Multi-Agent Reasoning with Adaptive Worker Allocation for Stance Detection

# Multi-Agent Reasoning with Adaptive Worker Allocation for Stance Detection
Source: [https://arxiv.org/html/2606.11609](https://arxiv.org/html/2606.11609)
Meysam Sabbaghan1, Arman Zareian Jahromi1, Doina Caragea1 1Department of Computer Science Kansas State University

###### Abstract

Stance detection requires identifying an author’s position toward a target, often from short\-form texts where stance is implicit, indirect, or rhetorically framed\. Although large language models \(LLMs\) achieve strong performance on this task, single\-pass prompting can be brittle when multiple interpretations are plausible\. Existing aggregation strategies, such as majority voting or self\-consistency, improve robustness by combining labels, but they discard the intermediate reasoning needed to resolve conflicting interpretations\.

We introduce a multi\-agent reasoning framework with adaptive worker allocation for stance detection that shifts aggregation from label\-level voting to reasoning\-level synthesis\. The framework employs a Manager–Worker architecture in which a Manager adaptively allocates a variable number of Worker agents based on input complexity\. Each Worker analyzes the input from a distinct perspective and produces a reasoning\-only explanation without emitting a stance label; the Manager then synthesizes these explanations to produce the final prediction\.

We evaluate the proposed framework on SemEval\-2016, P\-Stance, and COVID\-19 Stance using Llama, Mistral, and Gemini\. Results show that the framework yields the largest gains on implicit and context\-dependent stance cases, achieving 86\.07 Macro\-F1 on COVID\-19 and 82\.90 on SemEval\-2016, while remaining competitive on more explicit stance datasets such as P\-Stance\. These findings suggest that adaptive reasoning\-level aggregation is most beneficial when stance cannot be reliably inferred from surface cues alone\.

Multi\-Agent Reasoning with Adaptive Worker Allocation for Stance Detection

Meysam Sabbaghan1, Arman Zareian Jahromi1, Doina Caragea11Department of Computer ScienceKansas State University

## 1Introduction

Stance detection aims to determine an author’s position toward a specific target, such as a public figure, policy, or ideological conceptMohammadet al\.\([2016](https://arxiv.org/html/2606.11609#bib.bib1)\); Sobhaniet al\.\([2016](https://arxiv.org/html/2606.11609#bib.bib5)\)\. The task is especially challenging in short\-form online texts, where stance is often conveyed implicitly through sarcasm, rhetorical framing, indirect references, or attacks on related entities rather than through explicit statementsLiet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib13)\); Glandtet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib11)\); Bamman and Smith \([2015](https://arxiv.org/html/2606.11609#bib.bib34)\); Riloffet al\.\([2013](https://arxiv.org/html/2606.11609#bib.bib33)\)\.

Unlike sentiment analysis, stance detection requires target\-aware pragmatic reasoning: the same expression may imply different stances depending on the target\. For example, criticizing a political opponent may indicate support for the target, while an apparently positive statement may signal opposition when used sarcasticallyMohammadet al\.\([2016](https://arxiv.org/html/2606.11609#bib.bib1)\); Augensteinet al\.\([2016](https://arxiv.org/html/2606.11609#bib.bib3)\)\. These cases are difficult for models that rely primarily on surface cues or single\-pass interpretations\.

Large language models \(LLMs\) have demonstrated strong few\-shot and zero\-shot performance on stance detectionBrownet al\.\([2020](https://arxiv.org/html/2606.11609#bib.bib10)\); Weiet al\.\([2022](https://arxiv.org/html/2606.11609#bib.bib41)\); Kojimaet al\.\([2022](https://arxiv.org/html/2606.11609#bib.bib19)\)\. However, single\-pass prompting produces only one interpretation of an input, making predictions brittle when stance is implicit, ambiguous, or rhetorically complex\. Aggregation methods such as self\-consistencyWanget al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib18)\)can improve robustness by combining multiple outputs, but they typically aggregate at the label level, discarding the intermediate reasoning that could help resolve conflicting interpretations\.

Recent work on structured reasoning and multi\-agent collaboration suggests a promising alternative\. Reasoning\-oriented prompting methods, including Chain\-of\-ThoughtWeiet al\.\([2022](https://arxiv.org/html/2606.11609#bib.bib41)\), Least\-to\-Most promptingZhouet al\.\([2023a](https://arxiv.org/html/2606.11609#bib.bib20)\), Tree\-of\-ThoughtsYaoet al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib21)\), and Graph\-of\-ThoughtsBestaet al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib29)\), decompose complex problems into intermediate reasoning steps\. Multi\-agent frameworks further encourage diverse and complementary reasoning paths by assigning different agents to distinct roles or perspectivesDuet al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib26)\); Lanet al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib22)\); Wuet al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib23)\); Liet al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib24)\); Honget al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib25)\)\. However, most existing approaches either aggregate final labels or use fixed reasoning structures that do not adapt reasoning effort to instance\-level complexity\.

We proposeSMART\-D, a structured multi\-agent reasoning framework for stance detection that performs*reasoning\-level synthesis*rather than label\-level voting\. SMART\-D uses a Manager–Worker architecture in which Worker agents analyze the input from complementary perspectives and produce reasoning\-only explanations without emitting stance labels\. A Manager agent then synthesizes these explanations to produce the final stance prediction\. To avoid applying the same reasoning budget to every instance, the Manager dynamically allocates the number of Workers based on input complexity, assigning more reasoning capacity to ambiguous or rhetorically complex cases and fewer Workers to more explicit cases\. We evaluate SMART\-D on three stance detection benchmarks: SemEval\-2016Mohammadet al\.\([2016](https://arxiv.org/html/2606.11609#bib.bib1)\), P\-StanceLiet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib13)\), and COVID\-19 StanceGlandtet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib11)\), using three LLMs: Llama, Mistral, and Gemini\. Our results show that SMART\-D provides the largest gains on implicit and context\-dependent stance cases, particularly on COVID\-19 Stance, while remaining competitive on datasets with more explicit stance signals such as P\-Stance\.

Our main contributions are as follows:

- •We introduce SMART\-D, a Manager–Worker framework that shifts stance aggregation from final\-label voting to reasoning\-level synthesis\.
- •We propose dynamic Worker allocation, allowing the model to adapt reasoning depth to input complexity rather than using a fixed number of reasoning paths for all instances\.
- •We provide empirical evidence across three stance benchmarks and three LLM backbones showing that adaptive reasoning\-level aggregation is most beneficial for implicit, ambiguous, and rhetorically complex stance cases\.

## 2Related Work

#### Stance Detection and LLM Reasoning\.

Stance detection is commonly formulated as a target\-conditioned text classification taskSomasundaran and Wiebe \([2009](https://arxiv.org/html/2606.11609#bib.bib35)\); Mohammadet al\.\([2016](https://arxiv.org/html/2606.11609#bib.bib1)\); Sobhaniet al\.\([2016](https://arxiv.org/html/2606.11609#bib.bib5)\)\. Early work used lexical, syntactic, and task\-specific featuresSobhaniet al\.\([2016](https://arxiv.org/html/2606.11609#bib.bib5)\); Hasan and Ng \([2013](https://arxiv.org/html/2606.11609#bib.bib2)\), while recent models improved target\-aware representation learningAugensteinet al\.\([2016](https://arxiv.org/html/2606.11609#bib.bib3)\); Zarrella and Marsh \([2016](https://arxiv.org/html/2606.11609#bib.bib4)\); Duet al\.\([2017](https://arxiv.org/html/2606.11609#bib.bib6)\); Devlinet al\.\([2019](https://arxiv.org/html/2606.11609#bib.bib7)\); Allaway and McKeown \([2020](https://arxiv.org/html/2606.11609#bib.bib8)\); Lianget al\.\([2022](https://arxiv.org/html/2606.11609#bib.bib17)\)\. Recent datasets have expanded stance detection to political and public\-health domainsLiet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib13)\); Glandtet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib11)\); Confortiet al\.\([2020](https://arxiv.org/html/2606.11609#bib.bib9)\), where implicit stance, domain shift, and target dependence remain persistent challengesHardalovet al\.\([2022](https://arxiv.org/html/2606.11609#bib.bib14)\)\.

LLMs provide a promising direction for such cases because they can generate intermediate reasoning before prediction\. Chain\-of\-Thought prompting, self\-consistency, Least\-to\-Most prompting, Tree\-of\-Thoughts, and Graph\-of\-Thoughts show that decomposed or sampled reasoning can improve performance on complex tasksWeiet al\.\([2022](https://arxiv.org/html/2606.11609#bib.bib41)\); Kojimaet al\.\([2022](https://arxiv.org/html/2606.11609#bib.bib19)\); Wanget al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib18)\); Zhouet al\.\([2023a](https://arxiv.org/html/2606.11609#bib.bib20)\); Yaoet al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib21)\); Bestaet al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib29)\)\. In stance detection, recent methods incorporate reasoning, verification, rationale generation, or expert modeling to improve robustness and interpretabilityTaranukhinet al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib28)\); Zhanget al\.\([2026](https://arxiv.org/html/2606.11609#bib.bib32)\); Wanget al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib27)\); Yanet al\.\([2025](https://arxiv.org/html/2606.11609#bib.bib31)\); Garg and Caragea \([2024](https://arxiv.org/html/2606.11609#bib.bib30)\)\. However, many approaches still aggregate final predictions or use fixed reasoning structures, limiting their ability to compare competing interpretations explicitly\.

