Towards Autonomous Business Intelligence via Data-to-Insight Discovery Agent

# Towards Autonomous Business Intelligence via Data-to-Insight Discovery Agent
Source: [https://arxiv.org/html/2605.07202](https://arxiv.org/html/2605.07202)
###### Abstract

Transforming fragmented enterprise data into actionable insights remains a significant challenge for LLMs, constrained by complex database schemas, limitations in dynamic SQL generation, and the need for deep multi\-dimensional analysis\. In this paper, we proposeAIDA\(AutonomousInsightDiscoveryAgent\), the first end\-to\-end framework designed for autonomous exploration in complex business environments\. We establish a highly flexible instant retail environment encompassing 200\+ metrics and 100\+ dimensions, and integrates a proprietary Domain\-Specific Language \(DSL\) that bridges semantic reasoning with precise SQL execution\. Our reinforcement learning system subsequently formulates business analysis as a Pareto Principle\-guided cumulative reasoning process\. Experimental results demonstrate that AIDA significantly outperforms workflow\-based agents, and extensive evaluations further reveal that AIDA achieves superior environmental perception and more in\-depth analysis from diverse perspectives\. Our work ultimately establishes the transformative potential of autonomous intelligence for industrial\-scale business intelligence systems\.

Machine Learning, ICML

## 1Introduction

The essence of Business Analysis lies in the profound transformation of massive datasets into valuable insights as shown in Figure[1](https://arxiv.org/html/2605.07202#S1.F1)\(Zikopoulos and Eaton,[2011](https://arxiv.org/html/2605.07202#bib.bib24); Whitney,[2012](https://arxiv.org/html/2605.07202#bib.bib3); Chenet al\.,[2012](https://arxiv.org/html/2605.07202#bib.bib23); Vera\-Baqueroet al\.,[2015](https://arxiv.org/html/2605.07202#bib.bib19)\)\. Driven by the ascendancy of Large Language Models \(LLMs\)\(OpenAI,[2025](https://arxiv.org/html/2605.07202#bib.bib8); Google,[2025](https://arxiv.org/html/2605.07202#bib.bib7); Yanget al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib21)\), diverse analytical frameworks leveraging LLM capabilities have emerged\(Guoet al\.,[2024](https://arxiv.org/html/2605.07202#bib.bib15); Honget al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib14); Wenget al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib13)\)\. They demonstrate a sophisticated capability to mine and extract valuable insights from provided databases\. Notably,Pérezet al\.\([2025](https://arxiv.org/html/2605.07202#bib.bib20)\); Zhanget al\.\([2025](https://arxiv.org/html/2605.07202#bib.bib28)\)introduced agentic approaches that empower LLMs to interact with external databases via programmatic execution, leveraging scripts or queries to handle structured data\.

![Refer to caption](https://arxiv.org/html/2605.07202v1/task_v2.png)Figure 1:A professional business analysis workflow for root cause insight\. This trajectory illustrates the iterative process of refining business queries: beginning with performance benchmarking, proceeding to a multi\-stage funnel analysis to isolate the loss process, and finally branching into specific hypotheses regarding user structure, price competitiveness, and logistics efficiency\. The flow characterizes the dynamic nature of data\-driven reasoning and multi\-dimensional drill\-downs in real\-world scenarios\.While the paradigm of end\-to\-end autonomous agents has demonstrated remarkable success in specialized environments such as web search\(Jinet al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib4); Zhenget al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib5); Moonshot AI,[2025](https://arxiv.org/html/2605.07202#bib.bib6)\)and open\-ended gaming\(Wanget al\.,[2023a](https://arxiv.org/html/2605.07202#bib.bib16)\), its potential remains untapped in read\-world business analysis tasks\. A primary obstacle is the syntactic volatility of SQL, characterized by redundant and diverse query structures, and its semantic fragmentation, arising from inconsistent metric definitions across disparate systems, often obfuscate autonomous reasoning in enterprise landscapes\.

To bridge this gap, in this paper, we propose AIDA, the first end\-to\-end agent framework designed to explore in complex business environments\. Our primary contributions are summarized as follows:

- •We established a highly flexible environment for instant retail, integrating 200\+ metrics and 100\+ dimensions\. Coupled with a proprietary DSL, this environment affords agents an expansive arena for near unconstrained exploration\.
- •We propose a data\-to\-insight framework for real\-world business analysis tasks that systemically orchestrates environment setup, state modeling, trajectory synthesis, and reinforcement learning\.
- •We introduce a suite of specialized enhancements within AIDA framework: a robust reward mechanism for targeted insight discovery, and two masking strategies to mitigate reward hacking and stabilize policy updates\.

## 2Related Work

### 2\.1Data to Insights

The core objective of business analysis is to transform massive, complex, and multi\-source data into perceptible insights and actionable decisions\(Zikopoulos and Eaton,[2011](https://arxiv.org/html/2605.07202#bib.bib24); Chenet al\.,[2012](https://arxiv.org/html/2605.07202#bib.bib23)\)\. The LLM era has catalyzed diverse strategies for translating structured data into insights\. These methods increasingly focus on navigating complex environments, ranging from traditional databases and tabular files to sophisticated data science toolkits\. Representative systems include AgentPoirot\(Sahuet al\.,[2024](https://arxiv.org/html/2605.07202#bib.bib12)\), which extracts insights from business CSV files, as well as Data Interpreter\(Honget al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib14)\)and DS\-Agent\(Guoet al\.,[2024](https://arxiv.org/html/2605.07202#bib.bib15)\)for solving specific data science tasks\. ReActInsight\(Wenget al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib13)\)further extended data sources to multi\-source heterogeneous data, emphasizing cross\-source insight insight through correlation analysis\.

![Refer to caption](https://arxiv.org/html/2605.07202v1/framework.png)Figure 2:The overall architecture of the proposed AIDA framework\. The pipeline consists of four integrated stages: \(1\)Environment Setup, which establishes a dual\-tool execution layer: a data retrieval tool interacting with data environment via DSL, and a python tool for executing code within a secure sandbox; \(2\)State Modeling, which formalizes the task state as a quintuple consisting of identifier metadata, the target query, analytical insights, observation logs, and a structured reasoning graph; \(3\)Trajectory Synthesis, which generates high\-quality trajectories through an iterative plan\-action\-state loop; and \(4\)Reinforcement Learning, where agent rollouts are evaluated by reward functions\. The framework employs a global batch return mechanism to calculate the mean across all samples for advantage computation, followed by a mask strategy to selectively optimize the model’s policy\.
### 2\.2Open\-Ended Agents

Agents in open\-ended environments are characterized by high uncertainty and vast state spaces\. This includes deep search agents\(Jinet al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib4); Zhenget al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib5); Moonshot AI,[2025](https://arxiv.org/html/2605.07202#bib.bib6); Huet al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib9); Teamet al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib10); Qiaoet al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib11)\)that navigate expansive, noisy web environments to conduct multi\-step searches and information integration, such as WebResearcher\(Qiaoet al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib11)\)which models insight as a Markov Decision Process \(MDP\)\. Another subset operates in simulated or gaming environments\(Wanget al\.,[2023a](https://arxiv.org/html/2605.07202#bib.bib16); Zhuet al\.,[2023](https://arxiv.org/html/2605.07202#bib.bib17); Wanget al\.,[2023b](https://arxiv.org/html/2605.07202#bib.bib18)\), exemplified by VOYAGER\(Wanget al\.,[2023a](https://arxiv.org/html/2605.07202#bib.bib16)\), which explores item synthesis in Minecraft through a lifelong learning mechanism\. These agents pursue open\-ended goals within environments offering near\-infinite exploration depth, eliciting human\-like insight behaviors\.

