d-itlab at SemEval-2026 Task 12: Per-Option Surprisal and Multi-Stage Gating for Precision-Oriented Causal 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 d-itlab system, submitted to SemEval-2026 Task 12 (Abductive Event Reasoning), addresses the challenge of identifying the most plausible direct cause(s) for an observed event from candidate options, using reference documents for grounding. Their approach integrates three key components: a per-option multi-stage LLM inference process that independently evaluates each option with progressively stricter verification; surprisal-based features derived from teacher-forcing candidate sentences and measuring token-level negative log-likelihood; and an XGBoost ensemble model trained on these diverse features to generate a precision-oriented final prediction. This system achieved a score of 0.91 on the official test set, securing the third rank among 116 participating teams.

Key takeaway

For NLP Engineers developing precision-oriented causal reasoning systems, consider integrating a multi-stage LLM inference approach with surprisal-based features. Your systems can benefit from progressively stricter verification and token-level negative log-likelihood to quantify plausibility. Training an XGBoost ensemble on these heterogeneous features can significantly boost predictive precision, as demonstrated by d-itlab's third-place ranking at SemEval-2026 Task 12.

Key insights

The d-itlab system combines multi-stage LLM inference, surprisal features, and XGBoost for precision-oriented abductive reasoning, ranking third at SemEval-2026.

Principles

• Combining diverse features enhances precision.
• Staged verification improves LLM inference.
• Surprisal quantifies causal plausibility.

Method

The method involves per-option multi-stage LLM inference, generating surprisal-based features via token-level negative log-likelihood, and training an XGBoost ensemble on these heterogeneous features for precision-oriented causal prediction.

In practice

• Use multi-stage LLM inference for verification.
• Apply surprisal features for causal tasks.
• Employ XGBoost for feature ensemble.

Topics

• Abductive Reasoning
• Causal Reasoning
• Large Language Models
• Surprisal Features
• XGBoost
• Multi-stage Inference
• SemEval-2026

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

Related on AIssential

• STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability — Policy entropy collapse in LLM RL can be stabilized by reweighting token-level advantages based on surprisal and using a closed-loop entropy gate.
• From Correlation to Cause: A Five-Stage Methodology for Feature Analysis in Transformer Language Models — The five-stage methodology links correlation to causation in Transformer feature analysis, revealing partial causality and robustness gaps.
• 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.
• Sequential statistical inference for Large Language Models: Representation, validity, and monitoring — Sequential statistical inference enhances LLM trustworthiness by treating interactions as dependent stochastic processes.
• 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.
• Structured Abductive-Deductive-Inductive Reasoning for LLMs via Algebraic Invariants — A symbolic scaffold improves LLM logical reasoning by separating inference types and enforcing consistency via algebraic invariants.

Open in AIssential →

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