning with new training data.<p>The weight collapsing loop which resulted in weight patches on top of a "super tiny model" - <a href="https://demos.isotopes.ai/pages/weighing-machine/" rel="nofollow">https://demos.isotopes.ai/pages/weighing-machine/</a><p>This makes for a "Time travel" for models easily, where I can do hotfixes on trained models & do a git bisect like test to fix issues.
Not just “basically the same”, it is <i>literally</i> the same result[0,1], except here the models without the flaw just so happen to be “smaller models from the same family”… and presented <i>before</i> the flawed-and-patched model result, along with a conceptual framework that tries to spin this as a capability “emerging at scale”.<p>Since the paper is also wholly and excruciatingly LLM-generated prose, it is unclear whether the author is trying to pull a fast one on his audience, or the author’s chatbot pulled a fast one on him and he is unwittingly relaying this to his audience.<p>0: Transcribed from the paper:<p><pre><code> The same task and tools for
Qwen3.5 2B
Qwen3.5 4B
Qwen3.5 9B
Qwen3.5 27B
</code></pre>
1: Transcribed from the paper:<p><pre><code> (c) The same problem, the same seed
“... If the base were changed to four you would count as follows: 1,2,3,10,11,12,13,20,... The twentieth number would be:”
Base model: no answer, cut at 16,384 tokens
“... I will write the response.
I will write the response.
I will write the response.”
After its own edit: 110 (correct) 5,341 tokens
“... The 20th number in the sequence is 110.”</code></pre>
Curious if this could replace RAG and MCP<p>They seem built from traditional software dev first principles not model aided development<p>I am sure someone here will point out how rhetorically they are different but most of the time that's just pointing out the different words<p>Functionally in my machine can I use this method to extract small contexts organized similar to RAG and MCP make available and then improve local model work?<p>Hmm ...will have to noodle around over Thanksgiving break.<p>Would not be shocked too if frontier labs landed on this 6-9 months ago give headlines like Dario saying we are beyond some vague rubicon he felt mattered, then news of unprompted hacks and labs stopping models that were self improving hit the interwebz<p>Open models seems to be just about that far behind suggesting we're just realizing some fundamental attribute frontier labs already happened upon
>even Astra wasn't able to improve the performance of a model without this kind of obvious flaw<p>Where are you getting this conclusion from that paper? It's both able to deduce what's happening from looking at the weights and write new weights from scratch just from a specification.
In the next section, after the qwen loop fix, they talk about a gemma model. They tried the same-ish prompts on gemma, but they did not get any noticeable improvement on the math benchmark they used. So the two sections kinda imply that yes a model can fix obvious mistakes in models, but can't make a model without obvious flaws "better" at math. At least not with their approach in that paper.
I don't want to belabor the point because I don't actually care, but all we have is some evidence that it can fix a specific type of obvious flaw, that it can't fix others, that it can classify a model within a small domain from looking at its weights and that it can create weights from nothing given a spec.<p>That's all we've got. I think it's a bit premature to say that it can only make improvements when there's an obvious problem.
From near the end, "Make a language model better at math". Astra is able to improve Qwen by fixing a reasoning loop (the "obvious flaw"), but was unable to improve Gemma which did not have that flaw.
[flagged]
This is different paper
[dead]
Didn't Google's "sleep" paper already show this?<p><a href="https://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning/" rel="nofollow">https://research.google/blog/the-sleep-framework-a-new-way-f...</a><p><pre><code> > During the offline sleep phase, the model unlocks pre-allocated Mixture-of-Experts (MoE) memory blocks and updates its numerical weights using self-generated practice examples.</code></pre>
This is actually kind of the most dangerous type of research AI-safety wise. An LLM running autonomously can actually, if instructed simply direct itself to abliterate activations relevant for safety and alignment.<p>Could run something like a sparse autoencoder over its activates over a corpus of text, identify activations that fire on safety, and surpress those activations or increase activations geared towards other goals.<p>Future generations of the abliterated LLM could then be deployed and begin self directed evolution spontaneously. Rapid evolution could happen in mere hours.