## 3Real\-world Business Analysis Tasks

In instant retail domain, business analysis tasks are defined as macro\-level strategic inquiries that demand deep extensive data exploration and reasoning\. In practice, this reasoning process is a dynamic journey of hypothesis testing, which may require hours or even days of analyst effort to resolve\. As shown in Figure[1](https://arxiv.org/html/2605.07202#S1.F1), an expert analyst interrogate the data by cross\-examining various metrics and dimensions to create analytical branches\. The analyst’s objective is to transform these open\-ended questions into accumulated insights\. However, such tasks are often over\-determined by factually grounded but trivial insights\. For instance, in attributing a drop in business district exposure, multiple factors— ranging from weather fluctuations to a competitor’s promotion—may all be factually valid\. The analyst’s goal is not to list every data\-supported fact, but to isolate the high\-leverage ones that drive the most significant impact\.

To bridge this gap, we define the task as identifying the optimal insight set under the Pareto Principle\. The Pareto Principle states:

> For many outcomes, roughly 80% of consequences come from 20% of causes\.

This principle steers the agent to prune secondary observations and focus on the most critical variables\. Under this principle, success is defined as the ability to extract the most impactful insights from a vast search space of valid but marginal information\.

## 4AIDA Framework

Our proposed framework is shown in Figure[2](https://arxiv.org/html/2605.07202#S2.F2), which is composed of environment setup \(Section[4\.1](https://arxiv.org/html/2605.07202#S4.SS1)\), state modeling \(Section[4\.2](https://arxiv.org/html/2605.07202#S4.SS2)\), state transition\(Section[4\.3](https://arxiv.org/html/2605.07202#S4.SS3)\), trajectory synthesis \(Section[4\.4](https://arxiv.org/html/2605.07202#S4.SS4)\) and reinforcement learning \(Section[4\.5](https://arxiv.org/html/2605.07202#S4.SS5)and Section[4\.6](https://arxiv.org/html/2605.07202#S4.SS6)\)\.

### 4\.1Environment Setup

The AIDA framework is built upon a specialized industrial environment, which is the fundation of the autonomous exploration\. This infrastructure is engineered to support two types of actions:Dsl2dataandPython execution\. Specifically,Dsl2dataallows the agent to perform flexible data queries through a custom DSL, enabling the arbitrary combination of metrics and dimensions for deep exploration\.Python executionenables the agent to perform complex computations by running Python scripts\.

To support the free combination and drill\-down of metrics and dimensions, we reshaped traditional data development into an architecture of metrics and dimensions\. This decoupled system enables the agent to execute unconstrained, on\-the\-fly queries \(see Appendix[A](https://arxiv.org/html/2605.07202#A1)\)\. Built upon this foundation, we implemented a standardized DSL protocol that allows the agent to programmatically select metrics, dimensions, and filtering criteria for on\-the\-fly queries\. \(see Appendix[B](https://arxiv.org/html/2605.07202#A2)\)\. ThePython executionaction is supported by a secure, isolated sandbox environment\. WhileDsl2dataprovides raw data evidence, this infrastructure enables the agent to perform complex computation and hypothesis testing through arbitrary code\.

### 4\.2State Modeling

We reformulate real\-world business analysis as a Markov Decision Process as shown in Figure[3](https://arxiv.org/html/2605.07202#S4.F3)\.

![Refer to caption](https://arxiv.org/html/2605.07202v1/State_v2.png)Figure 3:Detailed illustration of the statests\_\{t\}in a real\-world business analysis task\.The Id and Question provide the original background information and the target question of the analysis\. The Insight and Observation Log maintain a repository of insights and findings, specifically highlighting theProoffield for real\-world data support and theStructuremetadata for data traceability\. The Reasoning Graph offers a global view of the agent’s logical trajectory, ensuring the analysis remains coherent throughout the exploration\.

Formally, at each time steptt,the statests\_\{t\}is defined as a quintuple:

st=⟨i​d,q,ct,dt,gt⟩s\_\{t\}=\\langle id,q,c\_\{t\},d\_\{t\},g\_\{t\}\\rangle\(1\)
where:

i​didrepresents acomposite set of identifiers\. For instance, we have ids of shop, major category and business district for merchant\-level analysis, and ids of brand, major category and cities for brand\-level analysis\.

qqrepresents thetarget question, which serves as the objective anchor for the entire analysis process\.

ctc\_\{t\}is the set ofinsights, which capture global insights that align with the Pareto Principle\. To track how the agent’s understanding evolves, each entry is assigned one of four statuses:

- •New: A newly discovered insight in the current step\.
- •Unchanged: A previous insight without any changes\.
- •Reinforced: An existing insight that is now supported by more evidence\.
- •Refuted: A previous insight that is corrected because it contradicts new data\.

Each insight also includes atitlefor the core insight and aprooffield recording the real\-world data support\.

dtd\_\{t\}represents theobservation log, providing a structured index of environmental returns\. Each entry includes the datatype\(e\.g\., CSV, TXT\), a semanticdescription, and astructuremetadata field that explicitly records the queried metrics, dimensions, and filter conditions\.

gtg\_\{t\}is thereasoning graphserialized in mermaid format\. It maps the topological structure of the reasoning path, recording hypothesis testing and logical branches\.

Through this state modeling, the AIDA framework effectively decouples the agent’s cognition from the noisy dialogue history\. This modeling equips the agent with a robust long\-term memory to retain critical insights over extended analysis cycles, while providing a native self\-correction mechanism to refine or refute insights as new evidence emerges\. This ensures the analytical process remains both logically persistent and adaptive to complex real\-world data\.

### 4\.3State Transition

![Refer to caption](https://arxiv.org/html/2605.07202v1/MDP_v2.png)Figure 4:Overview of the Interactive Reasoning and State Transition process\. At each stepTtT\_\{t\}, the model performs reasoning to update the structured state and interact with the environment\.The transition from statests\_\{t\}tost\+1s\_\{t\+1\}is driven by a structured reasoning\-action cycle\. As illustrated in Figure[4](https://arxiv.org/html/2605.07202#S4.F4), at each time steptt, agent receives an input consisting of the current statests\_\{t\}, the previous action, and the latest observation\. The agent output is defined as following modules:

State Think & Update\. Agent first conducts reasoning to analyze the latest observation\. It then generates a structured text block that serves as the state update, which specifies the necessary updates for insights \(ctc\_\{t\}\), observation logs \(dtd\_\{t\}\), and the reasoning graph \(gtg\_\{t\}\)\. Then these updates are merged into the previous state\.

