LIMI Framework: Efficient AI Agency

• LIMI framework formalizes AI agency as the system’s ability to autonomously identify issues, hypothesize solutions, and execute multi-step workflows.
• It employs a 'less is more' approach where only 78 curated training samples trigger advanced agentic behavior, outperforming models with exponentially larger datasets.
• Empirical evaluations on AgencyBench prove that strategic data curation leads to significant boosts in performance, underscoring the Agency Efficiency Principle.

The LIMI framework—formally introduced as “Less Is More for Intelligent Agency”—addresses the emergence and efficient cultivation of agency in artificial intelligence systems. Agency is concretely defined as the system’s capacity to serve as an autonomous agent that actively discovers problems, formulates hypotheses, and executes solutions through direct, self-directed engagements with its environment and available tools. This conception marks a transition from cognitive systems that only generate or reason ("think"), to productive agentic systems capable of executing and adapting ("work"). LIMI challenges the prevailing assumption of traditional scaling laws, which posit that greater data abundance is necessary to achieve advanced agency, and instead demonstrates that a small, strategically curated set of agentic demonstrations suffices to trigger sophisticated autonomous behavior. In the context of LIMI, agentic intelligence emerges from only 78 well-designed training samples, yielding substantially higher performance than prior work utilizing orders of magnitude more data (Xiao et al., 22 Sep 2025).

1. Conceptual Foundations: Defining Agency in AI

Agency, as formalized in the LIMI framework, entails an AI’s emergent ability to autonomously identify problems, hypothesize solutions, and carry out multi-step actions incorporating tool usage and environmental adaptation. This goes beyond passive completion of user prompts and encompasses the proactive execution of collaborative tasks. The operationalization of agency within LIMI integrates: (a) autonomous task execution, (b) structured reasoning, (c) coordinated tool invocation, and (d) adaptive collaboration analogous to human workflows in software engineering and scientific research.

This standpoint reorients AI progress from mere cognitive competencies to autonomous productivity, required for practical industrial deployment. Agency thus becomes a metric for differentiating passive cognitive models from agentic AI workers with functional value in real-world systems.

2. The “Less Is More” Principle and Strategic Data Curation

LIMI’s core hypothesis interrogates traditional scaling laws, which maintain that increasing training set size inexorably produces greater agency. Empirical results contradict this paradigm: LIMI demonstrates that high-fidelity agency arises from minimal datasets provided those demonstrations are strategically curated to maximize representational diversity and behavioral richness. Specifically, the framework uses only 78 agentic sample trajectories—each one a fully resolved instance of real collaborative software development or authentic scientific workflow.

The “Less Is More” approach is distinguished by selection and curation over sheer volume, showing that agency increases superlinearly with the informativeness and complexity of training data rather than with scale alone.

3. Data Formalization and Interaction Modeling

Training inputs to LIMI consist of queries $q_i$ naturally sourced from high-value collaborative contexts (e.g., GitHub Pull Requests, developer workflows, research automation pipelines). For each query, the collaborative interaction sequence is captured as a trajectory $\tau_i = \{a_{i,1}, a_{i,2}, ..., a_{i,n_i}\}$, where each atomic action $a_{i,j}$ represents a step in the workflow, subdivided as:

• $m_{i, j}$: Model’s internal reasoning output for step $j$
• $t_{i, j}$: Tool-calling/invocation at step $j$
• $o_{i, j}$: Environment feedback/observation at step $j$

The training sample pool is mathematically described as $\mathcal{Q} = \{q_1, q_2, ..., q_{78}\}$, and these tuples are collectively used to formulate the agentic learning objective.

This formalization encapsulates comprehensive agentic behavior, from problem identification through iterative adaptation and tool orchestration, with an emphasis on realistic and collaborative context.

4. Empirical Evaluation and Performance Metrics

LIMI’s empirical evaluation uses AgencyBench, a comprehensive benchmark targeting agentic intelligence in collaborative and scientific workflow settings. Using only the 78 curated interactions, LIMI achieves 73.5% performance, exceeding state-of-the-art models by substantial margins: GLM-4.5 (45.1%), Kimi-K2-Instruct (24.1%), DeepSeek-V3.1 (11.9%), and Qwen3-235B-A22B-Instruct (27.5%) (Xiao et al., 22 Sep 2025). LIMI further achieves a 53.7% improvement over models trained on 10,000 samples, underscoring the outsized impact of data quality and interaction diversity.

These results demonstrate that agency acquisition is fundamentally not governed by dataset size but by the strategic selection and richness of demonstrations—articulated as the “Agency Efficiency Principle” in the LIMI literature.

5. The Agency Efficiency Principle

Central to LIMI is the Agency Efficiency Principle: true machine autonomy is best cultivated via strategic curation of training samples that fully represent the complexity of agentic behavior. Rather than maximizing sample counts, researchers should focus on identifying instances that encode the “essence” of autonomous workflow—problem discovery, reasoning, multi-step execution, and real-world adaptation.

This principle is operationalized by:

• Assembling a diverse set of queries and trajectories from realistic workflows.
• Recording complete sequences (reasoning, tool-calling, observation) for each sample.
• Prioritizing demonstrations that cover collaborative, interactive, and adaptive agentic patterns.

A plausible implication is that further scaling beyond the demonstrated sample efficiency may be subject to diminishing returns unless sample diversity and behavioral coverage are expanded.

6. Applications, Workflow Integration, and Broader Impact

LIMI is applied in collaborative software development (e.g., end-to-end construction of a C++ chat system, frontend game synthesis), scientific research orchestration (data pipeline assembly, experimental robotics), and integrated task environments requiring multi-agent coordination. These are domains where classical language modeling falls short and agentic autonomy is mandatory for productive deployment.

Its impact lies in enabling AI systems to move from static, knowledge-generation roles to dynamic, productive agents capable of orchestrating complex workflows, self-improving through environmental feedback, and integrating human-like reasoning and tool usage for practical utility.

7. Mathematical Model Summary

The formal specification for LIMI’s agentic interaction training set is as follows:

• For every query $q_i$, trajectory $\tau_i$ records the agent’s collaborative workflow steps:

$\tau_i = \{ a_{i,1}, a_{i,2}, ..., a_{i,n_i} \}$

with

$a_{i,j} = (m_{i,j}, t_{i,j}, o_{i,j})$

• The query pool is denoted $\mathcal{Q} = \{q_1, q_2, ..., q_{78}\}$

This formalism enables rigorous sampling, benchmarking, and further research on agency emergence via minimal, strategically selected data.

---

LIMI formalizes, tests, and validates a data-efficient pathway for cultivating high-level agentic intelligence in AI systems, challenging the orthodoxy of data scaling and highlighting the unique requirements of agency emergence. The approach establishes that agency is unlocked not by quantitative expansion, but qualitative enrichment and structural diversity of the training set—a principle likely to influence future research on autonomous agents, AI agency, and efficient machine autonomy (Xiao et al., 22 Sep 2025).