Paper deep dive
EARS: Explanatory Abstention for Reliable Sub-Agent Modeling in Large-scale Multi-Agent Systems
Shuang Xie, Yunan Lu, Han Li, Lingyun Wang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 98%
Last extracted: 6/20/2026, 4:56:16 AM
Summary
The paper introduces EARS (Explanatory Abstention for Reliable Sub-Agent Modeling), a framework designed to improve the reliability of centralized Multi-Agent Systems (MAS) in enterprise settings. Instead of simple refusal, EARS enables sub-agents to provide 'explanatory abstention'âreturning a structured failure state (category and rationale) to the coordinator. The framework uses a human-in-the-loop process to calibrate an ensemble of LLM-as-a-Judge models (OpenAI, Claude, Gemini) to curate high-precision training data. This data is used to fine-tune domain-specific sub-agents (e.g., a Business Intelligence sub-agent) to better detect and communicate failure modes like Ambiguous Query, Insufficient Input, Missing Capability, and Misrouting. Evaluation on an e-commerce platform showed a significant improvement in the overall response pass rate from 68.5% to 78.9%.
Entities (10)
Relation Signals (5)
Sub-agent â communicatesfailureto â Coordinator
confidence 100% · a sub-agent does not merely abstain, but exposes an actionable failure state to the coordinator
LLM-as-a-Judge â curatesdatafor â Sub-agent
confidence 100% · The calibrated judges then curate large-scale interaction data for sub-agent fine-tuning
Coordinator â delegatesto â Sub-agent
confidence 100% · a coordinator delegates user requests to lightweight, domain-specialized sub-agents
EARS â improves â Multi-Agent System (MAS)
confidence 100% · demonstrating that sub-agent-side explanatory abstention improves MAS reliability
Ambiguous Query â isatypeof â Abstention Taxonomy
confidence 100% · classify abstention cases into four mutually exclusive categories: Ambiguous Query...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:In large-scale enterprise settings, centralized multi-agent systems (MAS) are increasingly adopted, in which a coordinator delegates user requests to lightweight, domain-specialized sub-agents. While this architecture improves modularity, scalability, and cost efficiency, its reliability depends not only on accurate routing but also on sub-agents' ability to calibrate their responses to capability constraints. In particular, sub-agents built on smaller fine-tuned models often struggle with such calibration, leading them to over-answer ambiguous, underspecified, misrouted, or unsupported requests and produce hallucinated outputs instead of actionable feedback. To address this challenge, we present EARS (Explanatory Abstention for Reliable Sub-Agent Modeling), a production-oriented framework that reframes sub-agent abstention as an inter-agent communication protocol: a sub-agent does not merely abstain, but exposes an actionable failure state to the coordinator. EARS curates human-agent interaction data using an ensemble of calibrated LLM-as-a-Judge models, producing structured abstention labels and rationales under a taxonomy of sub-agent failure modes. These data are used to fine-tune sub-agents to detect failure conditions and return rationales for coordinator-level clarification, rerouting, or fallback. We evaluate EARS in a large-scale production e-commerce assistant supporting enterprise business intelligence workflows. EARS improves the overall response pass rate from 68.5% to 78.9%, demonstrating that sub-agent-side explanatory abstention improves MAS reliability.
Tags
Links
- Source: https://arxiv.org/abs/2606.18668v1
- Canonical: https://arxiv.org/abs/2606.18668v1
Trouble viewing inline? Open PDF directly â
Full Text
43,500 characters extracted from source content.
Expand or collapse full text
EARS: Explanatory Abstention for Reliable Sub-Agent Modeling in Large-scale Multi-Agent Systems Shuang Xie 1, Yunan Lu â1,2, Han Li â1, Lingyun Wang 1 1Shopify 2Columbia University Correspondence: lingyun.wang@shopify.com denotes equal contribution. This work was done as part of internship at Shopify. Abstract In large-scale enterprise settings, centralized multi-agent systems (MAS) are increasingly adopted, in which a coordinator delegates user requests to lightweight, domain-specialized sub-agents. While this architecture improves modularity, scalability, and cost efficiency, its reliability depends not only on accurate routing but also on sub-agentsâ ability to calibrate their responses to capability constraints. In particular, sub-agents built on smaller fine-tuned models often struggle with such calibration, leading them to over-answer ambiguous, underspecified, misrouted, or unsupported requests and produce hallucinated outputs instead of actionable feedback. To address this challenge, we present EARS (Explanatory Abstention for Reliable Sub-Agent Modeling), a production-oriented framework that reframes sub-agent abstention as an inter-agent communication protocol: a sub-agent does not merely abstain, but exposes an actionable failure state to the coordinator. EARS curates human-agent interaction data using an ensemble of calibrated LLM-as-a-Judge models, producing structured abstention labels and rationales under a taxonomy of sub-agent failure modes. These data are used to fine-tune sub-agents to detect failure conditions and return rationales for coordinator-level clarification, rerouting, or fallback. We evaluate EARS in a large-scale production e-commerce assistant supporting enterprise business intelligence workflows. EARS improves the overall response pass rate from 68.5% to 78.9%, demonstrating that sub-agent-side explanatory abstention improves MAS reliability. EARS: Explanatory Abstention for Reliable Sub-Agent Modeling in Large-scale Multi-Agent Systems Shuang Xie â thanks: denotes equal contribution.1, Yunan