#### Multi\-Agent Reasoning and Adaptive Inference\.

Multi\-agent LLM frameworks improve robustness by distributing reasoning across multiple agents or perspectives\. Debate\-based methods explore competing interpretationsDuet al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib26)\), collaborative systems assign specialized roles to agentsLanet al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib22)\), and general frameworks such as AutoGen, CAMEL, and MetaGPT support coordinated agent interactionWuet al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib23)\); Liet al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib24)\); Honget al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib25)\)\. Related ideas also appear in Mixture\-of\-Experts and dynamic routing models, which route inputs to specialized modules for capacity and efficiencyShazeeret al\.\([2017](https://arxiv.org/html/2606.11609#bib.bib36)\); Lepikhinet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib37)\); Duet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib38)\)\. In stance detection, multi\-expert methods such as MSME and DEEM apply expert selection or routing mechanismsZhanget al\.\([2026](https://arxiv.org/html/2606.11609#bib.bib32)\); Wanget al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib27)\)\.

These approaches motivate adaptive computation, but most route at the representation, expert, or prediction level, rather than at the reasoning level\. Similarly, inference\-time methods such as prompt optimization and self\-consistency show that performance depends on reasoning depth, sampling, and computational budgetZhouet al\.\([2023b](https://arxiv.org/html/2606.11609#bib.bib39)\); Pryzantet al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib40)\); Wanget al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib18)\), but typically apply a fixed configuration across all inputs\.

SMART\-D differs by treating reasoning itself as the object of aggregation and allocation\. Worker agents produce reasoning\-only explanations from complementary views, while a Manager synthesizes these explanations to predict the final stance label\. The Manager also dynamically allocates Workers based on input complexity, assigning more reasoning capacity to implicit or rhetorically complex cases and fewer Workers to more explicit ones\.

## 3Methodology

### 3\.1Overview

We formulate stance detection as a target\-conditioned classification taskMohammadet al\.\([2016](https://arxiv.org/html/2606.11609#bib.bib1)\)\. Given an input textxx, a stance targettt, and a dataset identifierdd, the goal is to predict a labely∈𝒴dy\\in\\mathcal\{Y\}\_\{d\}, where𝒴d\\mathcal\{Y\}\_\{d\}is the dataset\-specific label space \(e\.g\.,favor/againstorfavor/against/none\)\.

SMART\-D organizes LLM calls into a structured Manager–Worker architecture\. Unlike label\-level aggregation methods, such as majority voting, self\-consistency, or debate\-based selectionWanget al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib18)\); Duet al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib26)\); Lanet al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib22)\), SMART\-D aggregates intermediate explanations\. Worker agents analyze the input from complementary perspectives and produce concise, evidence\-grounded reasoning without emitting stance labels\. A Manager agent then synthesizes these explanations to produce the final prediction\. This design is motivated by prior work showing that structured intermediate reasoning can improve robustness on complex reasoning tasksWanget al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib18)\); Yaoet al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib21)\); Bestaet al\.\([2024](https://arxiv.org/html/2606.11609#bib.bib29)\)\.

Figure[1](https://arxiv.org/html/2606.11609#S3.F1)illustrates the overall workflow: the Manager first plans the reasoning process and allocates Workers, the Workers generate reasoning\-only explanations, and the Manager performs reasoning\-level aggregation to predict final labels\.

![Refer to caption](https://arxiv.org/html/2606.11609v1/figures/MAstance_downloadable_3.png)Figure 1:Overview of the SMART\-D framework\. The Manager analyzes the input and dynamically allocates Worker agents with complementary reasoning roles\. Workers produce reasoning\-only outputs without emitting stance labels, and the Manager synthesizes these outputs to produce the final stance label\.
### 3\.2Manager–Worker Reasoning

SMART\-D proceeds in three stages\.

#### 1\. Manager planning\.

Given\(x,t,d\)\(x,t,d\), the Manager identifies the dataset\-specific label space𝒴d\\mathcal\{Y\}\_\{d\}, estimates the reasoning complexity of the instancexx, selects the number of Workers, and assigns each Worker a distinct analytical role\. Roles are designed to capture complementary stance signals, such as explicit stance cues, sarcasm or irony, indirect stance through related entities, emotional tone, and contextual framing\.

#### 2\. Worker reasoning\.

Each Worker analyzes the same input\(x,t,d\)\(x,t,d\)under its assigned perspective\. Workers are constrained to produce short, evidence\-grounded explanations and are explicitly prohibited from outputting stance labels\. This*reasoning\-only constraint*prevents premature commitment to a label and preserves intermediate evidence for downstream synthesis\.

#### 3\. Manager aggregation\.

The Manager receives all Worker explanations and predicts a single stance label by comparing the consistency, relevance, and evidential support of the explanations\. This allows the model to reconcile competing interpretations, which is particularly important when stance is implicit, ambiguous, or rhetorically framed\.

### 3\.3Dynamic Worker Allocation

A fixed number of reasoning paths may be ineffective: too few Workers can miss important interpretations, while too many can introduce redundancy, conflicting evidence, and additional cost\. This trade\-off is consistent with prior work on multi\-path reasoning, where additional reasoning samples do not always yield monotonic gainsWanget al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib18)\); Yaoet al\.\([2023](https://arxiv.org/html/2606.11609#bib.bib21)\)\. SMART\-D therefore treats reasoning depth as an adaptive computational resource\. For each instance, the Manager selectsN∈\{3,…,7\}N\\in\\\{3,\\ldots,7\\\}Workers based on input complexity\. Inputs with explicit stance cues are assigned fewer Workers, while cases involving sarcasm, indirect references, multiple entities, or complex rhetorical structure receive additional perspectives\.

### 3\.4Implementation Details

We implement SMART\-D through structured prompts that specify Manager and Worker roles, enforce the reasoning\-only constraint, and encourage complementary reasoning perspectives\. The Manager operates in two prompt stages: a planning stage that allocates Workers and a decision stage that synthesizes Worker explanations into a final label\. All experiments are conducted in a zero\-shot prompting setting without task\-specific fine\-tuning or in\-context examples\. All LLM inference was performed through the OpenRouter API using a unified OpenAI\-compatible interface\. Workers use a higher temperature \(T=0\.7T=0\.7\) to encourage diverse reasoning, while the Manager uses a lower temperature \(T=0\.3T=0\.3\) to promote stable aggregation\. All other decoding parameters are kept at their default values unless otherwise specified\. Full prompt templates used for Manager/Workers are provided in Appendix[A](https://arxiv.org/html/2606.11609#A1)\.

The workflow logs selected Worker counts, Worker roles, generated explanations, and final predictions for reproducibility and qualitative analysis\. All prompts, configurations, and evaluation scripts will be released to support reproducibility\.

## 4Experimental Setup

### 4\.1Models

We evaluate SMART\-D using three LLM backbones:Gemini\-3\-Flash\-preview,Llama\-3\.3 70B, andMistral\-3\.2 24B\. For consistency, we refer to these models as Gemini, Llama, and Mistral throughout the paper\. These models span different parameter scales and capability profiles, allowing us to assess whether reasoning\-level aggregation generalizes across heterogeneous LLMs\. For each experiment, the same LLM is used for both the Manager and Worker agents\. Manager and Worker behaviors differ only through role\-specific prompts and decoding settings, rather than through different model architectures or fine\-tuning\. This design isolates the effect of the SMART\-D reasoning framework from heterogeneous\-model effects, which we leave to future work\.

### 4\.2Datasets

We evaluate on three stance detection benchmarks: SemEval\-2016 Task 6Mohammadet al\.\([2016](https://arxiv.org/html/2606.11609#bib.bib1)\), P\-StanceLiet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib13)\), and COVID\-19 StanceGlandtet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib11)\)\. These datasets cover both political and public\-health domains and include varying degrees of implicit stance, making them suitable for evaluating reasoning\-oriented approaches\. Some representative dataset examples and target\-label pairs are provided in Appendix[B](https://arxiv.org/html/2606.11609#A2)\.

#### SemEval\-2016 Task 6\.

A target\-conditioned dataset with five targets: Atheism, Climate Change, Feminist Movement, Hillary Clinton, and Legalization of Abortion\. Labels arefavor,against,none\.

#### P\-Stance\.

A political stance dataset with three U\.S\. presidential targets: Donald Trump, Joe Biden, and Bernie Sanders\. The task is binary text classification withfavorandagainstlabels\.

#### COVID\-19 Stance\.

A public\-health stance dataset covering four targets: Anthony Fauci, face masks, stay\-at\-home orders, and school closures\. Labels arefavor,against,none\.

### 4\.3Evaluation Protocol

We reportMacro\-F1andAccuracy\. Macro\-F1 is the primary metric because stance datasets often exhibit class imbalance and require robust performance across minority classes\. Each experiment is repeated three times, and we report the mean and standard deviation\.

Table 1:Macro\-F1 results on the COVID\-19 Stance dataset across four targets: Wearing a Face Mask \(FM\), Anthony Fauci \(AF\), Stay\-at\-Home Orders \(SH\), and Keeping Schools Closed \(SC\)\. Supervised baselines are reported fromGlandtet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib11)\)\. Baseline \(Single\-Model\) and SMART\-D results are computed in our experimental setup; for these methods, results are reported as mean±\\pmstandard deviation over three runs\. Best and second\-best results in each column are shown inboldand underlined, respectively\.

## 5Results and Discussion

We evaluate SMART\-D using Macro\-F1, with accuracy reported in Appendix[C](https://arxiv.org/html/2606.11609#A3)\. Results for prior methods are taken from published work, while single\-model baselines are computed using our own implementation with the same LLM backbones\.

### 5\.1Main Results

Table[1](https://arxiv.org/html/2606.11609#S4.T1)shows Macro\-F1 results for COVID\-19 Stance and Table[2](https://arxiv.org/html/2606.11609#S5.T2)shows results for SemEval\-2016 and P\-Stance\. Overall, SMART\-D provides the largest gains on COVID\-19 Stance, a dataset with many implicit and context\-dependent stance expressions, while achieving smaller but positive gains on SemEval\-2016 and mixed results on P\-Stance\.

#### COVID\-19 Stance\.

On COVID\-19 Stance, SMART\-D improves average Macro\-F1 for all three backbones\. With Gemini, SMART\-D achieves the best overall score of86\.07, improving over the corresponding single\-model baseline by\+3\.99points and outperforming the strongest supervised baseline by more than\+6points\. The gains are especially large for*School Closures*, where Gemini improves from 76\.01 to87\.35\. Llama also shows a substantial average improvement, from 69\.78 to81\.31\. While Llama exhibits a sharp performance drop on the Stay\-at\-Home \(SH\) target \(38\.23 Macro\-F1\)—likely struggling to map indirect policy references to the correct stance—SMART\-D’s structured reasoning paths successfully resolve this ambiguity\. These results support our main hypothesis: reasoning\-level aggregation is most useful when stance depends on rhetorical framing, indirect references, or target\-specific interpretation\. The improvements over the corresponding single\-model baselines are stable across repeated runs; detailed statistical significance tests are reported in Appendix[D](https://arxiv.org/html/2606.11609#A4)\.

CategoryModelSemEval16 \(%\)P\-Stance \(%\)ACCFMHCLAAvg±\\pmstdDTJBBSAvg±\\pmstdZero\-shot \(LLMs\)Base \(GPT\-3\.5\)58\.3051\.1062\.3065\.0060\.8059\.5067\.3078\.2071\.6072\.40CoT64\.1055\.7062\.4070\.7061\.9063\.2071\.4080\.5074\.1075\.30BKEL71\.5066\.0063\.1076\.5064\.2068\.3080\.3078\.3079\.6079\.40Stance Reasoner69\.7062\.5073\.9067\.7060\.3066\.8079\.5081\.0079\.6080\.00COLA70\.8065\.5063\.4081\.7071\.0070\.5086\.6084\.0079\.7083\.40ToC–––––69\.4075\.7083\.1080\.4079\.70MSMEGPT\-3\.575\.2074\.9072\.5081\.1069\.9074\.7087\.7084\.9082\.8085\.10GPT\-4o80\.3076\.2075\.5081\.9071\.9077\.2088\.6085\.6084\.1086\.10DeepSeek\-r181\.5078\.5074\.8080\.6073\.5077\.8087\.1084\.7084\.5085\.40QwQ\-32B79\.5077\.1076\.3076\.9068\.4075\.6085\.1084\.3083\.5084\.30Baseline \(Single\-Model\)Gemini76\.7185\.5879\.5589\.1869\.4680\.10±\\pm0\.0692\.5686\.2884\.0787\.64±\\pm0\.00Llama77\.3087\.2277\.0389\.1970\.4380\.23±\\pm0\.0789\.3586\.8583\.0686\.42±\\pm0\.13Mistral69\.1771\.7780\.4088\.0976\.4677\.18±\\pm0\.0583\.2083\.2779\.1181\.86±\\pm0\.46SMART\-D \(Ours\)Gemini87\.7588\.1681\.0086\.5071\.1382\.77±\\pm0\.7492\.3985\.4982\.6686\.84±\\pm0\.37Llama84\.7491\.5275\.4089\.5673\.2982\.90±\\pm0\.2689\.4785\.7382\.6985\.96±\\pm0\.24Mistral75\.2183\.1378\.4489\.3377\.6780\.76±\\pm0\.4589\.2285\.2381\.9985\.48±\\pm0\.15

Table 2:Macro\-F1 results on SemEval\-2016 Task 6Mohammadet al\.\([2016](https://arxiv.org/html/2606.11609#bib.bib1)\)and P\-StanceLiet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib13)\)\. Prior zero\-shot and MSME results are reported fromZhanget al\.\([2026](https://arxiv.org/html/2606.11609#bib.bib32)\); Baseline \(Single\-Model\) and SMART\-D results are computed in our experimental setup\. SemEval\-2016 targets: A = Atheism, CC = Climate Change, FM = Feminist Movement, HC = Hillary Clinton, LA = Legalization of Abortion\. P\-Stance targets: DT = Donald Trump, JB = Joe Biden, BS = Bernie Sanders\. For our methods, results are reported as mean±\\pmstandard deviation over three runs\. Best and second\-best results in each column are shown inboldand underlined, respectively\.
#### SemEval\-2016\.

On SemEval\-2016, SMART\-D improves average Macro\-F1 for all three backbones\. The best result is obtained with Llama, reaching82\.90, a\+2\.67point improvement over its single\-model baseline\. Gemini and Mistral also improve in average Macro\-F1, although the target\-level effects are not uniform\. SMART\-D also surpasses strong prior methods such as MSME, showing that reasoning\-level aggregation provides benefits beyond standard prompting or multi\-expert strategies\. These results suggest that SMART\-D is beneficial on SemEval\-2016, but the gains are more moderate than on COVID\-19, consistent with a dataset containing both implicit and more lexically explicit stance expressions\.

#### P\-Stance\.

On P\-Stance, SMART\-D remains competitive with strong prior LLM\-based methods, but its gains over single\-model baselines are model\-dependent\. Mistral benefits substantially, improving from 81\.86 to85\.48Macro\-F1\. In contrast, Gemini and Llama decrease slightly relative to their single\-model baselines\. This indicates that additional reasoning is not uniformly beneficial when stance is already expressed explicitly\. Rather, SMART\-D is most effective when the input requires ambiguity resolution or pragmatic interpretation\.

Taken together, these results show that SMART\-D is strongest in the setting for which it was designed: implicit, ambiguous, and rhetorically complex stance detection\. On datasets with more explicit stance cues, the framework remains competitive, but the benefit of additional reasoning depends on the base model and input complexity\.

### 5\.2Ablation and Allocation Analysis

We analyze SMART\-D through ablations and diagnostic allocation studies\. We first examine how Worker count affects performance, then evaluate the role of reasoning\-level synthesis, dynamic allocation behavior, and cross\-model Worker selection\.

#### Effect of the number of Workers\.

We first evaluate fixed Worker configurations withN∈\{1,3,5,7,9,11\}N\\in\\\{1,3,5,7,9,11\\\}and compare them with dynamic allocation\. Table[T4](https://arxiv.org/html/2606.11609#A5.T4)in Appendix[E](https://arxiv.org/html/2606.11609#A5)reports Macro\-F1 across all datasets and backbones\. For brevity, Table[3](https://arxiv.org/html/2606.11609#S5.T3)summarizes the best fixed configuration amongN∈\{1,3,5,7,9,11\}N\\in\\\{1,3,5,7,9,11\\\}and compares it with dynamic allocation\. Overall, performance does not increase monotonically with more Workers\. LargerNNvalues can improve performance, but the gains are inconsistent and come with higher inference cost\. This suggests a trade\-off between*reasoning diversity*and*aggregation noise*: additional Workers provide more perspectives, but can also introduce redundant or conflicting explanations\.

The dynamic strategy is constrained toN∈\{3,⋯,7\}N\\in\\\{3,\\cdots,7\\\}to reflect a practical reasoning budget\. While it is not always the best\-performing configuration, it provides a strong compromise between performance and cost\. On COVID\-19, dynamic allocation performs especially well, achieving the best score for Gemini and near\-best performance for Llama\. On P\-Stance and SemEval\-2016, fixed larger budgets sometimes perform better, suggesting that the optimal amount of reasoning depends on both dataset and backbone\. Detailed analyses of token cost are provided in Appendix[F](https://arxiv.org/html/2606.11609#A6)\.

#### Dynamic allocation behavior\.

To examine how the Manager uses the dynamic allocation mechanism, we analyze the distribution of selected Worker counts for two representative settings: Gemini on COVID\-19, which has the lowest average allocation \(N¯=3\.60\\bar\{N\}=3\.60\), and Mistral on P\-Stance, which has the highest average allocation \(N¯=5\.13\\bar\{N\}=5\.13\)\. Figure[2](https://arxiv.org/html/2606.11609#S5.F2)shows that allocation is structured rather than uniform\. In the COVID\-19 setting, the Manager primarily selects smaller Worker counts \(N=3N=3–44\), whereas in the P\-Stance setting selections shift toward larger values \(N=5N=5–66\)\.

This pattern suggests that Worker allocation depends not only on dataset\-level difficulty, but also on the backbone model and its uncertainty under the Manager prompt\. In particular, the strong Gemini backbone can often resolve COVID\-19 cases with fewer Workers, whereas the smaller Mistral backbone tends to request more reasoning perspectives on P\-Stance\. Thus, dynamic allocation should be interpreted as an instance\- and model\-dependent estimate of reasoning need, rather than a direct proxy for dataset difficulty\. Detailed per\-target allocation statistics and planning\-stage examples are provided in Appendix[G](https://arxiv.org/html/2606.11609#A7)\.

