A sleep-like consolidation mechanism for LLMs
arxiv.org
arxiv.org
Essentially it goes "You know how your model can remember its training data? Well, what if you treated its recent context like more training data and updated (some of) the weights using (mostly) the same process used to train it?"
The end result is very good at remembering things but also really good at adapting to new unseen distributions.
Remember Microsoft Tay.
[1]: https://flann.cs.yale.edu
[2]: https://www.cs.toronto.edu/~hinton/csc2535/readings/ws.pdf
[3]: https://arxiv.org/abs/1711.02282
[4]: https://arxiv.org/abs/2006.08381
[5]: https://mural.maynoothuniversity.ie/id/eprint/1653/1/Hamilto...
Scaling test-time compute has emerged as a key ingredient for enabling large language models (LLMs) to solve difficult problems, but comes with high latency and inference cost. We introduce sleep-time compute, which allows models to "think" offline about contexts before queries are presented: by anticipating what queries users might ask and pre-computing useful quantities, we can significantly reduce the compute requirements at test-time. To demonstrate the efficacy of our method, we create modified versions of two reasoning tasks - Stateful GSM-Symbolic and Stateful AIME. We find that sleep-time compute can reduce the amount of test-time compute needed to achieve the same accuracy by ~ 5x on Stateful GSM-Symbolic and Stateful AIME and that by scaling sleep-time compute we can further increase accuracy by up to 13% on Stateful GSM-Symbolic and 18% on Stateful AIME. Furthermore, we introduce Multi-Query GSM-Symbolic, which extends GSM-Symbolic by including multiple related queries per context. By amortizing sleep-time compute across related queries about the same context using Multi-Query GSM-Symbolic, we can decrease the average cost per query by 2.5x. We then conduct additional analysis to understand when sleep-time compute is most effective, finding the predictability of the user query to be well correlated with the efficacy of sleep-time compute. Finally, we conduct a case-study of applying sleep-time compute to a realistic agentic SWE task.
Biologically humans do similar compression, so introducing a similar concept to an LLM also feels reasonable. Hardware isn't fast/cheap enough to do this on an ongoing basis, similar to how it's too expensive for our brains to do this while we're moving through the world.
All we have now most of the time in LLMs is "working memory" we're missing a lot of the functionality that allows for episodic memory and selective plasticity.
The more you read about how human brains work, the more you realize that we may have figured out a piece with LLMs, but it's certainly nothing approaching AGI. People insisting so are blowing smoke for investor hype or don't understand a big piece of the concepts involved.
That's already possible with LLMs. The challenge is that 1. it would allow permanently jail-breaking models and 2. there'd be no way for them to efficiently transfer what they'd learned to a new model generation.
Coincidentally the human brain is also jailbroken and nontransferable
This would create a three-layer memory system:
- Stable long-term memory (initial base weights)
- Mid-term memory built from the compactions and replay buffers
- Short-term memory (KV cache)
Sleeping would just be a fancy term for consolidating and transferring information from one memory layer to another during offline hours. Maybe that's also what the brain does while sleeping.
Also, we wouldn't train on the whole session. A separate critic module, like a reward model, would filter the KV cache to extract the high-value information, like a garbage collector before the LoRA.
That's just an idea though. Right now most research focuses on changing the architecture itself (TITAN, HOPE...) instead.
Felt like stating the obvious there? Greenwich being the center of everything after all.
We're talking about a step that updates weights based on say between 10k and 1M tokens.
A more appropriate title would have been something like "Offline Recurrent Memory Consolidation for Long-Context Language Models". This is supposed to be a research paper, not a story book. The title should give context to other researchers, and not be clearly engineered for clicks. If you don't think so, that's your prerogative, but you're objectively wrong.
Does a motor vehicle get "sleep" when it is serviced? When I reboot a computer, is that equivalent to a nap?
There is a strong, non-trivial connection here between what your brain does in sleep and what they are studying.
You wouldn't object to referring to robot eyes or robot legs.
At the very least, we know that sleep and dreaming do exist in biological brains. (Doesn't mean any of it is applicable to artificial neural nets, doesn't mean it'll work for our specific architectures etc. etc., but at least the idea requires fewer assumptions than a completely untested novel theory.)
Maybe someday we'll understand the way our minds work well enough to design from first principles but until then we've only got one template for how a thinking machine should look.