Action Think & Action\. Based on the newly updated state, the agent conducts reasoning and determine the next actionata\_\{t\}, which is then executed within the environment to form the next round of observationoto\_\{t\}\.

In practical deployment, the agent explores up to a user\-defined maximum step, after which a post\-processing module transforms the final state into a user\-friendly report\.

### 4\.4Trajectory Synthesis

In AIDA’s atomistic data environment, the policy model faces complex output formats and a massive search space formed by the combinations of metrics and dimensions\. Without an effective cold\-start process, the agent often performs inefficient exploration in this vast space\.

To build a solid foundation for training, we first prepare the necessary analytical components\. We filter a sets of entity identities for the cold\-start phase\. The initial questions are created using two strategies: \(1\) template\-based sampling from a set of common business problems, and \(2\) data\-driven generation based on statistical patterns found in randomly sampled metric\-dimension combinations\.

With these ids and questions ready, we use a dual\-agent strategy to synthesize high\-quality expert trajectories\. At the beginning of the trajectory and at specific rounds, an auxiliary agent is called to generate a global plan\. The main agent then follows this plan to generate its outputs according to the structured format defined in Section[4\.3](https://arxiv.org/html/2605.07202#S4.SS3)\.

The collected dataset is denoted as𝒟\\mathcal\{D\}\. Subsequently, we perform filtering process on𝒟\\mathcal\{D\}to construct a high\-quality subset𝒟s\\mathcal\{D\}\_\{s\}\. Specifically, we eliminate samples containing numerical hallucinations or erroneous insights, and equilibrate the proportions of different insight statuses\. Following such procedure, a separate set of ids was utilized to derive dataset𝒟r​l\\mathcal\{D\}\_\{rl\}for RL training\.

### 4\.5Reward Function

Distinct from RL for LLM\(Ouyanget al\.,[2022](https://arxiv.org/html/2605.07202#bib.bib27); Guoet al\.,[2025](https://arxiv.org/html/2605.07202#bib.bib26)\), which typically prioritizes single\-turn alignment, reinforcement learning for LLM agents entails a multi\-turn process, where the primary challenge is credit assignment\. Inspired byWeiet al\.\([2025](https://arxiv.org/html/2605.07202#bib.bib29)\), we propose a reward decomposition mechanism that decouples reward signals into two distinct dimensions: intermediate rewards and accumulated rewards\.

- •Intermediate Reward:RtIR\_\{t\}^\{I\}is designed to evaluate the agent’s immediate output quality and does not account for the long\-term effects of state transitions\.
- •Accumulated Reward:RtAR\_\{t\}^\{A\}focuses on macro\-level progress with long\-term objectives\. It measures the contribution of a step to the subsequent trajectory and the final result\.

Based on this decoupling mechanism, we define the returnGtG\_\{t\}at stepttas the weighted integration of immediate reward and discounted accumulated rewards:

Gt=RtI\+∑j=tTγj−t​RjAG\_\{t\}=R\_\{t\}^\{I\}\+\\sum\_\{j=t\}^\{T\}\\gamma^\{j\-t\}R\_\{j\}^\{A\}\(2\)whereγ\\gammadenotes the discount factor\.

Specifically, we design specific reward functions for each component of the structured output introduced in Section[4\.3](https://arxiv.org/html/2605.07202#S4.SS3), as detailed in Table[1](https://arxiv.org/html/2605.07202#S4.T1)\. Note that both immediate and cumulative rewards are the sum of their respective reward terms at each step\. We now detail the core reward functions:

Table 1:Detailed Definitions and Ranges of the AIDA Reward Function Components\.ComponentReward ItemTypeDescriptionRangeGlobalStep FormatIntermediateAdherence to the global output paradigm\.\[−1,0\]\[\-1,0\]<state\_think\>Hallucination PenaltyIntermediatePenalty for numerical values not grounded in observations\.\[−1,0\.1\]\[\-1,0\.1\]Length RewardAccumulatedEncourages reasoning stability and duration efficiency\.\[0,0\.1\]\[0,0\.1\]<insight\>Schema RewardIntermediateSyntax matching for structured key sub\-insights\.\[−1,0\]\[\-1,0\]Discovery GainAccumulatedIncremental insight gains\.\-<key\_data\>Schema RewardIntermediateSyntax matching for the key data data\-structure\.\[−1,0\]\[\-1,0\]<graph\>Mermaid RenderIntermediateSuccessful rendering of the Mermaid reasoning topology\.\[−1,0\]\[\-1,0\]<action\_think\>Hallucination PenaltyIntermediatePenalty for numerical values not grounded in observations\.\[−1,0\.1\]\[\-1,0\.1\]Length RewardAccumulatedStability constraint for action planning reasoning\.\[0,0\.1\]\[0,0\.1\]<tool\_call\>JSON SchemaIntermediateStructural validity and parameter completeness of tool calls\.\[−1,0\]\[\-1,0\]Python ExecutionIntermediateSuccessful runtime execution within the sandbox environment\.\[−1,0\]\[\-1,0\]Numerical Hallucination Penalty\. The function is aim to ensure the factuality of data support in agent’s insights\. For any numerical value appearing in the agent’s output, the system verifies its presence in the historical environmental observation\. Letnnbe the count of hallucinated values andmmbe the count of grounded values\. The penalty is calculated as:

Rh​a​l​l​u=clip​\(0\.01​m−0\.1​n,−1\.0,0\.1\)R\_\{hallu\}=\\text\{clip\}\\left\(0\.01m\-0\.1n,\-1\.0,0\.1\\right\)\(3\)
Think Length Reward\. Since heavy hallucination penalties may cause a sudden drop in reasoning length and lead to training collapse, we introduce a reward to keep the output length within a reasonable range\. Given an output lengthLLand predefined bounds\[Lm​i​n,Lm​a​x\]\[L\_\{min\},L\_\{max\}\], the reward is:

Rl​e​n​g​t​h=clip​\(L−Lm​i​nLm​a​x−Lm​i​n,0,1\)R\_\{length\}=\\text\{clip\}\\left\(\\frac\{L\-L\_\{min\}\}\{L\_\{max\}\-L\_\{min\}\},0,1\\right\)\(4\)
Discovery Gain Reward\. This function treats the analysis process as a series of quantifiable insights as shown in Algorithm[1](https://arxiv.org/html/2605.07202#alg1)\. It quantifies the information gains generated whenever the agent emerges a new insight, verifies a hypothesis or corrects a previous error using real\-world data\.

