#chain of thought reasoning

[jcmarchi]

Beyond Chain-of-Thought: How Thought Preference Optimization is Advancing LLMs

New Post has been published on https://thedigitalinsider.com/beyond-chain-of-thought-how-thought-preference-optimization-is-advancing-llms/

Beyond Chain-of-Thought: How Thought Preference Optimization is Advancing LLMs

A groundbreaking new technique, developed by a team of researchers from Meta, UC Berkeley, and NYU, promises to enhance how AI systems approach general tasks. Known as “Thought Preference Optimization” (TPO), this method aims to make large language models (LLMs) more thoughtful and deliberate in their responses.

The collaborative effort behind TPO brings together expertise from some of the leading institutions in AI research. 

The Mechanics of Thought Preference Optimization

At its core, TPO works by encouraging AI models to generate “thought steps” before producing a final answer. This process mimics human cognitive processes, where we often think through a problem or question before articulating our response. 

The technique involves several key steps:

The model is prompted to generate thought steps before answering a query.

Multiple outputs are created, each with its own set of thought steps and final answer.

An evaluator model assesses only the final answers, not the thought steps themselves.

The model is then trained through preference optimization based on these evaluations.

This approach differs significantly from previous techniques, such as Chain-of-Thought (CoT) prompting. While CoT has been primarily used for math and logic tasks, TPO is designed to have broader utility across various types of queries and instructions. Furthermore, TPO doesn’t require explicit supervision of the thought process, allowing the model to develop its own effective thinking strategies.

Another key difference is that TPO overcomes the challenge of limited training data containing human thought processes. By focusing the evaluation on the final output rather than the intermediate steps, TPO allows for more flexible and diverse thinking patterns to emerge.

Experimental Setup and Results

To test the effectiveness of TPO, the researchers conducted experiments using two prominent benchmarks in the field of AI language models: AlpacaEval and Arena-Hard. These benchmarks are designed to evaluate the general instruction-following capabilities of AI models across a wide range of tasks.

The experiments used Llama-3-8B-Instruct as a seed model, with different judge models employed for evaluation. This setup allowed the researchers to compare the performance of TPO against baseline models and assess its impact on various types of tasks.

The results of these experiments were promising, showing improvements in several categories:

Reasoning and problem-solving: As expected, TPO showed gains in tasks requiring logical thinking and analysis. 

General knowledge: Interestingly, the technique also improved performance on queries related to broad, factual information. 

Marketing: Perhaps surprisingly, TPO demonstrated enhanced capabilities in tasks related to marketing and sales. 

Creative tasks: The researchers noted potential benefits in areas such as creative writing, suggesting that “thinking” can aid in planning and structuring creative outputs.

These improvements were not limited to traditionally reasoning-heavy tasks, indicating that TPO has the potential to enhance AI performance across a broad spectrum of applications. The win rates on AlpacaEval and Arena-Hard benchmarks showed significant improvements over baseline models, with TPO achieving competitive results even when compared to much larger language models.

However, it’s important to note that the current implementation of TPO showed some limitations, particularly in mathematical tasks. The researchers observed that performance on math problems actually declined compared to the baseline model, suggesting that further refinement may be necessary to address specific domains.

Implications for AI Development

The success of TPO in improving performance across various categories opens up exciting possibilities for AI applications. Beyond traditional reasoning and problem-solving tasks, this technique could enhance AI capabilities in creative writing, language translation, and content generation. By allowing AI to “think” through complex processes before generating output, we could see more nuanced and context-aware results in these fields.

In customer service, TPO could lead to more thoughtful and comprehensive responses from chatbots and virtual assistants, potentially improving user satisfaction and reducing the need for human intervention. Additionally, in the realm of data analysis, this approach might enable AI to consider multiple perspectives and potential correlations before drawing conclusions from complex datasets, leading to more insightful and reliable analyses.

Despite its promising results, TPO faces several challenges in its current form. The observed decline in math-related tasks suggests that the technique may not be universally beneficial across all domains. This limitation highlights the need for domain-specific refinements to the TPO approach.

Another significant challenge is the potential increase in computational overhead. The process of generating and evaluating multiple thought paths could potentially increase processing time and resource requirements, which may limit TPO’s applicability in scenarios where rapid responses are crucial.

