Really fascinating how this works; it's basically context-aware decoding. From the paper:
> Code interleaves fork positions, where several continuations are genuinely plausible and may correspond to different solution approaches, with lock positions, where syntax and semantics leave little ambiguity but a low-probability distractor tail still remains… The best global decoding setting is therefore necessarily a compromise; we call this tension the precision-exploration conflict.
In other words, just like us, the model needs to shift from "exploration" in "fork" mode (divergent thinking to produce a creative solution) to "precision" in "lock" mode (producing syntactically correct code).
What this paper shows is that their simple technique (SSD) can improve the ranking of optimal tokens in both lock and fork positions, meaning the model is more likely to explore when it should be exploring, and more likely to be precise when it needs to be.
I love that we're still learning the emergent properties of LLMs!
> I love that we're still learning the emergent properties of LLMs!
TBH, this is (very much my opinion btw) the least surprising thing. LLMs (and especially their emergent properties) are still black boxes. Humans have been studying the human brain for millenia, and we are barely better at predicting how humans work (or for eg to what extent free will is a thing). Hell, emergent properties of traffic was not understood or properly given attention to, even when a researcher, as a driver, knows what a driver does. Right now, on the front page, is this post:
> 14. Claude Code Found a Linux Vulnerability Hidden for 23 Years (mtlynch.io)
So it's pretty cool we're learning new things about LLMs, sure, but it's barely surprising that we're still learning it.
(Sorry, mini grumpy man rant over. I just wish we knew more of the world but I know that's not realistic.)
I'm a psychiatry resident who finds LLM research fascinating because of how strongly it reminds me of our efforts to understand the human brain/mind.
I dare say that in some ways, we understand LLMs better than humans, or at least the interpretability tools are now superior. Awkward place to be, but an interesting one.
LLMs are orders of magnitude simpler than brains, and we literally designed them from scratch. Also, we have full control over their operation and we can trace every signal.
Are you surprised we understand them better than brains?
We've been studying brains a lot longer. LLMs are grown, not built. The part that is designed are the low-level architecture - but what it builds from that is incomprehensible and unplanned.
"Designed" is a bit strong. We "literally" couldn't design programs to do the interesting things LLMs can do. So we gave a giant for loop a bunch of data and a bunch of parameterized math functions and just kept updating the parameters until we got something we liked.... even on the architecture (ie, what math functions) people are just trying stuff and seeing if it works.
> We "literally" couldn't design programs to do the interesting things LLMs can do.
That's a bit of an overstatement.
The entire field of ML is aimed at problems where deterministic code would work just fine, but the amount of cases it would need to cover is too large to be practical (note, this has nothing to do with the impossibility of its design) AND there's a sufficient corpus of data that allows plausible enough models to be trained. So we accept the occasionally questionable precision of ML models over the huge time and money costs of engineering these kinds of systems the traditional way. LLMs are no different.
Saying ML is a field where deterministic code would work just fine conveniently leaves out the difficult part - writing the actual code.... Which we haven't been able to do for most of the tasks at hand.
Interestingly enough, for a while physics used to be studied by philosophers (and used to be put in the natural philosophy basket, together with biology and most other hard sciences).
The intersection of physics isnt psychology it is philosophy, and the same is true (at present) with LLM's
Much as Diogenes mocked Platos definition of a man with a plucked chicken, LLM's revealed what "real" ai would require: contigous learning. That isnt to diminish the power of LLM's (the are useful) but that limitation is a fairly hard one to over come if true AGI is your goal.
Sir Roger Penrose, on quantum consciousness (and there is some regret on his part here) -- OR -- Jacob Barandes for a much more current thinking on this sort of intersectional exploratory thinking.
Learning about the emergent properties of these black boxes is not surprising, but it's also not daily. I think every new insight is worth celebrating.
Oh I very much agree that it's great to see more research and findings and improvements in this field. I'm just a little puzzled by GP's tone (which suggested that it isn't completely expected to find new things about LLMs, a few years in).
Indeed. For me, it's also a good reminder that AI is here to stay as technology, that the hype and investment bubble don't actually matter (well, except to those that care about AI as investment vehicle, of which I'm not one). Even if all funding dried out today, even if all AI companies shut down tomorrow, and there are no more models being trained - we've barely begun exploring how to properly use the ones we have.
We have tons of low-hanging fruits across all fields of science and engineering to be picked, in form of different ways to apply and chain the models we have, different ways to interact with them, etc. - enough to fuel a good decade of continued progress in everything.
I mean... You could? AI comes in all kinds of forms. It's been around practically since Eliza. What is (not) here to stay are the techbros who think every problem can be solved with LLMs. I imagine that once the bubble bursts and the LLM hype is gone, AI will go back to exactly what it was before ChatGPT came along. After all, IMO it's quite true that the AIs nobody talks about are the AIs that are actually doing good or interesting things. All of those AIs have been pushed to the backseat because LLMs have taken the driver and passenger seats, but the AIs working on cures for cancer (assuming we don't already have said cure and it just isn't profitable enough to talk about/market) for example are still being advanced.