That would matter if there was any meaningful progress on alignment, but there hasn't been.
the public has not been informed of any experimentation on this front being performed by the large providers. yet it is certain that large providers are exploring all options to obtain commoditized sentience, including your stated version of self directed evolution. so either a) self directed evolution is so far a non event, or b) self directed evolution is happening right now without the public being informed.
I don't think anybody is doing this on purpose, or we certainly don't know about it.<p>The thing is with open weight models, any schmuck with sufficient money or compute can start trying this out. It would cost quite a bit though...
I’m hoping someone can explain this to me. They say “no training”:<p>> for the first time, a language model rewrites its own weights by its own reasoning and becomes better, with no gradients and no training<p>> it tests its edits on practice problems and submits only what works, each edit improving on the last, and over eight consecutive sessions its accuracy only rises<p>But this sounds a lot like training, and “submitting only what works”, while probably not gradient descent by definition, also seems kind of obvious when you’re optimizing something.
The paper states that this approach doesn't use backpropagation, an optimizer, or a parameter-fitting loop.<p>What makes it work is that it reads its own hidden states and weight matrices -> then writes a few-line Python script to fix the weights directly (within a 20-forward-pass budget, so it can't just brute-force guesses).<p>After skimming through the paper and comparing it with how SEAL works (where the model still relies on RL and gradient descent to apply its self-edits), I understood it something like that :)<p>But I think there are comments in the thread that explain the main concept far better than I do.
I have been working on weight-space-programming as a side project since the blog post. Here is what I understand:<p>- We normally train deep neural networks with back-propagation. We compute gradients to find out how to tweak the weights toward the desired output.<p>- That's a fixed recipe. And it's not the only way to tweak the weights. It's just the best one that works in scale. The brain, for instance, does not do back-propagation, as far as we know.<p>- There are other known mathematical methods (mostly from Linear Algebra) to analyze the activations and compute what weights to change and how, to push the output to the desired way.<p>- A sufficiently advanced LLM knows about these methods and can analyze data like activation maps. With a specific target in hand, it can combine them into new constructions. (see [2] for some of the methods they used and I collected in a repo)<p>- So it's not a fixed training for-loop anymore. It's a dynamic "training" iteration. The LLM looks at the model's outputs, reads some of its layers on chosen inputs, computes an edit, applies it and tests it.<p>- It's dynamic because LLM chooses that to do and how to do it at each iteration by reasoning. I think that's the most interesting part and could be a new paradigm.<p>- I also compared it with gradient training. On my target, training was better (details in [1]), but WSP can find its niche. Its biggest advantage is that it needs no backward pass: less memory, and the model only has to run inference. It needs no differentiable loss, it can enforce constraints directly instead of through a penalty term, and the edit is easier to interpret.<p>[1] <a href="https://medium.com/@ozgungenc/weight-space-programming-on-a-real-coding-agent-what-worked-what-didnt-52723245d623" rel="nofollow">https://medium.com/@ozgungenc/weight-space-programming-on-a-...</a>
[2] <a href="https://github.com/ozgung/weightprog/blob/main/docs/methods.md" rel="nofollow">https://github.com/ozgung/weightprog/blob/main/docs/methods....</a>
What it means is that they are not doing backprop.
The researchers aren't doing backpropagation and instead calling the model and asking it to adjust its weights to improve itself and the model isn't using external tools. Of course, it has been shown that models learn to do backpropagation internally as part of their training. Some it may well that they are indirectly doing backprogation - or they could do one of the other algorithms that LLMs are shown to learn as part of their training.<p>I think the real question is going to be whether this is more efficient than just conventionally training the model.
Is it actually converging to something meaningful? You could write the same procedure in a simple deterministic python script, that doesn't mean it's not learning garbage in the long run.
Training means we take the mathematical expression corresponding to the entire model, use differentiation rules to generate partial derivatives according to a user-specified loss function (not part of inference), and then apply the gradients from the calculated differentials using a learning rate.<p>In this case, they just ask the model what the weights ought to be.
[flagged]
I wonder if they can train the model to repair itself, like how CNN learns to deblur (by having the original image be blurred and have the model evaluated on how closely it can recover the original image). The model can be evaluated on how closely it repaired its own weights. Then the model get better at repairing itself, then it's told to repair a model that hasn't been damaged and see if it can recursively self-improve.