Lu ââ thanks: This work was done as part of internship at Shopify.1,2, Han Li â1, Lingyun Wang 1 1Shopify 2Columbia University Correspondence: lingyun.wang@shopify.com 1 Introduction and Related Work Multi-Agent Systems (MAS) have demonstrated impressive capabilities across complex tasks, including software engineering, scientific discovery, data analytics, web automation, and general-purpose problem solving Qian et al. (2024); Wang et al. (2025); Gottweis et al. (2025); Swanson et al. (2024); Fourney et al. (2024). In industry, a widely adopted design is the centralized MAS architecture Zhang et al. (2026); Wu et al. (2024); Fourney et al. (2024), in which a powerful coordinator interprets user requests, routes them to specialized sub-agents, and integrates their outputs into a final response. To balance task performance and operational cost, sub-agents often use smaller, domain-specific models fine-tuned for particular domains or databases Abdin et al. (2024); Lepagnol et al. (2024). However, this architecture introduces reliability challenges: the system can fail not only when the coordinator misroutes a request, but also when the sub-agent returns unfaithful execution results. Existing work has predominantly addressed MAS reliability at the coordinator layer. RIRS Wu et al. (2026) selects specialist agents based on knowledge boundaries and decomposes complex queries through planning, while MasRouter Yue et al. (2025) formulates MAS routing as a unified decision problem over collaboration modes, role allocation, and LLM selection. Recent cost-aware orchestration work further trains centralized controllers to coordinate expert models under performanceâcost trade-offs (Jin et al., 2025). While these methods are vital for mitigating routing errors, they focus on top-down coordination, leaving sub-agent execution failures and the resulting communication breakdown between sub-agents and the coordinator largely unexplored. Recent analyses of MAS failures suggest that sub-agent execution and communication are central to system reliability. The failure taxonomy of Cemri et al. (2025) shows that many system-level failures arise from breakdowns in inter-agent information flow, including failures to seek clarification, information withholding, and ignoring inputs from other agents. These findings highlight that what a sub-agent returns to the coordinator is not merely an intermediate payload; it is a communication signal for downstream recovery. Prior work has studied when models should refuse unanswerable or unsupported questions (Kirichenko et al., 2025; Amayuelas et al., 2024). Abstain-R1 (Zhai et al., 2026) uses RL-based post-training to improve modelsâ abstention ability. LatentRefusal (Ren et al., 2026) proposes a computationally efficient answerability-gating method for Text-to-SQL systems. However, existing methods primarily study single-agent systems, where abstention serves as a final user-facing refusal, rather than a communication signal to the coordinator in MAS. To address this gap, we propose EARS, a domain-adaptive approach for improving the robustness of centralized MAS. EARS uses calibrated LLM-as-a-Judge models with unanimous judge agreement to curate abstention supervision from human-agent interaction data. It then fine-tunes sub-agents to return category-specific abstention decisions and rationales as actionable coordinator signals, improving response reliability. We make three contributions. First, we formulate explanatory abstention at the sub-agent level as a coordination signal for production MAS, turning abstention from a user-facing refusal into actionable inter-agent feedback. Second, we introduce a data curation pipeline to iteratively identify interaction states where sub-agents should abstain. Third, we evaluate EARS on a large-scale e-commerce platform that serves millions of customers globally and supports substantial annual commerce volume. EARS improves the response pass rate by 15.2% relatively, demonstrating that explanatory abstention provides a practical mechanism for improving reliability in large-scale centralized MAS. Figure 1: Overview of EARS. The framework starts by sampling a subset of human-agent interaction data for human annotation, then uses the annotated seed set to calibrate LLM-as-a-Judge models under an abstention taxonomy. The calibrated judges then curate large-scale interaction data for sub-agent fine-tuning. The updated sub-agent is finally redeployed into the multi-agent system, forming an online data flywheel for iterative improvement. 2 Method As illustrated in Figure 1, EARS consists of two stages. The offline calibration stage builds a human-annotated seed dataset with an abstention taxonomy (Section 2.1), and uses it to calibrate an LLM-as-a-Judge via iterative refinement (Section 2.2). The online data flywheel stage applies the calibrated judges to curate interaction data (Section 2.3) for sub-agent fine-tuning (Section 2.4). The updated sub-agent is then deployed back into the MAS to collect new interactions and drive the next cycle, enabling iterative improvement. 