Furthermore, the current study focused on a specific model size, raising questions about how well TPO will scale to larger or smaller language models. There’s also the risk of “overthinking” – excessive “thinking” could lead to convoluted or overly complex responses for simple tasks. 

Balancing the depth of thought with the complexity of the task at hand will be a key area for future research and development.

Future Directions

One key area for future research is developing methods to control the length and depth of the AI’s thought processes. This could involve dynamic adjustment, allowing the model to adapt its thinking depth based on the complexity of the task at hand. Researchers might also explore user-defined parameters, enabling users to specify the desired level of thinking for different applications.

Efficiency optimization will be crucial in this area. Developing algorithms to find the sweet spot between thorough consideration and rapid response times could significantly enhance the practical applicability of TPO across various domains and use cases.

As AI models continue to grow in size and capability, exploring how TPO scales with model size will be crucial. Future research directions may include:

Testing TPO on state-of-the-art large language models to assess its impact on more advanced AI systems 

Investigating whether larger models require different approaches to thought generation and evaluation 

Exploring the potential for TPO to bridge the performance gap between smaller and larger models, potentially making more efficient use of computational resources

This research could lead to more sophisticated AI systems that can handle increasingly complex tasks while maintaining efficiency and accuracy.

The Bottom Line

Thought Preference Optimization represents a significant step forward in enhancing the capabilities of large language models. By encouraging AI systems to “think before they speak,” TPO has demonstrated improvements across a wide range of tasks, potentially revolutionizing how we approach AI development. 

As research in this area continues, we can expect to see further refinements to the technique, addressing current limitations and expanding its applications. The future of AI may well involve systems that not only process information but also engage in more human-like cognitive processes, leading to more nuanced, context-aware, and ultimately more useful artificial intelligence.

[protoslacker]

It will be possible to create better teachers out of ChatGPT tech. What led me to that is that it certainly could help people write better bug reports. It would have far more patience than a busy developer, whose job is to fix bugs, not decypher a user's (justifiably) imprecise understanding of how software works. But a chatbot could help, patiently asking the user questions. Then, of course, over time -- the user would learn how to do it themselves, and get this -- they would also learn how software works. There might be some great programmers out there who don't know they are. 😄

Dave Winer at Scripting News

[lotus-pear]

regret

[quixoticanarchy]

please. “there is no ethical consumption under late stage capitalism” does not mean that all consumption is EQUALLY unethical. it’s a structural critique of systems of production it’s not a get-out-of-guilt-free carte blanche to invoke while carelessly consuming whatever you want

[the0maski]

Not necessarily LU based, but in the overall “canon” -verse the Hero of Time doesn’t exist in two timelines, right? He died in the Downfall Timeline and never came to be a hero in the Child Timeline.

Legend and Hyrule probably only heard about the title Fallen Hero, never his true title. Same goes for Twilight, only knowing Time as the Hero’s Shade or Cursed Swordsman. Which means, only Wind knows about Time, and he is the only one with a legend about him.

Funny detail, if Hyrule Warriors would fall under the Child Timeline that would mean that Time was present during the whole timeline, it gets better if Mask was only dragged into the war, because the goddesses pulled a: You broke it, you fix it! Making it more fun, Time becomes just history’s biggest mystery and meme under historians. To the point where there are huge debates about him, because some records say that he lived after the Hyrulean civil war, but at the same time he is mention being at the War of Ages which was two whole Eras later! Was he ever a hero? Why are they no family records about him? Was he really just a forest spirit, was he even hylian?

Flora would absolutely have a field day, if the chain ever stepped only a foot in Wild’s Hyrule. Seeing how she is extremely interested in history.

For real: there is to little mention, in the fandom, that the Hero of Time is only known to one person (Or two if Mask had been in the war). Everyone else had never heard of him, less knew that there was a hero that came after Four. How had they all found each other? I know that, there is a fanfic troupe of the chain slowly forming, while hopping through portals ending in a new Link’s Hyrule. But in comic, the first time they all walked through a portal together, was after visiting Malon. Meaning they all met in Time’s Hyrule, in the Timeline were he is no “hero”. How did they find Time, since asking for a hero would not worked? What makes me also believe, that Time is only leader of the group, because he is the oldest and apparently has a high rank among Hyrule’s military. Maybe he showed the Triforce mark on his hand? But less likely since he hides it most times.

My money goes to Wind or Twilight. Wind talking randomly to this soldier, about the legend of his time, not knowing he is speaking directly to said person of the legend. The rancher only because, he got flashbacks of Shade, and he needed to find out more. Bonus would be Warriors lost at words, because that one deity that sometimes possessed his little brother, became the Milkman!

[mirensiart]

So, funny story, I just discovered KeyChain and I love them very dearly, saw a bunch of other people drawing cute pictures, wanted to get in on it-

The thing is, I saw that original post about them but not the reblog with bullet points, originally. I read that they were post Spirit Tracks (my beloved I love st so much omg) and went “oh. Trains exist. I need them to argue on top of a train right now.” Now, you might see the problem: train surfing does not fit a character who hates riding trains. Which I did not realize was the case until I’d already started sketching. I cannot finish this drawing.

But I already had Chain so I shopped him onto a surfboard instead. Does it make sense? No. But does he have a shark tooth necklace? Yes.

OK BUT THIS IS THE BEST THING EVER⁉️⁉️⁉️

The great sea is still around in this hyrule along with the trains, so surfer chain COULD BE POSSIBLE

Anyway this made my entire week LMAO THANK YOU SO MUCH !!!!!!!!

[necrotic-nephilim]

this is an invitation to ramble about slade/batboy ships: sladick, sladejay, sladetim, sladedami, and other batfam member/villain ships, especially jayroman and ra'stim :)

AAAAAA this is so delightful oh my god thank you. adding a read more just because this one is going to get Long to cover all the ships and all my opinions. because my god do i love Slade.

firstly, the original Robin/villain ship, SlaDick. Slade Wilson, literally created to be a Teen Titans villains, with the original Robin he cannot be normal about ever. i'm so sad there's not much interest in Slade aside from making him a generic Evil Guy who canonically likes teenagers because i think to just boil down his complex with Dick to 'weird attraction' robs them of SUCH nuance. Slade *trusts* Dick, he trusts Dick enough to ask Dick to train his own daughter Rose. and initially Slade's complex over Dick isn't sexual, it's seeing Dick as a replacement for his dead son, Grant. that's messy as hell and i love them for it. i don't think there's a single villain that has the respect for Dick that Slade has. i'm always of the opinion Dick's attraction to Slade is rooted in daddy issues and Slade's attraction to Dick is rooted in dead son issues. do i think they could end up as an old married couple? yes but only in a world where Dick is completely broken and feels alone. my favorite SlaDick flavor is post-Jason's death. Dick and Bruce are arguably at their worst during that era to begin with so Dick is pretty isolated and emotionally unstable. and Slade would take such advantage of that, swooping in to offer Dick emotional stability and fucked up sex to get out pent up emotions. (i'm big a big fan of Dick fucking out his feelings tbh) and Slade is just. this sort of bad habit Dick will kick for a year or two then come crawling back to. you can directly track how well Bruce and Dick are getting along based on how many times Dick has slept with Slade recently. and that's the prize, for Slade. knowing Dick will come back to him, eventually. it's all about patience. and if something really extreme happened to Dick (like Bruce's fake death) i think they'd even date briefly. it's not entirely impossible for Dick to date someone he disagrees with morally (see: his flings with Helena) and i think Dick would keep trying to 'save' Slade, using the upper hand he has of filling in this role of Slade's dead son to try to domesticate him. would it work? who knows but if anyone is going to try over and over, it's going to be Dick. it's practically self-harm for Dick yet the only thing keeping him sane. i love them.

