Hallucination as output-boundary misclassification: a composite abstention architecture for language models

A new research paper by Angelina Hintsanen, accepted to the ICLR 2026 Workshop on LLM Reasoning, reframes the persistent problem of AI 'hallucination' as an 'output-boundary misclassification.' The core idea is that language models incorrectly emit internally generated completions as if they were grounded in external evidence. To address this, the paper introduces a composite intervention architecture that merges two distinct control mechanisms.

The first mechanism is instruction-based refusal, where the model is prompted to decline answering if it lacks sufficient knowledge. The second is a novel structural abstention gate. This gate calculates a 'support deficit score' (St) by analyzing three black-box signals from the model's output: self-consistency (At), stability under paraphrasing (Pt), and citation coverage (Ct). If the score exceeds a threshold, the output is blocked.

In a controlled evaluation across 50 items, five epistemic scenarios, and three models (including GPT-3.5-turbo), neither mechanism alone was sufficient. Instruction prompting reduced hallucinations but was overly cautious and missed some errors. The structural gate preserved accuracy on answerable questions but failed on items with conflicting evidence. The composite architecture, however, successfully combined their strengths, achieving high overall accuracy with significantly reduced hallucination, though it inherited some over-caution from the instruction component.

A supplementary 100-item stress test derived from TruthfulQA confirmed that the structural gating provides a reliable, capability-independent 'abstention floor.' The research concludes that instruction-based refusal and structural gating have complementary failure modes, making their combination a promising path for more reliable and trustworthy language model deployments.