A new study by <a href="https://en.sjtu.edu.cn/">Shanghai Jiao Tong University</a> and <a href="https://plms.ai/">SII Generative AI Research Lab</a> (GAIR) shows that training large language models (LLMs) for complex, autonomous tasks does not require massive datasets. Their framework, LIMI (<a href="https://arxiv.org/abs/2509.17567">Less Is More for Intelligent Agency</a>), builds on similar work in other areas of LLM research and finds that “machine autonomy emerges not from data abundance but from strategic curation of high-quality agentic demonstrations.” In other words, it&#x27;s data quality, not quantity, that matters. In experiments, the researchers found that with a small, but carefully curated, dataset of just 78 examples, they could train LLMs to outperform models trained on thousands of examples by a considerable margin on key industry benchmarks. This discovery could have important implications for enterprise applications where data is scarce or expensive to collect.<h2>The challenge of building agents that work</h2>The researchers define agency as “the emergent capacity of AI systems to function as autonomous agents–actively discovering problems, formulating hypotheses, and executing solutions through self-directed engagement with environments and tools.” In other words, these are AI systems that “don’t just think, but work.” The problem is that current training frameworks assume that higher agentic intelligence requires a lot of data, as has been shown in the classic scaling laws of language modeling. The researchers argue that this approach leads to increasingly complex training pipelines and substantial resource requirements. Moreover, in many areas, data is not abundant, hard to obtain, and very expensive to curate.However, research in other domains suggests that you don’t necessarily require more data to achieve training objectives in LLM training. For example, <a href="https://arxiv.org/abs/2305.11206">LIMA</a>, a 2023 paper, showed a model could be effectively aligned with just 1,000 curated examples. More recently, <a href="https://venturebeat.com/ai/researchers-find-you-dont-need-a-ton-of-data-to-train-llms-for-reasoning-tasks">LIMO</a> demonstrated that complex mathematical reasoning could emerge from only 817 training samples. With LIMI, the researchers sought to apply the same “less is more” principle to the complex world of AI agents.<h2>How LIMI works</h2>The LIMI framework demonstrates that sophisticated agentic intelligence can emerge from minimal but strategically curated demonstrations of autonomous behavior. Key to the framework is a pipeline for collecting high-quality demonstrations of agentic tasks. Each demonstration consists of two parts: a query and a trajectory. A query is a natural language request from a user, such as a software development requirement or a scientific research goal. The trajectory is the series of steps the AI takes to address the query, including its internal reasoning, its calls to external tools like a code interpreter, and the observations it receives from the environment. For example, a query might be &quot;build a simple chat application,&quot; and the trajectory would include the agent’s internal reasoning and action plan, the code it writes and executes, and the resulting output or errors. The trajectory could include multiple iterations of planning, execution, and reflection until it achieves the desired objective.To build their dataset, the researchers started with 60 queries from real-world scenarios faced by professional developers and researchers. They then expanded this pool by using <a href="https://venturebeat.com/ai/openai-launches-gpt-5-not-agi-but-capable-of-generating-software-on-demand">GPT-5</a> to synthesize additional queries from GitHub Pull Requests. They employed a team of four computer science PhD students to vet the quality of these queries and choose 18 examples to create a high-quality set of 78 queries focused on software development and research workflows.To generate the trajectories, the same PhD students collaborated with a CLI coding agent powered by GPT-5 to complete the 78 tasks. They followed an iterative process, collecting the entire interaction sequence until each task was successfully completed, capturing the full arc of realistic human-AI collaboration, including back-and-forth communication and iterative refinement. For the more complex queries, the collected trajectories could extend to more than 152,000 tokens.“This approach guarantees that our models learn not only from successful outcomes but also from the complete problem-solving process, including how to adapt strategies and recover from failures during collaborative execution,” the researchers write.<h2>LIMI in action</h2>To test their framework, the team evaluated models on <a href="https://github.com/GAIR-NLP/AgencyBench">AgencyBench</a>, a benchmark designed for measuring agentic skills, as well as other established benchmarks for tool use and coding. They fine-tuned <a href="https://venturebeat.com/ai/chinese-startup-z-ai-launches-powerful-open-source-glm-4-5-model-family-with-powerpoint-creation">GLM-4.5</a>, a powerful open-source model, using their 78-sample dataset and compared its performance against several frontier models, including the base GLM-4.5, <a href="https://venturebeat.com/ai/moonshot-ais-kimi-k2-outperforms-gpt-4-in-key-benchmarks-and-its-free">Kimi-K2-Instruct</a>, and <a href="https://venturebeat.com/ai/deepseek-v3-1-just-dropped-and-it-might-be-the-most-powerful-open-ai-yet">DeepSeek-V3.1</a>. The LIMI-trained model achieved an average score of 73.5% on AgencyBench, significantly outperforming all baseline models, the best of which (GLM-4.5) scored 45.1%. This superiority extended to other benchmarks covering tool use, coding, and scientific computing, where LIMI also outperformed all baselines.More importantly, the study showed that the model trained on just 78 examples outperformed models trained with 10,000 samples from another dataset, delivering superior performance with 128 times less data. “This discovery fundamentally reshapes how we develop autonomous AI systems, suggesting that mastering agency requires understanding its essence, not scaling training data,” the researchers write. “As industries transition from thinking AI to working AI, LIMI provides a paradigm for sustainable cultivation of truly agentic intelligence.”The researchers have released <a href="https://github.com/GAIR-NLP/LIMI">the code</a> for the data synthesis and training and <a href="https://huggingface.co/GAIR/LIMI">model weights</a>. For the enterprise, this approach offers a practical path toward developing highly specialized AI agents. Instead of undertaking massive data collection projects, organizations can leverage their in-house talent and subject matter experts to create small, high-quality datasets for bespoke agentic tasks. This lowers the barrier to entry and enables businesses to build custom AI agents that can provide a competitive edge on the workflows that matter most to them.