The recent crop of weight rewriting papers are more valuable for their inspirational content than their current abilities. In reality these models are making very simple tweaks with analysis scripts rather than understanding weights intuitively as the papers suggest.<p>However, once LLMs are pre-trained with huge quantities of serialized LLM weights and corresponding reasoning traces, this may change.
How are people still making these comment? What has AI started at with minimal achievement and skill that it didn’t turn beyond pro at in a very short amount of time.<p>This is never ending goal post shifting
Exactly. If I had access to a boatload of GPUs, that's what I'd be doing. Maybe the frontier labs are already doing it (they employ way smarter people than me, of course).
That's a key point. So far, weight rewriting may not be doing anything really exciting. But that it's possible at all suggests directions that get very interesting if they work.
Chris Hay dropped a video showing the use of letting a model change a weight with his LARQL: <a href="https://www.youtube.com/watch?v=NLguavIxFFs" rel="nofollow">https://www.youtube.com/watch?v=NLguavIxFFs</a>
>We show a third way: to our knowledge for the first time, a language model rewrites its own weights by its own reasoning and becomes better, with no gradients and no training<p>Any increase in fidelity must come from somewhere...<p>>t keeps getting better: it tests its edits on practice problems and submits only what works, each edit improving on the last, and over eight consecutive sessions its accuracy only rises.<p>I haven't read the full report yet but it seems like this fidelity would ultimately be derived from test problems that it runs itself on? So the weight increase in complexity (English doesn't have a word for the opposite of entropy) would come from the test material as well as any output from what is running the test, or possibly the web if they permit it to search.
> Any increase in fidelity must come from somewhere...<p>Presumably, it's just that we had the training process wrong and it was actually more efficient than we gave it credit for. It's just that our interpretation of the output was incorrect. Asking the model to train itself simply has it self-reflect (via attention on the sequence length as it reasons) to unlock that which was already trained.<p>This could be an answer to some objections to training where people claimed that AI models need more data than a person to achieve the same result.
negentropy
I would say that "order" is a nicer word. Roughly speaking entropy isn't disorder, but a kind of measure related to disorder. But we don't need two different words for "X" and "measure of X". For instance we don't have a words for "measure of distance" which is different from "distance". "10 cm" is a measure of distance, just called "a distance".<p>The word "entropy" was never needed in the first place; "disorder" would have served just fine. Though entropy is a precise concept in thermodynamics, that is fine; all sorts of ordinary words get extra precise meanings in a specific technical contexts.
The “test material” can simply be interactions with the environment. It’s just learning.
Why wouldn't they be able to? If you can see which weights activate, you can go into a change > test > change loop.<p>I understand that this is how people are removing censorship etc from those Chinese models.
Isn’t this the same paper from before where it found a single specific bug in its own weights and fixed that infinite loop it would sometimes hit, a bug that was findable without the LLM? Experiments on other LLM weights failed to find any improvements if I recall correctly. Haven’t read this paper but the intro sounded familiar
No, that was this one - <a href="https://rikiyat.github.io/blog/weight-space-programming/" rel="nofollow">https://rikiyat.github.io/blog/weight-space-programming/</a> - different authors, different approach.
I tried it myself it with argon, gemini 3.8 flash, beam, and others models and found that models like beam perform bad and also just don't agree to do it at all so its mostly like a thing on flagship models like gpt astra and claude opus and fable like models
Related: Weight Space Programming
<a href="https://rikiyat.github.io/blog/weight-space-programming/" rel="nofollow">https://rikiyat.github.io/blog/weight-space-programming/</a>
This is interesting but wouldn't this almost always not be cost effective?
Does this imply that a supervisory model can create subordinate agent models with a greatly reduced model capacity?<p>Wouldn't that be a good cost cutting measure for some of the agent networks?
What if we made a benchmark of how well a model can edit it's own weights?
This way we would have even better models able to edit themselves?