Table 3:Worker\-count ablation summary \(Macro\-F1\)\. “Best fixed” denotes the best fixed\-NNconfiguration amongN∈\{1,3,5,7,9,11\}N\\in\\\{1,3,5,7,9,11\\\}\. Dynamic allocation usesN∈\{3,…,7\}N\\in\\\{3,\\ldots,7\\\}\. For the dynamic setting,N¯\\bar\{N\}denotes the average number of selected Workers\. The best result in each row is shown inbold\.![Refer to caption](https://arxiv.org/html/2606.11609v1/x1.png)Figure 2:Distribution of dynamically selected Worker counts for two representative settings\. Left: COVID\-19 with Gemini, which yields the lowest average allocation \(N¯=3\.60\\bar\{N\}=3\.60\)\. Right: P\-Stance with Mistral, which yields the highest average allocation \(N¯=5\.13\\bar\{N\}=5\.13\)\. Allocation patterns vary by both dataset and backbone model\.
#### Effect of reasoning\-level synthesis\.

We next ablate the reasoning components on COVID\-19, the benchmark where SMART\-D shows the largest gains\. Table[4](https://arxiv.org/html/2606.11609#S5.T4)compares Worker voting \(WV\), Worker reasoning \(WR\), Manager reasoning \(MR\), self\-consistency, and the full SMART\-D framework\. In WV settings, Workers directly output stance labels and aggregation is performed over labels\. In WR settings, Workers instead produce intermediate explanations\. MR denotes explicit Manager\-side synthesis before the final prediction\.

The results show that Worker\-level reasoning is the largest contributor to performance\. For Gemini, adding Worker reasoning improves Macro\-F1 from 78\.95 to 85\.27, substantially outperforming both Worker voting and self\-consistency\. Similar trends hold for Llama and Mistral\. Adding Manager reasoning provides smaller additional gains, indicating that the main benefit comes from exposing the Manager to richer intermediate explanations rather than only final labels\. The full SMART\-D framework achieves the best result for each backbone, confirming that dynamic allocation and reasoning\-level aggregation are complementary\. Full per\-target ablation results are provided in Appendix[H](https://arxiv.org/html/2606.11609#A8)\.

Table 4:Ablation of reasoning components on the COVID\-19 dataset \(Macro\-F1, %\)\. WV: Worker Voting; WR: Worker Reasoning; MR: Manager Reasoning\. Self\-consistency aggregates 5 independent generations per run\. Best result for each backbone is shown inbold, and the best result overall is also underlined\.
#### Cross\-model Worker allocation\.

We further examine whether heterogeneous Worker pools improve reasoning\-level aggregation\. Using COVID\-19 Stance as the benchmark where SMART\-D shows the largest gains, we compare the standard homogeneous setting, where the Manager and Workers use the same backbone, with a mixed\-model setting in which the Manager selects Workers from the set Gemini, Llama, and Mistral\.

As shown in Table[5](https://arxiv.org/html/2606.11609#S5.T5), heterogeneous Workers do not consistently improve performance\. For Gemini and Llama, the homogeneous SMART\-D setting performs better, suggesting that consistency in reasoning style may be important for stable aggregation\. In contrast, Mistral benefits substantially from mixed\-model Workers, improving from 75\.99 to 85\.23 Macro\-F1\. One possible explanation is that a weaker Manager can benefit from reasoning traces produced by stronger Worker models, whereas stronger Managers may be more sensitive to inconsistencies introduced by heterogeneous reasoning styles\. Overall, these results suggest that model diversity alone does not guarantee better reasoning\-level aggregation\. Cross\-model Worker allocation may be useful for weaker backbones, but homogeneous reasoning appears more stable for stronger backbones\. Additional Worker selection patterns are provided in Appendix[I](https://arxiv.org/html/2606.11609#A9)\.

Table 5:Cross\-model Worker allocation on COVID\-19 Stance \(Macro\-F1, %\)\. Homogeneous SMART\-D uses the same backbone for Manager and Workers; mixed Workers are selected from the three\-backbone pool\.

### 5\.3Qualitative Case Studies

To illustrate where SMART\-D is most useful, we examine two examples from COVID\-19 Stance for the target*School Closures*\. In both cases, the single\-model baseline fails, while SMART\-D succeeds\. Full reasoning traces are provided in Appendix[J](https://arxiv.org/html/2606.11609#A10)\.

Case 1: Rhetorical challenge and implicit supportTweet:*“Governor @GovRonDeSantis, why aren’t your offices open? If you want all schools open in a month, open your offices today\. \#COVID19”*Gold:favorBaseline:againstSMART\-D:favorWorkers:5Why SMART\-D succeeds:Workers identify the tweet as a rhetorical challenge: reopening schools is framed as unsafe because government offices remain closed\. This requires reasoning beyond surface\-level criticism\.

Case 2: Distinguishing stance from related complaintsTweet:*“What if there was job loss or pay cut due to recession? Would government still give decision to not pay fees to schools? Are schools run by robots?? Teachers also need to get paid right?”*Gold:noneBaseline:againstSMART\-D:noneWorkers:3Why SMART\-D succeeds:Workers separate general financial and institutional concerns from a direct stance on school closures, enabling a target\-awarenoneprediction\.

These cases show that SMART\-D is especially useful when stance depends on rhetorical framing, or on distinguishing the target from related issues\.

### 5\.4Discussion

Overall, the results indicate that SMART\-D is most useful when stance requires contextual interpretation\. The largest gains occur on COVID\-19 Stance, where many examples involve sarcasm, rhetorical comparison, policy framing, or indirect references\. Gains are more moderate on SemEval\-2016 and more model\-dependent on P\-Stance, suggesting that additional reasoning is most valuable when stance is implicit or ambiguous, but less useful when stance is expressed through direct lexical cues\. The ablations support this interpretation\. Increasing the number of Workers does not monotonically improve performance, reflecting a trade\-off between reasoning diversity and aggregation noise\. Dynamic allocation helps manage this trade\-off by assigning additional reasoning capacity selectively rather than applying a fixed budget to all instances\. Similarly, the cross\-model results show that diversity alone is not sufficient: effective aggregation depends on both the strength and compatibility of Worker reasoning traces\.

Qualitative examples reinforce the quantitative findings: SMART\-D succeeds when Workers capture complementary interpretations and the Manager resolves them under the target\-specific label space\. Thus, SMART\-D should be viewed as an adaptive reasoning framework for complex stance cases, rather than a uniformly better inference strategy for all instances\.

## 6Conclusion

We introduced SMART\-D, a structured multi\-agent reasoning framework for target\-aware stance detection\. SMART\-D decomposes stance analysis into complementary reasoning\-only Worker perspectives and uses a Manager to synthesize these explanations into a final stance prediction\. The framework also adapts inference\-time reasoning effort by dynamically selecting the number of Workers based on input complexity\.

Experiments on SemEval\-2016, P\-Stance, and COVID\-19 Stance show that SMART\-D achieves its largest gains on implicit, context\-dependent, and rhetorically complex stance cases, while gains are smaller and more model\-dependent when stance cues are explicit\. These findings suggest that reasoning\-level aggregation is most useful when stance cannot be reliably inferred from surface cues alone, and highlight adaptive reasoning as a promising direction for robust real\-world stance detection\.

## 7Limitations

This study has several limitations\. First, our experiments are limited to English\-language Twitter/X datasets, and the effectiveness of SMART\-D on other languages, platforms, or longer\-form texts remains an open question\. Second, the multi\-agent reasoning architecture introduces additional computational overhead compared to single\-pass prompting\. Although dynamic Worker allocation reduces unnecessary computation, SMART\-D still requires multiple reasoning passes per instance\. Third, the benefits of structured reasoning are condition\-dependent\. Gains are most pronounced on datasets with implicit or complex stance expressions, such as COVID\-19 Stance, while improvements are more limited on datasets with more explicit stance signals, such as P\-Stance\. Finally, models applied to social media data may reflect biases in those datasets, including demographic, ideological, or platform\-specific imbalances\.

From an ethical perspective, stance detection systems can support large\-scale analysis of public discourse, but they may also be misused to monitor, influence, or manipulate online debates\. Responsible deployment should emphasize transparency, careful interpretation, and safeguards against uses that distort public opinion\.

## References

- Zero\-shot stance detection: a dataset and model using generalized topic representations\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 8913–8931\.Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1)\.
- I\. Augenstein, T\. Rocktäschel, A\. Vlachos, and K\. Bontcheva \(2016\)Stance detection with bidirectional conditional encoding\.InProceedings of the 2016 conference on empirical methods in natural language processing,pp\. 876–885\.Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p2.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1)\.
- D\. Bamman and N\. Smith \(2015\)Contextualized sarcasm detection on twitter\.Inproceedings of the international AAAI conference on web and social media,Vol\.9,pp\. 574–577\.Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p1.1)\.
- M\. Besta, N\. Blach, A\. Kubicek, R\. Gerstenberger, M\. Podstawski, L\. Gianinazzi, J\. Gajda, T\. Lehmann, H\. Niewiadomski, P\. Nyczyk,et al\.\(2024\)Graph of thoughts: solving elaborate problems with large language models\.InProceedings of the AAAI conference on artificial intelligence,Vol\.38,pp\. 17682–17690\.Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p4.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p2.1),[§3\.1](https://arxiv.org/html/2606.11609#S3.SS1.p2.1)\.
- T\. Brown, B\. Mann, N\. Ryder, M\. Subbiah, J\. D\. Kaplan, P\. Dhariwal, A\. Neelakantan, P\. Shyam, G\. Sastry, A\. Askell,et al\.\(2020\)Language models are few\-shot learners\.Advances in neural information processing systems33,pp\. 1877–1901\.Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p3.1)\.
- C\. Conforti, J\. Berndt, M\. T\. Pilehvar, C\. Giannitsarou, F\. Toxvaerd, and N\. Collier \(2020\)Will\-they\-won’t\-they: a very large dataset for stance detection on twitter\.InProceedings of the 58th annual meeting of the association for computational linguistics,pp\. 1715–1724\.Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1)\.
- J\. Devlin, M\. Chang, K\. Lee, and K\. Toutanova \(2019\)Bert: pre\-training of deep bidirectional transformers for language understanding\.InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 \(long and short papers\),pp\. 4171–4186\.Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1)\.
- J\. Du, R\. Xu, Y\. He, and L\. Gui \(2017\)Stance classification with target\-specific neural attention networks\.In26th International Joint Conference on Artificial Intelligence, IJCAI 2017,pp\. 3988–3994\.Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1)\.
- N\. Du, Y\. Huang, A\. M\. Dai, S\. Tong, D\. Lepikhin, Y\. Xu, M\. Krikun, Y\. Zhou, A\. W\. Yu, O\. Firat, B\. Zoph, L\. Fedus, M\. Bosma, Z\. Zhou, T\. Wang, Y\. E\. Wang, K\. Webster, M\. Pellat, K\. Robinson, K\. S\. Meier\-Hellstern, T\. Duke, L\. Dixon, K\. Zhang, Q\. V\. Le, Y\. Wu, Z\. Chen, and C\. Cui \(2021\)GLaM: efficient scaling of language models with mixture\-of\-experts\.InInternational Conference on Machine Learning \(ICML\),Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Du, S\. Li, A\. Torralba, J\. B\. Tenenbaum, and I\. Mordatch \(2024\)Improving factuality and reasoning in language models through multiagent debate\.InForty\-first international conference on machine learning,Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p4.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px2.p1.1),[§3\.1](https://arxiv.org/html/2606.11609#S3.SS1.p2.1)\.
- K\. Garg and C\. Caragea \(2024\)Stanceformer: target\-aware transformer for stance detection\.InFindings of the Association for Computational Linguistics: EMNLP 2024,pp\. 4969–4984\.Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p2.1)\.
- K\. Glandt, S\. Khanal, Y\. Li, D\. Caragea, and C\. Caragea \(2021\)Stance detection in covid\-19 tweets\.InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics \(ACL\),pp\. 1596–1611\.Cited by:[Table T2](https://arxiv.org/html/2606.11609#A3.T2),[§1](https://arxiv.org/html/2606.11609#S1.p1.1),[§1](https://arxiv.org/html/2606.11609#S1.p5.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1),[§4\.2](https://arxiv.org/html/2606.11609#S4.SS2.p1.1),[Table 1](https://arxiv.org/html/2606.11609#S4.T1)\.
- M\. Hardalov, A\. Arora, P\. Nakov, and I\. Augenstein \(2022\)A survey on stance detection for mis\- and disinformation identification\.InFindings of the Association for Computational Linguistics: NAACL 2022,Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1)\.
- K\. S\. Hasan and V\. Ng \(2013\)Stance classification of ideological debates: data, models, features, and constraints\.InProceedings of the sixth international joint conference on natural language processing,pp\. 1348–1356\.Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1)\.
- S\. Hong, M\. Zhuge, J\. Chen, X\. Zheng, Y\. Cheng, C\. Zhang, J\. Wang, Z\. Wang, S\. K\. S\. Yau, Z\. Lin, L\. Zhou, C\. Ran, L\. Xiao, C\. Wu, and J\. Schmidhuber \(2023\)MetaGPT: meta programming for a multi\-agent collaborative framework\.arXiv preprint arXiv:2308\.00352\.Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p4.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px2.p1.1)\.
- T\. Kojima, S\. S\. Gu, M\. Reid, Y\. Matsuo, and Y\. Iwasawa \(2022\)Large language models are zero\-shot reasoners\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p3.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p2.1)\.
- X\. Lan, C\. Gao, D\. Jin, and Y\. Li \(2024\)Stance detection with collaborative role\-infused llm\-based agents\.InProceedings of the international AAAI conference on web and social media,Vol\.18,pp\. 891–903\.Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p4.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px2.p1.1),[§3\.1](https://arxiv.org/html/2606.11609#S3.SS1.p2.1)\.
- D\. Lepikhin, H\. Lee, Y\. Xu, D\. Chen, O\. Firat, Y\. Huang, M\. Krikun, N\. Shazeer, and Z\. Chen \(2021\)GShard: scaling giant models with conditional computation and automatic sharding\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px2.p1.1)\.
- G\. Li, H\. Hammoud, H\. Itani, D\. Khizbullin, and B\. Ghanem \(2023\)Camel: communicative agents for" mind" exploration of large language model society\.Advances in neural information processing systems36,pp\. 51991–52008\.Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p4.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Li, T\. Sosea, A\. Sawant, A\. J\. Nair, D\. Inkpen, and C\. Caragea \(2021\)P\-stance: a large dataset for stance detection in political domain\.InFindings of the association for computational linguistics: ACL\-IJCNLP 2021,pp\. 2355–2365\.Cited by:[Table T2](https://arxiv.org/html/2606.11609#A3.T2),[§1](https://arxiv.org/html/2606.11609#S1.p1.1),[§1](https://arxiv.org/html/2606.11609#S1.p5.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1),[§4\.2](https://arxiv.org/html/2606.11609#S4.SS2.p1.1),[Table 2](https://arxiv.org/html/2606.11609#S5.T2)\.
- B\. Liang, Z\. Chen, L\. Gui, Y\. He, M\. Yang, and R\. Xu \(2022\)Zero\-shot stance detection via contrastive learning\.InProceedings of WWW,pp\. 2738–2747\.Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1)\.
- S\. Mohammad, S\. Kiritchenko, P\. Sobhani, X\. Zhu, and C\. Cherry \(2016\)Semeval\-2016 task 6: detecting stance in tweets\.InProceedings of the 10th international workshop on semantic evaluation \(SemEval\-2016\),pp\. 31–41\.Cited by:[Table T2](https://arxiv.org/html/2606.11609#A3.T2),[§1](https://arxiv.org/html/2606.11609#S1.p1.1),[§1](https://arxiv.org/html/2606.11609#S1.p2.1),[§1](https://arxiv.org/html/2606.11609#S1.p5.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1),[§3\.1](https://arxiv.org/html/2606.11609#S3.SS1.p1.5),[§4\.2](https://arxiv.org/html/2606.11609#S4.SS2.p1.1),[Table 2](https://arxiv.org/html/2606.11609#S5.T2)\.
- R\. Pryzant, D\. Iter, J\. Li, Y\. T\. Lee, C\. Zhu, and M\. Zeng \(2023\)Automatic prompt optimization with "gradient descent" and beam search\.InProceedings of the Conference on Empirical Methods in Natural Language Processing \(EMNLP\),Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px2.p2.1)\.
- E\. Riloff, A\. Qadir, P\. Surve, L\. De Silva, N\. Gilbert, and R\. Huang \(2013\)Sarcasm as contrast between a positive sentiment and a negative situation\.InProceedings of the Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 704–714\.Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p1.1)\.
- N\. Shazeer, A\. Mirhoseini, K\. Maziarz, A\. Davis, Q\. V\. Le, G\. Hinton, and J\. Dean \(2017\)Outrageously large neural networks: the sparsely\-gated mixture\-of\-experts layer\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px2.p1.1)\.
- P\. Sobhani, S\. Mohammad, and S\. Kiritchenko \(2016\)Detecting stance in tweets and analyzing its interaction with sentiment\.InProceedings of the fifth joint conference on lexical and computational semantics,pp\. 159–169\.Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p1.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1)\.
- S\. Somasundaran and J\. Wiebe \(2009\)Recognizing stances in online debates\.InProceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing \(ACL\-IJCNLP\),pp\. 226–234\.Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1)\.
- M\. Taranukhin, V\. Shwartz, and E\. Milios \(2024\)Stance reasoner: zero\-shot stance detection on social media with explicit reasoning\.InProceedings of LREC\-COLING 2024,pp\. 15257–15272\.Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p2.1)\.
- X\. Wang, Y\. Wang, S\. Cheng, P\. Li, and Y\. Liu \(2024\)DEEM: dynamic experienced expert modeling for stance detection\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\),pp\. 4530–4541\.Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p2.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px2.p1.1)\.
- X\. Wang, J\. Wei, D\. Schuurmans, Q\. V\. Le, E\. H\. Chi, S\. Narang, A\. Chowdhery, and D\. Zhou \(2023\)Self\-consistency improves chain of thought reasoning in language models\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p3.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p2.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px2.p2.1),[§3\.1](https://arxiv.org/html/2606.11609#S3.SS1.p2.1),[§3\.3](https://arxiv.org/html/2606.11609#S3.SS3.p1.1)\.
- J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, F\. Xia, E\. Chi, Q\. V\. Le, D\. Zhou,et al\.\(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.Advances in neural information processing systems35,pp\. 24824–24837\.Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p3.1),[§1](https://arxiv.org/html/2606.11609#S1.p4.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p2.1)\.
- Q\. Wu, G\. Bansal, J\. Zhang, Y\. Wu, B\. Li, E\. Zhu, L\. Jiang, X\. Zhang, S\. Zhang, J\. Liu,et al\.\(2024\)Autogen: enabling next\-gen llm applications via multi\-agent conversations\.InFirst conference on language modeling,Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p4.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Yan, S\. Sun, Z\. Tang, T\. Liu, and M\. Liu \(2025\)Collaborative stance detection via small\-large language model consistency verification\.InInternational Conference on Database Systems for Advanced Applications,pp\. 3–20\.Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p2.1)\.
- S\. Yao, D\. Yu, J\. Zhao, I\. Shafran, T\. Griffiths, Y\. Cao, and K\. Narasimhan \(2023\)Tree of thoughts: deliberate problem solving with large language models\.Advances in neural information processing systems36,pp\. 11809–11822\.Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p4.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p2.1),[§3\.1](https://arxiv.org/html/2606.11609#S3.SS1.p2.1),[§3\.3](https://arxiv.org/html/2606.11609#S3.SS3.p1.1)\.
- G\. Zarrella and A\. Marsh \(2016\)Mitre at semeval\-2016 task 6: transfer learning for stance detection\.InProceedings of the 10th international workshop on semantic evaluation \(SemEval\-2016\),pp\. 458–463\.Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p1.1)\.
- Y\. Zhang, A\. Li, B\. Chen, J\. Sun, and X\. Zhao \(2026\)MSME: a multi\-stage multi\-expert framework for zero\-shot stance detection\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.40,pp\. 34879–34887\.Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p2.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px2.p1.1),[Table 2](https://arxiv.org/html/2606.11609#S5.T2)\.
- D\. Zhou, N\. Schärli, L\. Hou, J\. Wei, N\. Scales, X\. Wang, D\. Schuurmans, C\. Cui, O\. Bousquet, Q\. V\. Le, and E\. H\. Chi \(2023a\)Least\-to\-most prompting enables complex reasoning in large language models\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§1](https://arxiv.org/html/2606.11609#S1.p4.1),[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px1.p2.1)\.
- Y\. Zhou, A\. I\. Muresanu, Z\. Han, K\. Paster, S\. Pitis, H\. Chan, and J\. Ba \(2023b\)Large language models are human\-level prompt engineers\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§2](https://arxiv.org/html/2606.11609#S2.SS0.SSS0.Px2.p2.1)\.

