YNU-HPCC at SemEval-2026 Task 12: Retrieval-Guided Reasoning with Teacher Distillation for Abductive Event Reasoning

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

Summary

The YNU-HPCC system, developed for SemEval-2026 Task 12 on Abductive Event Reasoning (AER), addresses the challenge of identifying direct causes from multi-document retrieved evidence containing distractors. This task is complicated by sparse evidence in long documents and a boundary-sensitive set-level evaluation. The system employs a two-stage framework. Stage 1 utilizes a DeBERTa-v3-base student model with retrieval-guided evidence modeling, where documents are split into overlapping windows, BM25 ranks and filters candidates, and Top-K pooling aggregates scores. Stage 2 enhances generalization by distilling soft targets from a Qwen-14B teacher model, incorporating temperature scaling and high-confidence filtering to mitigate pseudo-label noise. The system achieved an official dev score of 0.9712 (micro-F1 0.9746, macro-F1 0.9745) and improved its test score from 0.46 to 0.73, securing 84th place among 221 submissions.

Key takeaway

For NLP Engineers tackling abductive reasoning or similar tasks with sparse, multi-document evidence, consider implementing a two-stage framework. Your system could benefit from an initial retrieval-guided stage using document windowing and BM25 filtering to manage long contexts, followed by a teacher distillation stage. This approach, leveraging a larger teacher model like Qwen-14B for soft targets and noise reduction, can significantly improve generalization and F1 scores, as demonstrated by the YNU-HPCC system's performance increase from 0.46 to 0.73.

Key insights

Retrieval-guided reasoning combined with teacher distillation effectively improves abductive event reasoning performance on sparse evidence.

Principles

• Address sparse evidence with windowing and retrieval.
• Aggregate evidence scores via Top-K pooling.
• Teacher distillation improves generalization and reduces noise.

Method

A two-stage framework: first, a DeBERTa-v3-base student uses BM25-filtered windows and Top-K pooling; second, a Qwen-14B teacher distills soft targets with temperature scaling and high-confidence filtering.

In practice

• Split long documents into overlapping windows.
• Use BM25 for initial evidence ranking.
• Apply temperature scaling in knowledge distillation.

Topics

• Abductive Event Reasoning
• Retrieval-Guided Reasoning
• Teacher Distillation
• DeBERTa-v3
• Qwen-14B
• SemEval-2026

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

Related on AIssential

• EGTR-Review: Efficient Evidence-Grounded Scientific Peer Review Generation via Multi-Agent Teacher Distillation — Multi-agent teacher distillation enables efficient, evidence-grounded, and traceable scientific peer review generation by training a lightweight student model.
• 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.
• ScholarSum: Student-Teacher Abstractive Summarization via Knowledge Graph Reasoning and Reflective Refinement — ScholarSum integrates hierarchical knowledge graphs and iterative student-teacher refinement for fluent, factually faithful scientific summarization.
• Rethinking Negative Sampling for Knowledge Distillation in Retrieval, A Distillation Recipe for Small Language Model Search Agents, and More! — Effective information retrieval advancements often stem from rethinking fundamental assumptions and optimizing data representation or model behavior.
• LoRi: Low-Rank Distillation for Implicit Reasoning — Hidden-state reasoning trajectories exhibit low-rank structure, enabling efficient distillation for implicit reasoning.
• MCompassRAG: Topic Metadata as a Semantic Compass for Paragraph-Level Retrieval — MCompassRAG uses topic metadata as a semantic compass to enhance RAG retrieval, improving efficiency and evidence quality by enriching chunk embeddings.

Open in AIssential →

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