I hate to "umm, akshually" but apparently we have been studying the brain for thousands of years. I wasn't talking about purely modern neuroscience (which ironically for our topic of emergence, (often till recently/still in most places) treats the brain as the sum of its parts - be them neurons or neurotransmitters).
> The earliest reference to the brain occurs in the Edwin Smith Surgical Papyrus, written in the 17th century BC.
I was actually thinking of ancient greeks when writing my comment, but I suppose Egyptians have even older records than them.
I've always thought that it is kinda weird that we spend exactly the same amount of compute to calculate both "fork" tokens and "lock" tokens.
I think that with grammar-aware sampling / constrained decoding [0][1] it is possible to sometimes skip calling the model altogether if only one token is allowed by grammar and just insert it, but I don't think that any of the current, widely used combinations of models/harnesses use it. And it only skips inference in rare edge cases.
I wonder if there is a more general solution that can make models spend more compute on making important choices, while making generation of the "obvious" tokens cheaper and faster.
We kinda have a little bit of it with some coding harnesses giving model access to LSP, but I think that we can insert this knowledge on a lower level if we find a clever way to somehow utilize it during sampling.
I think that there is a lot of low hanging fruit in this area.
And in general, I think that people try to use LLMs too much to solve problems that can be easily solved by cheaper (computationally), and, more importantly deterministic tools.
For example, back in the day when LLM-assisted coding just became a thing people very often complained about models generating syntactically incorrect code and inventing non-existent library methods.
Well, I, an experienced human programmer, probably would also be making syntax mistakes and inventing non-existent methods if you stripped me of my tools and made me write code in a bare text editor without syntax highlighting.
Thankfully, my IDE would autocomplete real syntax and actually existing library methods for me and immediately give me feedback if I make a mistake anyway. And all of it is achieved using reliable deterministic code without the inherent issues of statistical models.
I think that it is really inefficient to reach for an expensive and unreliable tool when a cheap and reliable tool will do.
In general these agents support LSPs, which is often as much information as your IDE will give you. They are also not required to output syntactically correct code token by token when running agentically, because the loop is:
1. code
2. syntax check / build / format / lint (details language dependent)
3. test
and they can hop between 1 and 2 however many times they want.
Doing a tool call for autocomplete is not going to make coding agents faster.
I do think there is some merit in a tool that dumps all namespaces and reachable symbols so the agent can do its own autocomplete without a round-trip.
> Give coding agents access to intellisense and syntax highlighting.
i once asked an LLM if it could ingest code from an interactive session more easily if it were in appropriately-typed markdown fences and it said absolutely yes, and that the syntax highlighting fed to it that way helps it immensely. i was downright shocked that syntax highlighting was anything more than noise for them.
> I wonder if there is a more general solution that can make models spend more compute on making important choices
There's a lot of work going on in various streams towards making it possible to vary compute per-token, dynamically, e.g. universal transformers. Maybe one day it'll work well enough to beat conventional techniques.
Another example of the mindf@#$ these systems are: I was doing some fine tuning to a small model, take data fields and make a sentence out of it. I was running into mode collapse (basically when the AI simplifies too much and always output the same thing).
I got unstuck by randomizing the field order for each row?!? At training, and now I'm thinking I should do the same at inference time...
the irony of modern software engineering: we spent decades perfecting deterministic algorithms, and now we're basically just shaking a black box and hoping the magic rocks align.
apparently you can straight up duplicate/add/rearrange layers without changing any of the weights and get better results as well - https://dnhkng.github.io/posts/rys/
> This is probably due to the way larger numbers are tokenised, as big numbers can be split up into arbitrary forms. Take the integer 123456789. A BPE tokenizer (e.g., GPT-style) might split it like: ‘123’ ‘456’ ‘789’ or: ‘12’ ‘345’ ‘67’ ‘89’
xVal basically says "tokenizing numbers is hard: what if instead of outputting tokens that combine to represent numbers, we just output the numbers themselves, right there in the output embedding?"
It works! Imagine you're discussing math with someone. Instead of saying "x is twenty five, which is large" in words, you'd say "x is", then switch to making a whistling noise in which the pitch of your whistle, in its position within your output frequency range, communicated the concept of 25.00 +/- epsilon. Then you'd resume speech and say "which is large".
I think the sentiment is that today's models are big and well-trained enough that receiving and delivering quantities as tokens representing numbers doesn't hurt capabilities much, but I'm still fascinated by xVal's much more elegant approach.
Seems like this is true for not just code but for all content being generated? Albeit for code it’s more well-defined, but the fork / lock mechanism works for a lot more problem domains.
That would seem intuitively true; it certainly applies to written language, where a clause could go off in another direction, but at other positions the correct grammar/syntax is unambiguous.
thinking -
well if we think of lock as happening in a narrative, then I think we can see there can be points where "everything you know is wrong" which essentially allows you to go back into a sort of fork mode and work towards another lock.