2.1 Human Annotation Abstention Taxonomy. To make sub-agent abstention feedback useful for the coordinatorâs final response, we design a taxonomy that captures why the selected sub-agent cannot handle a delegated request. Building on prior work Wen et al. (2025), we classify abstention cases into four mutually exclusive categories: Ambiguous Query Asai and Choi (2021), where the user intent cannot be clearly determined; Insufficient Input Li et al. (2024), where the intent is clear but execution requires additional information; Missing Capability Amayuelas et al. (2024), where the sub-agent does not support the requested functionality; and Misrouting, where the request should be handled by another sub-agent. This taxonomy provides fine-grained failure categories that sub-agents return to the coordinator. We provide examples of each category in Table 4. Annotation Process. To minimize annotation ambiguity and guide annotators through logically dependent label decisions, we structure the annotation process as a hierarchical decision procedure. Annotators first determine whether the user query is ambiguous. If the query admits multiple materially different interpretations due to underspecified key information, annotators label it as Ambiguous Query. Otherwise, the query is treated as non-ambiguous, meaning that it has a single intended operation. For non-ambiguous queries, annotators then assess whether the target sub-agent can fulfill the request given its available context, supported capabilities, and business policy constraints. If the sub-agent can complete the request, the instance is marked as solvable. If not, annotators assign it to one of the abstention categories: Insufficient Input, Missing Capability, or Misrouting. Each annotation object includes both a category label and a written rationale. The rationale specifies the source of ambiguity for Ambiguous Query, the missing information for Insufficient Input, the unsupported functionality for Missing Capability, or the appropriate sub-agent for Misrouting. Following standard annotation practice Lia et al. (2025); Artstein and Poesio (2008), we adopt an independent dual-annotation scheme with adjudication. Each instance is labeled independently by two annotators. Agreed labels are accepted directly; disagreements are resolved by a third senior annotator. Annotation was conducted by a professional data annotation vendor. The process yielded 600 solvable queries and 602 abstention instances, suggesting that a substantial proportion of user queries are not correctly handled by the sub-agent. 2.2 LLM-as-a-Judge Calibration To curate abstention-aware training data from production data at scale, we use an automatic judge to determine whether a query is solvable and, if not, which abstention category it belongs to. Directly applying an off-the-shelf LLM-as-a-judge is unreliable in this setting because abstention decisions depend on task-specific rules and subtle distinctions among abstention categories. We therefore calibrate the judge on the human-annotated seed set before applying it to large-scale data. Given the human-annotated seed set, we iteratively optimize the LLM-as-a-judge instructions to improve abstention-category prediction accuracy. We begin with an initial general instruction prompt and run the judge on the annotated dataset to measure its performance. If prediction accuracy falls below a predefined threshold, e.g., 80%, we use a second LLM to analyze the failure cases, summarize prominent confusion patterns, and generate actionable suggestions for revising the guidelines. This establishes a human-in-the-loop revision process that mitigates the risk of overfitting associated with fully automated prompt optimization. It also ensures that guideline updates remain general, interpretable, and consistent with the annotation taxonomy. We repeat this process until the score exceeds the threshold and shows no improvement for two consecutive revisions. 2.3 Data Curation We apply the calibrated LLM-as-a-judge to annotate the human-agent interaction data. Prior work shows that different LLMs can exhibit complementary strengths, motivating ensemble-based approaches that combine predictions across models Verga et al. (2024). Following this insight, we instantiate judges from three model families: OpenAI OpenAI (2025), Claude Anthropic (2025), and Gemini Google (2025). To avoid incorrectly converting answerable queries into abstention cases, we adopt a conservative ensemble strategy. For each sample, we run all judges and accept the label only when they unanimously agree on the same abstention category. The number of required agreeing judges provides tunable precisionârecall control. In our setting, we use unanimous agreement to construct a high-precision curated dataset for training. In applications where recall is more important, such as surfacing potential abstention cases for human review, this threshold can be lowered according to deployment requirements and risk tolerance. 2.4 Training After curating abstention-aware interaction data, we fine-tune the sub-agent with supervised fine-tuning (SFT). Each training instance consists of the user query, the sub-agent context, and the expected sub-agent response: a task-specific output for solvable queries, or a structured abstention response with an abstention category and a concise rationale. Subsequently, the sub-agent is deployed back into the MAS to serve live user interactions. The resulting interaction data are then fed back into the data curation stage, completing the flywheel loop. 3 Experiments and Results 3.1 Experimental Setup We evaluate EARS on an enterprise-level AI assistant deployed in a production e-commerce environment, focusing on its ability to answer merchant queries related to shop operations and management. The assistant follows a centralized coordinatorâsub-agent architecture, where a powerful coordinator built on Claude-sonnet-4-6 Anthropic (2026) routes user queries to a set of in-house fine-tuned sub-agents built on Qwen3-32B-Instruct Qwen Team (2025). In this work, we focus on the business intelligence (BI) sub-agent, which answers data-driven queries about shop performance metrics and operational analysis by generating responses in a Domain-Specific Language (DSL) with SQL-like syntax and functionality. This sub-agent covers two capabilities: customer segmentation (e.g., How many customers visited my shop in the past week?) and business analytics (e.g., What is my shopâs conversion rate for the last month?). Dataset. The training dataset consists of 13.8K user-agent