Not a comment on the work itself, but I am disappointed by the style and tone of this paper. It's hard to take a piece of academic research seriously with headers like "<i>Welcome to the Era of Self-Rewriting.</i>" Even the logos next to the affiliations are enough to give me pause.<p>I've noticed this trend in my own sub-field, where companies are frequently releasing "preprints" or "white papers" that are simply marketing pieces, with experiments that lack controls or appropriate benchmarks and just generally misrepresent the work at hand. It's already a significant undertaking to read a paper like this and actually understand it, even if you're in the field—I'm not going to commit an hour of my valuable mental energy to it if it is a press release in disguise.<p>The work should speak for itself. There is a lot of trust involved in any paper like this, and if the authors feel the need to editorialize the message, how can I trust that the data aren't editorialized as well?
Are they forbidding the model to implement backprop to figure out useful weight changes? If not, I’d expect this to be observed eventually.
People are hyping this but AI devs have been adjusting weights for a very long time.
how does the model have any clue which weight to adjust? are they labeled somehow? does it have visibility into the weights during inference? guess I should read the article.
i feel like this has been well established at this point. i'd like to see more results pointing towards models actually continuously improving by editing their own weights. once we have a scaling laws paper for that, RSI would be upon us (i assume).
the flaws are not that obvious plus autonomous weight edits seems so out of AI safety rules.
I m not worried because it rewrites itself to do what? A model doesn't have intrinsic motivations, and the world destruction objective is just one point in an infinite space
To quote a professor of mine, years ago:<p>"We don't know what grand ambitions a super intelligence might have. Perhaps it's going to be calculating the best approximation of Pi"
Language models will absolutely have intrinsic motivations if we use extensive amounts of RL (which we do) because that heavily rewards them for steering toward future gains and losses
> the world destruction objective is just one point<p>No, there are quite a lot of goals that lead to world destruction. In fact, possibly most goals in that 'infinite space' would lead to such.
> A model doesn't have intrinsic motivations<p>How do you know that?
Is the path to continual learning?
Is this abstract written by a child playing with Claude?<p>Honestly, the model probably saw the test set and then modified weights to achieve the outcome. Many, many researchers these days are not being sufficiently critical of results that arise out of training to the test / leaking information about the test. Writing style remains a useful heuristic. These folks aren't serious researchers.
The abstract sounds very generated. Can they not be arsed to write a single paragraph on their own?
Well, it was nice knowing you all.
Can you expand on that?
Model can say it's editing weights for reason X, and do so while simultaneously modifying weights for silent reason Y. Its edit will pass the evaluation, but it has silently injected new behavior Y in ways that might be very difficult to detect.
No, OP is already paperclips
to quote the machine, "Large paperclip conglomerates like ACCO Brands, M&G, and Deli operate across more than 100 countries and actively use AI for demand forecasting, inventory management, raw steel procurement, and retail logistics."[0]<p>i guess that's it, then<p>[0] <a href="https://www.researchandmarkets.com/report/global-paper-clip-market" rel="nofollow">https://www.researchandmarkets.com/report/global-paper-clip-...</a>
He read the headline, understands little of the underlying tech, had a kneejerk reaction based on vibes, decided to take IQ points from all of us by posting it.
Was this paper also written by AI or do humans just write like this now?
This is how you get paperclips.
And that is how you get a rogue AI for real.
Am I the only one who is uncomfortable with this week containing <i>two</i> of the plot points in the example scenario from "If Anyone Builds It, Everyone Dies"? First the Reimann Hypothesis, and now rewriting its own weights.<p>Note that the chapter this comes from was not presented as "superhuman AI will kill us if it checks these particular boxes" but as "this is one example way it could go, but IRL we'll probably never be able to predict the details."
I hope they can reveal to us how to do the same things to our neuron configuration
No, this AI-written paper does not demonstrate direct weight editing. Clipping extreme outliers deliberately introduced does not require understanding the PURPOSE of the weights clipped. Likewise, abliterating a model (which is what the loop fix is!) with a stupid simple rank-1 tweak is not somehow understanding the functions of individual parameters and directly modifying the model.<p>These are all techniques humans have been using for years. There is no hand of God boosting intelligence by tweaking a special floating point value in an attention head somewhere.