Completely artistic creation, creating something that does not exist and that cannot produce things out of itself, means that locking can be more diffuse, not as settled.
I think this seems similar to what Anthropic had been doing since the latest few Opus releases, which is interleaved thinking; CoT reasoning in the middle of a message. But they operate at different layers.
Could we not get the same with EAFT? Maybe that’s what it’s doing but definitely not the first to think “let’s lock in high probability solutions”
In nemotron the high perplexity solutions are selected for RL, in VLM training a few people are looking at the entropy distributions of the training set, etc
One relevant thing is that these forks are unnaturally narrow in all models, and rather resemble locks (not quite but close). From multiple possible continuations models tend to prefer just a couple, i.e. the model is a lot less random than it should be. That's why you're seeing annoying slop in writing and instantly recognizable color schemes in vibecoded sites. Lack of diversity probably limits the usefulness of this method as well.
>I love that we're still learning the emergent properties of LLMs!
it feels like the modern recurrence of the early 2010s bootstrap templates. we figured out how to automate building sites instantly, but at the cost of making the entire web look exactly the same.
I don't really understand the internal mechanics of of this, but my first thought was why not combine this with a linter/tests. So that it produces all the forks and only keeps the syntactically correct ones.
So... it's like a golfer who hits thousands of balls into an open field without ever once aiming for a hole. The relentless repetition flawlessly locks in their foundational muscle memory and basic swing mechanics, so when they finally step up to a real course, they don't have to waste a single thought on how to hold the club. Their basic swing is completely automatic - they can confidently take the creative, high-risk shot required to actually sink a hole-in-one.
After TurboQuant and Gemma 4, came across the following video[0] running Gemma on local machine at 50 token/second.
That already looks like Sonnet 3x and 4 level capabilities to me where the model in question (Gemma 4) set ups whole python project with a UI and installs python libraries using uv etc.
Add this Simple Self Distillation to the picture and by 2028 I see cheaper coding model providers with much more generous usage limits in the future and power users would be mostly running their own models anyway.
Anyone using these models as "non-deterministic transpilers" from natural language to code (experienced engineers who can write code themselves) would probably not be paying to any AI providers.
I always wonder how much smaller and faster models could be if they were only trained on the latest versions of the languages I use, so for me that is PHP, SQL, HTML, JS, CSS, Dutch, English, plus tool use for my OS of choice (MacOS).
Right now it feels like hammering a house onto a nail instead of the other way around.
Not very. LLMs derive a lot of their capability profile from the sheer scale.
LLMs have something that's not entirely unlike the "g factor" in humans - a broad "capability base" that spans domains. The best of the best "coding LLMs" need both good "in-domain training" for coding specifically and a high "capability base". And a lot of where that "base" comes from is: model size and the scale of data and compute used in pre-training.
Reducing the model scale and pruning the training data would result in a model with a lower "base". It would also hurt in-domain performance - because capabilities generalize and transfer, and pruning C code from the training data would "unteach" the model things that also apply to code in PHP.
Thus, the pursuit of "narrow specialist LLMs" is misguided, as a rule.
Unless you have a well defined set bar that, once cleared, makes the task solved, and there is no risk of scope adjustment, no benefit from any future capability improvements above that bar, and enough load to justify the engineering costs of training a purpose-specific model? A "strong generalist" LLM is typically a better bet than a "narrow specialist".
In practice, this is an incredibly rare set of conditions to be met.
It's more complicated than that. Small specialized LLMS are IMO better framed as "talking tools" than generalized intelligence. With that in mind, it's clear why something that can eg look at an image and describe things about it or accurately predict weather, then converse about it, is valuable.
There are hardware-based limitations in the size of LLMs you can feasibly train and serve, which imposes a limit in the amount of information you can pack into a single model's weights, and the amount of compute per second you can get out of that model at inference-time.
My company has been working on this specifically because even now most researchers don't seem to really understand that this is just as much an economics and knowledge problem (cf Hayek) as it is "intelligence"
It is much more efficient to strategically delegate specialized tasks, or ones that require a lot of tokens but not a lot of intelligence, to models that can be served more cheap. This is one of the things that Claude Code does very well. It's also the basis for MOE and some similar architectures with a smarter router model serving as a common base between the experts.
I seem to remember that's one of the first things they tried, but the general models tended to win out. Turns out there's more to learn from all code/discussions than from just JS.
Wouldn't that mean they're bad at migration tasks? I feel like for most languages, going from [old] to [current] is a fairly to very common usage scenario.
> power users would be mostly running their own models
...with a fair amount of supervision, while frontier models would be running circles around them using project-specific memory and on-demand training (or whatever we would have by then).