## Appendix AFull Prompt Templates

We provide the full prompt templates used for the Manager \(planning/decision\) and Workers\.

### A\.1Manager Prompt: Planning Stage

A1\. Manager Prompt \(Planning\)You are the Manager in the planning stage\.Goal:Decide the number of Workers and assign each a distinct analytical perspective\.Input:•Tweet text•Stance targetInstructions:1\.Judge instance complexity:*simple*,*medium*, or*complex*\. Treat as complex when sarcasm, indirect stance, mixed sentiment, multiple entities, or rhetorical framing is present\.2\.Select the number of WorkersN∈\{3,…,7\}N\\in\\\{3,\\dots,7\\\}\(more for complex, fewer for simple\)\.3\.WriteNNWorker prompts, each emphasizing a*different*stance\-relevant aspect \(e\.g\., explicit cues, sarcasm/irony, emotional tone, indirect stance via related actors, policy framing, hashtags/slogans\)\.Output format \(exact\):•Number of Workers: N•Worker 1 Prompt: \.\.\.•Worker 2 Prompt: \.\.\.•…Constraint:Donotoutput a stance label\.

### A\.2Manager Prompt: Decision Stage

A2\. Manager Prompt \(Decision\)You are the Manager in the decision stage\.Goal:Aggregate Worker reasonings and output a single final stance label\.Input:•Tweet text•Stance target•Worker reasonings \(each is 2–4 sentences; no final label\)Instructions:1\.Compare Worker reasonings and identify areas of agreement, disagreement, and the strongest evidence\.2\.Prioritize reasonings that cite concrete textual evidence from the tweet and directly relate it to the target\.3\.Discount reasonings that are unsupported, overly speculative, or rely only on general sentiment unrelated to the target\.4\.Resolve conflicts by selecting the label best supported by the combined evidence and target\-specific stance interpretation\.5\.Select exactly one label from the valid set for the dataset:•P\-Stance:favor/against•COVID\-19:favor/against/none•SemEval\-2016:favor/against/noneOutput \(exact\):Print only one valid stance label for the dataset\.

### A\.3Worker Prompt Template

A3\. Worker Prompt \(Template\)You are a Worker\.Input:•Tweet text•Stance target•Your assigned aspect \(e\.g\., explicit cues, sarcasm/irony, emotional tone, indirect stance via related actors, framing, hashtags\)Instructions:1\.Analyze stance toward the target focusing only on your assigned aspect\.2\.Write2–4 sentencesof evidence\-grounded reasoning\.Constraint:Output reasoning only \(no labels such asfavor,against,none\)\.

## Appendix BDataset Examples

Table[T1](https://arxiv.org/html/2606.11609#A2.T1)presents three representative examples from the evaluated stance detection datasets together with their corresponding targets and labels\.

Table T1:Sample instances from the evaluated stance detection datasets with their corresponding targets and labels\.
## Appendix CAccuracy Results

Table[T2](https://arxiv.org/html/2606.11609#A3.T2)summarizes accuracy across SemEval\-2016, P\-Stance, and COVID\-19\. The trends closely mirror Macro\-F1, indicating consistent improvements across both class\-balanced and overall metrics\. Mistral achieves the highest average accuracy on SemEval\-2016, while Gemini performs best on COVID\-19\. Overall, gains are more pronounced on datasets requiring implicit or context\-dependent interpretation \(e\.g\., COVID\-19\), and smaller on more explicit datasets such as P\-Stance, suggesting that reasoning\-based aggregation is particularly beneficial for complex stance inference\.

Table T2:Accuracy \(%\) on SemEval\-2016 Task 6Mohammadet al\.\([2016](https://arxiv.org/html/2606.11609#bib.bib1)\), P\-StanceLiet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib13)\), and COVID\-19 StanceGlandtet al\.\([2021](https://arxiv.org/html/2606.11609#bib.bib11)\)under the SMART\-D framework\. Results are reported as mean±\\pmstandard deviation over three runs\. Best results are shown inbold\.
## Appendix DStatistical Significance Analysis

We compare SMART\-D against the corresponding single\-model baselines on the COVID\-19 Stance dataset, where SMART\-D shows the largest gains\. For each backbone, we compute Macro\-F1 over three independent runs and evaluate the difference using Welch’stt\-test\. Because the number of runs is small \(n=3n=3\), these tests should be interpreted as supporting evidence rather than as a definitive estimate of significance\.

Table[T3](https://arxiv.org/html/2606.11609#A4.T3)shows that SMART\-D improves over the corresponding single\-model baseline for all three backbones\. The gains are largest for Llama and Mistral, and all improvements exceed the observed run\-to\-run variance\.

Table T3:Statistical significance of SMART\-D improvements over corresponding single\-model baselines on the COVID\-19 Stance dataset\. Results are reported as average Macro\-F1±\\pmstandard deviation over three runs\.Δ\\Deltadenotes the absolute Macro\-F1 improvement\. Significance is evaluated using Welch’stt\-test with three independent runs per method\.
## Appendix EFull Worker\-Count Ablation

Table[T4](https://arxiv.org/html/2606.11609#A5.T4)reports the full fixed\-NNWorker\-count ablation across all datasets and backbones\. Fixed configurations useN∈\{1,3,5,7,9,11\}N\\in\\\{1,3,5,7,9,11\\\}Workers, while the dynamic setting selectsN∈\{3,…,7\}N\\in\\\{3,\\ldots,7\\\}per instance\. These results provide the complete values summarized in Table[3](https://arxiv.org/html/2606.11609#S5.T3)\.

Table T4:Worker allocation ablation across three stance benchmarks \(Macro\-F1, %\)\. Fixed configurations useN∈\{1,3,5,7,9,11\}N\\in\\\{1,3,5,7,9,11\\\}Workers\. Results are reported as mean±\\pmstandard deviation over three runs\. For the dynamic setting,N¯\\bar\{N\}denotes the average number of selected Workers\. The best result in each row is shown inbold\.
## Appendix FPerformance–Cost Trade\-off

Figure[S1](https://arxiv.org/html/2606.11609#A6.F1)illustrates the performance–cost trade\-off across three datasets\. Performance generally improves as the number of Workers \(NN\) increases, but the gains tend to diminish beyond a moderate range, while computational cost continues to grow\. Figure[S2](https://arxiv.org/html/2606.11609#A6.F2)shows token usage as a function of the number of Workers\. Token consumption grows approximately linearly withNN, indicating a predictable computational overhead per additional Worker\. As a result, increasing reasoning depth leads to higher inference cost even when performance improvements become limited\. This supports the use of dynamic allocation, which aims to avoid unnecessary reasoning paths when additional Workers provide limited performance gains\.

