Singapore

Enterprise meeting environments demand AI assistants that can handle diverse operational tasks—from rapid fact-checking during live discussions to deep cross-meeting analysis for strategic planning—under strict latency, cost, and privacy constraints. Existing meeting benchmarks focus on simplified QA tasks and lack alignment with real-world enterprise workflows where questions arise organically from multi-stakeholder collaboration, span temporal contexts, and require tool-augmented reasoning. 

We address this gap through a grounded dataset construction and learned agent framework. First, we present MeetAll: a bilingual, multimodal corpus derived from 231 meetings (140 hours) with enterprise-informed question injection validated through domain expert review (kappa=0.82) and human discriminability studies (56% detection rate, near-random). Unlike purely synthetic benchmarks, our injection protocol anchors on four enterprise-critical dimensions—cognitive load, temporal context span, domain expertise, and actionable task execution—calibrated via interviews with 15 enterprise meeting stakeholders across finance, healthcare, and tech sectors. 

Second, we introduce MeetBench-XL: a multi-dimensional evaluation protocol calibrated against human judgment (Pearson r=0.78, p&lt;0.01) that measures factual fidelity, user intent alignment, response efficiency, structural clarity, and completeness—criteria prioritized by enterprise users in our needs assessment. 

Third, we propose MeetMaster-XL: a learned dual-policy agent that jointly optimizes routing (fast Talker vs. slow Planner) and tool invocation (retrieval, cross-meeting aggregation, web search) through offline policy learning on quality–latency–cost objectives. Unlike prior rule-based fast/slow systems, our lightweight classifier (1ms overhead) achieves 81.2% routing accuracy and unlocks a superior Pareto frontier: 30% latency reduction on simple queries and 12.4% quality gain on complex tasks over single-model baselines, while remaining deployable on consumer GPUs (RTX 4090, 22GB VRAM). 

We benchmark against commercial APIs (GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro) and provide ablations, robustness tests (noise, overlap, multilingual splits), and a field deployment case study. Resources: https://github.com/huyuelin/MeetBench.

AAAI 2026

MeetBench-XL: Calibrated Multi-Dimensional Evaluation and Learned Dual-Policy Agents for Real-Time Meetings

mas: multiagent learning，ml: evaluation and analysis，hai: human-ai collaboration / human-ai teaming

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

Enterprise meeting environments demand AI assistants that can handle diverse operational tasks—from rapid fact-checking during live discussions to deep cross-meeting analysis for strategic planning—under strict latency, cost, and privacy constraints. Existing meeting benchmarks focus on simplified QA tasks and lack alignment with real-world enterprise workflows where questions arise organically from multi-stakeholder collaboration, span temporal contexts, and require tool-augmented reasoning. 

We address this gap through a grounded dataset construction and learned agent framework. First, we present MeetAll: a bilingual, multimodal corpus derived from 231 meetings (140 hours) with enterprise-informed question injection validated through domain expert review (kappa=0.82) and human discriminability studies (56% detection rate, near-random). Unlike purely synthetic benchmarks, our injection protocol anchors on four enterprise-critical dimensions—cognitive load, temporal context span, domain expertise, and actionable task execution—calibrated via interviews with 15 enterprise meeting stakeholders across finance, healthcare, and tech sectors. 

Second, we introduce MeetBench-XL: a multi-dimensional evaluation protocol calibrated against human judgment (Pearson r=0.78, p<0.01) that measures factual fidelity, user intent alignment, response efficiency, structural clarity, and completeness—criteria prioritized by enterprise users in our needs assessment. 

Third, we propose MeetMaster-XL: a learned dual-policy agent that jointly optimizes routing (fast Talker vs. slow Planner) and tool invocation (retrieval, cross-meeting aggregation, web search) through offline policy learning on quality–latency–cost objectives. Unlike prior rule-based fast/slow systems, our lightweight classifier (1ms overhead) achieves 81.2% routing accuracy and unlocks a superior Pareto frontier: 30% latency reduction on simple queries and 12.4% quality gain on complex tasks over single-model baselines, while remaining deployable on consumer GPUs (RTX 4090, 22GB VRAM). 