Here the analogy isn't without reason.
> In animals, the transfer from short-term memory to long-term memory is thought to be supported by hippocampal replay [33], especially during sleep [41]; in this phase, short-term hippocampal memories are reactivated and consolidated into cortical synaptic weights. Sleep makes animals unable to respond to external stimuli, suggesting that it must provide enough cognitive benefit to justify this cost [41]. Inspired by these biological processes, we propose a method for transferring context-window memory into persistent weights. When the model’s context window becomes full during inference, the model enters a “sleep” in which it performs multiple forward passes over the accumulated context and recursively updates its fast weights via a learned local rule. As in animal sleep, the model receives no external input tokens during this phase. After consolidation, the context window is cleared, and the model resumes operation with updated fast weights. During training, the model is optimized end-to-end by backpropagating through the entire process to maximize task performance after sleep.
One thing we do know for certain is that it is necessary, it is needed in "dumb" animals as well as in you and I. If an animal can't sleep it will eventually die.
I don't think that applies to the activity described in the OP. Does their LLM "die" if it can't perform the function described?
If you don't periodically clean the context, an LLM effectively goes insane in terms of outputs.
If the LLM were fully controlling a physical system (like a robot body) that contained it the resulting insanity of an ever-growing, never cleaned context would likely result in some sort of death-like event.
It's still weak, though. An LLM without constant human input is likely more similar to a bicycle that starts to lose its gyroscopic balance as it moves more slowly, a human can however keep a stationary bicycle upright (while riding it).
Meaning: it might just provide a big advantage.
I don't want to overextend and assume that any advantage extends to LLMs. That rest-and-recuperate advantage might also extend to LLM-based AIs. Or maybe not, and the rest-and-recuperate is mainly useful for biology-based organisms. But there is some logic to it.
> The function of sleep in animals is largely obscure.
In my understanding, it's well-understood that sleep is used to consolidate and store long-term memories (amongst other functions, like cell and muscle repair). They've found this memory-consolidation-during-sleep even in relatively simple animals like bees.
You're talking about different things: biological necessity and evolutionary benefit.
You can find out about the former by preventing an animal from sleeping (but otherwise provide all other needed things), and seeing if it will eventually die.
That is actually almost impossible to do. The rat study was as close as we’ve ever come, and it’s still debated whether the rats died due to lack of sleep or some other mechanism, since the autopsy couldn’t confirm a cause of death. (It could have been due to the way the experiment ran, for example, not the lack of sleep.)
What is described in the OP is therefore not a specific characteristic of sleep. It may however be a "useful" rhetorical device.
I do however object to the extensive use of such rhetorical tricks in the conversations that surround LLMs. For example, why does a consumer-grade LLM display "thinking" while it is actually sending data from my computer to some datacentre, processing it, and sending the result back? Equally, why does it output human-emotive phrases such as "sorry" when such computation is revealed to be incorrect?
Such rhetorical tricks, and more, likely underlie to a large degree the popularity of LLMs, despite their actual performance being clearly below what the rhetoric implies.
The question was "why is it needed". In context the meaning is clearly to ask what it's doing that's essential and (it follows) why those things are essential.
The subsequent response did not (as you suggest) articulate some subset of nonessential things done during sleep. Rather it rattled off plausible (and widely understood) aspects of the process that could have led to the current situation. Even if it had listed concrete activities that would still not have made for a meaningful answer.
The first clue that something is wrong should be that the linked article is recent and prominent. Thus short a brand new groundbreaking development we can be reasonably certain that a random commenter on the internet will not be sensibly rebutting the claims (and certainly not in the span of ~2 sentences).
No that was not the question. Why precisely sleep is essential is a complete non-sequitur to the original question, which was "does something occur during sleep which resembles what is described in TFA such that it can justifiably be called sleep."
As a general rule of thumb, if you find someone's responses incoherent, it's good practice to check what is actually being discussed.
GTAGGA turning into GTACCA may make you sleep 8 fewer hours but also keep you from producing haemoglobin.
It's like leveraging a qsort implementation from your mp3 player to develop your OS scheduling algorithm.
His argument is that there is nothing essential about its apparent function(playing mp3s).
If he articulated a particular essential process and why it depends on sleep in an incidental manner that might make for a reasonable hypothesis. However it would not refute the earlier (cited) claim that there is no consensus.
As presented without any concrete information about the processes involved it doesn't even qualify as a hypothesis, merely empty handwaving. In context it's even worse, being an entirely baseless contradiction of a claim pulled from a prominent paper.
Also there being no consensus means most scientists who touch on the topic are FFA speculating except the person stating there is no consensus in the overview. It's not "settled science" but rather the opposite.
So "the reason it originally exists" and "what breaks if you take it away" aren't necessarily the same thing.
As with, say, digestion, or an major organ like the liver, it's reasonable to think that it does simple things in simple animals, and more complex things in more complex ones.
Take out an animal's liver, it's not one process that stops working, it's dozens. There's one or two that will kill it quicker, so those are the ones it dies of, but artificial livers are hard to build as they implement so many vital processes.
Take your liver example. We can largely answer that same question. I can't off the top of my head but the answer is fairly well established even if incomplete to varying degrees depending on the species.
There is widespread consensus on why a liver is needed for survival whereas there is not for sleep. That's particularly interesting when you consider that sleep is more common across the tree of life than dedicated livers are (at least AFAIK).
Older than bilateral symmetry even - jellyfish are thought to sleep, sponges however do not. Jellyfish don't have spinal columns, lungs, gills, livers, kidneys, hearts, guts or blood, but they do have nerves and they do seem to sleep.
There is widespread consensus as to which processes failing will kill you first in acute liver failure, but it governs dozens of processes that, medium term, are essential to life; not all are widely understood.
In the case of sleep, it seems to be nervous system dysregulation that kills. It's notable that comatose patients don't seem to suffer the ill-effects of sleep deprivation. But still, "the thing that kills an animal subject to extreme sleep deprivation" is not necessarily "the original process for which sleep was evolved".
Human brains do some fairly complicated vital things during sleep (REM, spindles, slow wave activity), but that can't be the original essential function - the simplest animals that sleep (jellyfish) don't really have brains, although they do have nervous systems.
Whereas those animals which lack nervous systems (sponges) can't be said to sleep, although it's reasonable to ask.. "how would you tell?", or ask whether the question itself makes sense for something that lacks the ability to sense, plan, act.
So another framing is "anything which can be awake, must also be asleep". But one might equally argue, we don't know why animals are awake.
We can go one step further and suppose that, in order for an animal to act, to exercise will, it must do so at above its average metabolic rate, and in doing so it necessarily incurs metabolic debt.
Once energy conserving "sleep time" exists, the genome can postpone or schedule for activity during these times of day, if it turns out to be more effective somehow.
Very few animals fail to eventually die even with as much sleep as they want.
But before death, there is a loss of cognitive function from sleep deprivation, and we observe this too with AI whose context windows get too full.
While we don't know very much about sleep, my understanding is that we do have a long list of things that we do during it, we just don't really understand if sleep is necessary for each of them or simply a convenient opportunity for it.
There's lots of things biology does in response to easy-to-detect proxy signals instead of the real thing they care about: Our sensation of needing to breathe more is based on too much carbonic acid in our blood, not lack of oxygen, which is why in general nobody is allowed in an elevator with a liquid nitrogen dewar; Our natural distaste for incest is based on who we grew up with, not our actual DNA; Get too cold and some people suddenly feel warm and want to (and some do) take all their clothes off even though that would just make them hypothermic even faster.
Being asleep may trigger the things we need to get done, but that doesn't mean sleep is *fundamentally* necessary for the things we need to get done. It could be just that it happens to be the way our biochemistry is wired, and we may find some other way to trigger those things.
The quotation given by djeastm would by my guess for what a dream is, and why we have them. But we don't spend all our time asleep, dreaming. And I'd be the first to say that my guess isn't worth much, as I'm not a brain scientist.
We now have evidence for REM sleep in spiders(1). Our last common ancestor with spiders predates the development of nervous systems. This strongly suggests that sleep (and specifically REM sleep) serves some function important enough that it has independently evolved in both protostomes and dueterostomes. (And probably multiple times within the protostomes, being present in both cephalopods and jumping spiders.)
There may be some commonality in the origin of the ion channels, but I'll lay money that the requirements for sleep are more of a result of general information processing requirements.
Also, there's different kinds/stages of sleep, which probably perform different functions.
For instance, REM may do something like the GP describes, consolidating memories and processing learning. Deep sleep may do something else (I vaguely recall some stage of sleep is used by neurons to clear certain waste products).