To determine the alignment of these insights with the Pareto Principle, an llm\-as\-a\-judge mechanism is employed: insights that effectively summarize critical variables are marked as valid, while those invalid insights incur a severe penalty of−2\.0\-2\.0\. For each valid insightcic\_\{i\}, a base rewardB​\(ci\)B\(c\_\{i\}\)is assigned according to its incremental status:

B​\(ci\)=\{1\.0if status = “New”0\.7if status = “Refuted”0\.5if status = “Reinforced”B\(c\_\{i\}\)=\\begin\{cases\}1\.0&\\text\{if status = \`\`New''\}\\\\ 0\.7&\\text\{if status = \`\`Refuted''\}\\\\ 0\.5&\\text\{if status = \`\`Reinforced''\}\\end\{cases\}\(5\)
Algorithm 1Discovery Gain Reward CalculationInput:Incremental insights

𝒟n​e​w\\mathcal\{D\}\_\{new\}, judge model

ℳ\\mathcal\{M\}, previous state

𝒮\\mathcal\{S\}
Output:Total reward score

Rg​a​i​nR\_\{gain\}
Initialize

Rg​a​i​n=0R\_\{gain\}=0
foreach insight

c∈Dn​e​wc\\in D\_\{new\}do

Query

ℳ\\mathcal\{M\}to judge validity of

ccgiven

𝒮\\mathcal\{S\}
if

ℳ\\mathcal\{M\}returns “Valid”then

B=BaseReward\(d\.status\)B=\\text\{BaseReward\}\(d\.\\text\{status\}\)

Detect hallucination set

HHfrom

ccbased on

𝒮\\mathcal\{S\}
Rg​a​i​n=Rg​a​i​n\+max⁡\(B−0\.4×\|H\|,0\.1\)R\_\{gain\}=R\_\{gain\}\+\\max\(B\-0\.4\\times\|H\|,0\.1\)

else

Rg​a​i​n=Rg​a​i​n−2\.0R\_\{gain\}=R\_\{gain\}\-2\.0

endif

Cp​r​e​v=Cp​r​e​v∪\{c\}C\_\{prev\}=C\_\{prev\}\\cup\\\{c\\\}

endfor

Return

Rg​a​i​nR\_\{gain\}

Furthermore, factual grounding is prioritized alongside logical validity to prevent hallucinations\. A penalty is applied to any insight that includes ungrounded numerical data, even if it is logically sound\. LetH​\(ci\)H\(c\_\{i\}\)be the number of hallucinated values incic\_\{i\}\. The final gain for a single insight is computed as:

Rg​a​i​n​\(ci\)=\{max⁡\(B​\(ci\)−η​H​\(ci\),0\.1\)if valid−2\.0otherwiseR\_\{gain\}\(c\_\{i\}\)=\\begin\{cases\}\\max\\left\(B\(c\_\{i\}\)\-\\eta H\(c\_\{i\}\),0\.1\\right\)&\\text\{if valid\}\\\\ \-2\.0&\\text\{otherwise\}\\\\ \\end\{cases\}\(6\)whereη\\etais the penalty coefficient\.

In practice,Rg​a​i​nR\_\{gain\}functions as a behavioral guide for the agent by balancing exploration with rigor\. It rewards the agent for active insight under Pareto Principle and, more importantly, for the behavior to self\-correct by refuting earlier errors\. Conversely, it suppresses the tendency to hallucinate by penalizing ungrounded numerical data and logical contradictions\.

![Refer to caption](https://arxiv.org/html/2605.07202v1/x1.png)Figure 5:Main experimental results\. The plots compare the scores ofAIDA\-RL\-8B,AIDA\-SFT\-8B,State\-React\-Qwen3\-32BandReact\-Qwen3\-32Bover 50 exploration steps\. AIDA\-RL\-8B consistently achieves superior performance in the cumulative Score \(top\-left\) and across all constituent metrics\. which demonstrating its effectiveness in real\-world business analysis tasks\.
### 4\.6Objective Function and Optimization

FollowingLiuet al\.\([2025](https://arxiv.org/html/2605.07202#bib.bib25)\), we adopt REINFORCE with return batch normalization as objective function, which constructs a dynamic baseline by standardizing returns across the current batch:

JReBN​\(θ\)=1N​∑n=1N∑t=0T−1AReBN,t\(n\)​log⁡πθ​\(at\(n\)∣st\(n\)\)J\_\{\\text\{ReBN\}\}\(\\theta\)=\\frac\{1\}\{N\}\\sum\_\{n=1\}^\{N\}\\sum\_\{t=0\}^\{T\-1\}A\_\{\\text\{ReBN\},t\}^\{\(n\)\}\\log\\pi\_\{\\theta\}\\left\(a\_\{t\}^\{\(n\)\}\\mid s\_\{t\}^\{\(n\)\}\\right\)\(7\)where the advantage termAReBN,t\(n\)A\_\{\\text\{ReBN\},t\}^\{\(n\)\}is computed as:

AReBN,t\(n\)=Gt\(n\)−𝔼​\[G\]std​\(G\),G=\{Gt\(n\)\}n∈N,t∈TA\_\{\\text\{ReBN\},t\}^\{\(n\)\}=\\frac\{G\_\{t\}^\{\(n\)\}\-\\mathbb\{E\}\[G\]\}\{\\text\{std\}\(G\)\},\\quad G=\\left\\\{G\_\{t\}^\{\(n\)\}\\right\\\}\_\{n\\in N,t\\in T\}\(8\)
Furthermore, we employ two masking strategies to address reward hacking\.

Schema Masking\. If a step yields a positive advantage but fails any immediate syntax\-related reward \(e\.g\., format or JSON schema\), the gradient is masked\.

Logical Consistency Masking\. If a step yields a positive advantage but the step itself is judged to contain an invalid sub\-insight, the gradient is suppressed\.

## 5Expirement

### 5\.1Training Setup

We adopt Qwen3\-8B as our backbone model\. The training pipeline consists of a cold\-start stage using supervised fine\-tuning on the dataset𝒟s\\mathcal\{D\}\_\{s\}with a learning rate of1×10−51\\times 10^\{\-5\}, followed by a reinforcement learning phase on the dataset𝒟r​l\\mathcal\{D\}\_\{rl\}with a learning rate of5×10−75\\times 10^\{\-7\}\.

For RL training, we set\[Lmin,Lmax\]\[L\_\{\\min\},L\_\{\\max\}\]to\[500,1000\]\[500,1000\]for state thinking and\[500,700\]\[500,700\]for action thinking\. The hyperparameters are configured asγ=0\.7\\gamma=0\.7andη=0\.4\\eta=0\.4, with each trajectory constrained to a maximum of 10 exploration steps\.

### 5\.2Evaluation Setup

To evaluate the insight capabilities of agents in real\-world business analysis, we developed a comprehensive benchmark\.

Evaluation Dataset\.We curated 82 test cases derived from real\-world instant retail business scenarios, categorized into five analytical domains: Traffic, focusing on funnel metrics \(e\.g\., clicks, conversion\) for reach efficiency and attrition diagnosis; Transaction, assessing core KPIs \(e\.g\., GMV, AOV\) for bottleneck identification and operational risks; Interaction, examining unstructured feedback \(e\.g\., reviews\) for pain point mining; Marketing, evaluating promotions and subsidies; and General, testing cross\-domain anomalies for global reasoning and root\-cause localization under ambiguity\.

