sutta at SemEval-2026 Task 12: A Multi-Perspective Retrieve-Verify-Aggregate Framework for Abductive Event Reasoning

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, medium

Summary

The system "sutta" for SemEval-2026 Task 12: Abductive Event Reasoning (AER) identifies direct causes of real-world events from multiple-choice options using retrieved documents. Instead of fine-tuning, this zero-shot system, built around the Qwen3-8B model, employs a "Retrieve-Verify-Aggregate" pipeline. It first isolates relevant evidence using BM25 and cross-encoder reranking. To evaluate causal links, the model is prompted with several distinct "personas," and their independent decisions are aggregated via majority voting. This approach achieved a score of 0.7614 on the official test set, demonstrating that strict retrieval combined with diverse reasoning prompts can enable compact open-source models to perform complex causal inference without task-specific training.

Key takeaway

For NLP Engineers developing causal reasoning systems, especially with compact models like Qwen3-8B, consider implementing a zero-shot "Retrieve-Verify-Aggregate" pipeline. This approach, using diverse reasoning prompts and majority voting, allows your models to perform complex causal inference effectively without extensive task-specific fine-tuning. You can achieve strong performance, like the 0.7614 score, by focusing on robust retrieval and multi-perspective verification.

Key insights

Zero-shot multi-perspective reasoning with retrieval enhances compact LLMs for complex causal inference without fine-tuning.

Principles

• Diverse prompts improve causal inference.
• Strict retrieval reduces irrelevant context.
• Aggregating multiple perspectives boosts decision accuracy.

Method

A "Retrieve-Verify-Aggregate" pipeline uses BM25 and cross-encoder reranking for evidence, then prompts Qwen3-8B with distinct "personas" for causal link evaluation, aggregating decisions via majority voting.

In practice

• Use BM25 and cross-encoders for retrieval.
• Employ "personas" for varied model reasoning.
• Implement majority voting for decision aggregation.

Topics

• Abductive Event Reasoning
• SemEval-2026
• Zero-shot Learning
• Retrieve-Verify-Aggregate
• Qwen3-8B
• Causal Inference

Best for: Research Scientist, AI Scientist, NLP Engineer, Prompt Engineer

Related on AIssential

• AILS-NTUA at SemEval-2026 Task 12: Graph-Based Retrieval and Reflective Prompting for Abductive Event Reasoning — AILS-NTUA's three-stage system won SemEval 2026 Task 12 for Abductive Event Reasoning, achieving a first-place rank with an accuracy score of 0.95.
• MARCH: Evaluating the Intersection of Ambiguity Interpretation and Multi-hop Inference — Multi-hop QA requires models to navigate layered ambiguity across complex reasoning paths.
• Predicting Causal Effects from Natural Language Queries using Structured Representations — Finetuning and structured representations significantly enhance LLM performance in predicting causal effects from natural language.
• IdealGPT: Iteratively Decomposing Vision and Language Reasoning via Large Language Models — IdealGPT iteratively decomposes vision-and-language reasoning tasks using LLMs and VLMs to improve zero-shot performance.
• Zero-Shot Goal Recognition with Large Language Models — LLM goal recognition competence varies significantly, with some models integrating evidence effectively while others rely on world-knowledge priors.
• FIDES: Faithful Inference via Deep Evidence Signals for Retrieval-Memory Conflict in RAG — FIDES resolves RAG retrieval-memory conflict by selectively applying contrastive decoding based on token-level conflict concentration.

Open in AIssential →

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