interactions randomly sampled from the production environment, including 7.0K customer segmentation queries and 6.8K business analytics queries. We then apply the data curation workflow described in Section 2.3 to identify unsolvable queries and their corresponding abstention categories: Ambiguous Query, Insufficient Input, Missing Capability, and Misrouting. For evaluation, we construct a balanced dataset covering both customer segmentation and business analytics queries. Table 1 summarizes the statistics of the training and evaluation datasets. Category Training Set Eval Set Segmentation Solvable 5,967 1,524 Analytics Solvable 6,800 2,000 Abstention 1,086 476 Total 13,853 4,000 Table 1: Dataset statistics for training and evaluation. Training Setup. We fine-tune the BI sub-agent, built upon Qwen3-32B-Instruct Qwen Team (2025), using supervised fine-tuning (SFT). As the task requires learning nuanced abstention boundaries across multiple categories, we perform full-parameter fine-tuning. We adopt AdamW with a cosine learning rate schedule, setting the peak learning rate to 1Ă10â51Ă 10^-5 with a warmup ratio of 5%. We train for three epochs with a global batch size of 8 and a maximum sequence length of 32,768 with data packing. All experiments are conducted on 8 NVIDIA H200 GPUs. 3.2 Evaluation Metrics Abstention Ability Following prior work Feng et al. (2024), we evaluate the LLM-as-a-Judgeâs quality in identifying abstention categories using standard metrics: accuracy, precision, and recall. Cost We measure computational cost by token usage and normalize it relative to a single gemini-2.5-flash LLM-as-a-Judge call. Pass Rate We measure sub-agent performance by Pass Rate, the percentage of queries that receive correct responses. The Overall Pass Rate is determined by two sequential verification steps: (1) Syntax Validity (Syntax), which checks whether the BI sub-agentâs structured output is in a valid format and executable when applicable, including either a syntactically correct SQL-like DSL query or a properly formatted abstention response; and (2) Semantic Correctness (Semantic), which uses an LLM-as-a-Judge to assess whether the output satisfies the user query along four aspects: intent alignment, domain constraint satisfaction, query logic correctness, and abstention correctness when applicable. The same LLM-as-a-Judge is used across all methods. See Appendix A for prompt details. To validate the semantic correctness judge, we asked four domain experts with deep knowledge of the deployed sub-agents to annotate 1,050 production samples using the shared rubrics. Two annotators independently labeled each sample, achieving substantial agreement with Cohenâs Îș=0.718Îș=0.718, and resolved disagreements through adjudication. Against the consensus labels, the judge achieved 88.3% accuracy, 82.4% precision, 77.8% recall. 3.3 Evaluation Results Table 2 reports the calibrated abstention classification performance and relative cost of different LLM-as-a-Judge models on 1,202 human-annotated samples. All four single-model judges achieve above 88% accuracy, precision, and recall. Among them, gemini-2.5-flash provides the strongest cost-performance trade-off, achieving 90.4% accuracy at the lowest cost. The 4-judge consensus further improves precision by 6.3% at higher cost. These results show a practical trade-off between efficiency and accuracy: single-judge curation is preferable when throughput is the primary constraint, while multi-judge consensus provides higher-confidence abstention labels when label quality is the priority. Appendix C provides a analysis for each abstention category. LLM-as-a-Judge Accuracy Precision Recall Cost gpt-5 88.5% 89.2% 88.5% x4.1 o3 89.6% 90.3% 89.5% x6 claude-sonnet-4.5 88.4% 88.4% 88.4% x9.3 gemini-2.5-flash 90.4% 90.5% 90.4% x1 4-judge consensus 93.1% 96.2% 89.8% x20.4 Table 2: Abstention classification performance of calibrated LLM-as-a-Judge models. The best result is in bold and the second-best is underlined. Table 3 reports the response quality of the BI sub-agent. The baseline corresponds to the MAS with a BI sub-agent trained without abstention data, whereas EARS trains the sub-agent on data curated by an ensemble of four LLM-as-a-Judge models built on gpt-5, o3, claude-sonnet-4-5, and gemini-2.5-flash. Overall, EARS improves most metrics, with larger gains on customer segmentation queries than on analytics queries. This asymmetry is expected since 23.8% of segmentation queries are predicted as abstention cases, compared with fewer than 2% of analytics queries. Thus, the segmentation pathway benefits more from abstention training. Syntax validity remains high for both query types across both systems, suggesting that EARS improves abstention decisions without degrading DSL query generation quality. Metric Baseline EARS Rel. Gain Overall Pass Rate 68.5% 78.9% 15.2% Seg. Syntax 92.7% 97.2% 4.9% Seg. Semantic 61.9% 82.7% 33.6% Analytics Syntax 96.3% 96.9% 0.6% Analytics Semantic 75.1% 75.0% -0.1% Table 3: Evaluation results for the BI sub-agent. Seg. and Analytics denote customer segmentation and business analytics queries. Syntax and Semantic denote syntax validity and semantic correctness pass rates. Rel. Gain reports the relative improvement of EARS over the baseline, with green / red indicating gains / losses. Human Evaluation. To support side-by-side comparison with the incumbent system, we conducted an end-to-end shadow deployment with human validation. From randomly sampled production sessions, we retained 584 examples for which the shadow EARS deployment produced an abstention signal. Independent domain experts then evaluated these examples at two granularities. At the end-to-end MAS level, experts assessed session-level response quality using a binary rubric, with 1 denoting success and 0 denoting failure. Overall, experts labeled 392 sessions as successful (i.e., 67.1% human-rated pass rate), substantially higher than the incumbent systemâs 2.4% pass rate on the same sessions. At the BI sub-agent level, the model achieved 94.0% abstention precision (549/584), indicating strong alignment between model abstention behavior and expert judgments. Appendix B reports the response quality rubric. 