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).
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!
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!
I've made an open version of Google's NotebookLM, and it shows the superiority of the open source tech task! 💪
The app's workflow is simple. Given a source PDF or URL, it extracts the content from it, then tasks Meta's Llama 3.3-70B with writing the podcast script, with a good prompt crafted by @gabrielchua ("two hosts, with lively discussion, fun notes, insightful question etc.") Then it hands off the text-to-speech conversion to Kokoro-82M, and there you go, you have two hosts discussion any article.
The generation is nearly instant, because: > Llama 3.3 70B is running at 1,000 tokens/seconds with Cerebras inference > The audio is generated in streaming mode by the tiny (yet powerful) Kokoro, generating voices faster than real-time.
And the audio generation runs for free on Zero GPUs, hosted by HF on H200s.
Overall, open source solutions rival the quality of closed-source solutions at close to no cost!
New king of open VLMs: InternVL3 takes Qwen 2.5's crown! 👑
InternVL have been a wildly successful series of model : and the latest iteration has just taken back their crown thanks to their superior, natively multimodal vision training pipeline.
➡️ Most of the vision language models (VLMs) these days are built like Frankenstein : take a good text-only Large Language Model (LLM) backbone, stitch a specific vision transformer (ViT) on top of it. Then the training is sequential 🔢 : 1. Freeze the LLM weights while you train the ViT only to work with the LLM part, then 2. Unfreeze all weights to train all weights in order to work together.
💫 The Shanghai Lab decided to challenge this paradigm and chose this approach that they call "native". For each of their model sizes, they still start from a good LLM (mostly Qwen-2.5 series, did I tell you I'm a huge fan of Qwen? ❤️), and stitch the ViT, but they don't freeze anything : they train all weights together with interleaved text and image understanding data in a single pre-training phase 🎨.
They claim it results in more seamless interactions between modalities. And the results prove them right: they took the crown of top VLMs, at nearly all sizes, from their Qwen-2.5 parents. 👑
🚀 DeepSeek R1 moment has come for GUI agents: Rule-based Reinforcement Learning gives better results than SFT with 500x smaller datasets!
Traditionally (by which I mean "in the last few months"), GUI agents have been trained with supervised fine-tuning (SFT). This meant, collecting huge datasets of screen captures from people using computers, and using these to fine-tune your model. 📚
👉 But last week, a new paper introduced UI-R1, applying DeepSeek's R1-style rule-based reinforcement learning (RL) specifically to GUI action prediction tasks. This is big news: with RL, maybe we could build good agents without the need for huge datasets.
UI-R1 uses a unified reward function that evaluates multiple responses from models, optimizing via policy algorithms like Group Relative Policy Optimization (GRPO).
Specifically, the reward function assesses: 🎯 Action type accuracy: Does the predicted action match the ground truth? 📍 Coordinate accuracy (specifically for clicks): Is the predicted click within the correct bounding box? 📑 Output format: Does the model clearly articulate both its reasoning and final action?
Using just 136 carefully selected mobile tasks—compared to 76,000 tasks for larger models like OS-Atlas—UI-R1 shows significant efficiency and improved performance: 📈 Boosted action prediction accuracy from 76% to 89% on AndroidControl. 🌐 Outperformed larger, SFT-trained models (e.g., OS-Atlas-7B), demonstrating superior results with vastly fewer data points (136 tasks vs. 76K). 🔍 Enhanced adaptability and generalization, excelling even in out-of-domain scenarios.
The paper tests this RL-based method only in low-level GUI tasks. Could it generalize to more complex interactions? 🧐
As one of the most popular local inference solutions, the community had been asking us to integrate vLLM: after a heavy refactoring of our LLM classes, we've just released smolagents 1.11.0, with a brand new VLLMModel class.
If you ever asked which LLM is best for powering agents, we've just made a leaderboard that ranks them all! Built with @albertvillanova, this ranks LLMs powering a smolagents CodeAgent on subsets of various benchmarks. ✅
🏆 GPT-4.5 comes on top, even beating reasoning models like DeepSeek-R1 or o1. And Claude-3.7-Sonnet is a close second!
The leaderboard also allows you to show the scores of vanilla LLMs (without any agentic setup) on the same benchmarks: this shows the huge improvements brought by agentic setups. 💪
(Note that results will be added manually, so the leaderboard might not always have the latest LLMs)
We now have a Deep Research for academia: SurveyX automatically writes academic surveys nearly indistinguishable from human-written ones 🔥
Researchers from Beijing and Shanghai just published the first application of a deep research system to academia: their algorithm, given a question, can give you a survey of all papers on the subject.
