Disentangling Topology and Diversity in Multi-Agent LLMs for Multilingual Low-Resource Emotion Detection

1 Introduction

• Prior multi-agent comparisons often confound call topology, diversity source, and model strength, obscuring what accounts for gains under fixed inference budgets.The paper studies this distinction in fine-grained emotion detection, where calls can gather different evidence while producing a fixed multi-label decision.Source: §1 · p. 1
• The comparison uses multilingual low-resource emotion detection because lexical, emoji, sarcastic, negated, pragmatic, and co-occurring emotional cues make evidence gathering and decision aggregation separable.A prior multi-agent emotion-analysis study used one coordination architecture rather than isolating the design choices.Source: §1 · p. 1
• The study crosses parallel aggregation and sequential refinement with stochastic, prompted, and learned diversity on Qwen2.5-14B-Instruct and Llama-3.1-8B-Instruct.Within each backbone, the model, call budget, output schema, and reasoning protocol are held fixed.Source: §1 · pp. 1–2
• The paper compares topology and diversity under fixed compute and examines when sequential refinement helps or hurts, including model-dependent learned Width–Depth ordering.Its contribution also separates pooled supervised adaptation from specialist heterogeneity.Source: §1 · p. 2

Multi-agent LLM inference

• Prior multi-agent systems use parallel sampling and aggregation, debate or mixture designs, and sequential feedback, but do not isolate topology, diversity, or model strength (Wang et al., 2023; Du et al., 2024; Wang et al., 2025; Madaan et al., 2023; Shinn et al., 2023).Later calls can also degrade performance when feedback is weak or uninformative (Huang et al., 2024).Source: §2 · p. 2

Sources of diversity

• Multi-call diversity can come from stochastic decoding, prompts or roles, or trained parameters such as LoRA/QLoRA adapters (Hu et al., 2022; Dettmers et al., 2023; Kong et al., 2024; Salewski et al., 2023).The paper's learned condition combines pooled supervised specialization with parameter heterogeneity, so gains may reflect changes in both competence and diversity.Source: §2 · p. 2

Multilingual low-resource emotion detection

• Multilingual low-resource emotion detection is shaped by typology, script overlap, pretraining coverage, and resource level, making it a setting for testing whether calls add complementary evidence or repeat cross-lingual failures.Emotional interpretation can depend on lexical and pragmatic context, sarcasm, script, and culturally specific expressions.Source: §2 · pp. 2–3

Multi-agent emotion reasoning and our gap

• The related-work gap is a controlled test of whether stochastic, prompted, and learned diversity should be aggregated in parallel or chained sequentially under the same inference budget.Closest emotion-reasoning work uses a fixed coordination design; prompting and adapter studies typically vary only one relevant factor.Source: §2 · p. 3

3.1 Task and Evaluation Setting

• Macro-F1 averages emotion F1 uniformly over languages and emotions, and the evaluation uses nine BRIGHTER languages with 8,597 held-out test instances.Uniform language averaging prevents larger test sets from dominating and makes language-level failures visible; L3 adapters train on pooled public data without per-target-language selection.Source: §3.1 · pp. 3–4

3.2 Topology–Diversity Matrix

• The six-cell matrix crosses Width, which independently predicts and majority-votes per label, with Depth, which sequentially refines and reads out its final state.Each cell uses exactly three calls per input with the same shared input package and within-backbone output schema and decision protocol.Source: §3.2 · p. 4

3.3 Diversity Sources

• Stochastic diversity varies decoding, designed prompts assign analytical lenses or sequential roles, and learned specialization uses three QLoRA adapters per frozen backbone.Because learned adapters add pooled supervised adaptation as well as inter-agent heterogeneity, the raw L3–L1/L2 difference cannot be attributed to diversity alone.Source: §3.3 · p. 4

3.4 Shared Protocol and Ablations

• All six cells share reasoning banks, precision-prior decisions, JSON schema, and a three-call budget; ablations test adapter, temperature, prompt, specialist, aggregation, uncertainty, backbone, and depth choices.The full ablation suite is reported in Appendix G.Source: §3.4 · p. 5

4 Experiments

• The six-cell matrix is evaluated on the same nine-language BRIGHTER development and held-out test splits for Qwen2.5-14B-Instruct and Llama-3.1-8B-Instruct.Topology, prompts, reasoning banks, adapter objectives, output schema, and three-call budget remain unchanged across backbones.Source: §4 · p. 5
• Qwen comparisons include zero-shot, few-shot, single-pass CoT, and seven-call CoT self-consistency baselines; larger Qwen and multilingual encoder results are contextual references.The contextual references are not matrix cells.Source: §4 · p. 5
• Key contrasts use 95% paired bootstrap intervals over 2,000 shared resamples, while learned Width–Depth comparisons also use a Wilcoxon test across nine paired language scores.Bootstrap intervals quantify evaluation-sample uncertainty; targeted W-L3 seed perturbations separately assess sensitivity to training and inference stochasticity.Source: §4 · p. 5

5.1 Topology and Diversity Effects

• Depth beats Width for stochastic and prompted diversity, but learned Width leads on Qwen by 2.83 points and is nearly tied on Llama at 0.17 points.On Qwen, confidence-max aggregation raises W-L2 to 45.12%, 4.41 points above majority vote, showing prompt signal can be underused by the canonical readout.Source: §5.1 · p. 5

5.2 Cross-Backbone Replication

• W-L3 exceeds stochastic and prompted Width by 12.17 and 12.12 points on Qwen, and by 8.39 and 7.85 points on Llama.Sequential conditioning also improves the stochastic cell on both backbones; paired tests do not establish a universal learned Width–Depth winner.Source: §5.2 · pp. 5–6

5.1 Topology and Diversity Effects

• W-L3 reaches 52.83% test Macro-F1 on Qwen and 52.94% on Llama; learned specialization drives the largest shifts in both matrices.On Qwen, W-L3 exceeds designed-prompt Width W-L2 by 12.12 points.Source: §5.1 · p. 5, Table 1 · p. 6

5.3 Comparison with Reference Systems

• With three calls, Qwen W-L3 exceeds same-backbone zero-shot by 8.74 points, few-shot by 8.60, CoT by 4.36, and seven-call CoT+self-consistency by 4.02.These averaged nine-language comparisons contextualize the gain; they do not establish uniform dominance for every language or comparator.Source: Table 2 · p. 6, §5.3 · p. 6

5.4 Per-Language Patterns

• On Qwen, W-L3 improves Emakhuwa from 2.02% to 19.16% and Zulu from 13.69% to 27.05%, while Marathi remains below the contextual RoBERTa reference.The gain over W-L1 appears on seven of nine Qwen languages and all nine Llama languages; the nine-language competence association is supporting, not definitive evidence.Source: §5.4 · p. 7

5.5 Attributing the W-L3 Gain

• Shared adaptation contributes the largest W-L3 increment: +9.38 Qwen points and +4.53 Llama points, with smaller specialist-differentiation and temperature increments.The authors interpret L3 as learned specialization rather than a pure diversity intervention.Source: Table 3 · p. 7, §5.5 · p. 7
• Individual specialists are already strong, so aggregation adds modest, backbone-dependent value rather than explaining the full W-L3 gain.Qwen specialist scores are 52.42, 50.06, and 51.75 versus 52.83 for the ensemble; Llama scores are 52.33, 47.50, and 53.52 versus 52.94.Source: §5.5 · pp. 7–8

5.6 Learned Depth vs. Learned Width

• Qwen’s learned Depth chain falls from 52.11% at Layer 1 to 50.00% overall, while Llama’s rises from 52.33% to 52.77%.Later specialists break more Layer-1 label decisions than they fix on both backbones, but the aggregate Macro-F1 effect depends on model and language distribution; the authors suggest preservation-aware calibration.Source: §5.6 · p. 8

• 12.12 Macro-F1 points on Qwen and 7.85 points on Llama separate learned QLoRA specialists from designed prompts.The conclusion identifies this as a larger effect than changing topology.Source: §6 · p. 8
• Sequential depth helps stochastic and prompted regimes, but the learned Width–Depth ordering is clear on Qwen and nearly tied on Llama.The learned Width advantage shrinks from 2.83 points on Qwen to 0.17 on Llama.Source: §6 · p. 8
• The L3 advantage is driven primarily by pooled supervised adaptation, with a smaller additional contribution from specialist heterogeneity.The conclusion recommends evaluating topology jointly with specialization rather than treating Width or Depth as universally preferable.Source: §6 · p. 8

• Two decoder-only backbones and one BRIGHTER emotion-detection domain do not establish model-independent design rules or guarantee transfer to unrelated agent tasks.The language-competence variation is diagnostically useful, but topology–specialization interactions may not transfer unchanged.Source: Limitations · p. 8
• L3 uses pooled public training data from seven Track-A languages, unlike L1 and L2.Source: Limitations · p. 8
• Paired bootstrap intervals quantify evaluation-sample uncertainty, not training-run variance; targeted W-L3 seed perturbations shift test Macro-F1 by 0.55 points on Qwen and 0.68 on Llama.Source: Limitations · p. 9
• Emakhuwa remains at 19.16% Macro-F1 even in the best Qwen cell, so multi-agent inference cannot replace missing target-language competence.Source: Limitations · p. 9

Ethical considerations

• Emotion detection may enable surveillance, profiling, or unfair moderation, so high-stakes use should not make the system an individual decision-maker and requires human review and language-community validation.The authors use public BRIGHTER annotations, report aggregate results, and reproduce no identifiable examples.Source: Ethical considerations · p. 9

A Extended Related Work

• Parallel methods aggregate candidate answers, while sequential methods refine a trajectory; prior systems often vary topology, role structure, and diversity source together.Source: A Extended Related Work · p. 11
• Parallel aggregation may cancel independent errors, whereas sequential calls may add evidence or correct mistakes; their behavior can differ with sampling, prompt, or trained-adapter diversity.Source: A Extended Related Work · p. 11
• Aggregation requires members that are both competent and meaningfully diverse; shared systematic errors can persist in multilingual low-resource inference.QLoRA specialists are framed as both a diversity and competence intervention while retaining a shared base model.Source: A Extended Related Work · p. 11
• Surface, Context, and Structure prompts bias attention toward different evidence without instructing agents to ignore other evidence, aiming to avoid quality loss from restrictive roles.Source: A Extended Related Work · p. 11
• Low-resource multilingual emotion detection tests whether calls contribute complementary evidence or repeat cross-lingual failure modes, motivating learned adapters and per-language reporting.Source: A Extended Related Work · p. 12

B Dataset, Split, and Reasoning-Bank Details

• The nine-language BRIGHTER subset uses 11,214 public training, 1,975 development, and 8,597 held-out test instances; adapters train on seven Track-A languages and evaluate on all nine.Zulu and Indonesian are Track-C languages with no target-language training rows.Source: B Dataset, Split, and Reasoning-Bank… · p. 12

B.1 Design evolution

• Version 4 uses four cue-first reasoning seeds per language as a coverage-versus-brevity compromise, reused unchanged across all six matrix cells.Source: §B.1 · p. 12

B.2 Bank JSON schema

• Table 4 documents the reasoning-bank file schema, with the same file format used for all nine languages.One file per target language is rendered into the CoT block in Table 15.Source: §B.2 · p. 12, Table 4 · p. 12

B.3 Family/proxy transfer policy

• The target-excluded reasoning bank uses same-family sources when available, falling back to script overlap, sociolinguistic contact, pretraining coverage, or affect anchors such as emoji.Chinese and Tatar proxy cases are bounded approximations, not typological matches.Source: §B.3 · pp. 12–13

C Experimental Details

• The complete six-cell matrix uses the same development and held-out test splits for both backbones, with shared banks, output schema, decision policy, and three-call budget within each backbone.Llama serves as the second-backbone replication.Source: C Experimental Details · p. 13

C.1 Comparison groups

• Same-backbone Qwen baselines include zero-shot, few-shot, single-pass CoT, and seven-call CoT with self-consistency; larger-model and cross-lingual systems provide contextual references.The primary baselines control for backbone and output protocol, while SemEval comparisons are separate because systems may use target-language supervision, per-language tuning, or larger ensembles.Source: §C.1 · p. 13

C.2 Statistical reporting

• Key contrasts use 95% paired percentile bootstrap intervals from 2,000 shared resamples, with a Wilcoxon signed-rank test across nine language-level scores for learned Width–Depth comparisons.The resampling recomputes language-uniform Macro-F1 differences using the same within-language draws for both systems.Source: §C.2 · p. 13

C.3 Compute reporting

• All matrix cells issue the same number of requests, but token counts and wall-clock time vary slightly with topology and prompt length.Qwen inference totals approximately 73.6 GPU-hours on development and 155 GPU-hours on test, excluding one-time adapter training.Source: §C.3 · p. 13

D Per-Language Matrix Detail

• Tables 7 and 8 expand the nine-language test Macro-F1 matrix results, showing learned-specialization gains are not driven by one language while revealing substantial remaining cross-language variation.The averages are uniform means over the nine languages.Source: D Per-Language Matrix Detail · p. 14

E Seed Robustness

• Alternative W-L3 seed runs remain close to canonical test Macro-F1: Qwen reaches 53.38 versus 52.83, and Llama reaches 53.62 versus 52.94.The Qwen perturbation changes the inference seed; the Llama run changes the QLoRA training seed, so these are targeted sensitivity checks rather than full-matrix variance estimates.Source: E Seed Robustness · p. 14, Table 9 · p. 15

