FARM: Field-Aware Resolution Model for Intelligent Trigger-Action Automation

TL;DR

The work tackles the challenge of generating executable Trigger-Action applets from natural language by addressing the function-level binding gap that plagues service-level TAP approaches. It introduces FARM, a two-stage framework that first retrieves high-quality trigger and action candidates using contrastive, schema-enriched dual encoders with layer freezing, then employs a four-agent, LLM-based pipeline to analyze intent, select a trigger and an action, and generate precise ingredient-to-field bindings. Quantitative results show Stage 1 recall at rank 5 exceeding 92% for both encoders and a Stage 2 joint accuracy of 81% on Gold data, outperforming baselines by significant margins and yielding executable applets with bindings. The approach demonstrates robustness across Gold, Noisy, and One-Shot settings, with strong domain alignment (Topic Adherence) and moderate schema grounding (Faithfulness), and offers practical efficiency through a staged design and minimal trainable parameters. Overall, FARM advances TAP automation by delivering complete, executable configurations at function-level scale, enabling scalable and user-intent-aligned WoT automations without heavy manual intervention.

Abstract

Trigger-Action Programming (TAP) platforms such as IFTTT and Zapier enable Web of Things (WoT) automation by composing event-driven rules across heterogeneous services. A TAP applet links a trigger to an action and must bind trigger outputs (ingredients) to action inputs (fields) to be executable. Prior work largely treats TAP as service-level prediction from natural language, which often yields non-executable applets that still require manual configuration. We study the function-level configuration problem: generating complete applets with correct ingredient-to-field bindings. We propose FARM (Field-Aware Resolution Model), a two-stage architecture for automated applet generation with full configuration. Stage 1 trains contrastive dual encoders with selective layer freezing over schema-enriched representations, retrieving candidates from 1,724 trigger functions and 1,287 action functions (2.2M possible trigger-action pairs). Stage 2 performs selection and configuration using an LLM-based multi-agent pipeline. It includes intent analysis, trigger selection, action selection via cross-schema scoring, and configuration verification. Agents coordinate through shared state and agreement-based selection. FARM achieves 81% joint accuracy on Gold (62% Noisy, 70% One-shot) at the function level, where both trigger and action functions must match the ground truth. For comparison with service-level baselines, we map functions to their parent services and evaluate at the service level. FARM reaches 81% joint accuracy and improves over TARGE by 23 percentage points. FARM also generates ingredient-to-field bindings, producing executable automation configurations.

Figures (14)

• Figure 1: An example of an TAP applet interface for manual configuration and our proposal for an automated applet generation.
• Figure 2: IFTTT applet execution pipeline with credential boundaries.
• Figure 3: FARM two-stage architecture. Stage 1: Dual contrastive encoders retrieve top-$k$ trigger and action candidates; cross-score matrix ranks $k \times k$ pairs by compatibility scores (70% ingredient-field coverage, 30% retrieval quality). Stage 2: Field-aware selection pipeline performs deep ingredient-to-field analysis—Intent Analyzer decomposes query intent, Trigger Selector identifies optimal trigger $t^*$ and extracts its ingredient schema $I_t$ (available output fields), Action Selector evaluates field compatibility and generates bindings $\beta$ mapping ingredients to action input fields, and Verifier scores binding completeness and semantic coherence. Fallback retries with next candidate pair when verification score $s < \theta$.
• Figure 4: Contrastive training pipeline for dual encoders. (a) Training data construction: query-function pairs extracted from IFTTT dataset. (b) Dual encoder architecture with layer freezing—trainable layers (orange, 18%) adapt to domain-specific patterns while frozen layers (blue-gray, 82%) preserve pretrained semantics. (c) In-batch negative sampling: diagonal entries are positive pairs, off-diagonal entries serve as hard negatives. (d) InfoNCE objective in embedding space: queries are pulled toward positive documents and pushed away from negatives.
• Figure 5: Stage 2 multi-agent selection components. (a) Agent structure showing input query, LLM processing with retrieved candidates $C_k$, and structured JSON output. (b) State-based pipeline where Intent Analyzer, Trigger Selector, Action Selector, and Verifier communicate through shared state. (c) Ingredient-to-field binding illustrating how trigger outputs (Subject, Body) dynamically populate action inputs, with static user-provided values for required fields.