SladeJay is... an interesting one for me. because i like the *potential*. but they have no significant interactions pre-Flashpoint. and while usually i can forgive New-52 and Rebirth for their grievances if it has ship fodder i just... can't do that for Jason. Judd Winick's Jason is the only Jason that exists to me so even Slade and Jason's canon interactions matter little to me because it's not the version of Jason i care for. the upside of that though, is it's more of a sandbox to explore what they could be and there are no limitations. i can just run wild. which is fun bc. you're telling me Slade wouldn't be so drawn in by the idea of a dead Robin who's come back and is now the antithesis of Bruce's morality? i think at some point Slade would want to poke the bear, really see what Red Hood is made of. do i see them working long-term? no but i do think Jason would have zero qualms working with Slade if he got something out of it. and if he could fuck with Bruce or Dick by having a short, fucked up relationship with Slade? that's even better. i don't think Slade could ever truly respect Jason, at the end of the day the Dick Grayson standard is too high and Slade would sneer at the idea of a legacy who fucked it up so bad he got blown up. but, he'd see that as Bruce's failure more than Jason's. and for Jason to have someone look him in the eye and say that Bruce *failed* him? i think that'd just *do* something to Jason. and Slade has lost a son, he knows what that loss feels like, how you feel you failed as a father. would he have interest in being fatherly to Jason? no but i think he'd have fun momentarily manipulating Jason and seeing what reactions he gets out of what jeers. Jason's been calling himself a failure this whole time, so to have someone else say it is no real big deal, but to have someone else say it's Bruce's fault and voice Jason's feelings? they'd have the most fucked up sex with the most unhealthy dirty talk that's both gentle and degrading. i don't think Jason would ever let himself get too close, he's far too emotionally guarded. but for a second, i think he'd fantasize about having even *half* the amount of attention that Slade gives Dick. bc what has Jason always been, but in Dick's shadow.

SladeTim. my two blorbos. in one place. somewhere in my drafts i have a half-started longfic about SladeTim that's one half really fucked up porn and one half slowburn feelings. arguably Tim and Slade don't have many canon interactions, but it's fun to me that when they do, Slade always seems sort of startled by how well Tim fights back and Tim's willingness to fight dirty in a way even Dick doesn't. and to me, that's the crux of this ship. as far as Robins go, Tim should sort of slip under the radar for Slade. he's not the dead one turned villain, he's not the grandson of Ra's al Ghul, hell he's not even the child of a second-rate villain like Steph, he's not *the* Dick Grayson, he's just... the other one. grew up pretty rich and normal and fell for all of Bruce's wax poetic nonsense. so when Tim puts himself on the map as a hero, makes himself a worthy opponent against Slade that's interesting. even to Tim, Slade isn't a particularly remarkable villain since Slade cares to stay more on Dick's radar. so when they cross paths there's a lot of unexpected. neither of them have thought about the other too hard. so there's this interest and intrigue about it i love. i'm a big fan of the idea Tim is a massive masochist, both physically and emotionally and Slade is The Sadist Ever so. i like them falling into bed together and having the most fucked up sex. like Tim just being a Weird Little Freak so fucked up even Slade raises an eyebrow. because this isn't what you *expect* of a kid like Tim, who's had a pretty easy life before tangling with vigilantes. he should be like a fish out of water, but instead he's matching Slade's energy in ways even Dick doesn't. and of course, how smart he is, that's an asset. it takes a special kind of kid to have the audacity to poison Lady Shiva with hotel chocolates and pull it *off* no less. it earns a begrudging respect, and it's rare to get Slade to respect someone. i really like the idea of Tim seeking Slade out only for fucked up sex and somehow Slade falls for this weird little freak who's cold and clinical outside of sex and keeps him guessing.

i'll be honest i've only considered SladeDami in the context of seeing antis say 'omg Slade has been predatory toward Damian ewww' and going 'no the fuck he hasn't but if you want that so bad i'll ship it just to spite you all' but their canon interactions do fascinate me. a lot of how they interact is predicated on Slade as a father, even more so than SlaDick. like Slade will fight Damian and then be like 'hey be good to your old man fathers need their sons' and fucking dip. and then with the whole Respawn thing and Shadow War? that was extra crunchy. for a brief moment Slade had a son who was a brother to Damian and then he goes and *dies*? talk about the complex that would give him with Damian, the spitting image of Respawn. Make Slade Weird About Batkids That Remind Him of His Son 2024. Damian holds an utter contempt for Slade that is simply unmatched. so Slade not leaving that kid alone because of his weird issues, making sure that Bruce doesn't screw up with Damian the way he screwed up with Respawn is very fun. and Damian slowly building up a tolerance to Slade's annoying antics could be fun. Damian is, at his core, still just a kid who needs the approval of something father-shaped and he will Take What He Can Get. are they ever healthy or long lasting? no but i do think Damian would cling to Slade during his teen years for something incredibly fucked up and codependent until either Slade dumped him or he forced himself to get over it.

