VARH-AI at SemEval-2026 Task 10: Exploiting Architectural Diversity with Transformer-SSM Ensembles and Confidence-Based Iterative Refinement for Conspiracy Detection

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

VARH-AI's system for SemEval 2026 Task 10 (PsyCoMark) Subtask 2 achieved a weighted F1 score of 0.78, ranking 4th on the CodaBench leaderboard for binary conspiracy classification in Reddit submission statements. The system employs a heterogeneous ensemble approach, integrating Transformer-based models like DeBERTa and RoBERTa with State-Space Models (SSMs) such as Mamba, to exploit architectural diversity for enhanced generalization. Key innovations include Bidirectional Mamba (BiMamba), which adapts SSMs for bidirectional document classification, and a safety-switched multi-task training setup that applies marker supervision exclusively to gold-annotated samples, preventing noisy pseudo-labels from impacting span extraction. Additionally, Confidence-Based Iterative Refinement utilizes committee voting to generate high-quality pseudo-labels. Ablation studies confirm the complementary contributions of these architectural components.

Key takeaway

For NLP engineers developing robust text classification systems, particularly for sensitive tasks like conspiracy detection, you should explore heterogeneous ensembles combining Transformers with State-Space Models like BiMamba. This approach, coupled with confidence-based iterative pseudo-label refinement, can significantly boost your model's generalization and F1 score, as demonstrated by the 0.78 F1 score on SemEval 2026 Task 10. Consider implementing safety-switched multi-task training to prevent noisy data from degrading performance.

Key insights

Heterogeneous ensembles of Transformers and State-Space Models, combined with iterative refinement, enhance conspiracy detection performance.

Principles

• Architectural diversity improves generalization.
• Bidirectional adaptation enhances sequence models.
• Confidence-based voting refines pseudo-label quality.

Method

The system uses a heterogeneous ensemble of Transformer and Mamba models, applies safety-switched multi-task training, and refines pseudo-labels via confidence-based iterative committee voting.

In practice

• Integrate BiMamba for bidirectional text classification.
• Use safety-switched multi-task training.
• Implement committee voting for pseudo-labeling.

Topics

• Conspiracy Detection
• SemEval 2026
• Transformer Ensembles
• State-Space Models
• Bidirectional Mamba
• Pseudo-labeling
• Multi-task Training

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• From TF-IDF to Transformers: A Comparative and Ensemble Approach to Sentiment Classification — A soft voting ensemble of diverse machine learning and transformer models significantly enhances sentiment classification accuracy, particularly for nuanced text.
• Learning Uncertainty from Sequential Internal Dispersion in Large Language Models — Uncertainty in LLMs can be reliably inferred from the dispersion of hidden states across layers and token sequences.
• InfoMamba: An Attention-Free Hybrid Mamba-Transformer Model — InfoMamba combines linear filtering and selective SSMs via information-maximizing fusion for efficient, high-performance sequence modeling.
• Short-form Text Rewriting with Phi Silica — Targeted adaptation of Small Language Models (SLMs) can significantly improve their performance for precision-critical short-form text rewriting tasks.
• DiZiNER: Disagreement-guided Instruction Refinement via Pilot Annotation Simulation for Zero-shot Named Entity Recognition — Simulating human pilot annotation with LLMs to refine instructions significantly boosts zero-shot NER performance.
• LLM-based Detection of Manipulative Political Narratives — LLM-driven pipelines can effectively identify and cluster manipulative political narratives in raw social media data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.