That turns out to be un-settled science. No human has ever died from lack of sleep.
People point to “fatal familial insomnia” as a counterexample. But they die to the disease, not the lack of sleep.
In a series of controlled experiments, rats and fruit flies did die from lack of sleep. But no one has yet proven that it holds true for vertebrates except for rats.
In other words, it could be true that “among vertebrates, only rats die of sleep deprivation.”
So “if an animal can’t sleep, it will eventually die” is actually quite hard to prove, and depending on how you look at it, somewhat easy to disprove by the fact that rats and fruit flies were so difficult to kill from sleep depravation alone.
Personally I’m skeptical of the rat study too. Claude amends this:
> What they did not establish: the mechanism. On autopsy, “no anatomical cause of death was identified.” The rats showed weight loss despite eating more, body temperature problems, and skin lesions, but nothing that pointed to a clean cause. So no, they could not say a rat “died from sleep deprivation alone” in the sense of identifying what sleep loss did to the body to kill it. They showed a strong association under tight controls, not a proven causal pathway.
It’s not a small quibble to point out that the central argument (“animals need sleep or they’ll die”) may be mistaken.
As far as I understand it, there is a disease that destroys your brain's ability to produce sleep. Once you have it, you suffer total, progressive insomnia and die within roughly 6–18 months. Scientists debate whether it's the underlying brain damage or the sleeplessness itself that causes death, but the two are inseparable in practice, and sleep deprivation is considered the leading candidate.
Separately, the longest anyone has stayed awake under controlled conditions was 11 days, which produced severe cognitive impairment, paranoia, and hallucinations; suggesting the body deteriorates rapidly without sleep.
It's probably not wise to state your original claim as established fact.
> People point to “fatal familial insomnia” as a counterexample. But they die to the disease, not the lack of sleep.
It’s a prion disease. It’s established fact that they don’t die from the lack of sleep.
Jeez. People here are really stretching to defend their false “we die without sleep” claim.
> You're using absence of evidence as evidence of absence — which is a weak foundation when the evidence is genuinely hard to capture. You can't ethically deprive humans of sleep to death in a lab, and FFI affects only a handful of families worldwide.
> On the prion disease specifically: researchers haven't dismissed the role of sleep deprivation they've actively attempted to treat the insomnia in FFI patients on the hypothesis that it contributes to decline. That's not how a field behaves when it considers something a settled, irrelevant symptom.
> More broadly, "no human has ever died from lack of sleep" is an extraordinarily strong claim. To support it you'd need to rule out sleep deprivation as a factor in every candidate case and have a complete understanding of the mechanism. We have neither. The honest position is "we don't know" — not confident assertion in either direction.
There is quite literally no prion disease that isn't fatal.
Sleep does a lot of very important things that we probably wouldn't live long without, but it really is unclear to what extent sleep is necessary for them. If we had enough knowledge, could we trigger all the things sleep does without invoking sleep itself ? Perhaps sleep is just a very convenient mechanism.
For something so incredibly difficult to do (die from lack of sleep) it’s frankly crazy that most people here are saying it like it’s fact.
Lack of sleep doesn't kill you / does kill you in the same sense.
I do not believe this analogy really confused you. No one is saying they're the same and you're well aware of that.
As to the factual nature of the argument, I'll let you argue with Harvard Brain Institute, as I have no interest in this debate. https://brain.harvard.edu/hbi_news/why-severe-sleep-deprivat...
Why do you think I'm arguing something? Again, smacks of Pauli's not even wrong. You're confusing your headspace with everyone else's and rushed to copy pasta an AI you're browbeating into disavowing things you're skeptical of to...win an argument, I guess? Based on this post? Unclear to me what the argument is, or, if I'm understanding correctly, due to the narrow focus while being self-absorbed.
We know more generally that people who get decreased amount of sleep suffer increased rates of physical and mental health issues.
It is not a very big leap from "causes permanent damage" to "enough permanent damage can cause death" and of course, keeping someone awake until they are hurt or killed is deeply unethical, so even if it could be proven in other species, you'd still be here arguing that 'they aren't humans".
"Sleep" is just used in their context to describe a non-interactive mode and they didn't lean heavily into zoomorphic - I think you mean - parallels.
You're grinding an axe on a single term. What is your broader hangup with them using the term "sleep"?