تظهر دراسة رائدة من جامعة جياو تونغ في شنغهاي ومختبر أبحاث الذكاء الاصطناعي التوليدي SII أن تدريب نماذج اللغة الكبيرة على مهام معقدة يمكن أن يتم باستخدام 78 مثالًا فقط، مما يتحدى الفكرة القائلة بأن مجموعات البيانات الضخمة ضرورية. هذا الإطار المبتكر، المعروف باسم LIMI، لا يبسط فقط عملية التدريب، بل يعزز أيضًا استقلالية الآلات، مما يمثل تقدمًا كبيرًا في أبحاث الذكاء الاصطناعي. يمكن أن يؤدي هذا التطور إلى أنظمة ذكاء اصطناعي أكثر كفاءة تتطلب بيانات أقل، مما يسرع من نشر الوكلاء البرمجيين الأذكياء عبر مختلف الصناعات.

Un estudio innovador de la Universidad Jiao Tong de Shanghái y el SII Generative AI Research Lab muestra que entrenar grandes modelos de lenguaje para tareas complejas se puede lograr con solo 78 ejemplos, desafiando la noción de que se necesitan conjuntos de datos masivos. Este marco, conocido como LIMI, no solo simplifica el proceso de entrenamiento, sino que también mejora la autonomía de las máquinas, lo que representa un avance significativo en la investigación de IA. Este desarrollo podría llevar a sistemas de IA más eficientes que requieren menos datos, acelerando así el despliegue de agentes de software inteligentes en diversas industrias.

Une étude révolutionnaire de l'Université Jiao Tong de Shanghai et du SII Generative AI Research Lab révèle que la formation de grands modèles de langage pour des tâches complexes peut être réalisée avec seulement 78 exemples, remettant en question l'idée selon laquelle d'énormes ensembles de données sont nécessaires. Ce cadre innovant, connu sous le nom de LIMI, rationalise non seulement le processus de formation, mais améliore également l'autonomie des machines, ce qui constitue une avancée significative dans la recherche en IA. Ce développement pourrait conduire à des systèmes d'IA plus efficaces nécessitant moins de données, accélérant ainsi le déploiement d'agents logiciels intelligents dans divers secteurs.

A groundbreaking study from Shanghai Jiao Tong University and SII Generative AI Research Lab reveals that training large language models for complex tasks can be achieved with just 78 examples, challenging the notion that vast datasets are necessary. This innovative framework, known as LIMI, not only streamlines the training process but also enhances machine autonomy, making it a significant advancement in AI research. This development could lead to more efficient AI systems that require less data, ultimately accelerating the deployment of intelligent software agents across various industries.

New AI training method creates powerful software agents with just 78 examples

Researchers say results underline large language models’ poor grasp of humour, empathy and cultural nuanceComedians who rely on clever wordplay and writers of witty headlines can rest a little easier, for the moment at least, research on AI suggests.Experts from universities in the UK and Italy have been investigating whether large language models (LLMs) understand puns – and found them wanting. <a href="https://www.theguardian.com/technology/2025/nov/24/ai-doesnt-get-puns-study-finds">Continue reading...</a>

وجد الباحثون من جامعات في المملكة المتحدة وإيطاليا أن نماذج اللغة الكبيرة (LLMs) تواجه صعوبة في فهم الألعاب اللغوية، مما يبرز قيودها في فهم الفكاهة والتعاطف والفروق الثقافية. تشير هذه الدراسة إلى أن قدرة الذكاء الاصطناعي على التفاعل مع العناصر اللغوية المعقدة لا تزال غير كافية.

Investigadores de universidades en el Reino Unido e Italia han encontrado que los modelos de lenguaje de gran tamaño (LLMs) tienen dificultades para entender los juegos de palabras, lo que resalta sus limitaciones en la comprensión del humor, la empatía y las sutilezas culturales. Este estudio sugiere que la capacidad de la IA para interactuar con elementos lingüísticos complejos sigue siendo inadecuada.

Des chercheurs d'universités au Royaume-Uni et en Italie ont découvert que les modèles de langage de grande taille (LLMs) ont du mal à comprendre les jeux de mots, mettant en évidence leurs limites dans la compréhension de l'humour, de l'empathie et des nuances culturelles. Cette étude suggère que la capacité de l'IA à interagir avec des éléments linguistiques complexes reste inadéquate.

Researchers from universities in the UK and Italy have found that large language models (LLMs) struggle to understand puns, highlighting their limitations in grasping humor, empathy, and cultural nuances. This study suggests that AI's ability to engage with complex linguistic elements remains inadequate.

New AI training method creates powerful software agents with just 78 examples

Was this article worth reading? Share it

Dyad

Chattermate

LucidQuery AI