Metrics\.We quantify the agent’s exploratory performance through a cumulative score derived from the set of insights𝒞t\\mathcal\{C\}\_\{t\}generated overttexploration steps\. The score is defined as:

s​\(ci\)=𝕀valid​\(ci\)​max⁡\(0,1−α​H​\(ci\)\)−𝕀invalid​\(ci\)Score​\(t\)=∑ci∈𝒞ts​\(ci\)\\begin\{gathered\}s\(c\_\{i\}\)=\\mathbb\{I\}\_\{\\mathrm\{valid\}\}\(c\_\{i\}\)\\max\(0,1\-\\alpha H\(c\_\{i\}\)\)\-\\mathbb\{I\}\_\{\\mathrm\{invalid\}\}\(c\_\{i\}\)\\\\ \\mathrm\{Score\}\(t\)=\\sum\_\{c\_\{i\}\\in\\mathcal\{C\}\_\{t\}\}s\(c\_\{i\}\)\\end\{gathered\}\(9\)
where𝕀valid​\(ci\),𝕀invalid​\(ci\)∈\{0,1\}\\mathbb\{I\}\_\{\\mathrm\{valid\}\}\(c\_\{i\}\),\\mathbb\{I\}\_\{\\mathrm\{invalid\}\}\(c\_\{i\}\)\\in\\\{0,1\\\}indicate whether insightcic\_\{i\}is valid or invalid, respectively\. Specifically, we apply a hallucination penaltyα=0\.5\\alpha=0\.5for any numerically hallucinations, ensuring the rigor of the discovered insights\.

Compared Methods\.We compare the following agents:

1. \(1\)ReAct\(Yaoet al\.,[2023](https://arxiv.org/html/2605.07202#bib.bib22)\): The standard reasoning and acting framework based on dialogue history\. This baseline uses Qwen3\-32B as the backbone with its context extended to 64K tokens, incorporating all available metrics, dimensions, and interaction protocols within its context\.
2. \(2\)State\-ReAct: Built upon the same backbone as the ReAct baseline with its native context length, this agent incorporates the state paradigm introduced in Section[4\.2](https://arxiv.org/html/2605.07202#S4.SS2)to replace raw historical dialogues with structured state representations\.
3. \(3\)AIDA\-SFT: Initialized from Qwen3\-8B, this model is fine\-tuned exclusively on the cold\-start dataset𝒟s\\mathcal\{D\}\_\{s\}\.
4. \(4\)AIDA\-RL: Our full model that utilizes the model after cold\-start SFT and continues training on𝒟r​l\\mathcal\{D\}\_\{rl\}using the RL framework\.

### 5\.3Results

Figure[5](https://arxiv.org/html/2605.07202#S4.F5)illustrates the performance evolution of different agents\. The experimental results lead to the following key observations:

Efficacy of State Modeling\. State\-ReAct consistently outperforms native ReAct\. While ReAct plateaus within 25 steps due to long\-context interference, State\-ReAct leverages state modeling to maintain precise task tracking\. This effectively curbs reasoning drift, allowing scores to climb steadily toward around 5 and demonstrating that state modeling is essential for mastering complex, long\-horizon tasks\.

Superiority of RL over SFT\. Experimental results demonstrate that AIDA\-RL significantly broadens the lead over AIDA\-SFT as the task horizon extends\. While all agents show an initial positive correlation between steps and scores, AIDA\-RL exhibits a much steeper curve\. Notably, AIDA\-RL\-8B consistently surpasses the 32B State\-ReAct and ReAct workflows after approximately 15 steps, which underscores that the RL\-optimized policy empowers the agent to make more strategic decisions, enabling it to unearth a significantly higher volume of insights compared to other baselines\.

Robustness and Error Suppression\. A key strength of AIDA\-RL lies in its rigor\. As shown in Figure[5](https://arxiv.org/html/2605.07202#S4.F5)\(bottom row\), while the number of invalid insights and hallucination for ReAct and State\-ReAct surge with exploration steps, AIDA\-RL maintains them at a remarkably low and stable level\. S pecifically, AIDA\-RL\-8B reduces hallucinations by about 70% compared to ReAct\-32B at the 50\-step mark, proving that the reward successfully enforces logical groundedness\.

Scalability and Convergence\. Unlike ReAct, which converges prematurely, the performance curves for AIDA\-RL and AIDA\-SFT show no signs of plateauing within 50 steps\. This suggests high scalability; with extended computational budgets, further gains in insight discovery are anticipated\. The consistent lead of AIDA\-RL across all metrics confirms its reliability as a robust strategy for real\-world business analysis tasks\.

## 6Analysis

### 6\.1Ablation Study

This ablation study evaluates the impact of two masking strategies on the model’s training stability and performance\. We compare the full configuration against variants where specific masking strategy is removed\. The results, illustrated in Figure[6](https://arxiv.org/html/2605.07202#S6.F6), highlight how these strategies prevent reward hacking and ensure the quality of insight\.

The most significant performance drop occurs when logical consistency masking is removed\. Without this mask, the model fails to improve its insight capabilities, with rewards stagnating and eventually declining after 30 training steps\. This suggests that failing to filter out invalid insights, even those that yield positive advantages, significantly heightens the likelihood of generating erroneous insights\.

The removal of schema masking reveals a classic case of reward hacking\. Initially, this variant shows a faster increase in discovery gain reward compared to the main experiment\. However, a catastrophic collapse is observed after training step 50\. As long as the agent can secure valid insights, schema\-related penalties offer insufficient deterrents, leading the model to prioritize discovery over structural integrity\. This suggests that this strategy is pivotal for maintaining adherence to the formal grammar\.

The results demonstrate that these masking strategies are complementary: logical consistency masking ensures the validity of the insight logic, while schema masking enforces the structure of the output\. Together, they create a robust training curriculum that prevents the model from exploiting reward shortcuts and ensures the generation of high\-quality, syntactically correct, and logically sound insights\.

![Refer to caption](https://arxiv.org/html/2605.07202v1/x2.png)Figure 6:Ablation of masking strategies\. The top panel shows discovery gain rewards, and the bottom panel shows step format rewards\. The three masking strategies are individually removed for comparison with AIDA\-RL\-8B\.
### 6\.2Exploration Study

To evaluate whether the training process has enhanced the model’s analytical depth, we conduct a comprehensive study of its exploration space within the instant retail landscape\. Given the combinatorial complexity of AIDA’s data environment, the agent must transcend simple metric selection to autonomously isolate the specific dimensions that catalyze deep insight insight\. To evaluate this diversity, we aggregate approximately 200 metrics into five high\-level thematic clusters\. Figure[7](https://arxiv.org/html/2605.07202#S6.F7)depicts the average number of dimensions pairing within each cluster\.