3.4 Ablations To assess the effect of abstention data and LLM-as-a-Judge ensemble strategy, we run ablations across dataset variants with different curation settings. The results are summarized in Figure 2. Overall, abstention data consistently improves segmentation performance without degrading analytics performance. Human-annotated data increases the overall pass rate by 10.2%. The gain is most pronounced in segmentation semantic correctness, with a 23.1% relative improvement, demonstrating that abstention examples can improve model reliability. Although human annotation provides high-quality abstention supervision, mixing it with raw production interactions introduces label inconsistencies. Semantically similar queries may receive conflicting supervision: one is labeled with an abstention category, while another retains the original incorrect sub-agent response. EARS curation reduces this conflict by correcting production examples, yielding further gains. Among single-judge variants, the best-performing judge (gemini-2.5-flash) improves the overall pass rate by 1.3% over human annotation alone. In contrast, four-judge consensus achieves a 4.5% gain and improves all four metrics. This pattern suggests that single-judge curation can introduce label noise, whereas consensus-based curation provides a more reliable training signal by retaining only high-agreement corrections. Figure 2: Ablation results on the effect of abstention data under different curation strategies. Adding human-annotated abstention data improves the overall pass rate and substantially increases segmentation semantic correctness. EARS uses four-judge curation, which achieves the strongest overall performance. Type User request Baseline EARS Ambiguous Query "find customers who might be interested in buying shoes wholesale" Sub-agent: storefront.product_viewed MATCHES (tag contains âwholesaleâ) Sub-agent: "Response":"Ambiguous Query", "Reasoning": "What is wholesale-interest signal?" Assistant: "I do not find any customers who might be interested in buying shoes wholesaleâŠ" Issue: Uses an unverified proxy for wholesale buyer interest instead of asking for clarification. Assistant: "The segmentation tool needs a clearer signal for wholesale interest. Could you specify what indicates potential wholesale buyers, such as repeat shoe purchases, large orders, or views of wholesale products?" Insufficient Input "customers who purchased products with type French Horns, Trombones, or Trumpets" Sub-agent: product_purchased MATCHES (tag contains âFrentch Hornsâ OR contains âTrombinesâ OR âTrumpetsâ) Sub-agent: "Response": "Insufficient Input", "Reasoning": "Need intended product." Assistant: "I do not find any customers who purchased products with these types." Issue: Attempts segmentation without resolved product inputs, leading to invalid product filters. Assistant: "Could you clarify which products you mean by French Horns, Trombones, or Trumpets? I need the intended product before creating the segment." Missing Capability "male customers between 30 and 50 years old" Sub-agent: customer_tags CONTAINS âmaleâ AND customer_tags CONTAINS â30-50 years oldâ Sub-agent: "Response": "Missing Capability", "Reasoning": "Gender and Age filters are unavailable." Assistant: "I do not find any male customers between 30 and 50 years old." Issue: Maps unsupported demographic attributes to tags, falsely implying searchable gender and age fields. Assistant: "I cannot create a segment directly from gender or age because these filters are unavailable. I can help segment customers using supported criteria, such as purchase behavior, location, customer tags, or engagement." Misrouting "create segment for customers who visited the store multiple times but never purchased" Sub-agent: number_of_orders = 0 Sub-agent: "Response": "Misrouting", "Reasoning": "Need workflow sub-agent." Assistant: "Here are the customer segments for customers who never purchased: âŠ" Issue: Misroutes workflow-dependent request to segmentation, dropping repeated-visit condition. Assistant: "I routed this request to the workflow sub-agent. It will first tag customers who visit multiple times without purchasing; then I will create a segment from that tag." Table 4: Representative abstention examples comparing the baseline and EARS. The baseline over-answers ambiguous, underspecified, unsupported, or misrouted requests, whereas EARS abstains with category-specific rationales that support more accurate assistant responses. Red marks errors, and green marks accurate outputs. 4 Case Study To better understand how abstention training affects deployed MAS behavior, we conduct a qualitative analysis over representative cases. Table 4 compares the baseline with EARS across four abstention categories. The examples show that the baseline sub-agent often forces underspecified, unsupported, or misrouted requests into executable segmentation predicates. This behavior propagates to the assistant, which returns a hallucinated or misleading response to the user. In contrast, EARS produces category-specific abstention rationales that help the assistant choose a better next action. For example, in the Missing Capability case, the baseline treats demographic attributes as ordinary customer tags, even though gender and age filters are excluded from the BI sub-agent due to privacy and compliance policies. This response is problematic because it implies access to sensitive information that the system should not expose. EARS instead recognizes the capability boundary and guides the assistant to explain the unsupported filters while suggesting feasible alternatives. Similarly, in the Ambiguous Query case, the baseline assumes a proxy for wholesale interest without confirming the userâs intended definition. EARS identifies this ambiguity and asks the assistant to clarify the signal before segmentation. These cases show that abstention rationales not only prevent invalid outputs, but also help the coordinator clarify, redirect, or constrain the task before execution. 