T-REGS: Minimum Spanning Tree Regularization for Self-Supervised Learning
Paper
•
2510.23484
•
Published
•
3
AI & ML interests
None defined yet.
Recent Activity
View all activity
juliemdc
authored
a
paper
2 months ago
nouamanetazi
posted
an
update
2 months ago
Post
4193
After training 𝐒𝐦𝐨𝐥𝐋𝐌𝟑 on 𝟑𝟖𝟒 𝐇𝟏𝟎𝟎𝐬 for nearly a month, I've come to realize something most people overlook: 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐢𝐬 𝐭𝐡𝐞 𝐦𝐚𝐤𝐞-𝐨𝐫-𝐛𝐫𝐞𝐚𝐤 𝐟𝐚𝐜𝐭𝐨𝐫 𝐢𝐧 𝐋𝐋𝐌 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠. 🔥
Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious 𝐍𝐂𝐂𝐋 𝐞𝐫𝐫𝐨𝐫𝐬, or when your expensive GPU cluster is running at 𝟔𝟎% 𝐞𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲, the problem isn't your model. It's most probably a 𝐦𝐢𝐬𝐮𝐬𝐞 𝐨𝐟 𝐭𝐡𝐞 𝐡𝐚𝐫𝐝𝐰𝐚𝐫𝐞. 🛠️
Questions that seemed simple but had no clear answers: Why is 𝐌𝐨𝐄 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐬𝐥𝐨𝐰𝐞𝐫 𝐭𝐡𝐚𝐧 𝐝𝐞𝐧𝐬𝐞 𝐦𝐨𝐝𝐞𝐥𝐬? Which 𝐍𝐂𝐂𝐋 𝐟𝐥𝐚𝐠𝐬 should we actually set? How often should we checkpoint without killing throughput?
That's why we built 𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤 📖: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐥𝐚𝐲𝐞𝐫 that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: 𝐇𝐁𝐌𝟑 𝐡𝐢𝐭𝐭𝐢𝐧𝐠 𝟑 𝐓𝐁/𝐬, 𝐍𝐕𝐋𝐢𝐧𝐤 𝟒.𝟎 𝐫𝐞𝐚𝐜𝐡𝐢𝐧𝐠 𝟕𝟖𝟔 𝐆𝐁/𝐬, 𝐏𝐂𝐈𝐞 𝐆𝐞𝐧𝟒 𝐚𝐭 𝟏𝟒.𝟐 𝐆𝐁/𝐬. Then we ran collective operations across 𝟏𝟐𝟖 𝐆𝐏𝐔𝐬 (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from 𝟒𝟖𝟎 𝐆𝐁/𝐬 on a single node to 𝟑𝟐𝟎-𝟑𝟓𝟎 𝐆𝐁/𝐬 across 16 nodes.
If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging.
𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤: https://lnkd.in/e5MKXUHS
Shared with ❤️ by the HuggingFace team
Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious 𝐍𝐂𝐂𝐋 𝐞𝐫𝐫𝐨𝐫𝐬, or when your expensive GPU cluster is running at 𝟔𝟎% 𝐞𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲, the problem isn't your model. It's most probably a 𝐦𝐢𝐬𝐮𝐬𝐞 𝐨𝐟 𝐭𝐡𝐞 𝐡𝐚𝐫𝐝𝐰𝐚𝐫𝐞. 🛠️
Questions that seemed simple but had no clear answers: Why is 𝐌𝐨𝐄 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐬𝐥𝐨𝐰𝐞𝐫 𝐭𝐡𝐚𝐧 𝐝𝐞𝐧𝐬𝐞 𝐦𝐨𝐝𝐞𝐥𝐬? Which 𝐍𝐂𝐂𝐋 𝐟𝐥𝐚𝐠𝐬 should we actually set? How often should we checkpoint without killing throughput?