Honestly right now it's mainly stagnation in frontiere model capabilities. Most of the recent afvancemdnts are towards generation speed, compression and tool usage. The quality of the models are not improving at the same rate as before. I doubt this big gap will continue, given that open source and especially chinese labs keep pushing well documented frontiere papers.
Those will be great for projects that look just like everybody else's. That's not a knock. We'll see plenty of new systems built by anyone who needs one.
If you're building something groundbreaking and new, the advantage will be slim to none.
Their explanation for why their idea (SSD) might work - precision-exploration conflict hypothesis - is something adaptive decoding also tries to solve.
Incredible, will translate to better coding models in the near future.
We really need to develop better tools to understand what's happening inside these NNs. Working with high-D spaces is not something we're good at, and we're basically throwing stuff at it and seeing if it sticks.
Haven't read the paper yet, but it is interesting how seemingly simple many breakthroughs in ML are. Even transformers are like that. Maybe it's hindsight bias.
I suppose we just don't have a deeper underlying theory to lean on and help us 'design' anything.
A lot of discoveries are like that. In fact, simplicity is often the hallmark of correctness, and complexity is often a sign that our understanding is incomplete and we’re still stumbling towards the right model. Not always, but often. It’s been a good rule of thumb in my programming career.
Crypto punishes that fast.
Simple code can still be wrong, because the missing parts often come back as undocumented contracts between services and operators who now have to guess where the edge cases went.
You can wrap ugly internals in an elegant API for months.
Blind faith in minimalism gets you the same bugs with nicer naming and a codebase that only looks clean untl load shows up.
A designer knows he has achieved perfection not when there is nothing left to add, but when there is nothing left to take away. -- Antoine de Saint-Exupery
IME human developers also span a spectrum on this. On one end, you have devs who might meditate half a day on different solutions before writing a line of code. On the other end are devs who run full speed ahead with the first working solution that comes to mind. LLMs in their current form are mostly the latter.
I'm working on a tool to determine which portions of an LLM process can be optimized, and how to measure that optimization and check whether it's optimizable at all. The shaping pattern that they talk about here is directly relevant and makes a whole lot more processes potentially optimizable by looking at the pattern rather than if the metrics just go up or down.
Maybe not the thing I should be focusing on, but I was surprised this paper came from apple. I was under the impression that apples ai/LLM research was far behind the curve. I get that research is a rising tides lifts all boats situation, I just thought that I had seen lots of negative news about apples progress in the front, and heuristically haven’t seen many (any?) apple research papers make it the front page of hacker news. Wondering if anyone more familiar with apple/ai research could comment on this?
> Our method, simple self-distillation (SSD), is embarrassingly simple: sample solutions from the base model with specified temperature and truncation, then fine-tune on those raw, unverified samples via standard cross-entropy loss.
So you prompt the base model for answer and then rerun the prompt with the answer from the first run?
They use self-distillation to shift the output distribution of the model towards that of the same model, but running with different temperature/truncation settings in sampling.
This effectively "folds" the logit tail truncation behavior into the model itself.
Not entirely unlike a few "model controlled sampling settings" things I've seen in what it does, but different in execution.
You use the outputs from the first run (right or wrong) as answers for the second training run, and repeat. Magically it works. That's what's so surprising.
I guess a theory is because there are so many diverse ways to be wrong that they don't accumulate error... still seems surprising and would be interesting to see if it works in other domains.
I’d like to understand AI research better and I recall some posts a while back where someone collected all the key papers that one should read, but I don’t remember enough to be able to find it. Does anyone know what I’m talking about and could link me to that post?
This might sound paradoxical -- but any decent LLM will be happy to explain all the papers to you at great depth, and read new ones, and translate the math into simpler concepts and such. It'll also happily recommend relevant math to study, or give training problems, or whatever you want.
"SSD improves Qwen3-30B-Instruct from 42.4% to 55.3% pass@1 on LiveCodeBench v6"
I know virtually nothing about this area but my naive take is that something that means it still only passes tests around half the time doesn't seem like a particularly big jump forwards.
There's no shortage of benchmarks (coding or otherwise) that any competent coding model will now pass with ~100%.
But no-one quotes those any more because if everyone passes them, they don't serve any useful purpose in discriminating between different models or identifying advancements
So people switch to new benchmarks which either have more difficult tasks or some other artificial constraints that make them in some way harder to pass, until the scores are low enough that they're actually discriminating between models. and a 50% score is in some sense ideal for that - there's lots of room for variance around 50%.
(whether the thing they're measuring is something that well correlates to real coding performance is another question)
So you can't infer anything in isolation from a given benchmark score being only 50% other than that benchmarks are calibrated to make such scores the likely outcome
Can anyone help clarify these doubts - I didn't see any information about how different the test/benchmark set is from the training set. It feels like an important gap to not fill in a ML paper. What if there is an overlap between the problems in the test set and the training set?? What is the decontamination strategy of going from LCBv5 to LCBv6 ?
you're probably overcomplicating it; as the paper says, it's embarrassingly simple: given a problem set, generate a response for each problem with a fixed temperature and truncation - then fine tune the model on the generations.
Their hypothesis as to why this works requires a bit more knowledge about model architecture, but basically when a model generates code some positions have only one right answer and some have many valid options - but the model has to use one global confidence setting for both. Sampling with a specific temperature + a garbage-token filter, then training on those outputs, teaches the model to internalize 'be precise where there's one answer, stay open-minded where there are several' — without anyone labeling which is which.
Note that there's a lot more nuance to this and I simplified a lot.
You teach the machine by asking it to solve some problems, and then whatever answer it gives you say "That's exactly right. Now we train on those answers YOU just gave me" (even if they are wrong) and repeat. Somehow THAT works over time.
if the probability mass is on a single token, its a precise answer like `1 + 1 = `
if next token predicted shares probability with other token, then there are multiple answers like `position: `
you can generate and train answers by exploring on varying the length of the code generated
I'm excited for the long tail of techniques like this that are going to be discovered over the next several decades that's going to make this technology eventually run on a toaster!
It’s an interesting claim, and the reported benchmark gains are large, but it is still an April 1, 2026 arXiv preprint, so I’d treat it as promising rather than settled.
I think the analogy is actually pretty specific to this paper, not just self-distillation in general.
During sleep your brain replays experiences but noisy and distorted. The replays are often incoherent as narratives (dreams are weird). But the consolidation still works because the value isn't in the narrative coherence, it's in the activation patterns at each moment. Important pathways get strengthened, weak ones get pruned. Section 4.4 of this paper is what makes the connection click. They cranked training temperature to 2.0 with no truncation. 62% of the sampled outputs had no extractable code. Coherent Python that devolves into multilingual gibberish halfway through. The model still improved (+5.7pp pass@1).
This makes no sense if you think the model is learning from good code examples. But it makes a lot of sense if you think of it as the model replaying its own knowledge back to itself in a noisy/distorted form, and the replay process strengthening what matters (sharp distributions at "lock" positions where one token is correct, broad distributions at "fork" positions where multiple approaches work) while pruning what doesn't (distractor tails). The model doesn't learn anything new. It just wakes up performing better because what it already knew got cleaned up.
Self-distillation shifts the behavior of the model towards that of the model + steering. As such, you don't strictly "need" the tokens to be in-domain for it to work. The logits are a vessel for transferring the steering into the model's internals.
The tokens can be gibberish. What transfers isn't whether they're gibberish or not, but how the flavor of model predictions, if given gibberish, differs from that of an unsteered version of itself.
In this specific case, the behavioral difference comes from the "temperature-shifted, truncated samples" in the "teacher" sampling strategy, and it is that difference that is internalized by the "student" model.
Very cool. An evolutionary biologist would say: Welcome to the party!
Mutation rate modulation is the AI engineers’ heat. And selection does the trimming of the outliers.
Some more serious biomorphic thinking and we may get to the next big insight courtesy of 3+ billion years of evolution—- evolution that enabled a great ape species to write a paper like this and build LMM’s like Gemma4 that totally rock on a 3.5 pound MacBookPro M5 Max with 128 GB of RAM.
Another potentially usable trick is the following: based on the observation that longer token budget improves model performances, one could generate solutions using a lot of thinking budget, then ask the LLM to turn the trace into a more compact one, and later SFT on that. That said, I have the feeling the result of the paper will likely be hard to apply in practice without affecting other capabilities, and/or not superior to other techniques that provide similar improvement in sampling.
If you sample from the base model with T=1.6, top_k=20, top_p=0.8, i.e, the decode settings used for the distillation's ground truth, does it match the SSD'd model + some decoding? Performance wise.
Their sweep is missing this. And only covers "standard" decoding settings.
I definitely pay more attention to papers affiliated with Chinese companies; the economics seem to be more conducive to doing good academic work and publishing it. I would say the same for companies like Apple (where TFA came from).
But to filter based on author's names sounds pretty darn racist.
I used to invent TLAs on the spot for fun, and when someone asked what it was, would respond, "It's a PUA", eventually revealing that meant "previously unknown acronym". It was even more annoying that it sounds.
Lol, you're probably not wrong. But have you ever noticed that the most important papers tend to be on the clear and readable side of things? It's as if researchers understand that being understood is important, but deemphasize that when the paper itself isn't important in the first place. (Maybe if they're only publishing to not perish, not being understood is actually a goof thing from their perspective?)
> Code interleaves fork positions, where several continuations are genuinely plausible and may correspond to different solution approaches, with lock positions, where syntax and semantics leave little ambiguity but a low-probability distractor tail still remains… The best global decoding setting is therefore necessarily a compromise; we call this tension the precision-exploration conflict.
In other words, just like us, the model needs to shift from "exploration" in "fork" mode (divergent thinking to produce a creative solution) to "precision" in "lock" mode (producing syntactically correct code).