JayRoman. i will not lie love these two but i don't think i've read many Black Mask comics when he's not interacting with Jason. which is funny because my entire conception of Roman is him just getting humiliated by Jason and really what more is there to know about the man. Jason is so unserious in how he handles Roman and the best part is you can tell it's truly because he doesn't see Roman as a threat. Roman's just a pawn in the game of getting Bruce's attention and sure, Jason is aiming to kill Roman by the end of it, but he'll always have bigger fish to fry. and that's so *infuriating* for Roman. this new guy who's *clearly* a fucking teenager shows up, owns you so badly it shatters your empire, and then you only live bc he seems to have gotten bored of you. JayRoman is my particular favorite ship for the flavor of 'the sub in bed is in control of every other aspect of their relationship and their submission is a gift that can be revoked at any time' which we don't get enough. fucked up power dynamics always have the sub being the one lacking control. and whilst i enjoy when Roman is able to absolutely control and manipulate Jason through various means, i think in canon, it makes far more sense he's pathetic and begging Jason for even a *chance*. and Jason very specifically picking who he subs for based on someone who he could kill or destroy at the drop of the hat if he needed to is a very Jason thing to do. there will never be trust between these two. they will fuck nasty and Roman will be in love with Jason. but they are both carrying a gun during sex. the gun is probably involved during the sex.

Ra'sTim. my everything. Red Robin (2009) you will always be famous to me. what *don't* they have. forced proximity. enemies to lovers. forced partnership. one-sided obsession. ridiculously large age gap. deep unforgivable betrayal. i will never evacuate these two from my brain dear god. Ra's is another one of those villains who gets painted with one broad stroke of being cartoonishly evil with no exploration of his interesting nuance. making him nothing but a villain is boring. where is the Ra's who loves so deeply and fully and has to lose his loved ones over and over and will not let that happen to Tim. he wants to consume Tim in a 'cannibalism as a metaphor for love but also probably literal cannibalism' way. the amount of trust put in Ra's in order for Tim to be able to betray him as spectacularly as he did? that's glorious. Tim had full unfiltered access to Ra's' computers even when he was advised against trusting Tim so much. and then Tim wins against Ra's and willingly lets Ra's kill him. (obviously Dick saves him, but I'm of the opinion Tim was just committed to dying in that moment and he was Okay With That) 'i will betray you if it's the last thing i do' as an act of love. Tim is to Ra's what Dick is to Slade. you will never convince me Tim and Ra's didn't hatefuck at least once during RR (2009) with a questionable level of consent. i'm so serious i will never shut up about them. the way Tim talks about working with Ra's as if he's making a deal with the devil and Ra's talks about Tim like he's the precious, once in a life time thing, one of the only people worthy to produce an heir for Ra's. how's that not gay. what other ship involved one of them literally trying to have the other's baby to raise as an heir. Ra's would probably carry the baby himself if he could. memes aside they're just so. they're so it. i love when Tim is forced into a Situation where he has to work with Ra's and confronts the darker aspects of himself that Ra's wants to bring out but Tim wants to squash. it is The corruption kink. whether Ra's succeeds or not in corrupting Tim doesn't even matter because the real crux of this ship is the chase. it's the way the heart pounds when they reach out for each other and you don't know if it's for a kiss or a killing blow. it's very Hannigram to me, in that i don't even need or want them to kiss to know they're in love. love to them is not true love's kiss, it's the thoughtful place they decide to stab the other in. be the sheath to my dagger type ship. hold all this bloody violence i know you're capable of inside of you. let me cut the violence out of you ship. what more can you ask for from a ship. Ra's would tie Tim down and torture him both as foreplay and as a love language and Tim would be too fucked up and self-sacrificial to stop him. always playing the dangerous game of how far will the other let them go until someone tries to die or kill. listen i think i lost the plot here but my point is they're unwell about each other. Tim will make Ra's regret the day he met Tim Drake not just for the betrayal but because Ra's can never go back to a time Before Tim. before knowing what the chase felt like. they're so. them.

