HCMUS RepeatedGames at SemEval-2026 Task 12: CausalRAG: Synergizing Causal Graph Retrieval and Extended LoRA for Abductive Reasoning

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

Summary

The HCMUS RepeatedGames team developed CausalRAG for SemEval-2026 Task 12: Abductive Event Reasoning (AER), a shared task focused on identifying the most plausible cause of real-world events from multiple-choice options, supported by retrieved evidence. Their system employs a hybrid retrieval approach, integrating BM25 keyword matching with dense semantic search to effectively capture causal keywords. CausalRAG also utilizes extended LoRA fine-tuning, which trains both attention and MLP layers of a 32-billion parameter language model using only 0.81% of trainable parameters. A final refinement step involves development set fine-tuning before inference. The system achieved a score of 0.90 on the official test set, securing a tie for fifth place among participating teams and demonstrating a +0.27 improvement over their baseline.

Key takeaway

For NLP Engineers developing abductive reasoning systems, consider integrating hybrid retrieval and efficient fine-tuning techniques. Your approach could benefit from combining BM25 with dense semantic search to improve evidence capture. Additionally, applying extended LoRA to large language models, even with minimal trainable parameters like 0.81%, can yield significant performance gains, as demonstrated by the 0.90 score on SemEval-2026 Task 12.

Key insights

CausalRAG combines hybrid retrieval and extended LoRA fine-tuning for abductive event reasoning, achieving competitive performance.

Principles

• Hybrid retrieval enhances causal keyword capture.
• Extended LoRA can efficiently fine-tune large LMs.
• Development set fine-tuning refines inference.

Method

The system uses hybrid retrieval (BM25 + dense semantic search), followed by extended LoRA fine-tuning on a 32-billion parameter LM (0.81% trainable parameters), and concludes with development set fine-tuning for refinement.

In practice

• Combine BM25 with semantic search for causal tasks.
• Apply extended LoRA for efficient large LM adaptation.
• Use validation data for final model refinement.

Topics

• Abductive Reasoning
• SemEval-2026
• Causal Graph Retrieval
• Extended LoRA
• Hybrid Retrieval
• Large Language Models

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP 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.
• System Report for CCL25-Eval Task 5: New Dataset and LoRA-Fine-Tuned Qwen2.5 — Domain-specific datasets and LoRA fine-tuning significantly enhance LLM performance in specialized tasks like classical poetry analysis.
• A2RAG: Adaptive Agentic Graph Retrieval for Cost-Aware and Reliable Reasoning — A2RAG uses adaptive control and agentic retrieval to provide cost-aware, reliable GraphRAG by progressively acquiring evidence and recovering provenance.
• Learning to Reason by Analogy via Retrieval-Augmented Reinforcement Fine-Tuning — Reasoning-aware retrieval, guided by a judge model, enables language models to learn analogical problem-solving via reinforcement fine-tuning.
• Causal Scaffolding for Physical Reasoning: A Benchmark for Causally-Informed Physical World Understanding in VLMs — VLMs struggle with causal physical reasoning; CausalPhys benchmark and Causal Rationale-informed Fine-Tuning (CRFT) offer evaluation and improvement.
• Statistical Parsing for Logical Information Retrieval — Integrating LLMs with formal grammars enables robust statistical parsing for logical information retrieval.

Open in AIssential →

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