> Does their LLM "die" if it can't perform the function described?
We're reaching an age where LMGTFY should now be Let Me LLM That For You. Have you tried asking an LLM this question about the article? I believe it answers it very well.
So, whether the LLM "dies" in any sense may or may not be important for what "sleep" is defined to be in this article. It's quite possible that sleep also affects endocrine system in animals or hormones etc... and that's what's causing death, not necessarily anything to do with how brain functions.
It dies in terms of usefulness if it can't stay up to date with new knowledge. That is, it will no longer be used and thus effectively die off.
This just obscures the conversation IMO. We know a great deal of functions that sleeping performs. We don't know everything. For some reason this prevents us for using this word in a computing context? What do you think about sleep(1) in Unix?
Also..
> Does their LLM "die" if it can't perform the function described?
what "death" means in the context of a computer program?
i feel like its confusing to reuse the word for a process that aims to deliberately change state of the machine / process
I'm autistic and just went through massive changes that basically keep locking up my brain and I freeze.
I learned that sleeping helps, even if just 20 minutes. It helps that I can fall asleep "while awake". It's as if I relinquish control of responding to stimuli, which instantly brings so much rest to my mind. It is odd trying to move a limb and the brain basically responds with noop. But it works.
Afterwards I can generally make a decision and perform it.
So in a sense it seems similar to what you describe the model would have to do. I forget short term concerns that overwhelm and refocus on the long term goal.
I mean, you do put your computer into "sleep" mode and then "wake" it.
Analogies are useful. I think we need to learn how to continue to benefit from them despite the risk of anthropomorphication.
it is very non-helpful—or worse—to use this language, this way.
One might as well say "need neural plasticity" which is as much an analogy and equally misleading and counterproductive in shaping the right model of the system.
One might even call this pernicious, what it encourages is already a social problem; and it doesn't aid understanding, it confounds it.
Clicking through, that’s exactly what it is. Seems like “sleep” is an excellent term to use here.
That's more like a doctor visit and a workout. The sleep will be the part of the duty cycle when it's not operating.
> When I reboot a computer, is that equivalent to a nap?
Yes, it wakes up completely refreshed and in good working order, usually, and if there's still a problem you know you need a technician.
Also, even when something is "specific" to humans, it might not be anthropomorphizing to observe it in something else, it could just be an emergent pattern of high intelligence.
I agree we need to be mindful of our metaphores, but they do help both with inspiration for developing techniques as well as for naming things. The onus of keeping bias in check when using metaphores is on the reader, authors can't really do that for you. However once bias is in check you can have a very productive debate in terms of these namings given that everyone is aware of their ontology.
Also keep in mind that most if not all devices with a chip have had a function called "sleep" for many years, without this argument.
This is not anything new, its just a word that fits the function.
One of the mayors of New York in the 80's (Koch?) famously doubled the city's bus fleet for zero cost by running them 24 hours, instead of letting them rest at the end of their shifts, as was the previous policy.
Second, explicitly avoiding things that sound like anthropomorphisation is equally not helpful- why avoid a metaphor that works?
Third, it's really a pity that this pointless nitpicking is dominating the thread.
What is dominating the thread are claims that the LLM operation in question is analogous to the function of sleep in humans. It obviously is not.
The anthropomorphization of LLMs has reached ridiculous proportions. Applying the same standards as used in this field to others would result in claims that laundry machines "hallucinated" that they had had sufficient water when they failed due to the faucet being turned off.
Do androids dream of electric sheep?
I do think it points at something bigger than just attention architecture: "memory" isn't just storage, and merely longer context isn't the same thing as having a better understanding of the source data.
I'm looking at this through the "personal AI" lens, where I think the missing "memory" layer seems to be consolidation & prioritization. It's not enough to just pattern match and grab the right emails, notes, etc, stuff them into the context window & hope, but instead it's useful to consider offline processing and turn events into durable state: clusters of observed data becomes episodes, assumptions, contradictions and power confidence for suggestions.
That also pushes up the need for provenance & inspectability. It's going to be interesting to see what kind of memory consolidation strategies are required for each domain use case.
Also not too sure about provenance and inspectability - it is part of memory. If the source is deemed 'important' it will survive forgetting. If not, then maybe not. And its ok. I am sure you dont know the exact source who told you that the capital of France is Paris. You forgot, and its no big deal.