[holland-vosijk-antari]

massive lack of sleep-induced ramble incoming:

been thinking more about the bonds between antari (from reading the telepathy between holland and ojka while osaron is in them) which makes me think about if the antari could telepathically communicate. firstly, holland would chose to Not Do That.... i can imagine a young kell who has learnt in a book that antari are meant to be able to talk across a bond and so he tries it and gets no response, he wonders if its him doing it incorrectly or if he's not strong enough yet so he keeps trying until eventually giving up. holland is very glad when he gives up because he would be trying to close his mind to the suffering he is experiencing only to be brought back to the surface by a small eager voice in his head saying "hello :) can you hear me?" and he knows if he responds whether it be kindly or harshly he will never have peace from it

on the other hand though, during agos before kell knows that lila is antari he thinks of her so often that she hers his voie in her head, lila of course has her suspicions that she may be antari but has no idea that such a connection is possible, so she considers these thoughts that sound alarmingly like kell as just a lingering attachment to him that she's desperately tried to sever. when they do figure it out though, people question how those two pirates (ahem, privateers) seem to always know what the other is thinking or what the other is about to do...

[sundial-bee-scribbles]

trying to psych myself up to finally do oc refs by doing fandom-related refs instead: volume 1

wanted to update my yuma from whatever tf this au is so he was a bit more unique... takes inspo from a lot of different things while also trying to be its own sorta thing? which is fitting given the au ;)

bonus chibi now that i'm also figuring out how tf to do chibis lol:

[bleue-flora]

Ok, given that the original courthouse had chains in the jail cell, we know they were well aware they existed as items. So, I think the reason they didn’t use them in Pandora’s Vault must be because the mining fatigue made them too hard to break. So, the implication is they used rope (leads) instead…

Like the ccs went through the effort of having a noticeably low durability’s netherite axe, shears named: Warden’s Torment, and you’re telling me they didn’t think of just putting some chains in their hot bar to give us brain rot?

[jcmarchi]

Meta’s COCONUT: The AI Method That Thinks Without Language

New Post has been published on https://thedigitalinsider.com/metas-coconut-the-ai-method-that-thinks-without-language/

Meta’s COCONUT: The AI Method That Thinks Without Language

Understanding COCONUT’s Innovation

Picture the difference between speaking your thoughts out loud and the actual mental process happening in your brain. That gap – between verbalized thoughts and neural activity – is exactly what Meta’s researchers tapped into with COCONUT.

The real breakthrough of COCONUT lies in how it lets AI models think in two distinct ways, much like how humans do. Think about when you’re solving a complex puzzle – you don’t narrate every possible move in your head, right? Instead, you:

Absorb the Problem: You take in the information (like reading the puzzle rules)

Think Silently: Your brain explores multiple possibilities without putting them into words

Share the Solution: Only then do you explain your thinking to others

COCONUT gives AI models this same natural flexibility. Instead of forcing them to “speak” every thought out loud (like traditional methods do), it lets them think in their natural neural space – what researchers call the “latent space.”

The model smoothly switches between two modes:

When it needs to understand questions or give answers, it uses regular language

But for the actual thinking process? It uses pure neural patterns, free from the constraints of words

Image: Meta

The Training Journey

One of the most fascinating aspects of COCONUT is its training curriculum. What makes this one special is how it mirrors natural learning progression. Think about how we teach complex skills – you don’t throw someone into the deep end immediately. You build up gradually, adding complexity as they master each level.

The researchers took this exact approach with COCONUT:

Stage 1: The Foundation

First, the model learns like any other AI – through traditional chain-of-thought reasoning. This gives it a solid base understanding.

Stage 2: The Transition

Here is where it gets interesting. Gradually, those written-out reasoning steps get replaced with continuous thoughts. Imagine slowly removing the training wheels, letting the model develop its own internal thinking patterns.

Stage 3: The Balance

Finally, the model learns to seamlessly switch between deep thinking in latent space and communicating its insights in clear language.

During training, the model developed abilities nobody explicitly programmed – like considering multiple reasoning paths simultaneously. This emergent behavior is particularly exciting because it suggests we might be getting closer to more natural forms of AI reasoning. It is these unexpected developments that often lead to the biggest breakthroughs.

Remember those neuroimaging studies I mentioned earlier? They showed that human brains often process complex reasoning tasks without heavily engaging language centers. COCONUT seems to be developing similar patterns – thinking deeply in its native neural space and only converting to language when needed for communication.

The Numbers Tell a Story

A few more key findings stand out from the research:

Math Word Problems (GSM8k): Here, COCONUT achieved 34.1% accuracy. While this falls below traditional Chain-of-Thought (42.9%), it’s significantly better than baseline approaches.