What this paper shows is that their simple technique (SSD) can improve the ranking of optimal tokens in both lock and fork positions, meaning the model is more likely to explore when it should be exploring, and more likely to be precise when it needs to be.
I love that we're still learning the emergent properties of LLMs!
TBH, this is (very much my opinion btw) the least surprising thing. LLMs (and especially their emergent properties) are still black boxes. Humans have been studying the human brain for millenia, and we are barely better at predicting how humans work (or for eg to what extent free will is a thing). Hell, emergent properties of traffic was not understood or properly given attention to, even when a researcher, as a driver, knows what a driver does. Right now, on the front page, is this post:
> 14. Claude Code Found a Linux Vulnerability Hidden for 23 Years (mtlynch.io)
So it's pretty cool we're learning new things about LLMs, sure, but it's barely surprising that we're still learning it.
(Sorry, mini grumpy man rant over. I just wish we knew more of the world but I know that's not realistic.)
I dare say that in some ways, we understand LLMs better than humans, or at least the interpretability tools are now superior. Awkward place to be, but an interesting one.
Are you surprised we understand them better than brains?
That's a bit of an overstatement.
The entire field of ML is aimed at problems where deterministic code would work just fine, but the amount of cases it would need to cover is too large to be practical (note, this has nothing to do with the impossibility of its design) AND there's a sufficient corpus of data that allows plausible enough models to be trained. So we accept the occasionally questionable precision of ML models over the huge time and money costs of engineering these kinds of systems the traditional way. LLMs are no different.
What you are saying is fantasy nonsense.
So it doesn't work.
Very, monsieur Laplace.
Much as Diogenes mocked Platos definition of a man with a plucked chicken, LLM's revealed what "real" ai would require: contigous learning. That isnt to diminish the power of LLM's (the are useful) but that limitation is a fairly hard one to over come if true AGI is your goal.
From what I understand, a living neural network learns several orders of magnitude more efficiently than an artificial one.
I'm not sure where that difference comes from. But my brain probably isn't doing back propagation, it's probably doing something very different.
The intersection of what with physics?
Sir Roger Penrose, on quantum consciousness (and there is some regret on his part here) -- OR -- Jacob Barandes for a much more current thinking on this sort of intersectional exploratory thinking.
We have tons of low-hanging fruits across all fields of science and engineering to be picked, in form of different ways to apply and chain the models we have, different ways to interact with them, etc. - enough to fuel a good decade of continued progress in everything.
> The earliest reference to the brain occurs in the Edwin Smith Surgical Papyrus, written in the 17th century BC.
I was actually thinking of ancient greeks when writing my comment, but I suppose Egyptians have even older records than them.
From https://en.wikipedia.org/wiki/History_of_neuroscience
I think that with grammar-aware sampling / constrained decoding [0][1] it is possible to sometimes skip calling the model altogether if only one token is allowed by grammar and just insert it, but I don't think that any of the current, widely used combinations of models/harnesses use it. And it only skips inference in rare edge cases.
I wonder if there is a more general solution that can make models spend more compute on making important choices, while making generation of the "obvious" tokens cheaper and faster.
[0] https://github.com/ggml-org/llama.cpp/blob/master/grammars/R...
[1] https://developers.redhat.com/articles/2025/06/03/structured...
Making coding agents spit out syntactically correct code token by token is like asking a human to code on a whiteboard.
We kinda have a little bit of it with some coding harnesses giving model access to LSP, but I think that we can insert this knowledge on a lower level if we find a clever way to somehow utilize it during sampling.
I think that there is a lot of low hanging fruit in this area.
And in general, I think that people try to use LLMs too much to solve problems that can be easily solved by cheaper (computationally), and, more importantly deterministic tools.
For example, back in the day when LLM-assisted coding just became a thing people very often complained about models generating syntactically incorrect code and inventing non-existent library methods.
Well, I, an experienced human programmer, probably would also be making syntax mistakes and inventing non-existent methods if you stripped me of my tools and made me write code in a bare text editor without syntax highlighting.
Thankfully, my IDE would autocomplete real syntax and actually existing library methods for me and immediately give me feedback if I make a mistake anyway. And all of it is achieved using reliable deterministic code without the inherent issues of statistical models.
I think that it is really inefficient to reach for an expensive and unreliable tool when a cheap and reliable tool will do.
1. code
2. syntax check / build / format / lint (details language dependent)
3. test
and they can hop between 1 and 2 however many times they want.
I do think there is some merit in a tool that dumps all namespaces and reachable symbols so the agent can do its own autocomplete without a round-trip.
i once asked an LLM if it could ingest code from an interactive session more easily if it were in appropriately-typed markdown fences and it said absolutely yes, and that the syntax highlighting fed to it that way helps it immensely. i was downright shocked that syntax highlighting was anything more than noise for them.
There's a lot of work going on in various streams towards making it possible to vary compute per-token, dynamically, e.g. universal transformers. Maybe one day it'll work well enough to beat conventional techniques.