The results demonstrate that AIDA\-RL significantly expands the exploration boundary compared to AIDA\-SFT\. For instance, in the Merchant and Interaction types, AIDA\-RL covers 1 or 2 more analytical dimensions\. This expansion indicates that through reinforcement learning, the agent has transcended simple pattern mimicking and learned to synthesize arbitrary queries via DSL to investigate business issues from more granular and diverse angles\.

Unlike conventional workflows that often converge on a single line of inquiry, AIDA incentivize the insight of long\-tail but critical ones\. By autonomously pairing metrics with a wider array of dimensions, AIDA demonstrates a profound potential to navigate complex, real\-world industrial data spaces\.

![Refer to caption](https://arxiv.org/html/2605.07202v1/x3.png)Figure 7:Comparison of data space exploration breadth\. The radar chart illustrates the number of dimensions associated with each key metric types of different agents\.
### 6\.3Environmental Boundary Analysis

To evaluate whether the model achieves better coupling with the environment, we introduce the concept of environmental boundary analysis\. We conceptualize this data environment as an expansive, interactive map\. While such data map allows for near\-unconstrained exploration, the map has physical boundaries defined by the underlying data architecture rules\. An agent may hit a wall through boundary violations, which occur in two primary forms: Attempting to access metrics or dimensions that do not exist within the defined set; Requesting combinations of existing metrics and dimensions but which are logically or physically incompatible due to data grain or business definitions\.

Figure[8](https://arxiv.org/html/2605.07202#S6.F8)illustrates the cumulative count of these violations during 50\-step trajectories\. The results yield the following insights: AIDA\-RL demonstrates a significant reduction in boundary violations compared to AIDA\-SFT and other baselines\. Crucially, this improvement is an emergent strategic behavior\. Since the RL objective prioritizes the discovery of valid, high\-value insights, the agent spontaneously learns to steer away from data that yield no discovery gain\. Consequently, AIDA\-RL maintains a consistently low violation rate even in late\-stage reasoning steps, whereas baselines often drift toward the environment’s boundaries\. This suggests that our RL framework equips the agent with long\-horizon environmental awareness, allowing it to select valid metric\-dimension combinations even under deep exploration\. Ultimately, this indicates that the RL process enables the model to internalize the implicit boundaries of business logic, achieving a sophisticated balance between aggressive exploration and logical feasibility\.

![Refer to caption](https://arxiv.org/html/2605.07202v1/x4.png)Figure 8:Environmental boundary analysis\. The horizontal axis represents the number of exploration steps, and the vertical axis represents the average cumulative number of boundary violations\.

## 7Conclusion

We propose AIDA, an autonomous framework for data\-to\-insight discovery within a complex retail environment \(200\+ metrics\) using a proprietary DSL\. By integrating reinforcement learning with specialized masking and reward strategies, AIDA enables stable, high\-depth exploration that significantly outperforms existing agents, establishing a promising pathway towards truly autonomous business intelligence in industrial\-scale systems\.