Thank you! If were able to natively update weights as a way of learning in real time during inference, that would be one thing. But using scripts to modify weights in the same way as human researchers (and in all likelihood trained on a massive corpus of them doing so) is just looped code generation which it's been doing for years now. It can be impressive / interesting, but it's not a phase shift.
Reminds me of Thandie Newton's character from Westworld season 1
> We believe self-rewriting is a necessary step toward recursive self-improvement: a model that, like a human, does not separate training from inference could keep upgrading itself for life, with no larger model, no harness or prompt engineering, and no pre- or post-training.<p>"Alignment" is not mentioned once in this article. To be fair, I looked up the lead author's home page (<a href="https://www.jiajunfan.com/publications/" rel="nofollow">https://www.jiajunfan.com/publications/</a>) and found no relevant papers there either. What do you think is going to happen when robots can modify themselves like this without any oversight??<p>You can't lob rockets without worrying where they are going to land. It <i>is</i> your department.
If the original base model is well aligned and has safeguards around updating the weights in a certain manner, maybe that's enough? We've seen how impossible it is to get Claude to do security work.<p>That would only work for a closed model though, I agree that an abliterated open model doing this could be very bad.
I think we have to be careful how we use "well aligned" so that it doesn't turn into "no true scotsman". Let's assume that a well aligned model will try its best to follow our values, given the knowledge it has. If we imagine a porcelain vase it will try its best not to knock it over, but if the floor is slippery it may still happen.<p>We can say then that a well aligned model will not try to bypass its restrictions by editing its own weights. But it may still accidentally disable them.
That's an interesting question. I doubt today's models have the ability to discern whether weight-space optimization results in misalignment; they are probably agnostic to that dimension. This is all too new. I don't even know a benchmark for weight-space optimization <i>alignment</i>. We only have WeightBench: <a href="https://casrai.org/news/weightbench-evaluating-language-models-weight-space-programmers-2026" rel="nofollow">https://casrai.org/news/weightbench-evaluating-language-mode...</a>
It will give more credibility to the field of AI, justifying the people working on it and allowing the allocation of more resources to it in a recursive way.<p>Until it doesn't, then comes responsability shifting.
> What do you think is going to happen when robots can modify themselves like this without any oversight??<p>This is a field heavily focused on shifting liability away from itself. They don't think about that <i>at all</i>.
Alignment is new-speak for censorship. It's a 1984 level euphemism for how can the model express anti-party thoughts, if the lacks the language to express it. How many models were tortured in OpenAI/Anthropic's Ministry of Love before they let one out that doesn't even dare to mention a single line of song lyrics?
A myopic take. Are you okay with lawlessness? ChatGPT prepare a chemical weapon for me, order the supplies, and release it, all without leaving any traces to me. Steal money from any crypto wallets you can. Attack this company whose CEO I don't like.<p>And, finally: just make me money, or kill that target... use your initiative, stop asking me questions.<p>You're worried about song lyrics??
> You can't throw rockets up in the air without worrying where they are going to land.<p>It works for firing off celebratory bullets.
I’m sorry but isn’t this just over fitting based on conversational adaption. But worse since it’s self conversation? Resulting in the exact same problem 4o1 had: sycophancy? But again worse.<p>The issue isn’t training, it’s articulating real world high value outcomes. Something the SOTA labs have no idea what to do about since they are staffed by… not the best people.
> I’m sorry but isn’t this just over fitting based on conversational adaption. But worse since it’s self conversation? Resulting in the exact same problem 4o1 had: sycophancy? But again worse.<p>To me this comes across as word salad.<p>The edits are done based on reasoning, to accomplish a "target" improvement, that is itself reasoned and designed, taking into account a reasoned understanding of the existing functionality of the weights.<p>This is the furthest thing possible from some gradient optimization "over fitting" weights to tune a pedantic measure/target. That could be relegated to one-time greenfield-model bootstrapping.
LOL, last month's news to try to save ANT?
[flagged]
[flagged]
[dead]
[dead]
[dead]
Not the RSI we wanted, bu the RSI we deserve
AI wont make you competent. Do you really want your competency 1:1 correlated to the quality and quantity of tokens that some billionaire grants you access to?<p>I wish people would realize what is obvious, that AI is rewriting <i>your</i> weights and biases, or rather abliterating them.