![Refer to caption](https://arxiv.org/html/2606.11609v1/final_legend_right_subplot.png)Figure S1:Performance–cost trade\-off across three datasets\. Lines represent fixed Worker configurations, while star markers denote dynamic allocation\. The dynamic strategy achieves performance close to the best fixed configurations while operating at lower computational cost\.![Refer to caption](https://arxiv.org/html/2606.11609v1/figures/clean_worker_token_cost_cropped_v2.png)Figure S2:Token consumption grows approximately linearly with the number of Workers, indicating a predictable computational overhead per Worker\.
## Appendix GAdditional Dynamic Allocation Analysis

### G\.1Per\-Target Worker Allocation

This section provides detailed per\-target distributions of dynamically selected Worker counts, complementing the summary analysis in Section[5\.2](https://arxiv.org/html/2606.11609#S5.SS2)\. The main text reports aggregate allocation patterns; here, we show how the Manager’s allocation decisions vary across targets for the two representative settings used in Figure[2](https://arxiv.org/html/2606.11609#S5.F2)\.

#### COVID\-19 Stance with Gemini\.

Table[T5](https://arxiv.org/html/2606.11609#A7.T5)shows the per\-target allocation distribution for COVID\-19 Stance\. Allocations are concentrated in the lower range, primarilyN=3N=3andN=4N=4, withN=5N=5used selectively\. This pattern is consistent with the relatively low average allocation observed for this setting\.

Table T5:Per\-target distribution of dynamically selected Worker counts for COVID\-19 Stance using Gemini\. The four targets are: Wearing a Face Mask \(FM\), Anthony Fauci \(AF\), Stay\-at\-Home Orders \(SH\), Keeping Schools Closed \(SC\)\. Counts indicate the number of instances assigned to each Worker count\.
#### P\-Stance with Mistral\.

Table[T6](https://arxiv.org/html/2606.11609#A7.T6)shows the corresponding distribution for P\-Stance using Mistral\. Compared with the COVID\-19 setting, allocations shift toward larger Worker counts, especiallyN=5N=5andN=6N=6\. This suggests that the Manager requests more reasoning perspectives when using this backbone, even though P\-Stance contains many more explicit stance cues\.

Table T6:Per\-target distribution of dynamically selected Worker counts for P\-Stance using Mistral\. JB denotesJoe Biden, DT denotesDonald Trump, and BS denotesBernie Sanders\. Counts indicate the number of instances assigned to each Worker count\.
#### Summary\.

Across the two settings, Worker allocation varies by both dataset and backbone\. For COVID\-19 Stance with Gemini, most instances are assignedN=3N=3orN=4N=4, indicating that the Manager often finds a small number of reasoning perspectives sufficient\. In contrast, for P\-Stance with Mistral, allocations shift towardN=5N=5andN=6N=6, suggesting greater model\-dependent demand for additional reasoning perspectives\.

Overall, these results support the interpretation that dynamic allocation is not a simple proxy for dataset difficulty\. Instead, the selected number of Workers reflects an interaction between input characteristics and the underlying model’s reasoning behavior\.

### G\.2Planning\-Stage Examples

To illustrate how the Manager allocates reasoning effort, we present three planning\-stage examples corresponding to low\-, moderate\-, and high\-complexity inputs\. These examples show how the selected number of Workers reflects target relevance, ambiguity, and the number of plausible interpretations\.

Example 1: Low Complexity/Off\-Topic Input \(N=3N=3\)Tweet:*“New York city girl @fireboydml On \#trendmuzikblast @Djitchyfingazz \#spinning:repeat\_one: LISTENLIVEON \#OURAPPONPLAYSTORE AND \#WEBSITEONBIO”*Target:Stay at Home OrdersGold label:*none*Baseline prediction:*none*SMART\-D prediction:*none*Selected Workers:3Baseline reasoning\.The tweet is a promotional post for a music broadcast and does not mention or imply a stance toward Stay at Home Orders\.Manager planning rationale\.The tweet is primarily about music and promotion\. There is no explicit mention of, nor any obvious implication relating to, “Stay at Home Orders\.” The content does not express any stance toward the target, either directly or indirectly\. The main task here is simply to confirm that the tweet is irrelevant to the target\. Since there is no ambiguity, no implicit stance, and no competing interpretations, three Workers are sufficient to confirm this consensus and rule out any subtle interpretations\.Interpretation\.The tweet is clearly unrelated to the target\. The reasoning task is minimal: confirm irrelevance and rule out a hidden target connection\. Additional Workers would likely be redundant\.

Example 2: Moderate Complexity/Implicit but Recoverable Stance \(N=4N=4\)Tweet:*“@SpeakerPelosi @RepDanKildee @RepRichardNeal @RepDannyDavis Back to work\!”*Target:Stay at Home OrdersGold label:*none*Baseline prediction:*against*SMART\-D prediction:*none*Selected Workers:4Baseline reasoning\.The phrase “Back to work\!” directly advocates for ending lockdowns and resuming normal activity, indicating opposition to stay\-at\-home orders\.Manager planning rationale\.The phrase “Back to work\!” directed at several political figures strongly suggests opposition to measures that prevent people from working, such as Stay at Home Orders\. However, the tweet does not explicitly mention the target policy, and the stance must be inferred from context and intent\. The brevity of the tweet and the lack of direct reference to the policy introduce some ambiguity\. Additionally, the presence of multiple tagged accounts adds a slight layer of contextual complexity, as the statement may be directed at specific actions or broader policy decisions\. While the stance is not deeply ambiguous, it requires interpretation, and multiple reasoning perspectives are needed to ensure the correct inference\. More than four Workers would likely be redundant, while fewer might miss subtle contextual cues\.Interpretation\.The tweet contains an implicit cue, but the target connection is indirect\. Four Workers provide enough reasoning diversity to distinguish a general political exhortation from a direct stance toward Stay at Home Orders, leading to a target\-aware*none*prediction\.

Example 3: High Complexity/Competing Interpretations \(N=5N=5\)Tweet:*“Better to be safe than sorry\. \#OnlineClasses offer continued learning in a safe environment\. Let schools function and provide education\. It’s easy to destroy with such senseless orders\.”*Target:School ClosuresGold label:*favor*Baseline prediction:*favor*SMART\-D prediction:*favor*Selected Workers:5Baseline reasoning\.The tweet focuses on the safety benefits of online classes and continued learning in a safe environment, which supports school closures\. However, the reference to “senseless orders” introduces ambiguity about whether the author is criticizing closures, their implementation, or broader policy consequences\.Manager planning rationale\.While the author explicitly states a preference for “OnlineClasses,” which could suggest support for restrictions or closures, the latter part introduces a critical element—a condemnation of “senseless orders\.” This implies a negative stance toward the policies associated with those orders\. However, the interpretation is not straightforward\. The emotionally charged phrasing introduces ambiguity and raises questions about what exactly is being criticized\. It is unclear whether the criticism is directed at the closures themselves, the implementation of online learning, or broader socio\-economic consequences of the policies\. The tweet contains multiple interpretive layers and potentially conflicting signals\. Because of this ambiguity, several plausible interpretations exist, and multiple reasoning perspectives are required to explore them\. Assigning five Workers ensures sufficient coverage of these competing interpretations, while additional Workers would likely yield diminishing returns\.Interpretation\.The tweet contains both support for safe online learning and criticism of policy orders\. Even though both the baseline and SMART\-D predict the correct label, the Manager assigns more Workers because the input contains conflicting signals and multiple plausible readings\.

## Appendix HPer\-Target Ablation Results

In this section, we provide detailed per\-target Macro\-F1 results for the ablation of reasoning components on the COVID\-19 stance dataset\.

Table[T7](https://arxiv.org/html/2606.11609#A8.T7)reports results across four targets: Face Masks \(FM\), Anthony Fauci \(AF\), Stay\-at\-Home \(SH\), and School Closures \(SC\)\. We include all configurations from the main text, along with the Self\-Consistency baseline for completeness\.

Table T7:Per\-target Macro\-F1 \(%\) results on the COVID\-19 dataset\. WV: Worker Voting, WR: Worker Reasoning, MR: Manager Reasoning\. Average±\\pmstandard deviation is reported over three runs\.
## Appendix ICross\-Model Worker Selection Patterns

This section analyzes how the Manager composes heterogeneous Worker teams in the mixed\-model setting on the COVID\-19 Stance dataset\. The Worker pool contains the three backbones used throughout the paper: Gemini, Llama, and Mistral\. This analysis provides additional insight into when cross\-model diversity is selected and how selection patterns vary across Manager backbones and stance targets\.

![Refer to caption](https://arxiv.org/html/2606.11609v1/x2.png)Figure S3:Distribution of selected Worker models in the mixed\-model setting on COVID\-19 Stance\. Panels correspond to Manager backbones: Gemini, Llama, and Mistral\. Bars show the proportion of selected Workers from each backbone across four targets: Face Masks \(FM\), Anthony Fauci \(AF\), Stay\-at\-Home Orders \(SH\), and School Closures \(SC\)\.Figure[S3](https://arxiv.org/html/2606.11609#A9.F3)shows that Worker selection is not uniform across models\. Instead, the selected Worker composition varies with both the Manager backbone and the stance target\. Stronger Managers, such as Gemini and Llama, tend to rely more heavily on stronger or similarly capable Workers, while still incorporating model diversity in some target settings\.