We benchmark against commercial APIs (GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro) and provide ablations, robustness tests (noise, overlap, multilingual splits), and a field deployment case study. Resources: https://github.com/huyuelin/MeetBench.

workshop paper

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-26 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to AAAI 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://aaai.org/conference/aaai/aaai-26/registration/) first.

**Online Registration Form**: https://aaai.getregistered.net/conference-2026 

Registration Required

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

Performance evaluation of Retrieval-Augmented Generation (RAG) systems within enterprise environments is governed by multi-dimensional and composite factors extending far beyond simple final accuracy checks. These factors include reasoning complexity, retrieval difficulty, the diverse structure of documents, and stringent requirements for operational explainability. Existing academic benchmarks fail to systematically diagnose these interlocking challenges, resulting in a critical gap where models achieving high performance scores fail to meet the expected reliability in practical deployment. 

To bridge this discrepancy, this research proposes a multi-dimensional diagnostic framework by defining a four-axis difficulty taxonomy and integrating it into an enterprise RAG benchmark to diagnose potential system weaknesses.

Overcoming the‘Impracticality’of RAG: Proposing a Real-World Benchmark and Multi-Dimensional Diagnostic Framework

Despite recent advances, autonomous agents often struggle to solve complex tasks in enterprise domains that require coordinating multiple tools and processing diverse data sources. This struggle is driven by two main limitations. First, single-agent architectures enforce a monolithic plan-execute loop, which directly causes trajectory instability. Second, the requirement to use local open-weight models for data privacy introduces smaller context windows leading to the rapid consumption of context from large tool outputs. To solve this problem we introduce RP-ReAct (Reasoner Planner-ReAct), a novel multi-agent approach that fundamentally decouples strategic planning from low-level execution to achieve superior reliability and efficiency. RP-ReAct consists of a Reasoner Planner Agent (RPA), responsible for planning each sub-step, continuously analysing the execution results using the strong reasoning capabilities of a Large Reasoning Model, and one or multiple Proxy-Execution Agent (PEA) that translates sub-steps into concrete tool interactions using a ReAct approach. Crucially, we incorporate a context-saving strategy within the PEA to mitigate context window overflow by managing large tool outputs via external storage and on-demand access. We evaluate RP-ReAct, on the challenging, multi-domain ToolQA benchmark using a diverse set of six open-weight reasoning models. Our empirical results show that RP-ReAct achieves superior performance and improved generalization ability over state-of-the-art baselines when addressing diverse complex tasks across the evaluated domains. Furthermore we establish the enhanced robustness and stability of our approach across different model scales, paving the way for effective and deployable agentic solutions for enterprises.

Reason-Plan-ReAct: A Reasoner-Planner Supervising a ReAct Executor for Complex Enterprise Tasks

Modern enterprise systems exhibit complex interdependencies that make observability and incident response increasingly challenging. Manual alert triage, which typically involves log inspection, API verification, and cross-referencing operational knowledge bases, remains a major bottleneck in reducing mean recovery time (MTTR). This paper presents an agentic observability framework deployed within Adobe’s e-commerce infrastructure that autonomously performs alert triage using a ReAct paradigm. 

Upon alert detection, the agent dynamically identifies the affected service, retrieves and analyzes correlated logs across distributed systems, and plans context-dependent actions such as handbook consultation, runbook execution, or retrieval-augmented analysis of recently deployed code. Empirical results from production deployment indicate a 90% reduction in mean time to insight compared to manual triage, while maintaining comparable diagnostic accuracy. 

Our results show that agentic AI enables an order-of-magnitude reduction in triage latency and a step-change in resolution accuracy, marking a pivotal shift toward autonomous observability in enterprise operations.