I got unstuck by randomizing the field order for each row?!? At training, and now I'm thinking I should do the same at inference time...
> This is probably due to the way larger numbers are tokenised, as big numbers can be split up into arbitrary forms. Take the integer 123456789. A BPE tokenizer (e.g., GPT-style) might split it like: ‘123’ ‘456’ ‘789’ or: ‘12’ ‘345’ ‘67’ ‘89’
One of the craziest LLM hacks that doesn't get love is https://polymathic-ai.org/blog/xval/
xVal basically says "tokenizing numbers is hard: what if instead of outputting tokens that combine to represent numbers, we just output the numbers themselves, right there in the output embedding?"
It works! Imagine you're discussing math with someone. Instead of saying "x is twenty five, which is large" in words, you'd say "x is", then switch to making a whistling noise in which the pitch of your whistle, in its position within your output frequency range, communicated the concept of 25.00 +/- epsilon. Then you'd resume speech and say "which is large".
I think the sentiment is that today's models are big and well-trained enough that receiving and delivering quantities as tokens representing numbers doesn't hurt capabilities much, but I'm still fascinated by xVal's much more elegant approach.
Completely artistic creation, creating something that does not exist and that cannot produce things out of itself, means that locking can be more diffuse, not as settled.
In nemotron the high perplexity solutions are selected for RL, in VLM training a few people are looking at the entropy distributions of the training set, etc
>I love that we're still learning the emergent properties of LLMs!
There are tons of low-hanging fruits there.
"Simple Self-Distillation". We had an acronym for Solid-State Drive. Don't know about that technique but the naming sure sound.. Simple?
That already looks like Sonnet 3x and 4 level capabilities to me where the model in question (Gemma 4) set ups whole python project with a UI and installs python libraries using uv etc.
Add this Simple Self Distillation to the picture and by 2028 I see cheaper coding model providers with much more generous usage limits in the future and power users would be mostly running their own models anyway.
Anyone using these models as "non-deterministic transpilers" from natural language to code (experienced engineers who can write code themselves) would probably not be paying to any AI providers.
[0] https://www.youtube.com/watch?v=-_hC-C_Drcw
Right now it feels like hammering a house onto a nail instead of the other way around.
LLMs have something that's not entirely unlike the "g factor" in humans - a broad "capability base" that spans domains. The best of the best "coding LLMs" need both good "in-domain training" for coding specifically and a high "capability base". And a lot of where that "base" comes from is: model size and the scale of data and compute used in pre-training.
Reducing the model scale and pruning the training data would result in a model with a lower "base". It would also hurt in-domain performance - because capabilities generalize and transfer, and pruning C code from the training data would "unteach" the model things that also apply to code in PHP.
Thus, the pursuit of "narrow specialist LLMs" is misguided, as a rule.
Unless you have a well defined set bar that, once cleared, makes the task solved, and there is no risk of scope adjustment, no benefit from any future capability improvements above that bar, and enough load to justify the engineering costs of training a purpose-specific model? A "strong generalist" LLM is typically a better bet than a "narrow specialist".
In practice, this is an incredibly rare set of conditions to be met.
There are hardware-based limitations in the size of LLMs you can feasibly train and serve, which imposes a limit in the amount of information you can pack into a single model's weights, and the amount of compute per second you can get out of that model at inference-time.
My company has been working on this specifically because even now most researchers don't seem to really understand that this is just as much an economics and knowledge problem (cf Hayek) as it is "intelligence"
It is much more efficient to strategically delegate specialized tasks, or ones that require a lot of tokens but not a lot of intelligence, to models that can be served more cheap. This is one of the things that Claude Code does very well. It's also the basis for MOE and some similar architectures with a smarter router model serving as a common base between the experts.
...with a fair amount of supervision, while frontier models would be running circles around them using project-specific memory and on-demand training (or whatever we would have by then).
If you're building something groundbreaking and new, the advantage will be slim to none.
https://ai.meta.com/research/publications/adaptive-decoding-...
We really need to develop better tools to understand what's happening inside these NNs. Working with high-D spaces is not something we're good at, and we're basically throwing stuff at it and seeing if it sticks.
It's the first thing anyone would think of (like a self-hosted compiler) but everything I've read said "it doesn't work."
EDIT: For context:
I suppose we just don't have a deeper underlying theory to lean on and help us 'design' anything.
You can wrap ugly internals in an elegant API for months. Blind faith in minimalism gets you the same bugs with nicer naming and a codebase that only looks clean untl load shows up.
I often find, if I've got a complicated solution, it’s because I haven’t fully examined the problem.
(Not fine tuning, but interesting none the less. If a model can so easily find a more elegant solution, why didn't it pick that in the first place?)
[0] https://news.ycombinator.com/item?id=46117802
[1] https://news.ycombinator.com/item?id=47107974
So you prompt the base model for answer and then rerun the prompt with the answer from the first run?