To make a research survey, you generally follow two steps, preparation (collect and organize papers) and writing (outline creation, writing, polishing). Researchers followed the same two steps and automated them.
🎯 For the preparation part, a key part is find all the important references on the given subject. Researchers first cast a wide net of all relevant papers. But then finding the really important ones is like distilling knowledge from a haystack of information. To solve this challenge, they built an “AttributeTree” object that structures key information from citations. Ablating these AttributeTrees significantly decreased structure and synthesis scores, so they were really useful!
📝 For the writing part, key was to get a synthesis that's both short and true. This is not easy to get with LLMs! So they used methods like LLM-based deduplication to shorten the too verbose listings made by LLMs, and RAG to grab original quotes instead of made-up ones.
As a result, their system outperforms previous approaches by far!
As assessed by LLM-judges, the quality score os SurveyX even approaches this of human experts, with 4.59/5 vs 4.75/5 🏆
Less is More for Reasoning (LIMO): a 32B model fine-tuned with 817 examples can beat o1-preview on math reasoning! 🤯
Do we really need o1's huge RL procedure to see reasoning emerge? It seems not. Researchers from Shanghai Jiaotong University just demonstrated that carefully selected examples can boost math performance in large language models using SFT —no huge datasets or RL procedures needed.
Their procedure allows Qwen2.5-32B-Instruct to jump from 6.5% to 57% on AIME and from 59% to 95% on MATH, while using only 1% of the data in previous approaches.
⚡ The Less-is-More Reasoning Hypothesis: ‣ Minimal but precise examples that showcase optimal reasoning patterns matter more than sheer quantity ‣ Pre-training knowledge plus sufficient computational resources at inference levels up math skills
➡️ Core techniques: ‣ High-quality reasoning chains with self-verification steps ‣ 817 handpicked problems that encourage deeper reasoning ‣ Enough inference-time computation to allow extended reasoning
💪 Efficiency gains: ‣ Only 817 examples instead of 100k+ ‣ 40.5% absolute improvement across 10 diverse benchmarks, outperforming models trained on 100x more data
This really challenges the notion that SFT leads to memorization rather than generalization! And opens up reasoning to GPU-poor researchers 🚀
𝗚𝗿𝗲𝗮𝘁 𝗳𝗲𝗮𝘁𝘂𝗿𝗲 𝗮𝗹𝗲𝗿𝘁: you can now share agents to the Hub! 🥳🥳
And any agent pushed to Hub get a cool Space interface to directly chat with it.
This was a real technical challenge: for instance, serializing tools to export them meant that you needed to get all the source code for a tool, verify that it was standalone (not relying on external variables), and gathering all the packages required to make it run.
➡️ How well do reasoning models perform on agentic tasks? Until now, all indicators seemed to show that they worked really well. On our recent reproduction of Deep Search, OpenAI's o1 was by far the best model to power an agentic system.
So when our partner Adyen built a huge benchmark of 450 data science tasks, and built data agents with smolagents to test different models, I expected reasoning models like o1 or DeepSeek-R1 to destroy the tasks at hand.
👎 But they really missed the mark. DeepSeek-R1 only got 1 or 2 out of 10 questions correct. Similarly, o1 was only at ~13% correct answers.
🧐 These results really surprised us. We thoroughly checked them, we even thought our APIs for DeepSeek were broken and colleagues Leandro Anton helped me start custom instances of R1 on our own H100s to make sure it worked well. But there seemed to be no mistake. Reasoning LLMs actually did not seem that smart. Often, these models made basic mistakes, like forgetting the content of a folder that they had just explored, misspelling file names, or hallucinating data. Even though they do great at exploring webpages through several steps, the same level of multi-step planning seemed much harder to achieve when reasoning over files and data.
It seems like there's still lots of work to do in the Agents x Data space. Congrats to Adyen for this great benchmark, looking forward to see people proposing better agents! 🚀
OpenAI's latest agentic app Deep Research seems really good... But it's closed, as usual.
⏱️ So with a team of cracked colleagues, we set ourselves a 24hours deadline to replicate and open-source Deep Research! ⏱️
➡️ We built open-Deep-Research, an entirely open agent that can: navigate the web autonomously, scroll and search through pages, download and manipulate files, run calculation on data...
We aimed for the best performance: are the agent's answers really rigorous?
