Singapore

As organizations scale adoption of generative AI, model cost optimization and operational efficiency have emerged as critical factors determining sustainability and accessibility. While Large Language Models (LLMs) demonstrate impressive capabilities across diverse tasks, their extensive computational requirements make them cost prohibitive for routine enterprise use. This limitation motivates the exploration of Small Language Models (SLMs), which can deliver comparable performance in targeted applications while drastically reducing infrastructure overhead (Irugalbandara et al.,2023). In this work, we investigate the feasibility of replacing LLM-driven workflows with optimized SLMs. We trained a domain-adapted SLM to execute representative tasks traditionally handled by LLMs, such as document summarization, query answering, and structured data interpretation. As part of the experiment, we investigated the fine-tuning of facebook opt-350m model (single epoch only) using the Hugging Face TRL (Transformer Reinforcement Learning), specifically the Supervised Fine Tuning (SFT) trainer. The OPT-350M model was released by Meta AI in 2022 as part of the OPT (Open Pretrained Transformer) family of models. Similar studies demonstrate that even models at the 350M parameter scale can meaningfully contribute to instruction tuning pipelines (Mekala et al., 2024). Experimental results demonstrated that our fine-tuned SLM achieves exceptional performance with a 77.55% pass rate on ToolBench evaluation, significantly outperforming all baseline models including ChatGPT-CoT (26.00%), ToolLLaMA-DFS (30.18%), and ToolLLaMA-CoT (16.27%). These benchmarks, first introduced in ToolLLM (Qin et al., 2023) and later stabilized by follow-up efforts (Zhang et al., 2024), have become the standard for evaluating tool-augmented reasoning. Recent work has also extended ToolBench traces to preference-based optimization (Zeng et al., 2024) and designed alternative multi API corpora for tool-use robustness (Liu et al., 2024). These findings emphasize that thoughtful design and targeted training of SLMs can significantly lower barriers to adoption, enabling cost-effective, large-scale integration of generative AI into production systems.

AAAI 2026

Small Language Models for Efficient Agentic Tool Calling: Outperforming Large Models with Targeted Fine-tuning

machine learning

transfer

domain adaptation

multi-task learning

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.<br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>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.

workshop paper

Computer-Aided Design (CAD) documentation is fundamental to ensuring engineering integrity. However, its verification currently depends on the manual inspection of specifications, drawings, and 3D models. This practice is inherently prone to inefficiency, inconsistency, and undetected errors. To address this gap, we present the CAD Inspection Assistant, a tool-augmented agentic system designed to automate CAD-related document specification verification. At its core, the assistant leverages an agentic Large Language Model (LLM) as a dynamic planner, decomposing complex inspection tasks into executable steps. It then invokes specialized Python tools to extract and analyze data from CAD artifacts. The LLM synthesizes tool outputs to evaluate compliance with project requirements, generating actionable reports that highlight discrepancies. To demonstrate the effect of our work, we collect real-world data and conduct experiments, setting a foundation for intelligent agentic systems in manufacturing workflows.

CAD Inspection Assistant: Tool-Augmented Agentic CAD Inspection Solution

Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in cross-modal understanding, yet they remain prone to hallucinations—particularly those driven by overreliance on language and knowledge priors while neglecting visual evidence. We identify and formalize a critical subclass of such hallucinations, termed EvidenCe-prior Hallucination Observation (ECHO), which arises when models generate answers consistent with prior knowledge but inconsistent with manipulated visual inputs. To address the lack of fine-grained, scalable benchmarks for this phenomenon, we introduce the first dedicated ECHO evaluation framework, featuring a triplet-structured dataset of 414 samples (1,242 QA records) across seven domains, constructed via systematic filtering, adversarial image editing, and rigorous annotation. Our benchmark enables precise diagnosis of language-prior dependency through three controlled conditions: Text-QA, Raw-VQA, and Edit-VQA. Evaluations on both titan-class (e.g., GPT-4o, Command A Vision) and workhorse-class (e.g., Qwen2.5-VL) models reveal that stronger language priors correlate with higher ECHO susceptibility. Furthermore, we propose and validate two model-agnostic mitigation strategies—attention-guided evidence region enhancement and reinforcement learning–activated test-time reasoning—which significantly reduce ECHO rates (e.g., ECHO-$\phi$ drops from 6.3\% to 4.8\% in Qwen2.5-VL-3B). Our work provides a reproducible benchmark, interpretable metrics (ECHO--$\Delta$, ECHO--$\phi$, ECHO-$F$), and practical pathways toward more faithful multimodal reasoning.

ECHO: EvidenCe-prior Hallucination Observation

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.

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

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.

Next from AAAI 2026

CAD Inspection Assistant: Tool-Augmented Agentic CAD Inspection Solution

Similar lecture

Scale Regularization for Stable Low-Rank Adaptation