uir-cis at SemEval-2026 Task 12: Mitigating Prior-Induced Hallucinations in Retrieval-Augmented Reasoning via Precision-Oriented Decoding

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

Summary

The uir-cis system, developed for the SemEval-2026 Task 12 on Abductive Event Reasoning (AER), addresses "over-selection" hallucinations in Instruction-tuned Large Language Models (LLMs). This pathology causes models to incorrectly prioritize semantic priors over retrieved evidence. The system employs a 32-billion parameter Qwen2.5 foundational model, fine-tuned using Low-Rank Adaptation (LoRA) and evaluated in a Zero-shot Chain-of-Thought (CoT) setting. To combat epistemic noise, the researchers propose a Precision-Oriented Decoding (POD) strategy, combining low-temperature sampling (T=0.45) with scaled majority voting (K=9). After a three-stage empirical development, the system achieved a score of 0.802 on the official test set. Findings indicate that suppressing epistemic noise is more effective than heuristic recall compensation in causal reasoning tasks with strict penalties for incorrect predictions.

Key takeaway

For Machine Learning Engineers developing LLM-based causal reasoning systems, you should prioritize strategies that suppress epistemic noise over those focused on recall compensation. Implement Precision-Oriented Decoding by combining low-temperature sampling (e.g., T=0.45) with scaled majority voting (e.g., K=9) to reduce "over-selection" hallucinations where models rely on priors instead of evidence. This approach demonstrably improves precision in tasks with strict penalization for incorrect predictions.

Key insights

Precision-Oriented Decoding effectively mitigates prior-induced hallucinations in LLMs for abductive event reasoning.

Principles

• Epistemic noise suppression outperforms heuristic recall compensation in penalized causal reasoning.
• LLMs can erroneously prioritize semantic priors over retrieved evidence.

Method

A Precision-Oriented Decoding (POD) strategy combines low-temperature sampling (T=0.45) with scaled majority voting (K=9) on a LoRA-adapted Qwen2.5 model in a Zero-shot CoT setting.

In practice

• Implement low-temperature sampling (T=0.45) for decoding.
• Apply scaled majority voting (K=9) to model outputs.
• Utilize LoRA for fine-tuning large foundational models.

Topics

• Abductive Event Reasoning
• LLM Hallucinations
• Precision-Oriented Decoding
• Retrieval-Augmented Reasoning
• Low-Rank Adaptation
• Qwen2.5

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

Related on AIssential

• Hallucinations in LLMs: A Deep Technical Dive into Causes, Detection, and Mitigation — LLM hallucinations are inherent to their probabilistic nature, driven by training objectives that prioritize fluency over factual accuracy.
• Thinking to recall: How reasoning unlocks parametric knowledge in LLMs — Reasoning in LLMs unlocks parametric knowledge through latent computation and factual priming, even for simple facts.
• ADAPT: Attention Dynamics Alignment with Preference Tuning for Faithful MLLMs — MLLM hallucination stems from progressive text-to-image cross-attention degradation, which ADAPT mitigates by aligning attention dynamics.
• Mitigating Hallucinations in Large Language Models Via Decoder Layer Skipping — DeLask mitigates LLM hallucinations by dynamically skipping deeper decoder layers identified via gradient driftance.
• Enhancing Video Representations with Spatiotemporal-Semantic Residual to Mitigate Hallucinations in Video Large Multimodal Models — SSCD mitigates VideoLLM hallucinations by using deliberately disrupted spatiotemporal and semantic negative features in contrastive decoding.
• From Flat Facts to Sharp Hallucinations: Detecting Stubborn Errors via Gradient Sensitivity — EPGS detects confident LLM hallucinations by probing loss landscape curvature via input embedding perturbations.

Open in AIssential →

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