On GAIA benchmark, Deep Research had 67% accuracy on the validation set. ➡️ open Deep Research is at 55% (powered by o1), it is: - the best pass@1 solution submitted - the best open solution 💪💪
And it's only getting started ! Please jump in, drop PRs, and let's bring it to the top !
Now you can launch a code agent directly from your terminal! ✨ 𝚜𝚖𝚘𝚕𝚊𝚐𝚎𝚗𝚝 "𝚈𝚘𝚞𝚛 𝚝𝚊𝚜𝚔" directly launches a CodeAgent ▶️ This also works with web agents (replace 𝚜𝚖𝚘𝚕𝚊𝚐𝚎𝚗𝚝 with 𝚠𝚎𝚋𝚊𝚐𝚎𝚗𝚝) thanks to @merve !
💾 Another treat from smolagents release 1.7.0: Now agents have a memory mechanism, enabling many possibilities like replaying the last run with 𝚊𝚐𝚎𝚗𝚝.𝚛𝚎𝚙𝚕𝚊𝚢(), thank you @clefourrier !
✅ Hosting our own inference was not enough: now the Hub 4 new inference providers: fal, Replicate, SambaNova Systems, & Together AI.
Check model cards on the Hub: you can now, in 1 click, use inference from various providers (cf video demo)
Their inference can also be used through our Inference API client. There, you can use either your custom provider key, or your HF token, then billing will be handled directly on your HF account, as a way to centralize all expenses.
💸 Also, PRO users get 2$ inference credits per month!
Today we make the biggest release in smolagents so far: 𝘄𝗲 𝗲𝗻𝗮𝗯𝗹𝗲 𝘃𝗶𝘀𝗶𝗼𝗻 𝗺𝗼𝗱𝗲𝗹𝘀, 𝘄𝗵𝗶𝗰𝗵 𝗮𝗹𝗹𝗼𝘄𝘀 𝘁𝗼 𝗯𝘂𝗶𝗹𝗱 𝗽𝗼𝘄𝗲𝗿𝗳𝘂𝗹 𝘄𝗲𝗯 𝗯𝗿𝗼𝘄𝘀𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁𝘀! 🥳
Our agents can now casually open up a web browser, and navigate on it by scrolling, clicking elements on the webpage, going back, just like a user would.
The demo below shows Claude-3.5-Sonnet browsing GitHub for task: "Find how many commits the author of the current top trending repo did over last year." Hi @mlabonne !
Go try it out, it's the most cracked agentic stuff I've seen in a while 🤯 (well, along with OpenAI's Operator who beat us by one day)
This work from Chinese startup @MiniMax-AI introduces a novel architecture that achieves state-of-the-art performance while handling context windows up to 4 million tokens - roughly 20x longer than current models. The key was combining lightning attention, mixture of experts (MoE), and a careful hybrid approach.
𝗞𝗲𝘆 𝗶𝗻𝘀𝗶𝗴𝗵𝘁𝘀:
🏗️ MoE with novel hybrid attention: ‣ Mixture of Experts with 456B total parameters (45.9B activated per token) ‣ Combines Lightning attention (linear complexity) for most layers and traditional softmax attention every 8 layers
🏆 Outperforms leading models across benchmarks while offering vastly longer context: ‣ Competitive with GPT-4/Claude-3.5-Sonnet on most tasks ‣ Can efficiently handle 4M token contexts (vs 256K for most other LLMs)
🔬 Technical innovations enable efficient scaling: ‣ Novel expert parallel and tensor parallel strategies cut communication overhead in half ‣ Improved linear attention sequence parallelism, multi-level padding and other optimizations achieve 75% GPU utilization (that's really high, generally utilization is around 50%)
🎯 Thorough training strategy: ‣ Careful data curation and quality control by using a smaller preliminary version of their LLM as a judge!
Overall, not only is the model impressive, but the technical paper is also really interesting! 📝 It has lots of insights including a great comparison showing how a 2B MoE (24B total) far outperforms a 7B model for the same amount of FLOPs.
𝗪𝗲'𝘃𝗲 𝗷𝘂𝘀𝘁 𝗿𝗲𝗹𝗲𝗮𝘀𝗲𝗱 𝘀𝗺𝗼𝗹𝗮𝗴𝗲𝗻𝘁𝘀 𝘃𝟭.𝟯.𝟬 🚀, and it comes with a major feature: you can now log agent runs using OpenTelemetry to inspect them afterwards! 📊
This interactive format is IMO much easier to inspect big multi-step runs than endless console logs.