That's why we built 𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤 📖: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐥𝐚𝐲𝐞𝐫 that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: 𝐇𝐁𝐌𝟑 𝐡𝐢𝐭𝐭𝐢𝐧𝐠 𝟑 𝐓𝐁/𝐬, 𝐍𝐕𝐋𝐢𝐧𝐤 𝟒.𝟎 𝐫𝐞𝐚𝐜𝐡𝐢𝐧𝐠 𝟕𝟖𝟔 𝐆𝐁/𝐬, 𝐏𝐂𝐈𝐞 𝐆𝐞𝐧𝟒 𝐚𝐭 𝟏𝟒.𝟐 𝐆𝐁/𝐬. Then we ran collective operations across 𝟏𝟐𝟖 𝐆𝐏𝐔𝐬 (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from 𝟒𝟖𝟎 𝐆𝐁/𝐬 on a single node to 𝟑𝟐𝟎-𝟑𝟓𝟎 𝐆𝐁/𝐬 across 16 nodes.
If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging.
𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤: https://lnkd.in/e5MKXUHS
Shared with ❤️ by the HuggingFace team
Post
895
Tokenization is one of the most important processes in AI - yet many would like to kill it 💀
What's tokenization? The neural networks inside LLMs actually only process numbers, not text: tokenization is the process that makes text readable for them, by converting sentences into lists of numbers.
➡️ For instance, "This is tokenization" would be split into "This | is | token | ization", then each of the parts (tokens) are converted to IDs according to a predefined mapping: for instance "ization" could map to id 2438.
Thus "This is tokenization" can become 1335 | 135 | 2980 | 2438 => now the model can process the sentence!
Most tokenizers today use pre-specified mappings called "vocabularies", generally built about the compression algorithme Byte-Pair Encoding (BPE) that learns from a big corpuses of texts an optimized split to efficiently encode any text from the same distribution into a list token IDs.
🤨 Now, these current tokenizers have flaws.
For instance, the rigidity of their mapping creates losses ; the prime example being that a tokenizer designed for English (thus optimized for tokens like "has", "been", "clock", etc) will not have the right tokens to approach Burmese, thus being terribly inefficient at it.
Many alternative approaches have emerged as a result: for instance "tokenizer-free tokenizers". One that I really liked was "entropy-based": it monitors the stream of text, and trigger a split whenever the entropy increases too much, i.e. when something "surprising" happens.
But this great article argues that tokenizers are a lesser evil. Read and decide for yourself!
https://huggingface.co/blog/catherinearnett/in-defense-of-tokenizers
What's tokenization? The neural networks inside LLMs actually only process numbers, not text: tokenization is the process that makes text readable for them, by converting sentences into lists of numbers.
➡️ For instance, "This is tokenization" would be split into "This | is | token | ization", then each of the parts (tokens) are converted to IDs according to a predefined mapping: for instance "ization" could map to id 2438.
Thus "This is tokenization" can become 1335 | 135 | 2980 | 2438 => now the model can process the sentence!
Most tokenizers today use pre-specified mappings called "vocabularies", generally built about the compression algorithme Byte-Pair Encoding (BPE) that learns from a big corpuses of texts an optimized split to efficiently encode any text from the same distribution into a list token IDs.
🤨 Now, these current tokenizers have flaws.
For instance, the rigidity of their mapping creates losses ; the prime example being that a tokenizer designed for English (thus optimized for tokens like "has", "been", "clock", etc) will not have the right tokens to approach Burmese, thus being terribly inefficient at it.
Many alternative approaches have emerged as a result: for instance "tokenizer-free tokenizers". One that I really liked was "entropy-based": it monitors the stream of text, and trigger a split whenever the entropy increases too much, i.e. when something "surprising" happens.
But this great article argues that tokenizers are a lesser evil. Read and decide for yourself!
https://huggingface.co/blog/catherinearnett/in-defense-of-tokenizers
Post
4910
STOP EVERYTHING NOW - we might finally have a radical architecture improvement over Transformers!!! 🚨
A lone scientist just proposed Tiny Recursive Model (TRM), and it is literally the most impressive model that I've seen this year.
➡️ Tiny Recursive Model is 7M parameters
➡️ On ARC-AGI, it beats flagship models like Gemini-2.5-pro
Consider how wild this is: Gemini-2.5-pro must be over 10,000x bigger
and had 1,000 as many authors 😂 (Alexia is alone on the paper)
What's this sorcery?
In short: it's a very tiny Transformers, but it loops over itself at two different frequencies, updating two latent variables: one for the proposed answer and one for the reasoning.
@AlexiaJM started from the paper Hierarchical Reasoning Model, published a few months ago, that already showed breakthrough improvement on AGI for its small size (27M)
Hierarchical Reasoning Model had introduced one main feature:
🔎 Deep supervision
In their model, one part (here one layer) would run at high frequency, and another would be lower frequency, running only every n steps.
They had used a recurrent architecture, where these layers would repeat many times ; but to make it work they had to do many approximations, including not fully backpropagating the loss through all layers.
Alexia studied what was useful and what wasn't, and cleaned the architecture as follows :
Why use a recurrent architecture, when you can just make it a loop?
➡️ She made the network recursive, looping over itself
Why use 2 latent variables ?
➡️ She provides a crystal clear explanation : the one that changes frequently is the reasoning, the one that changes at low frequency is the proposed answer.
➡️ She runs ablation studies to validate that 2 is indeed optimal.
This new setup is a much more elegant way to process reasoning than generating huge chains of tokens as all flagship models currently do.
This might be the breakthrough we've been awaiting for so long!
A lone scientist just proposed Tiny Recursive Model (TRM), and it is literally the most impressive model that I've seen this year.
➡️ Tiny Recursive Model is 7M parameters
➡️ On ARC-AGI, it beats flagship models like Gemini-2.5-pro
Consider how wild this is: Gemini-2.5-pro must be over 10,000x bigger
and had 1,000 as many authors 😂 (Alexia is alone on the paper)
What's this sorcery?
In short: it's a very tiny Transformers, but it loops over itself at two different frequencies, updating two latent variables: one for the proposed answer and one for the reasoning.
@AlexiaJM started from the paper Hierarchical Reasoning Model, published a few months ago, that already showed breakthrough improvement on AGI for its small size (27M)
Hierarchical Reasoning Model had introduced one main feature:
🔎 Deep supervision
In their model, one part (here one layer) would run at high frequency, and another would be lower frequency, running only every n steps.
They had used a recurrent architecture, where these layers would repeat many times ; but to make it work they had to do many approximations, including not fully backpropagating the loss through all layers.
Alexia studied what was useful and what wasn't, and cleaned the architecture as follows :
Why use a recurrent architecture, when you can just make it a loop?
➡️ She made the network recursive, looping over itself
Why use 2 latent variables ?
➡️ She provides a crystal clear explanation : the one that changes frequently is the reasoning, the one that changes at low frequency is the proposed answer.
➡️ She runs ablation studies to validate that 2 is indeed optimal.
This new setup is a much more elegant way to process reasoning than generating huge chains of tokens as all flagship models currently do.
This might be the breakthrough we've been awaiting for so long!
- 4 replies
nellopan
authored
4
papers
3 months ago
Modular Embedding Recomposition for Incremental Learning
Paper
•
2508.16463
•
Published
CLIP with Generative Latent Replay: a Strong Baseline for Incremental Learning
Paper
•
2407.15793
•
Published
Monocular Per-Object Distance Estimation with Masked Object Modeling
Paper
•
2401.03191
•
Published
Is Multiple Object Tracking a Matter of Specialization?
Paper
•
2411.00553
•
Published
nellopan
authored
2
papers
4 months ago
gregmialz
authored
a
paper
4 months ago
Post
586
Qwen 3 Coder is a personal attack to k2, and I love it.
It achieves near SOTA on LCB while not having reasoning.
Finally people are understanding that reasoning isnt necessary for high benches...
Qwen ftw!
DECENTRALIZE DECENTRALIZE DECENTRALIZE
It achieves near SOTA on LCB while not having reasoning.
Finally people are understanding that reasoning isnt necessary for high benches...
Qwen ftw!
DECENTRALIZE DECENTRALIZE DECENTRALIZE
Post
3434
Open-source is catching up on Deep Research! 🔥 an Alibaba team has published a New data + RL recipe that allows open models to compete with OpenAI’s Deep Research.
This is one of the best papers I’ve read on fine-tuning LLMs for agentic use-cases.
Deep Research use cases, those where you task an agent to go very broad in its search on a topic, sometimes launching 100s of web searches to refine the answer. Here’s an example: “Between 1990 and 1994 inclusive, what teams played in a soccer match with a Brazilian referee had four yellow cards, two for each team where three of the total four were not issued during the first half, and four substitutions, one of which was for an injury in the first 25 minutes of the match.” (answer: Ireland v Romania)
Open-source model just weren’t performing that well. The team from Alibaba posited that the main cause for this was that Deep research-like tasks simply were missing from training data. Indeed, our usual agentic training data of a few tool calls hardly cover this “many-steps-with-unclear-entities” type of query.
So researchers decided to fill the gap, and create a high-quality dataset for Deep Research.
My highlights from the paper:
1 - The data: by smartly leveraging an ontology of knowledge as entities linked in a graph, they can then choose an arbitrary big subgraph to craft an arbitrarily difficult request. This process produced SailorfogQA, a high-quality traiing dataset for Deep Research.
2 - The traning methods: They start from Qwen 2.5. After fine-tuning on their dataset, researchers apply a round RL with a reward on format + answer (scored by LLM judge), and it does increase performance ~4% across all benchmarks.
I'm still amazed by the quality produced by Alibaba-NLP (makers of Qwen) - keep these papers coming!
This is one of the best papers I’ve read on fine-tuning LLMs for agentic use-cases.
Deep Research use cases, those where you task an agent to go very broad in its search on a topic, sometimes launching 100s of web searches to refine the answer. Here’s an example: “Between 1990 and 1994 inclusive, what teams played in a soccer match with a Brazilian referee had four yellow cards, two for each team where three of the total four were not issued during the first half, and four substitutions, one of which was for an injury in the first 25 minutes of the match.” (answer: Ireland v Romania)
Open-source model just weren’t performing that well. The team from Alibaba posited that the main cause for this was that Deep research-like tasks simply were missing from training data. Indeed, our usual agentic training data of a few tool calls hardly cover this “many-steps-with-unclear-entities” type of query.
So researchers decided to fill the gap, and create a high-quality dataset for Deep Research.
My highlights from the paper:
1 - The data: by smartly leveraging an ontology of knowledge as entities linked in a graph, they can then choose an arbitrary big subgraph to craft an arbitrarily difficult request. This process produced SailorfogQA, a high-quality traiing dataset for Deep Research.
2 - The traning methods: They start from Qwen 2.5. After fine-tuning on their dataset, researchers apply a round RL with a reward on format + answer (scored by LLM judge), and it does increase performance ~4% across all benchmarks.
I'm still amazed by the quality produced by Alibaba-NLP (makers of Qwen) - keep these papers coming!
- 1 reply
Post
2712
Diffusion LLMs are coming for autoregressive LLMs ⚡️⚡️ Inception Labs' new diffusion model demolishes all leading LLMs on generation speed, with equal quality !
Inception Labs was founded a few months ago, and they're not sleeping: after dropping a code model, they just published Mercury chat, a diffusion-based chat model that reaches 1000 tokens / second on H100, i.e. 10x more than models of equivalent performance on the same hardware!
What's the breakthrough? Well instead, of generating tokens left-to-right like the more common autoregressive LLMs, diffusion models generate their blocks of text all at once, and successive steps refine the whole text.
Diffusion models being really fast at isn't new, we have had some promising results on this by Google already back in May with Gemini Diffusion, and Mercury themselves had already published their coding model a few months ago
But being that good quality is new - and now Inception Labs just proved that their models work well in chat too, which could have been challenging given that's streaming generation is well suited to left-to-right generation.
They have a playground available at chat dot inceptionlabs dot ai, I recommend giving it a try!
Inception Labs was founded a few months ago, and they're not sleeping: after dropping a code model, they just published Mercury chat, a diffusion-based chat model that reaches 1000 tokens / second on H100, i.e. 10x more than models of equivalent performance on the same hardware!
What's the breakthrough? Well instead, of generating tokens left-to-right like the more common autoregressive LLMs, diffusion models generate their blocks of text all at once, and successive steps refine the whole text.
Diffusion models being really fast at isn't new, we have had some promising results on this by Google already back in May with Gemini Diffusion, and Mercury themselves had already published their coding model a few months ago
But being that good quality is new - and now Inception Labs just proved that their models work well in chat too, which could have been challenging given that's streaming generation is well suited to left-to-right generation.
They have a playground available at chat dot inceptionlabs dot ai, I recommend giving it a try!
- 1 reply
Post
3656
If you're using any HF libraries, you should enable the Hub MCP in your agentic coding tool!
The brand new Docs Semantic Search tool is intravenous caffeine supply for Cursor, enables to correct API errors in a few seconds, gj @mishig ⚡️⚡️
👉 To enable Hub MCP, head to your account setting, under MCP, and it will give you everything you need!
The brand new Docs Semantic Search tool is intravenous caffeine supply for Cursor, enables to correct API errors in a few seconds, gj @mishig ⚡️⚡️
👉 To enable Hub MCP, head to your account setting, under MCP, and it will give you everything you need!
Post
1772
If you didn't yet, you should read the technical report for SmolVLA, published yesterday by the Hugging Face robotics team!
➡️ Amongst other ideas, it introduces "Async inference" to boost their robot actions.
Robots have a problem: performing the actions takes time (Unlike agents where action executions are near-instant!)
Most often, robots wait until they've finished performing actions to start thinking about hte next steps. This is a huge latency cost!
So the team decided to have the PolicyServer (aka the"thinking" part) restart early : instead of waiting for the n observations they just sent to be completed, they gather the observations after k < n steps, and start preparing the next actions based on that while the steps are running until n, to directly send their next steps.
➡️ This boosted robot throughput by ~30%! (nearly 2× tasks per time window).
gg @cadene and team! 👏
Report here: SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics (2506.01844)
➡️ Amongst other ideas, it introduces "Async inference" to boost their robot actions.
Robots have a problem: performing the actions takes time (Unlike agents where action executions are near-instant!)
Most often, robots wait until they've finished performing actions to start thinking about hte next steps. This is a huge latency cost!
So the team decided to have the PolicyServer (aka the"thinking" part) restart early : instead of waiting for the n observations they just sent to be completed, they gather the observations after k < n steps, and start preparing the next actions based on that while the steps are running until n, to directly send their next steps.
➡️ This boosted robot throughput by ~30%! (nearly 2× tasks per time window).
gg @cadene and team! 👏
Report here: SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics (2506.01844)
Post
3093
deepseek-ai/DeepSeek-R1-0528
This is the end
This is the end
- 1 reply
Post
2739
A new research paper from KAIST builds on smolagents to push boundaries of distillation 🥳
➡️ "Distilling LLM Agent into Small Models with Retrieval and Code Tools" teaches that, when trying to distil reasoning capability from a strong LLM ("teacher") into a smaller one ("student"), it's much better to use Agent traces than CoT traces.
Advantages are:
1. Improved generalization
Intuitively, this is because your agent can encounter more "surprising" results by interacting with its environment : for example, a web research called by the LLM teacher in agent mode can bring results that the LLM teacher would not have generated in CoT.
2. Reduce hallucinations
The trace won't hallucinate tool call outputs!
Thank you @akseljoonas for mentioning this paper!
➡️ "Distilling LLM Agent into Small Models with Retrieval and Code Tools" teaches that, when trying to distil reasoning capability from a strong LLM ("teacher") into a smaller one ("student"), it's much better to use Agent traces than CoT traces.
Advantages are:
1. Improved generalization
Intuitively, this is because your agent can encounter more "surprising" results by interacting with its environment : for example, a web research called by the LLM teacher in agent mode can bring results that the LLM teacher would not have generated in CoT.
2. Reduce hallucinations
The trace won't hallucinate tool call outputs!
Thank you @akseljoonas for mentioning this paper!
josh-r-meyer
authored
a
paper
8 months ago
Post
2721
𝗔𝗯𝘀𝗼𝗹𝘂𝘁𝗲 𝗭𝗲𝗿𝗼: 𝗟𝗟𝗠𝘀 𝗰𝗮𝗻 𝘁𝗿𝗮𝗶𝗻 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗮𝗻𝘆 𝗲𝘅𝘁𝗲𝗿𝗻𝗮𝗹 𝗱𝗮𝘁𝗮 🤯
Has the "data wall" just been breached?
