Teaching Models to Abstain via Unanswerable Queries

19 concepts16 papers1 proposal

When a question-answering (QA) model is trained exclusively on pairs where the answer is guaranteed to exist within the provided text, it learns a dangerous implicit bias: always extract or generate a response. In real-world deployments, such as Retrieval-Augmented Generation (RAG) or clinical assistants, this bias manifests as hallucination. When the provided context lacks the necessary information, the model forces a match anyway, fabricating a plausible but incorrect answer.

As demonstrated by Sophie Henning et al. in their evaluation framework, modern large language models (LLMs) frequently struggle to recognize when a question is unanswerable based on a text, often hallucinating evidence instead of admitting an information void. To counteract this, a prominent research direction focuses on unanswerable query training (or negative constraint training). By intentionally integrating samples where the context cannot satisfy the query, researchers force models to learn a critical behavioral boundary: verifying the existence of evidence before attempting to formulate an answer.

The Mechanics of Information Voids

The core mechanism of this approach is straightforward: during training, a significant portion of the dataset is populated with negative samples—queries paired with documents or sensory inputs that are contextually relevant but do not contain the actual answer.

Early reading comprehension datasets recognized this need. For instance, the NewsQA dataset introduced by Adam Trischler et al. allowed crowd-workers to mark a "null span" when the source text lacked the required information. This design choice explicitly trains models to recognize the limits of their provided context.

This concept was later elevated to a pre-training primitive. Michael R. Glass et al. introduced Span Selection Pre-Training (SSPT), which incorporates a subset of unanswerable instances directly into the self-supervised pre-training phase. Rather than allowing the model to rely on its internal parametric memory to guess an answer, SSPT forces the model to evaluate the text's sufficiency, shifting its fundamental objective from fact memorization to strict reading comprehension.

Why It Works: Breaking Superficial Matching

Why does exposing a model to unanswerable questions fundamentally change its behavior?

Standard deep learning models are notorious for exploiting shortcuts, such as superficial keyword matching. If a question shares several words with a paragraph, the model's attention mechanisms naturally align with that paragraph, prompting it to extract nearby text as the answer.

Standard Training:
[Query: "When was the treaty signed?"] + [Context: "The treaty was signed in 1919..."] 
  --> Target: "1919" (Encourages keyword matching)

Unanswerable Query Training:
[Query: "When was the treaty signed?"] + [Context: "The treaty discussed borders..."] 
  --> Target: [NULL / ABSTAIN] (Forces evidence verification)

As argued by Jiaqi Li et al., integrating unanswerable questions breaks this reliance on surface-level patterns. When a model is penalized for extracting an answer from a contextually related but factually insufficient passage, it learns that keyword overlap is an unreliable signal. Instead, it must verify the logical entailment between the query and the context.

This acts as a "negative constraint," a concept utilized by Pei Guo et al. to prevent chatbots from forcing incorrect intent matches when the required metadata is absent. The model learns to map an information gap to an explicit token or phrase of abstention (such as "not mentioned" or a clarification request) rather than fabricating a response.

Scaling to Generative and RAG Systems

In modern generative architectures and RAG pipelines, this training strategy is essential for factual safety. For example, the ChatQA framework developed by Zihan Liu et al. demonstrates that incorporating even a small proportion of unanswerable samples during fine-tuning drastically reduces hallucinations. It teaches the generative model to explicitly acknowledge missing information rather than generating a plausible but false response.

Similarly, Evgeniia Razumovskaia et al. utilize "adequacy training" in dialogue systems, exposing models to instances where the knowledge base contains no relevant information. This behavioral fine-tuning teaches the system to ignore irrelevant "distractor" facts and confidently state its ignorance. Avi Sil et al. apply a similar philosophy to long-form RAG systems via the CLAPNQ benchmark, proving that high proportions of unanswerable questions are required to build robust abstention capabilities in abstractive models.

To avoid the high cost of human annotation for these edge cases, researchers have turned to automated synthesis. Yihang Wang et al. introduced MDPO, a framework that uses an unsupervised metric-based sampler to automatically generate high-quality negative samples. By exposing the model to these synthetic unanswerable queries during preference optimization, the system learns to identify information gaps autonomously.

Cross-Modal and Domain-Specific Robustness

The intuition behind unanswerable query grounding extends far beyond standard text documents; it is highly effective across complex layouts, specialized domains, and multiple modalities.

Document and Visual Understanding

When navigating multi-page documents or mobile user interfaces, models face severe spatial and structural noise. Document understanding benchmarks like DUDE, presented by Jordy Van Landeghem et al., and JDocQA, introduced by Eri Onami et al., explicitly use unanswerable queries to evaluate whether a model can recognize information gaps across complex multi-page layouts. In the visual domain, ScreenQA (Yu-Chung Hsiao et al.) includes questions lacking answers within mobile screenshots, training models to signal information absence directly from raw pixels.

Breaking Agreement Bias in Vision and Audio

Vision-Language Models (VLMs) often suffer from "agreement bias," where they blindly agree with a user's prompt even if it contains a false premise. To combat this, the HaloQuest dataset by Zhecan James Wang et al. trains models on questions that lack sufficient visual context or contain false premises, teaching the system to resist the prompt's assumptions and admit uncertainty. This principle applies to audio as well: Yuan Gong et al. demonstrate that integrating unanswerable questions into audio-linguistic models teaches them to recognize when sensory evidence is insufficient, significantly lowering audio-based hallucination rates.

High-Stakes Domains

In specialized fields like medicine, the cost of a hallucination can be catastrophic. Amin Dada et al. applied unanswerable question-answer pairs to fine-tune models reading unstructured radiological reports. By learning to output a null response when a query (e.g., "Is the hemorrhage growing?") cannot be answered by the text, the model achieves the high precision required for clinical safety.

Turning the Mechanism into an Evaluation Metric

Beyond improving generation, the robustness gained from unanswerable query training can be inverted to create better evaluation metrics.

The QuestEval framework, developed by Thomas Scialom et al., trains an internal QA model using synthetic negative samples (questions that cannot be answered by a reference text). Because this internal QA model has been rigorously trained to recognize when a question is unanswerable, it can be used to evaluate other text generation systems.

If a summarization model generates a summary containing a fact not present in the source document, the trained QA model will correctly flag the question corresponding to that fact as unanswerable, successfully exposing a neural hallucination without needing a human-written gold standard.

Key Foundational Work

Span Selection Pre-training for Question Answering (2020) — contributed 1 concept elements to this direction.
NewsQA: A Machine Comprehension Dataset (2016) — contributed 1 concept elements to this direction.
Molweni: A Challenge Multiparty Dialogues-based Machine Reading Comprehension Dataset with Discourse Structure (2020) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_hallucination_mitigation_67b52d11,
  title = {hallucination mitigation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hallucination-mitigation/rdr_67b52d11c6774a6ab8f1ba4773290b0d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_hallucination_mitigation_67b52d11_v1,
  title = {hallucination mitigation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hallucination-mitigation/rdr_67b52d11c6774a6ab8f1ba4773290b0d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake