TL;DR
Entity Resolution (ER) has traditionally been treated as a passive, one-shot classification problem, which fails to handle the messy, uncertain nature of modern data lakes. This vision paper proposes Agentic ER, a paradigm shift that frames ER as a sequential decision-making process where autonomous agents dynamically plan actions, gather external evidence, and optimize trade-offs between accuracy, cost, and latency.
What Problem Does This Paper Address?
Entity Resolution—the task of identifying different records that refer to the same real-world entity—is a cornerstone of data integration and cleaning. However, decades of ER research have relied on a "passive paradigm." As shown in Figure 1, traditional systems execute a rigid, one-shot pipeline (Blocking $\rightarrow$ Meta-blocking $\rightarrow$ Matching $\rightarrow$ Clustering) based on static similarity functions or static machine learning classifiers.
This passive approach breaks down in modern data lakes, where data arrives continuously in highly heterogeneous formats (CSV, JSON, RDF, free text) and contains high levels of noise or incomplete information. The authors identify four critical flaws in Passive ER:
- Lack of Sequential Reasoning: It treats ER as a simple classification task, ignoring the fact that resolving highly ambiguous entities often requires multi-step reasoning.
- No Adaptive Evidence Acquisition: It assumes a closed-world setting where all data is available upfront, lacking mechanisms to query external knowledge bases when local information is insufficient.
- No Runtime Interaction: It cannot dynamically ask clarification questions or integrate human feedback during inference.
- Lack of Cost Awareness: It focuses strictly on maximizing accuracy, ignoring the computational, monetary, or latency costs of heavy LLM queries or human interventions.
Methods
To bridge this gap, the authors introduce Agentic ER, formalizing it as a decision-theoretic problem governed by a tool-selection policy $\pi(a_t | s_t)$. Instead of a fixed pipeline, an autonomous agent runs a continuous Plan-Act-Observe-Update loop to resolve entity ambiguities under resource constraints.
To operationalize this paradigm, the paper proposes a modular reference architecture (Figure 3) consisting of five interacting components:
- Planner (Agent Core): The central decision-maker (potentially powered by an LLM or reinforcement learning) that determines the next best action based on the current state.
- Retriever: An open-world connector that queries external APIs, web search engines, or knowledge graphs to pull in missing context.
- Pipeline Executor: A selective runner that executes specific ER tasks (like running a cheap rule-based matcher or a heavy LLM matcher) on designated subsets of data.
- Memory Module: A structured repository that maintains the agent's evolving state, past action history, and intermediate beliefs (effectively acting as a dynamic similarity graph).
- Interaction Module: A first-class interface enabling the agent to dynamically escalate highly uncertain pairs to human annotators or domain experts during inference.
Datasets
Because this is a vision paper, the authors do not evaluate their framework on specific datasets. Instead, they argue that existing ER benchmarks are fundamentally inadequate for evaluating agentic behavior because they are static and closed-world.
They call on the community to build a new generation of benchmarks featuring:
- Multi-hop matching: Tasks requiring sequential reasoning across multiple pieces of evidence.
- Open-world settings: Scenarios where key identifying information is intentionally omitted from the profiles, forcing agents to retrieve external data.
- Dynamic and streaming settings: Environments where data arrives incrementally, requiring real-time updates to blocks and clusters.
Key Insights and Ideas
- ER as Meta-Decision Making: Agentic ER shifts the focus from pure similarity computation to goal-directed reasoning. The agent treats the choice of ER tools, thresholds, and human escalations as an optimization problem balancing accuracy against a strict resource budget.
- Granular Escalation: Instead of processing all record pairs uniformly, an agent can use a cheap attribute-overlap function for 90% of the data, and selectively escalate the remaining ambiguous 10% to an expensive LLM or a human expert.
- Dynamic Memory as a Graph: The memory module acts as an evolving similarity graph. As the agent uncovers new evidence, it updates its internal beliefs, allowing downstream clustering algorithms to adapt on the fly rather than waiting for a final, static batch of match scores.
Key Results
Quantitative results are not discussed in the paper, as it focuses entirely on introducing the conceptual framework, formalizing the problem space, and outlining the system architecture. The primary contribution is establishing a new research direction at the intersection of data management and autonomous intelligent agents.
Limitations and Future Work
The authors candidly lay out several open research challenges that must be solved to make Agentic ER practical:
- Planning Scalability: Action spaces spanning blocking, matching, retrieval, and human queries can become massive. Finding efficient planning policies that don't introduce crippling latency is a major hurdle.
- Evidence Management: Agents must learn when and how much to retrieve to avoid "over-retrieval," which introduces both financial costs and distracting noise.
- Uncertainty Propagation: Errors made early in an agent's sequential execution can compound downstream. Tracking how uncertainty propagates across adaptive blocking and matching steps remains an open problem.
- Explainability and Trust: Because an agent's final decision might be the result of a long, convoluted path of web searches, human queries, and model switches, generating a clean, auditable "reasoning trace" is essential for user trust.
Summary updated