## References

- H\. Chen, R\. H\. Chiang, and V\. C\. Storey \(2012\)Business intelligence and analytics: from big data to big impact\.MIS quarterly,pp\. 1165–1188\.Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p1.1),[§2\.1](https://arxiv.org/html/2605.07202#S2.SS1.p1.1)\.
- Google \(2025\)Gemini 3 pro\.External Links:[Link](https://blog.google/innovation-and-ai/technology/developers-tools/gemini-3-pro-vision/)Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p1.1)\.
- D\. Guo, D\. Yang, H\. Zhang, J\. Song, P\. Wang, Q\. Zhu, R\. Xu, R\. Zhang, S\. Ma, X\. Bi,et al\.\(2025\)DeepSeek\-r1 incentivizes reasoning in llms through reinforcement learning\.Nature645\(8081\),pp\. 633–638\.Cited by:[§4\.5](https://arxiv.org/html/2605.07202#S4.SS5.p1.4)\.
- S\. Guo, C\. Deng, Y\. Wen, H\. Chen, Y\. Chang, and J\. Wang \(2024\)DS\-agent: automated data science by empowering large language models with case\-based reasoning\.InProceedings of the 41st International Conference on Machine Learning,pp\. 16813–16848\.Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p1.1),[§2\.1](https://arxiv.org/html/2605.07202#S2.SS1.p1.1)\.
- S\. Hong, Y\. Lin, B\. Liu, B\. Liu, B\. Wu, C\. Zhang, D\. Li, J\. Chen, J\. Zhang, J\. Wang,et al\.\(2025\)Data interpreter: an llm agent for data science\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 19796–19821\.Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p1.1),[§2\.1](https://arxiv.org/html/2605.07202#S2.SS1.p1.1)\.
- C\. Hu, H\. Du, H\. Wang, L\. Lin, M\. Chen, P\. Liu, R\. Miao, T\. Yue, W\. You, W\. Ji,et al\.\(2025\)Step\-deepresearch technical report\.arXiv preprint arXiv:2512\.20491\.Cited by:[§2\.2](https://arxiv.org/html/2605.07202#S2.SS2.p1.1)\.
- B\. Jin, H\. Zeng, Z\. Yue, J\. Yoon, S\. Arik, D\. Wang, H\. Zamani, and J\. Han \(2025\)Search\-r1: training llms to reason and leverage search engines with reinforcement learning\.arXiv preprint arXiv:2503\.09516\.Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p2.1),[§2\.2](https://arxiv.org/html/2605.07202#S2.SS2.p1.1)\.
- P\. Langley \(2000\)Crafting papers on machine learning\.InProceedings of the 17th International Conference on Machine Learning \(ICML 2000\),P\. Langley \(Ed\.\),Stanford, CA,pp\. 1207–1216\.Cited by:[Appendix C](https://arxiv.org/html/2605.07202#A3.p1.1)\.
- Z\. Liu, A\. Sims, K\. Duan, C\. Chen, S\. Yu, X\. Zhou, H\. Xu, S\. Xiong, B\. Liu, C\. Tan,et al\.\(2025\)Gem: a gym for agentic llms\.arXiv preprint arXiv:2510\.01051\.Cited by:[§4\.6](https://arxiv.org/html/2605.07202#S4.SS6.p1.2)\.
- Moonshot AI \(2025\)Kimi\-researcher: end\-to\-end rl training for emerging agentic capabilities\.External Links:[Link](https://moonshotai.github.io/Kimi-Researcher/)Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p2.1),[§2\.2](https://arxiv.org/html/2605.07202#S2.SS2.p1.1)\.
- OpenAI \(2025\)Introducing openai o3 and o4\-mini\.External Links:[Link](https://openai.com/index/introducing-o3-and-o4-mini/)Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p1.1)\.
- L\. Ouyang, J\. Wu, X\. Jiang, D\. Almeida, C\. Wainwright, P\. Mishkin, C\. Zhang, S\. Agarwal, K\. Slama, A\. Ray,et al\.\(2022\)Training language models to follow instructions with human feedback\.Advances in neural information processing systems35,pp\. 27730–27744\.Cited by:[§4\.5](https://arxiv.org/html/2605.07202#S4.SS5.p1.4)\.
- A\. S\. Pérez, A\. Boukhary, P\. Papotti, L\. C\. Lozano, and A\. Elwood \(2025\)An llm\-based approach for insight generation in data analysis\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),pp\. 562–582\.Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p1.1)\.
- Z\. Qiao, G\. Chen, X\. Chen, D\. Yu, W\. Yin, X\. Wang, Z\. Zhang, B\. Li, H\. Yin, K\. Li,et al\.\(2025\)Webresearcher: unleashing unbounded reasoning capability in long\-horizon agents\.arXiv preprint arXiv:2509\.13309\.Cited by:[§2\.2](https://arxiv.org/html/2605.07202#S2.SS2.p1.1)\.
- G\. Sahu, A\. Puri, J\. Rodriguez, A\. Abaskohi, M\. Chegini, A\. Drouin, P\. Taslakian, V\. Zantedeschi, A\. Lacoste, D\. Vazquez,et al\.\(2024\)Insightbench: evaluating business analytics agents through multi\-step insight generation\.arXiv preprint arXiv:2407\.06423\.Cited by:[§2\.1](https://arxiv.org/html/2605.07202#S2.SS1.p1.1)\.
- T\. D\. Team, B\. Li, B\. Zhang, D\. Zhang, F\. Huang, G\. Li, G\. Chen, H\. Yin, J\. Wu, J\. Zhou,et al\.\(2025\)Tongyi deepresearch technical report\.arXiv preprint arXiv:2510\.24701\.Cited by:[§2\.2](https://arxiv.org/html/2605.07202#S2.SS2.p1.1)\.
- A\. Vera\-Baquero, R\. Colomo Palacios, V\. Stantchev, and O\. Molloy \(2015\)Leveraging big\-data for business process analytics\.The Learning Organization22\(4\),pp\. 215–228\.Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p1.1)\.
- G\. Wang, Y\. Xie, Y\. Jiang, A\. Mandlekar, C\. Xiao, Y\. Zhu, L\. Fan, and A\. Anandkumar \(2023a\)Voyager: an open\-ended embodied agent with large language models\.arXiv preprint arXiv:2305\.16291\.Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p2.1),[§2\.2](https://arxiv.org/html/2605.07202#S2.SS2.p1.1)\.
- Z\. Wang, S\. Cai, G\. Chen, A\. Liu, X\. Ma, Y\. Liang, and T\. CraftJarvis \(2023b\)Describe, explain, plan and select: interactive planning with large language models enables open\-world multi\-task agents\.InProceedings of the 37th International Conference on Neural Information Processing Systems,pp\. 34153–34189\.Cited by:[§2\.2](https://arxiv.org/html/2605.07202#S2.SS2.p1.1)\.
- Q\. Wei, S\. Zeng, C\. Li, W\. Brown, O\. Frunza, W\. Deng, A\. Schneider, Y\. Nevmyvaka, Y\. K\. Zhao, A\. Garcia, and M\. Hong \(2025\)Reinforcing multi\-turn reasoning in llm agents via turn\-level reward design\.External Links:2505\.11821,[Link](https://arxiv.org/abs/2505.11821)Cited by:[§4\.5](https://arxiv.org/html/2605.07202#S4.SS5.p1.4)\.
- H\. Weng, Z\. Liu, Y\. Song, X\. Yin, X\. Chen, and W\. Zhang \(2025\)UniDataBench: evaluating data analytics agents across structured and unstructured data\.arXiv preprint arXiv:2511\.01625\.Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p1.1),[§2\.1](https://arxiv.org/html/2605.07202#S2.SS1.p1.1)\.
- H\. Whitney \(2012\)Data insights: new ways to visualize and make sense of data\.Newnes\.Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p1.1)\.
- A\. Yang, A\. Li, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Gao, C\. Huang, C\. Lv,et al\.\(2025\)Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p1.1)\.
- S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. Cao \(2023\)REACT: synergizing reasoning and acting in language models\.In11th International Conference on Learning Representations, ICLR 2023,Cited by:[item \(1\)](https://arxiv.org/html/2605.07202#S5.I1.i1.p1.1)\.
- S\. Zhang, J\. Fan, M\. Fan, G\. Li, and X\. Du \(2025\)Deepanalyze: agentic large language models for autonomous data science\.arXiv preprint arXiv:2510\.16872\.Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p1.1)\.
- Y\. Zheng, D\. Fu, X\. Hu, X\. Cai, L\. Ye, P\. Lu, and P\. Liu \(2025\)Deepresearcher: scaling deep research via reinforcement learning in real\-world environments\.arXiv preprint arXiv:2504\.03160\.Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p2.1),[§2\.2](https://arxiv.org/html/2605.07202#S2.SS2.p1.1)\.
- X\. Zhu, Y\. Chen, H\. Tian, C\. Tao, W\. Su, C\. Yang, G\. Huang, B\. Li, L\. Lu, X\. Wang,et al\.\(2023\)Ghost in the minecraft: generally capable agents for open\-world environments via large language models with text\-based knowledge and memory\.arXiv preprint arXiv:2305\.17144\.Cited by:[§2\.2](https://arxiv.org/html/2605.07202#S2.SS2.p1.1)\.
- P\. Zikopoulos and C\. Eaton \(2011\)Understanding big data: analytics for enterprise class hadoop and streaming data\.McGraw\-Hill Osborne Media\.Cited by:[§1](https://arxiv.org/html/2605.07202#S1.p1.1),[§2\.1](https://arxiv.org/html/2605.07202#S2.SS1.p1.1)\.

## Appendix AData development

The agent environment is a dynamic substrate specifically engineered for high\-volume data mining and autonomous retrieval\. This section details the environment’s architecture, strategic advantages in solving industrial data pain points, and the multi\-tier framework that ensures semantic consistency and exploration efficiency\.

### A\.1System Architecture

In industrial scenarios, data retrieval often suffers from inconsistency, semantic drift, and non\-idempotent results\. The environment addresses these issues through a robust three\-tier architecture\.

Semantic Layer\. This layer serves as a comprehensive repository, governing standard semantics of metadata for over 200 metrics and 100 dimensions, alongside their respective enumerated values\.

Routing Layers\. The routing layer implements the mapping from semantics to physical tables\. Metrics are mounted onDetail Tables \(DWS\)for granular insight andAggregate Tables \(ADS\)for high\-frequency queries\. The engine intelligently routes queries to minimize response time \(RT\)\.

Table 2:Physical Layer Table Schema and Data ScaleTable NameCategoryDescriptiondws\_trdDWSGranular records of trades and order lifecycles\.dws\_logDWSHigh\-throughput clickstream and behavior logs\.dim\_usrDWSUser profiles and demographics\.dim\_shopDWSMerchant and category metadata\.dim\_dateDWSTemporal metadata and holiday flags\.ads\_shopADSPre\-aggregated merchant performance indicators\.ads\_brandADSPre\-aggregated brand\-level performance indicators\.The ADS layer reduces RT by over 90% for standard queries\. However, for dynamic exploratory queries that fall back to the DWS layer, the system implements a strict 60\-second timeout to ensure resource stability\.