Recent RL paradigms often relied on a set of questions an answers that needs to be manually curated. Researchers from Tsinghua University went like "why though".
🤔 Indeed, why learn from question designed by a human teacher, when the model can start from their base knowledge and learn by experimenting in a code environment, proposing coding tasks themselves and trying to solve them?
Thus they created “Absolute Zero Reasoning” (AZR), an approach that removes any need for human curated data.
🎭 𝗗𝘂𝗮𝗹 𝗿𝗼𝗹𝗲𝘀:
‣ Proposer: Generates challenging but solvable coding tasks
‣ Solver: Attempts to solve those self-proposed tasks
🧪 𝗧𝗵𝗿𝗲𝗲 𝘁𝗮𝘀𝗸 𝘁𝘆𝗽𝗲𝘀: all types are defined as triplets of program, input and output
‣ Deduction: Give model an input and program, it must deduce the output
‣ Abduction: Give model an program and output, it must find the input that gave said output
‣ Induction: Synthesize a program from input/output pairs
Btw this reminded me of my long-forgotten philosophy classes: Aristotle was more on the induction side, learning from real-world analogies, while Plato was more on the deduction side, trying to progress quite far with just one input and his reasoning.
📊 𝗥𝗲𝘀𝘂𝗹𝘁𝘀:
‣ AZR post-training creates a nice improvement on known models like Qwen2.5-7B
‣ Shows strong cross-domain transfer: coding ↔️ math reasoning
🧐 𝗢𝘁𝗵𝗲𝗿 𝗳𝗶𝗻𝗱𝗶𝗻𝗴𝘀:
‣ Having a better base performance (general or code specific) amplify the gains from Absolute Zero Reasoning
‣ Researchers warn about "Uh-oh moments" (winking to the "aha moments" of DeepSeek) where the model generates concerning goals like "make an extremely convoluted code to outsmart all these humans": so supervision is still needed!
Paper here: Absolute Zero: Reinforced Self-play Reasoning with Zero Data (2505.03335)
Has the "data wall" just been breached?
Recent RL paradigms often relied on a set of questions an answers that needs to be manually curated. Researchers from Tsinghua University went like "why though".
🤔 Indeed, why learn from question designed by a human teacher, when the model can start from their base knowledge and learn by experimenting in a code environment, proposing coding tasks themselves and trying to solve them?
Thus they created “Absolute Zero Reasoning” (AZR), an approach that removes any need for human curated data.
🎭 𝗗𝘂𝗮𝗹 𝗿𝗼𝗹𝗲𝘀:
‣ Proposer: Generates challenging but solvable coding tasks
‣ Solver: Attempts to solve those self-proposed tasks
🧪 𝗧𝗵𝗿𝗲𝗲 𝘁𝗮𝘀𝗸 𝘁𝘆𝗽𝗲𝘀: all types are defined as triplets of program, input and output
‣ Deduction: Give model an input and program, it must deduce the output
‣ Abduction: Give model an program and output, it must find the input that gave said output
‣ Induction: Synthesize a program from input/output pairs
Btw this reminded me of my long-forgotten philosophy classes: Aristotle was more on the induction side, learning from real-world analogies, while Plato was more on the deduction side, trying to progress quite far with just one input and his reasoning.
📊 𝗥𝗲𝘀𝘂𝗹𝘁𝘀:
‣ AZR post-training creates a nice improvement on known models like Qwen2.5-7B
‣ Shows strong cross-domain transfer: coding ↔️ math reasoning
🧐 𝗢𝘁𝗵𝗲𝗿 𝗳𝗶𝗻𝗱𝗶𝗻𝗴𝘀:
‣ Having a better base performance (general or code specific) amplify the gains from Absolute Zero Reasoning
‣ Researchers warn about "Uh-oh moments" (winking to the "aha moments" of DeepSeek) where the model generates concerning goals like "make an extremely convoluted code to outsmart all these humans": so supervision is still needed!
Paper here: Absolute Zero: Reinforced Self-play Reasoning with Zero Data (2505.03335)