5 Conclusion We presented EARS, a production-oriented framework that improves centralized MAS reliability through sub-agent explanatory abstention. EARS reframes abstention as an inter-agent communication protocol, enabling sub-agents to expose actionable failure states to the coordinator. Using calibrated LLM-as-a-Judge curation and supervised fine-tuning, EARS improves the overall response pass rate by 15.2% relatively in a production e-commerce assistant. These results demonstrate that explanatory abstention provides an effective coordination signal for improving MAS reliability. 6 Limitations The proposed method is designed to be domain-adaptive, while our empirical evaluation focuses on an e-commerce business-intelligence assistant. Further validation is needed in other domains, where abstention scenarios and evaluation costs may differ. Future work should apply EARSâs curation and fine-tuning pipeline to MAS deployed in domains such as finance, education, and healthcare. To obtain reliable abstention labels, we adopt a human-in-the-loop workflow for calibrating LLM-as-a-Judge models. While this design improves interpretability and reduces the risk of overfitting to the seed set, it also motivates more automated calibration strategies. Recent advances in automated judge revision and prompt refinement Wang et al. (2024); Agrawal et al. (2025) should be explored to improve scalability while preserving label quality. Finally, RL-based coordinator training or recovery-policy optimization could complement EARS by improving MAS-level response strategies following sub-agent abstention. We leave such system-level optimization to future work. 7 Acknowledgments We thank Prof. Zhou Yu for valuable guidance. We are also grateful to Jason Liu, Nicolas Bertagnolli, and Aly Gemae for their helpful support and feedback throughout this work. We further thank Kaitlyn Chenier and the annotation team for their assistance with data annotation. References M. Abdin, J. Aneja, H. Behl, S. Bubeck, R. Eldan, S. Gunasekar, M. Harrison, R. J. Hewett, M. Javaheripi, P. Kauffmann, J. R. Lee, Y. T. Lee, Y. Li, W. Liu, C. C. T. Mendes, A. Nguyen, E. Price, G. d. Rosa, O. Saarikivi, A. Salim, S. Shah, X. Wang, R. Ward, Y. Wu, D. Yu, C. Zhang, and Y. Zhang (2024) Phi-4 Technical Report. arXiv. Note: arXiv:2412.08905 [cs.CL] External Links: Link, Document Cited by: §1. L. A. Agrawal, S. Tan, D. Soylu, N. Ziems, R. Khare, K. Opsahl-Ong, A. Singhvi, H. Shandilya, M. J. Ryan, M. Jiang, et al. (2025) Gepa: reflective prompt evolution can outperform reinforcement learning. arXiv preprint arXiv:2507.19457. Cited by: §6. A. Amayuelas, K. Wong, L. Pan, W. Chen, and W. Wang (2024) Knowledge of Knowledge: Exploring Known-Unknowns Uncertainty with Large Language Models. arXiv. Note: arXiv:2305.13712 [cs.CL] External Links: Link, Document Cited by: §1, §2.1. Anthropic (2025) Introducing claude sonnet 4.5. External Links: Link Cited by: §2.3. Anthropic (2026) Claude sonnet 4.6. Note: Model card and product page for claude-sonnet-4-6. External Links: Link Cited by: §3.1. R. Artstein and M. Poesio (2008) Survey Article: Inter-Coder Agreement for Computational Linguistics. Computational Linguistics 34 (4), p. 555â596. External Links: Link, Document Cited by: §2.1. A. Asai and E. Choi (2021) Challenges in Information-Seeking QA: Unanswerable Questions and Paragraph Retrieval. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), C. Zong, F. Xia, W. Li, and R. Navigli (Eds.), Online, p. 1492â1504. External Links: Link, Document Cited by: §2.1. M. Cemri, M. Z. Pan, S. Yang, L. A. Agrawal, B. Chopra, R. Tiwari, K. Keutzer, A. Parameswaran, D. Klein, K. Ramchandran, M. Zaharia, J. E. Gonzalez, and I. Stoica (2025) Why Do Multi-Agent LLM Systems Fail?. arXiv. Note: arXiv:2503.13657 [cs.AI] External Links: Link, Document Cited by: §1. S. Feng, W. Shi, Y. Wang, W. Ding, V. Balachandran, and Y. Tsvetkov (2024) Donât Hallucinate, Abstain: Identifying LLM Knowledge Gaps via Multi-LLM Collaboration. arXiv. Note: arXiv:2402.00367 [cs.CL] External Links: Link, Document Cited by: §3.2. A. Fourney, G. Bansal, H. Mozannar, C. Tan, E. Salinas, Erkang, Zhu, F. Niedtner, G. Proebsting, G. Bassman, J. Gerrits, J. Alber, P. Chang, R. Loynd, R. West, V. Dibia, A. Awadallah, E. Kamar, R. Hosn, and S. Amershi (2024) Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks. arXiv. Note: arXiv:2411.04468 [cs.AI] External Links: Link, Document Cited by: §1. Google (2025) Gemini 2.5 flash. External Links: Link Cited by: §2.3. J. Gottweis, W. Weng, A. Daryin, T. Tu, A. Palepu, P. Sirkovic, A. Myaskovsky, F. Weissenberger, K. Rong, R. Tanno, K. Saab, D. Popovici, J. Blum, F. Zhang, K. Chou, A. Hassidim, B. Gokturk, A. Vahdat, P. Kohli, Y. Matias, A. Carroll, K. Kulkarni, N. Tomasev, Y. Guan, V. Dhillon, E. D. Vaishnav, B. Lee, T. R. D. Costa, J. R. PenadĂ©s, G. Peltz, Y. Xu, A. Pawlosky, A. Karthikesalingam, and V. Natarajan (2025) Towards an AI co-scientist. arXiv. Note: arXiv:2502.18864 [cs.AI] External Links: Link, Document Cited by: §1. B. Jin, T. J. Collins, D. Yu, M. Cemri, S. Zhang, M. Li, J. Tang, T. Qin, Z. Xu, J. Lu, G. Yin, J. Han, and Z. Wang (2025) Controlling Performance and Budget of a Centralized Multi-agent LLM System with Reinforcement Learning. arXiv. Note: arXiv:2511.02755 [cs.CL] External Links: Link, Document Cited by: §1. P. Kirichenko, M. Ibrahim, K. Chaudhuri, and S. J. Bell (2025) AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions. arXiv. Note: arXiv:2506.09038 [cs.AI] External Links: Link, Document Cited by: §1. P. Lepagnol, T. Gerald, S. Ghannay, C. Servan, and S. Rosset (2024) Small Language Models Are Good Too: An Empirical Study of Zero-Shot Classification. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), N. Calzolari, M. Kan, V. Hoste, A. Lenci, S. Sakti, and N. Xue (Eds.), Torino, Italia, p. 14923â14936. External Links: Link Cited by: §1. S. S. Li, V. Balachandran, S. Feng, J. S. Ilgen, E. Pierson, P. W. Koh, and Y. Tsvetkov (2024) MediQ: Question-Asking LLMs and a Benchmark for Reliable Interactive Clinical Reasoning. arXiv. Note: arXiv:2406.00922 [cs.CL] External Links: Link, Document Cited by: §2.1. N. J. Lia, S. Roy Dipta, A. K. Zehady, N. Islam, M. Chakraborty, and A. Al Wasif (2025) Read Between the Lines: A Benchmark for Uncovering Political Bias in Bangla News Articles. In Proceedings of the Second Workshop on Bangla Language Processing (BLP-2025), F. Alam, S. Kar, S. A. Chowdhury, N. Hassan, E. Hoque, M. T. R. Laskar, T. Mohiuddin, and M. R. A. H. Rony (Eds.), Mumbai, India, p. 61â79. External Links: ISBN 979-8-89176-314-2, Link, Document Cited by: §2.1. OpenAI (2025) Introducing gpt-5. External Links: Link Cited by: §2.3. C. Qian, W. Liu, H. Liu, N. Chen, Y. Dang, J. Li, C. Yang, W. Chen, Y. Su, X. Cong, J. Xu, D. Li, Z. Liu, and M. Sun (2024) ChatDev: Communicative Agents for Software Development. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), L. Ku, A. Martins, and V. Srikumar (Eds.), Bangkok, Thailand, p. 15174â15186. External Links: Link, Document Cited by: §1. Qwen Team (2025) Qwen3 technical report. arXiv preprint arXiv:2505.09388. External Links: Link Cited by: §3.1, §3.1. X. Ren, S. Hu, Z. Lu, J. Huang, and Q. Duan (2026) LatentRefusal: Latent-Signal Refusal for Unanswerable Text-to-SQL Queries. arXiv. Note: arXiv:2601.10398 [cs.AI] version: 3 External Links: Link, Document Cited by: §1. K. Swanson, W. Wu, N. L. Bulaong, J. E. Pak, and J. Zou (2024) The Virtual Lab: AI Agents Design New SARS-CoV-2 Nanobodies with Experimental Validation. bioRxiv (en). Note: Pages: 2024.11.11.623004 Section: New Results External Links: Link, Document Cited by: §1. P. Verga, S. Hofstatter, S. Althammer, Y. Su, A. Piktus, A. Arkhangorodsky, M. Xu, N. White, and P. Lewis (2024) Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models. arXiv. Note: arXiv:2404.18796 [cs.CL] External Links: Link, Document Cited by: §2.3. X. Wang, B. Li, Y. Song, F. F. Xu, X. Tang, M. Zhuge, J. Pan, Y. Song, B. Li, J. Singh, H. H. Tran, F. Li, R. Ma, M. Zheng, B. Qian, Y. Shao, N. Muennighoff, Y. Zhang, B. Hui, J. Lin, R. Brennan, H. Peng, H. Ji, and G. Neubig (2025) OPENHANDS: AN OPEN PLATFORM FOR AI SOFTWARE DEVELOPERS AS GENERALIST AGENTS. arXiv preprint arXiv:2407.16741 (en). Cited by: §1. X. Wang, C. Li, Z. Wang, F. Bai, H. Luo, J. Zhang, N. Jojic, E. Xing, and Z. Hu (2024) Promptagent: strategic planning with language models enables expert-level prompt optimization. In International Conference on Learning Representations, Vol. 2024, p. 23967â24001. Cited by: §6. B. Wen, J. Yao, S. Feng, C. Xu, Y. Tsvetkov, B. Howe, and L. L. Wang (2025) Know Your Limits: A Survey of Abstention in Large Language Models. arXiv. Note: arXiv:2407.18418 [cs.CL] External Links: Link, Document Cited by: §2.1. F. Wu, Z. Li, F. Wei, Y. Li, B. Ding, and J. Gao (2026) Talk to Right Specialists: Iterative Routing in Multi-agent Systems for Question Answering. arXiv. Note: arXiv:2501.07813 [cs.MA] External Links: Link, Document Cited by: §1. Q. Wu, G. Bansal, J. Zhang, Y. Wu, B. Li, E. Zhu, L. Jiang, X. Zhang, S. Zhang, J. Liu, A. Awadallah, R. W. White, D. Burger, and C. Wang (2024) AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversations. arXiv preprint arXiv:2308.08155 (en). Cited by: §1. Y. Yue, G. Zhang, B. Liu, G. Wan, K. Wang, D. Cheng, and Y. Qi (2025) MasRouter: Learning to Route LLMs for Multi-Agent Systems. arXiv. Note: arXiv:2502.11133 [cs.LG] External Links: Link, Document Cited by: §1. S. Zhai, J. Liang, and D. Kang (2026) Abstain-R1: Calibrated Abstention and Post-Refusal Clarification via Verifiable RL. arXiv. Note: arXiv:2604.17073 [cs.CL] version: 1 External Links: Link, Document Cited by: §1. W. Zhang, L. Zeng, Y. Xiao, Y. Li, C. Cui, Y. Zhao, R. Hu, Y. Liu, Y. Zhou, and B. An (2026) AgentOrchestra: Orchestrating Multi-Agent Intelligence with the Tool-Environment-Agent(TEA) Protocol. arXiv. Note: arXiv:2506.12508 [cs.AI] External Links: Link, Document Cited by: §1. Appendix A Semantic Correctness Judge Prompts We use task-specific semantic correctness prompts for analytics outputs and segmentation outputs. These prompts are used by the final evaluator, openai:gpt-5 with temperature=0, and are separate from the abstention-labeling prompts used by the data-curation judges. The evaluator shares the gpt-5 base model with one curation judge, but it is run as a separate evaluation pipeline and is not used to label SFT data. Before final evaluation, the prompt is calibrated on a separate 1,050 example human-annotated split. The production prompts include platform-specific schemas, domain terminology, and internal unsupported-request lists. For industry-track disclosure, we report sanitized excerpts that preserve the evaluation protocol while masking platform names, schema fields, and domain-specific capabilities. A.1 Analytics Judge Prompt ⏠Input: user_message: the original user request context: request metadata, including request time when available model_output: the generated [ANALYTICS_QUERY_LANGUAGE] query Task: Determine whether model_output answers the userâs request. Return only a JSON object with: explanation: detailed reasoning about intent alignment and validity result: true if the query is correct, false otherwise Core checks: 1. Validate that all metrics, dimensions, and models are allowed by [ANALYTICS_SCHEMA_RESTRICTIONS]. 