Agentic Observability: Automated Alert Triage for Adobe E-Commerce

This paper presents a multi-agent architecture for adaptive skill assessment using large language models (LLMs). The proposed system integrates three coordinated components-agent generation, dialog-based interaction, and multi-criteria evaluation-to enable personalized, interpretable, and scalable assessments through conversational testing. A central innovation is the use of meta-agents that construct contextualized, persona-driven profiles of virtual examiners and judges according to test objectives and evaluation criteria. Unlike conventional single-agent systems, our architecture distributes responsibilities across LLMs of different scales: large reasoning-oriented models perform context generation and final scoring, while lightweight models handle real-time user interaction. Experiments demonstrate measurable improvements in dialogue coherence, assessment stability, and interpretability compared to baseline systems lacking pre-generated context. The results suggest that multi-agent coordination and hybrid LLM reasoning can substantially advance transparent, efficient, and adaptive AI-based training and evaluation.

Multi-Agent AI Trainer: Adaptive Skill Evaluation via Persona-Driven Examiners and Multi-Criteria Judging

In industrial environments, time series sensor data are a primary resource for monitoring equipment health and performance, with alerts often triggered by manually defined heuristic rules. Deploying these rules in production can be a labour-intensive and error-prone process. We present an agentic system that assists reliability engineers by autonomously translating expert-defined monitoring rules, expressed in natural language, into runnable code and executing that code over the target data streams. Evaluations on 41 expert-authored rules from 30 industrial sites show that model performance generally improves with scale, but that consistency, efficiency, and accuracy are significantly enhanced when agents are more highly structured. These results position LLM-based agents as a promising direction for industrial monitoring systems.

Agentic Code Generation for Heuristic Rules in Equipment Monitoring

Enterprise back-office workflows require agentic systems that are auditable, policy-aligned, and operationally predictable, capabilities that generic multi-agent setups often fail to deliver. We present POLARIS (Policy-Aware LLM Agentic Reasoning for Integrated Systems), a governed orchestration framework that treats automation as typed plan synthesis and validated execution over LLM agents. A planner proposes structurally diverse, type-checked directed acyclic graphs (DAGs); a rubric-guided reasoning module selects a single compliant plan; and execution is guarded by validatorgated checks, a bounded repair loop, and compiled policy guardrails that block or route side effects before they occur. Applied to document-centric finance tasks, POLARIS produces decision-grade artifacts and full execution traces while reducing human intervention. Empirically, POLARIS achieves a micro-F1 of 0.81 on the SROIE dataset and, on a controlled synthetic suite, achieves 0.95–1.00 precision for anomaly routing with preserved audit trails. These evaluations constitute an initial benchmark for governed Agentic AI. POLARIS provides a methodological and benchmark reference for policy-aligned Agentic AI.

POLARIS: Typed Planning and Governed Execution for Agentic AI in Back-Office Automation

In today’s fast-paced environment, the ability to swiftly access, understand, and act on data is no longer optional—it is essential. Yet most organizations remain data-rich but insight-poor, constrained by the complexity of querying, interpreting, and explaining enterprise-scale information. We present POLARIS, a supervisor-led multi-agent framework for conversational enterprise analytics that bridges this gap. POLARIS introduces Dynamic Task Coordination (DTC), a decision-theoretic orchestration layer that models agent–task assignment as adaptive bipartite matching, enabling real-time coordination, recovery, and optimization across specialized agents for querying, visualization, and reasoning. By coupling DTC with reason-first, ReAct-style agents, POLARIS transforms natural language queries into coherent analytical workflows that not only retrieve and visualize data but also explain the underlying “why.” Evaluation on structured enterprise datasets demonstrates high semantic fidelity and answer relevancy, underscoring the potential of multi-agent orchestration to deliver trustworthy, end-to-end business intelligence at scale.

Polaris : Multi Agentic System for Conversational Enterprise Analytics

Tool calling enables Large Language Models (LLMs) to interact with external environments via tool invocation, providing a practical mechanism to overcome the inherent limitations of pretraining. However, the effectiveness of tool use depends critically on the quality of associated documentation and knowledge base context, which are typically authored for human users and often misaligned with LLMs' interpretive needs. This issue is further amplified in industrial settings, where hundreds of tools with overlapping functionalities introduce challenges of scalability, variability, and ambiguity. 