They use self-distillation to shift the output distribution of the model towards that of the same model, but running with different temperature/truncation settings in sampling.
This effectively "folds" the logit tail truncation behavior into the model itself.
Not entirely unlike a few "model controlled sampling settings" things I've seen in what it does, but different in execution.
You use the outputs from the first run (right or wrong) as answers for the second training run, and repeat. Magically it works. That's what's so surprising.
I guess a theory is because there are so many diverse ways to be wrong that they don't accumulate error... still seems surprising and would be interesting to see if it works in other domains.
I know virtually nothing about this area but my naive take is that something that means it still only passes tests around half the time doesn't seem like a particularly big jump forwards.
What am I missing?
But no-one quotes those any more because if everyone passes them, they don't serve any useful purpose in discriminating between different models or identifying advancements
So people switch to new benchmarks which either have more difficult tasks or some other artificial constraints that make them in some way harder to pass, until the scores are low enough that they're actually discriminating between models. and a 50% score is in some sense ideal for that - there's lots of room for variance around 50%.
(whether the thing they're measuring is something that well correlates to real coding performance is another question)
So you can't infer anything in isolation from a given benchmark score being only 50% other than that benchmarks are calibrated to make such scores the likely outcome
Their hypothesis as to why this works requires a bit more knowledge about model architecture, but basically when a model generates code some positions have only one right answer and some have many valid options - but the model has to use one global confidence setting for both. Sampling with a specific temperature + a garbage-token filter, then training on those outputs, teaches the model to internalize 'be precise where there's one answer, stay open-minded where there are several' — without anyone labeling which is which.
Note that there's a lot more nuance to this and I simplified a lot.
You teach the machine by asking it to solve some problems, and then whatever answer it gives you say "That's exactly right. Now we train on those answers YOU just gave me" (even if they are wrong) and repeat. Somehow THAT works over time.
you can generate and train answers by exploring on varying the length of the code generated
This feels eerily similar to sleep consolidation or synaptic pruning
I think the analogy is actually pretty specific to this paper, not just self-distillation in general.
During sleep your brain replays experiences but noisy and distorted. The replays are often incoherent as narratives (dreams are weird). But the consolidation still works because the value isn't in the narrative coherence, it's in the activation patterns at each moment. Important pathways get strengthened, weak ones get pruned. Section 4.4 of this paper is what makes the connection click. They cranked training temperature to 2.0 with no truncation. 62% of the sampled outputs had no extractable code. Coherent Python that devolves into multilingual gibberish halfway through. The model still improved (+5.7pp pass@1).
This makes no sense if you think the model is learning from good code examples. But it makes a lot of sense if you think of it as the model replaying its own knowledge back to itself in a noisy/distorted form, and the replay process strengthening what matters (sharp distributions at "lock" positions where one token is correct, broad distributions at "fork" positions where multiple approaches work) while pruning what doesn't (distractor tails). The model doesn't learn anything new. It just wakes up performing better because what it already knew got cleaned up.
How is this comment not at number 1??
Self-distillation shifts the behavior of the model towards that of the model + steering. As such, you don't strictly "need" the tokens to be in-domain for it to work. The logits are a vessel for transferring the steering into the model's internals.
The tokens can be gibberish. What transfers isn't whether they're gibberish or not, but how the flavor of model predictions, if given gibberish, differs from that of an unsteered version of itself.
In this specific case, the behavioral difference comes from the "temperature-shifted, truncated samples" in the "teacher" sampling strategy, and it is that difference that is internalized by the "student" model.
Mutation rate modulation is the AI engineers’ heat. And selection does the trimming of the outliers.
Some more serious biomorphic thinking and we may get to the next big insight courtesy of 3+ billion years of evolution—- evolution that enabled a great ape species to write a paper like this and build LMM’s like Gemma4 that totally rock on a 3.5 pound MacBookPro M5 Max with 128 GB of RAM.
If you sample from the base model with T=1.6, top_k=20, top_p=0.8, i.e, the decode settings used for the distillation's ground truth, does it match the SSD'd model + some decoding? Performance wise.
Their sweep is missing this. And only covers "standard" decoding settings.
But to filter based on author's names sounds pretty darn racist.
They seemed like they had to be churning out papers and any little adaptation to existing research triggered a new publication.
But it may have changed now.
"Made in China, designed by Apple in California"
should be:
"Made in China, designed by Chinese people in California"?
Sorry apple, SSD is already taken, you can't use that acronym.
Consistency Preservation Update (CPU)
Guided Probability Update (GPU)
History-aware Distillation Driving (HDD)
Probability Smoothing Update (PSU)
Title should be: Simple Self-Distillation Improves Code Generation
https://en.wikipedia.org/wiki/Embarrassingly_parallel
Many computer science paper titles allude to past titles in other CS papers.
Calling it “cringe worthy” is unnecessarily mean. There is context and history you don’t understand.
There are two distinct billions. https://en.wikipedia.org/wiki/Billion