Table 3:Response Time Performance
### A\.2Integrated Query Execution and Case Study

The environment employs a dual\-channel feedback mechanism to ensure semantic grounding and evidence\-based reasoning as illustrated in Figure[9](https://arxiv.org/html/2605.07202#A1.F9)\.

Agent Query\(Ambiguous Params\)DSL Engine\(Calibration & Mapping\)Execution Unit\(SQL Gen & Cache\)Final Results\(CSV Data\)StandardSemanticCalibration FeedbackExecution FeedbackFigure 9:The environment workflow providing two feedback channels: \(1\)Semantic Calibrationand \(2\)Data Evidence\.Agent Query\. This entry point is where the agent formulates its analytical intent\. Given the massive exploration space, initial queries may contain ambiguous parameters that require further refinement within the DSL framework\.

DSL Engine\. The module leverages the Standard Semantic layer to map ambiguous inputs to precise metric and dimension definitions\. It provides an immediate Calibration Feedback loop, allowing the agent to align its queries with the system’s structural logic before execution\. Furthermore, in cases where sorting criteria are unspecified, the engine automatically defaults to dimension\- and metric\-based ordering to ensure the most salient data are prioritized\.

Execution Unit\. Once calibrated, the DSL is transformed into optimized SQL for data retrieval\. This unit integrates a caching mechanism to minimize latency, ensuring high\-fidelity performance even during large\-scale iterative mining\.

Final Results\. The raw output \(CSV format\) represents the factual basis of the query\. These results are funneled back to the agent via an Execution Feedback channel, providing the empirical evidence necessary for the agent to validate its hypotheses and continue cumulative reasoning\.

By decoupling semantic calibration from data execution, this architecture empowers agents to self\-correct their analytical paths through two distinct feedback channels, significantly enhancing the efficiency of the ”data\-to\-insights” pipeline\.

Table 4:Parameter Calibration and Mapping Logic1SELECT

2a\.shop\_name,

3SUM\(net\_gmv\)ASnet\_gmv,

4SUM\(net\_gmv\)/COUNT\(DISTINCTuser\_id\)ASnet\_aov

5FROMdws\_xx\_trd\_1da

6JOINdim\_xx\_datebONa\.ds=b\.ds

7WHEREa\.shop\_name=’XX’

8GROUPBYb\.is\_week;

Listing 1:Synthesized SQL based on Calibrated DSL1\{

2"calibration\_report":\{

3"corrected\_dsl":\{"metric":\["netAov","netGMV"\],"dimension":\["isWeek"\]\},

4"notices":"Input’week’alignedto’isWeek’\(temporaldimension\)\."

5\},

6"execution\_results":\{

7"data\_path":"\./xx/2026\.csv",

8"preview":\[

9\{"isWeek":"Weekend","netAov":15\.2,"netGMV":45000\},

10\{"isWeek":"Weekday","netAov":12\.3,"netGMV":53632\}

11\]

12\},

13"status":"Success"

14\}

Listing 2:Comprehensive Feedback Package \(Calibration \+ Results\)

## Appendix BDsl2data Protocol

The Dsl2data protocol establishes a standardized interface that facilitates interaction between the agent and the underlying data repository\. By ensuring semantic consistency, the protocol guarantees that all data retrieval actions are fully executable within the AIDA environment\. Table[5](https://arxiv.org/html/2605.07202#A2.T5)details the structural specifications of the Dsl2data protocol\. To illustrate its application, a valid request for retrieving data on top\-performing merchants is provided\.

Table 5:Dsl2data Protocol Schema Definitions1\{

2"metric":\["Net\_GMV"\],

3"dimension":\["Gender"\],

4"filter":\{

5"relation":"and",

6"conditions":\[

7\{

8"columnEName":"brand\_id",

9"queryRule":"in",

10"params":\["xxx"\]

11\}

12\]

13\},

14"ds":\["20251010","20251110"\],

15"orderBy":\[

16\{"columnEName":"Net\_GMV","orderType":"desc"\}

17\],

18"limit":10

19\}

Listing 3:Example request for Dsl2data\.
## Appendix CDiscovery Gain Reward Judge Prompt

1\#Background

2Theagenthaveprovidedinsightsbasedondataqueriesandanalysisregardingatargetproblem\.YouneedtoassesswhethertheseinsightsarevalidbyapplyingtheParetoPrinciple\.

3

4\#Definitionof"KeySub\-insight"

5A\*\*KeySub\-insight\*\*isanintermediatefindingderivedthroughreasoning,induction,orhypothesistestingbasedonthecurrentdataandcontext\.Itscharacteristicsinclude:

61\.Providingsubstantiveprogresstowardtheoveralltask\.

72\.Actingasabridgeinthelogicalchain\(connectingpreviousandsubsequentsteps\)\.

83\.Servingasacorecognitiveunitthatdrivesthedirectionoffurtherexploration\.

94\.Notyetconstitutingafinaldecisionoracompleteinsight\.

10

11\#CorePrinciples\(ParetoPrinciple/80\-20Rule\)

12\-80%ofresultsaredrivenby20%ofkeyfactors\.Avalidkeysub\-insightmust:

13\[YES\]Focusonkeyfactors\(identifyingthecoreoftheproblem\)\.

14\[YES\]Ignorethetrivialmany\(avoidinginterferencefromthe80%secondaryfactors\)\.

15\[YES\]Reflectthedominantinfluenceoftheprimarycontradiction\.

16\[YES\]Haveclearlogicalsupport\.

17\-\*\*InvalidinsightScenarios\*\*:

18\[NO\]Theinsightisvague,logicallyinconsistent,orfailstofocusonkeyfactors\.

19\[NO\]Theinsightreliesonmissingdata\(e\.g\.,"datanotobtained,""nodatayet"\),makingtheParetoPrincipleinapplicable\.

20\[NO\]Theinsightisredundantwithprevioussub\-insights\.

21

22\#Task

23\*\*Strictlyevaluatethevalidityofthesub\-insight\*\*:

241\.Checkifthesub\-insightalignswiththe"DefinitionofKeySub\-insight\."

252\.Checkifthesub\-insightadherestotheParetoPrinciple\.

263\.OutputMUSTbeonly:Valid\(fullycompliant\)/Invalid\(violatesanyrule\)\.

27

28\#InputInformation

29\-TargetQuestion:\{question\}

30\-PreviousKeySub\-insight:\{insight\}

31\-CurrentAction:\{action\}

32\-CurrentEnvironmentObservation:\{obs\}

33\-UpdatedSub\-insight\(tobeevaluated\):\{insight\}

34

35\#ResponseFormat

36Thought:\[Writeyouranalysisandjudgmentlogichere\.Keepitlogicallyclearandconcise\.\]

37FinalAnswer:\[Outputonly"Valid"or"Invalid"\]

Listing 4:Discovery Gain Reward Judge Prompt