F SemEval Reference Detail

• W-L3 improves over the RoBERTa reference on eight of nine languages but trails the per-language Top-1 systems on most languages.These comparisons are contextual because SemEval systems may use target-language supervision, per-language model selection, and larger ensembles.Source: F SemEval Reference Detail · p. 14

G Ablation Details

• The ablations target whether W-L3 gains reflect adapter weights rather than prompt or temperature choices, whether D-L3 deficits reflect learned-specialist interaction rather than depth, and whether aggregation or uncertainty changes interpretation..Source: G Ablation Details · p. 14, Table 11 · p. 17

I Aggregation and Per-Emotion Detail

• Post-hoc aggregation variants test whether Width predictions contain signal beyond majority vote, while per-emotion F1 documents that W-L3 performance remains uneven across emotion labels.The any 3 rule is mildly aggregation-sensitive for W-L3.Source: I Aggregation and Per-Emotion Detail · p. 14

J Prompt and Reasoning Protocol

• The six matrix cells reuse the same reasoning directive, language-specific bank, decision policy, JSON schema, and exactly three calls per sample; comparison baselines retain their own call counts.The comparison with seven-call CoT+self-consistency is therefore conservative with respect to call budget.Source: J Prompt and Reasoning Protocol · pp. 14–15, §J.2 · p. 15, §J.3 · p. 15

J.1 Shared-call notation

• Each response carries a binary emotion vector, confidence vector, and short reasoning field; only the binary vector is scored, while confidence supports post-hoc aggregation analyses.The shared call package includes the raw text, language identifier, and that language’s rendered reasoning bank.Source: §J.1 · p. 15

K Width Topology Prompt Designs

• Width-L1 varies decoding stochastically, Width-L2 adds three analytical lenses, and Width-L3 uses distinct QLoRA adapters; Depth passes earlier JSON outputs to later calls under a descending temperature schedule.Later Depth layers may agree, modify, or fully revise prior predictions, rather than being forced to dissent.Source: K Width Topology Prompt Designs · p. 15, L Depth Topology Inter-Layer Prompts · p. 16

H Additional Figures

• Depth truncation shows sequential refinement improves D-L1 and D-L2 over Layer 1 alone, whereas D-L3 performs best at Layer 1, implicating learned-specialist interaction rather than depth itself.The depth dynamics figure shows D-L1 and D-L2 gaining across layers and D-L3 eroding after its Layer 1 peak.Source: Table 12 · p. 17

M.1 Qwen shared training recipe

• Qwen adapters use QLoRA over a frozen Qwen2.5-14B-Instruct backbone, with assistant-turn loss and pooled public training rows; Zulu and Indonesian have no Track-C training rows.The shared recipe uses NF4 quantization, rank 8, α = 16, dropout 0.05, AdamW, bf16 compute, and development-loss early stopping.Source: §M.1 · pp. 17–19

M.2 Width-L3 specialists

• Width-L3 combines general, ambiguity, and contrastive specialists; removing them reduces test Macro-F1 by 5.29, 2.01, and 4.78 points, respectively.The general adapter trains uniformly on pooled data, ambiguity upweights multi-label and hard examples, and contrastive training aims to separate present from absent emotion categories.Source: §M.2 · pp. 19–20

M.3 Depth-L3 specialists

• Depth-L3 assigns position-specific adapters to analyst, critic, and calibrator roles, with the critic trained on Layer 1 errors and the calibrator on Layer 2 errors using weaker correction weighting.The calibrator applies a 0.2 scaling to the Layer 2 error upweight, yielding an additional error weight of 0.6.Source: §M.3 · pp. 20–21
• D-L3 Layer 1 alone reaches 52.11 test Macro-F1, while the full learned chain reaches 50.00, indicating that later learned correction can be miscalibrated after a strong first specialist.The truncation results contrast with D-L1 and D-L2, which improve through sequential layers.Source: §M.3 · p. 21

M.4 Multi-LoRA serving

• Multi-LoRA serving hot-swaps adapters on one loaded frozen base model, preserving Width-L3’s three-call budget without serving three separate full models.Width predictions use the same per-label majority vote as other Width cells; Depth-L3 applies position-specific adapters sequentially.Source: §M.4 · p. 21

N Source Code and Configuration Map

• The reproducibility package provides source code, prompts, configurations, evaluation scripts, experiment utilities, and minimum software requirements.A source-code map connects the main experimental components to their implementation paths.Source: N Source Code and Configuration Map · pp. 21–22, Table 25 · p. 23