We propose Verification-Guided Context Optimization (VGCO), a framework that employs LLMs-as-editors to automatically refine tool-related documentation and knowledge base context. VGCO operates in two stages: (1) Evaluation, which collects real-world failure cases and identifies tool-context mismatches; and (2) Optimization, which performs hierarchical editing through offline learning with structure-aware, in-context optimization. 

The novelty of our LLM editors lies in three key aspects: (1) a hierarchical structure that integrates naturally into the tool-calling workflow; (2) a state-aware, action-specific, and verification-guided design that constrains the search space for efficient, targeted improvements; and (3) the potential for cost-efficient sub-task specialization through either prompt engineering large editor models or post-training smaller editor models. Unlike prior work emphasizing multi-turn reasoning, VGCO targets the single-turn large scale tool-calling problem, achieving significant gains in accuracy, robustness, and generalization across LLMs.

Verification-Guided Context Optimization for Tool Calling via Hierarchical LLMs-as-editors

As multimodal LLMs become tool-using agents, the field still lacks a standardized metric for translating visual inputs into correct tool invocations. We introduce MFCL Vision, the first large-scale benchmark for vision-based function calling, comprising 250 expert-verified tasks across five image domains (Places, Events, Media, Sports, Shopping) and five query types (Locate, Temporal, Select, Identify, Quantify). Each task comprises (1) a textual user query, (2) an accompanying image, (3) a ground-truth answer obtained from the web, and (4) a human-produced reasoning trace for comparative error analysis. To constrain the task, we expose a singular web-search tool to each model. To examine the robustness of LLM’s perception-to-tool-use pipeline, we introduce controlled visual perturbations, including crops, resizes, and color channel removal. Our automatic grader computes exact-match scores on model final answers, removing dependence on brittle and potentially biased LLM judges. We evaluate leading models and present a taxonomy of failure modes, including visual reasoning, assumption bias, keyword selection, and tool avoidance errors. By releasing MFCL Vision’s dataset, taxonomy, and diagnostics, we aim to accelerate progress towards versatile multimodal agents capable of intelligent tool usage in complex visual contexts.

MFCL Vision: Benchmarking Tool Use in Multimodal Large Language Models for Visual Reasoning Tasks

Evaluating generative AI (GenAI) products for compliance in high-stakes domains such as healthcare is difficult across multi-turn conversations and diverse contexts. This is especially true and pressing for enterprises evaluating human-facing AI Chatbots which can output life-threatening content, such as responses supporting self-harm. Existing approaches for automated compliance focus on single-agent approaches (e.g. prompt tuning). These are difficult to adopt in practice because they diverge from established enterprise compliance workflows and lack mechanisms for uncertainty quantification (UQ). This paper introduces a system for converting any compliance workflow into a human-in-the-loop, multi-agent framework. The approach maps process roles to specialized agents (e.g., content, business, legal), enables escalation between agents analogous to human review procedures, and enables incorporating human input for uncertain classifications. Methodologically, this system differs from more conventional single-orchestrator/supervisor designs in multi-agent settings by formulating the process as a constrained Markov Decision Process (MDP) and using Monte Carlo simulations to estimate empirical agent uncertainty. We demonstrate this framework on an open-source safety benchmark for self-harm detection. Results show improvements over a single-agent baseline in accuracy (up to 19%), reduction in required human review (up to 85×), and, in some configurations, less processing time. The framework also identified several mislabeled items in the open source benchmarks. We reported suspected label errors to the maintainers, several were acknowledged as likely errors and candidates for correction, emphasizing the timeliness and promise of this approach.

Next from AAAI 2026

Overcoming the‘Impracticality’of RAG: Proposing a Real-World Benchmark and Multi-Dimensional Diagnostic Framework

Similar lecture

Can LLMs Help You at Work? A Sandbox for Evaluating LLM Agents in Enterprise Environments