Logical Deduction (ProntoQA): COCONUT hit 99.8% accuracy, edging out traditional Chain-of-Thought’s 98.8%. But here’s the kicker – it did this while using just 9 tokens compared to CoT’s 92.5.

Complex Planning (ProsQA): The most impressive results came from this advanced reasoning test. COCONUT achieved 97% accuracy while traditional methods only reached 77.5%. And again, it did this with remarkable efficiency – 14.2 tokens versus 49.4.

What makes these results promising is not just the raw numbers – it is what they reveal about different types of thinking. While COCONUT may still be finding its footing with mathematical reasoning, it excels at tasks requiring complex logical planning and deduction.

COCONUT represents a fundamental rethinking of how AI systems can reason, and it moves us closer to more natural, efficient, and powerful forms of artificial intelligence. The journey from language-based reasoning to continuous thought is a step toward more capable and efficient AI systems.

[tobionigiris]

*narrator voice* yuppp that’s them. you’re probably wondering how they ended up in this situation...

[purecommemasolitude]

Nine people i want to get to know better tag game

But it’s a new post because the other one was far too long

LAST SONG? - Currently listening to Die Together by Amanda Tenfjord, but the *last* song I listened to was the original version (according to the Spotify title) of John Denver's Take Me Home, Country Roads

FAVORITE COLOR? - Purple!!

CURRENTLY WATCHING? - Theoretically Star Trek: The Animated Series but also it’s been months since I’ve watched an episode so. You know. Other than that I’ve generally been getting into vintage film recently — discovered Kanopy has a load of Keaton & Chaplin’s work so I am going to make my way through those shortly

LAST MOVIE? - Buster Keaton’s The Haunted House if we count it, Last Suspect (2023) if we don’t. It was the last movie I watched on the plane back to Canada from Taiwan, and while I enjoyed it I did prefer the other Chinese crime/legal proceedings movie I watched on the flight (The Invisible Guest from 2023). The plot twists from TIG felt really clever and purposeful within the context of the genre, whereas my internal monologue was more "WHAT IS GOING ON" by the time they got to the FUCKING MEAT GRINDER??? in Last Suspect. It was still a solid movie that I would rewatch, but again the meat grinder really comes out of nowhere.

SWEET/SPICY/SAVORY? - Savoury! I’d like to get my spice tolerance up, but since coming back from Taiwan I feel like my sweetness tolerance has gone down. Beverages there truly are the exact amount of sugary for me (someone who thinks Brisk Lemonade is way sweeter than it should be)

RELATIONSHIP STATUS? - Single and far too busy & also uninvested in the possibility to mingle

CURRENT OBSESSIONS? - In case it isn’t obvious, The Outsiders has been living in my mind for the past few months. Within that subset, Japanese-American Paul Holden specifically has me in a vice grip — Darry’s still my favourite character, but as a member of East Asian diaspora myself I am not immune to the implications Japanese-American Paul Holden has.

LAST THING YOU GOOGLED? - "try guys phoning it in" because i was trying to figure out how many episodes they were/what order they were in (update actually I had to stop answering to do something else and now it's "can you do police checks on minors" because they need to do that before I can volunteer somewhere but as I'm still a minor for a bit I don't know if I still have to get a police check or if I get to wait until I'm not)

This tagging list is a combination of people who actually fit the theme and people who do fit the title but whom I think I’ve already talked with too much to count for OP’s vision but who I am chickening out and putting anyway. Which is a really wordy way of saying I'm sometimes falling back on people I've interacted with before because I'm scared. Feel no pressure to participate, but if you do pretty please make a new post because the other reason I made a new post for this was because I fear what OP's notifs look like

@asexual-juliet @partiallypearl @granma-sweetie @sondheim-girly @hhuta @mishapen-moth @pesticine @mar-the-magician aaand a cheating blanket tag to anyone who wants to!

[synthshenanigans]

Jashtober Day 31- Loop

//bright colors & glitches

I had Dream(OfC) on loop again can you tell?

Separate/isolated images below v

[ratwars]

End of an era. I donated my instruments today.

[hazellvsq]

yeahhhh this is a copout and i would actually like to know the answer. what happens when they disagree on something. if hazel fucks up does frank have to make her run laps or some shit. if frank fucks up does hazel have to oust him. would this relationship not cause a massive conflict of interest.