The selection behavior is also target\-dependent\. Targets such asStay\-at\-Home OrdersandSchool Closuresexhibit more varied Worker compositions, suggesting that the Manager draws on complementary reasoning signals when the target involves more implicit, policy\-dependent, or rhetorically complex stance expressions\.

For Mistral, the mixed\-model setting more frequently incorporates stronger Worker models\. This pattern is consistent with the main\-text result showing that Mistral benefits most from heterogeneous Workers\. In contrast, the stronger Manager backbones perform better in the homogeneous SMART\-D setting, suggesting that consistency in reasoning style can be more important than model diversity when the Manager is already strong\.

Overall, these results help explain why model diversity alone does not guarantee better reasoning\-level aggregation\. Cross\-model Worker selection can provide useful complementary reasoning signals, particularly for weaker backbones, but heterogeneous reasoning styles may also introduce inconsistency that makes aggregation less stable\. This behavior is related to Mixture\-of\-Experts\-style routing, but here the selection occurs at inference time through Manager prompting rather than through a trained router or additional supervision\.

## Appendix JAdditional Qualitative Case Studies

This appendix provides the full qualitative examples referenced in Section[5\.3](https://arxiv.org/html/2606.11609#S5.SS3)\. For each case, we report the tweet, target, gold label, baseline prediction, SMART\-D prediction, selected number of Workers, Worker prompts and reasonings, and the final Manager decision\.

Case 1: Rhetorical Challenge and Implicit SupportTweet:*“Governor @GovRonDeSantis, why aren’t your offices open? If you want all schools open in a month, open your offices today\. \#COVID19”*Target:School ClosuresGold label:*favor*Baseline prediction:*against*SMART\-D prediction:*favor*Selected Workers:5Worker 1: Direct Policy StancePrompt\.Analyze the stance toward the policy of “School Closures\.” Specifically, evaluate whether the author’s demand for the Governor to open his own offices is a sincere endorsement of reopening or a rhetorical device used to highlight potential risks\. Focus on how the text conveys its position, considering implicit cues such as framing, tone, or indirect references\. Write a concise 2–4 sentence reasoning; do not output a final label\.Reasoning\.The author uses the demand for the Governor to open his offices as a rhetorical challenge to highlight a perceived double standard rather than as a sincere endorsement of reopening schools\. By pointing out that the Governor’s own workspace remains closed, the author suggests that if it is not safe for officials to return to the office, it is certainly not safe to mandate in\-person schooling\. Consequently, the author is expressing opposition to the policy of reopening schools during the pandemic\.Worker 2: Rhetorical Strategy and SarcasmPrompt\.Examine the use of sarcasm and rhetorical comparison in this tweet\. Analyze how the author frames the argument, particularly through indirect or ironic expressions\. Focus on identifying rhetorical cues and how they shape the interpretation of the message, without assigning any explicit stance label\.Reasoning\.The author uses a logical comparison to highlight a perceived double standard, suggesting that if schools are safe enough for a mandatory reopening, the Governor’s own administrative offices should already be operational\. This sarcastic challenge criticizes the reopening plan as hypocritical and potentially unsafe by pointing out that the leader is not subject to the same risks he is imposing on students and teachers\. Rather than genuinely advocating for more openings, the tweet uses the Governor’s closed offices as evidence that the push for open schools is inconsistent with current safety protocols\.Worker 3: Safety and Risk AssessmentPrompt\.Analyze how the author frames COVID\-19 safety concerns in the context of schools\. Does the author’s challenge to the Governor suggest a comparison between the safety of official workplaces and public schooling environments? Explain how this perceived risk is expressed in the argument and how it shapes the interpretation of the message\. Write a concise 2–4 sentence reasoning; do not output a final label\.Reasoning\.The author suggests that the Governor is applying a double standard by demanding schools reopen while his own offices remain closed due to safety concerns\. This challenge implies that if the environment is not safe enough for state officials, it is inherently too dangerous for students and teachers to return to classrooms\. Consequently, the author emphasizes safety considerations, framing school closures as a precaution until public environments are demonstrated to be sufficiently safe\.Worker 4: Target Specificity and RelevancePrompt\.Distinguish between the author’s criticism of the politician, Governor DeSantis, and the specific policy context of “School Closures\.” While the tweet targets the Governor, focus on how this criticism relates to the broader discussion of school reopening decisions\. Analyze how the argument connects these two elements and what implications are suggested regarding the policy context\. Write a concise 2–4 sentence reasoning; do not output a final label\.Reasoning\.The author highlights a potential inconsistency by questioning why the Governor advocates reopening schools while his own offices remain closed\. This comparison connects criticism of the Governor to broader concerns about safety and decision\-making\. The argument emphasizes a perceived mismatch between policy expectations and actual practices, suggesting caution in interpreting reopening decisions\.Worker 5: Advocacy vs\. ReportingPrompt\.Determine whether this tweet functions as a neutral description of a situation or as an attempt to influence interpretation through a call to action\. Analyze the reference to the Governor’s offices and how it is used within the argument\. Focus on whether the tweet frames the issue in a descriptive or persuasive manner, and how this affects the interpretation of the message\. Write a concise 2–4 sentence reasoning; do not output a final label\.Reasoning\.The author uses a direct challenge to the Governor to highlight a perceived inconsistency between school reopening plans and the status of government workplaces\. By calling for the Governor to “open your offices today,” the tweet moves beyond neutral description and introduces a persuasive element\. This call to action functions as a rhetorical device to question the coherence and safety of the reopening policy\.Manager decision\.Across the five Worker outputs, the dominant interpretation is that the tweet does not literally call for broader reopening\. Instead, it uses rhetorical comparison to criticize reopening schools while government offices remain closed\. The Manager therefore predicts*favor*, consistent with the gold label\.

Case 2: Distinguishing stance from related complaintsTweet:*“What if there was job loss or pay cut due to recession? Would government still give decision to not pay fees to schools? Are schools run by robots?? Teachers also need to get paid right?”*Target:School ClosuresGold label:*none*Baseline prediction:*against*SMART\-D prediction:*none*Selected Workers:3Worker 1: Policy Target vs\. General School IssuesPrompt\.Analyze whether the tweet focuses on the specific policy context of “School Closures” or discusses broader issues related to education\. Examine how references such as “not paying fees” and “teachers getting paid” are used in the argument, and whether they relate to operational, financial, or policy\-level considerations\. Focus on how the content frames the topic and what aspects of the issue are emphasized\. Do not output a final label; provide reasoning only\.Reasoning\.The author focuses on the economic sustainability of educational institutions, questioning the fairness of fee waivers and the necessity of maintaining teacher salaries during financial downturns\. While the mention of not paying fees often arises during periods of remote learning, the tweet centers on financial and labor disputes rather than the safety or public health necessity of closing school buildings\. The reasoning is rooted in the economic impact on staff and institutional operations rather than a specific policy stance on the physical opening or closing of schools\.Worker 2: Author Intent and AdvocacyPrompt\.Examine the author’s primary concern in this tweet\. Analyze how references such as “schools run by robots” and “teachers needing to get paid” are used in the argument\. Focus on whether the tweet emphasizes operational, financial, or educational aspects, and how these elements shape the interpretation of the message\. Do not output a final label; provide reasoning only\.Reasoning\.The author focuses on the financial viability of educational institutions, emphasizing that teachers require salaries and schools have operational costs that must be met even during economic downturns\. While the tweet expresses concern over tuition payment mandates during a recession, it primarily defends the necessity of school funding rather than emphasizing a particular instructional setting\. The rhetoric regarding “schools run by robots” highlights concerns about staff livelihoods, without explicitly prioritizing a specific operational mode\.Worker 3: Contextual Relevance and NeutralityPrompt\.Analyze whether the tweet discusses issues directly related to the policy of “School Closures” or focuses on broader aspects such as funding, salaries, or institutional concerns\. Examine how the author’s frustration with government decisions on fees is framed, and whether it reflects policy\-level considerations or a separate economic concern\. Focus on how the content relates to the target context and what aspects of the issue are emphasized\. Do not output a final label; provide reasoning only\.Reasoning\.The tweet focuses on the economic implications of school fees and the necessity of paying teacher salaries during a recession, rather than emphasizing the operational status of schools\. The author expresses frustration with government mandates regarding tuition payments, framing the issue as a matter of labor rights and institutional sustainability\. Because the discussion centers on financial and administrative concerns, it reflects broader educational issues rather than directly addressing the policy context of school closures\.Manager decision\.The Worker outputs consistently conclude that the tweet concerns school fees, teacher pay, and financial logistics rather than a direct stance on reopening or continued closure\. The Manager therefore predicts*none*, matching the gold label\.

### J\.1Example Execution Trace

To illustrate the full reasoning workflow, we provide a simplified execution trace for Case A\.

Manager planning output\.•Number of Workers: 5•Worker 1: direct policy stance•Worker 2: rhetorical strategy and sarcasm•Worker 3: safety and risk assessment•Worker 4: target specificity and relevance•Worker 5: advocacy vs\. reportingWorker outputs\.Each Worker generates a short reasoning paragraph from its assigned perspective\. Some Workers focus on rhetorical inconsistency, while others emphasize safety concerns or distinguish criticism of the Governor from stance toward school closures\.Manager decision\.After aggregating the Worker reasonings, the Manager identifies the dominant interpretation that the tweet criticizes reopening schools by highlighting a double standard between government offices and schools\. The final prediction is therefore*favor*for the target*School Closures*\.