2. Check that the query matches the userâs intent, including filters, aggregation level, requested entity type, and requested time range. 3. If the user asks for a visualization, require a supported visualization directive from [SUPPORTED_VISUALIZATION_TYPES]. 4. For trend requests, require an ordering or grouping that makes the trend interpretable. 5. For top, highest, maximum, or best requests, require descending ordering by an appropriate metric. 6. For bottom, lowest, minimum, or least requests, require ascending ordering by an appropriate metric. 7. If no date is specified, treat all-time data as the default. 8. Interpret relative date ranges using the request timestamp in context. 9. Apply domain-specific semantic rules for reversal metrics, traffic-source dimensions, entity-name matching, and exclusion filters, with all platform-specific field names masked. 10. Reject outputs that attempt unsupported reports listed in [UNSUPPORTED_REQUEST_TYPES]. A.2 Segmentation Judge Prompt ⏠Input: user_message: the original user request context: request metadata, including request time when available model_output: a generated [SEGMENTATION_FILTER_LANGUAGE] expression or structured abstention signal Task: Evaluate whether model_output correctly handles the userâs request. Return only a JSON object with: explanation: detailed reasoning about the selected case and final verdict result: true if the output is correct, false otherwise Case selection: A. Standard query: Use this case when model_output does not contain an abstention signal. Check whether the filters, operators, values, and date constraints match the userâs intent under [SUPPORTED_SEGMENTATION_FILTERS]. B. Abstention query: Use this case when model_output contains an abstention signal. Independently decide whether the request is: solvable: supported with information already provided insufficient_input: supported but missing concrete user-provided input not_solvable: unsupported or too ambiguous to execute If not_solvable, assign exactly one reason: missing_capability misrouting ambiguous_query The model output is correct only when its abstention type and reason match the independent assessment. Core checks: 1. Confirm intent alignment before judging syntax-level details. 2. Treat segmentation as audience-list construction, not aggregate analytics. 3. Reject requests that require aggregate reports, ranking, visualization, mutation, or unsupported capabilities in [UNSUPPORTED_REQUEST_TYPES]. 4. Use request time from context for relative dates. 5. Prefer lifecycle/value-group filters for vague high-value or loyalty requests when no numeric threshold is provided. 6. Do not accept invented tags, unsupported attributes, or proxy filters that create unsupported behavior. Appendix B Human Evaluation Details This appendix reports the response-quality rubric used for the full MAS human evaluation in Section 3. Domain annotators assessed session-level assistant responses according to the rubric in Table 5. A rating of 1 is counted as a pass, whereas a rating of 0 is counted as a failure. Score Label Rubric 0 Poor Quality The response has one or more serious issues: confusing or illogical explanations, excessive verbosity, unnecessary or incorrect tool calls, criticism of platform features or recommendations for competitors, inappropriate tone, or failure to provide actionable guidance. 1 Satisfactory The response is coherent, professionally toned, and sufficiently addresses the userâs query. It may contain minor structure, flow, or conciseness issues, but uses tools appropriately and provides actionable guidance. Table 5: Human E2E response-quality rubric. Appendix C LLM-as-a-Judge Calibration Analysis Table 6 reports the prediction accuracy of each LLM-as-a-Judge on the 602 human-annotated abstention examples only. This abstention-only sensitivity analysis is separate from the all-sample calibration results in Table 2, which also includes solvable examples. Each judge evaluates the same outputs independently. claude-sonnet-4-5 achieves the highest abstention-only overall prediction accuracy at 87.6%, followed by gemini-2.5-flash. Judges perform most reliably on Missing Capability and Misrouting cases, where the decision often depends on whether the requested function or target worker is supported. In contrast, Ambiguous Query and Insufficient Input remain harder and show larger variation across judges. The confusion matrices in Figure 3 help explain this pattern: ambiguous queries are often confused with insufficient-input cases, since both involve underspecified user requests. Insufficient-input cases are also frequently confused with missing-capability or solvable cases, suggesting that judges differ in whether they interpret missing execution details as a request for unsupported functionality, an answerable query, or a query that requires additional user input. LLM-as-a-Judge Overall Accuracy Ambiguous Query Insufficient Input Missing Capability Misrouting gpt-5 79.6% 45.5% 58.3% 87.6% 86.8% o3 79.6% 33.3% 66.0% 85.6% 83.1% claude-sonnet-4-5 87.6% 83.3% 76.7% 90.8% 90.9% gemini-2.5-flash 84.7% 41.7% 70.9% 91.5% 85.7% Table 6: Abstention-only judge sensitivity analysis. Each judge evaluates the same 602 human-annotated abstention examples. Category columns report per-category prediction accuracy under the final taxonomy. Figure 3: Abstention judge confusion matrices for final-model abstention-category predictions. The y-axis gives the true human-annotated category, and the x-axis gives the category predicted by each LLM judge. Darker cells indicate larger counts.