Towards Reliable Long-Context Reasoning: Detecting Unanswerable Questions via FactGuard

The gist The FactGuard framework introduces an innovative multi-agent collaborative methodology to autonomously generate answerable and unanswerable questions, developing the FactGuard-Bench dataset to evaluate long-context LLM extraction capabilities.

How it works

The core of the method is a multi-agent collaboration framework designed for automated data augmentation, which operates through three primary stages: preparation, QA generation, and negative example generation <ref:2504.05607#pg2> The agent console is responsible for aggregating the opinions of each agent and making the final data synthesis decision <ref:2504.05607#pg4>.

The preparation stage involves selecting short text segments from extensive documents, refining this process through Quality Scoring and Topic Selection <ref:2504.05607#pg5>.

** Quality Scoring utilizes quality agents to assign a quality score to each segment, mapping it into five quality intervals scorei in [1, 5] <ref:2504.05607#pg5>. The Topic Selection employs topic agents to select diverse topics from the text segments, ensuring coverage of categories such as time, numerical values, locations, persons, organizations, events, and objects <ref:2504.05607#pg5>. The agent console aggregates these inputs to make a final data synthesis decision <ref:2504.05607#pg4>. (Page 3) **

The QA Generation Stage leverages QA generation agents to produce tuples in the form of (Fragment, Question, Answer, Evidence) <ref:2504.05607#pg5>. This step ensures that each question is firmly grounded in the provided context <ref:2504.05607#pg5>. After generating these tuples, agents trigger a quality judgment mechanism to filter out low-quality QA pairs <ref:2504.05607#pg4>. (Page 3)

The Negative Example Generation Stage focuses on creating unanswerable questions by manipulating the previously generated tuples, mimicking real-world Negative Rejection scenarios <ref:2504.05607#pg5>. This involves two distinct approaches:

1. Contextually Missing Negative Example Generation: This is achieved by simply removing the evidence from the text, thus making the question unanswerable due to lack of information <ref:2504.05607#pg5>.

2. Misleading Negative Example Generation: Question rewriting agents perform entity substitutions, impossible condition insertions, and other types of false assumptions <ref:2504.05607#pg5>.

The process is formalized in Algorithm 1, which details the construction of FactGuard-Bench <ref:2504.05607#pg5>. This algorithm shows how segments are scored, filtered by topic, and then processed through Stage 2 (QA Generation) or Stage 3 (Negative Example Generation) depending on the condition <ref:2504.05607#pg5>.

Benchmark Construction and Evaluation

FactGuard-Bench is a comprehensive benchmark comprising 25,220 data examples generated from 16,742 texts <ref:2504.05607#pg4>. The dataset includes English (en) and Chinese (zh) across two domains: law and books <ref:2504.05607#pg4>. Unanswerable questions are categorized into two types: Contextually Missing Negative Examples and Misleading Negative Examples <ref:2504.05607#pg4>. The example lengths range from 8K to 128K tokens <ref:2504.05607#pg4>.

The data generation process involves dynamic creation of questions by leveraging a multi-agent collaboration process applied to raw, lengthy texts <ref:2504.05607#pg4>. The efficiency of the data synthesis process is managed by configurable parameters like alpha (alpha=1 in this experiment) which defines the ratio between raw textual data and selected segments <ref:2504.05607#pg5>. Post-processing review procedures are applied after initial agent processing to remove questions that fail requirements, such as those with conflicting answers or context-independent common sense, resulting in a reduction of approximately 21% in the number of examples <ref:2504.05607#pg5>.

Experimental Results and Findings

Experimental evaluations on seven popular LLMs reveal critical shortcomings in current models <ref:2504.05607#pg4>. Even the best-performing model achieves an overall accuracy of 61.79% <ref:2504.05607#pg4>. The evaluation tasks focus on two aspects: Answer Consistency Evaluation and Reasoning Ability for Unanswerable Questions <ref:2504.05607#pg4>.

For Answer Consistency Evaluation, the highest overall accuracy observed is 82.39%, achieved by the model augmented with both SFT and DPO <ref:2504.05607#pg4>. The results show that while baseline models perform well on answerable questions, their performance on unanswerable questions is suboptimal <ref:2504.05607#pg4>. For instance, Qwen2.5-72B achieves an 86.25% accuracy on answerable questions but only manages 63.34% and 63.16% on lack of and misleading evidence, respectively <ref:2504.05607#pg4>.

For Reasoning Ability for Unanswerable Questions, the use of SFT and DPO significantly improved performance, with the best-performing model variant reaching an overall performance enhancement from 41.21% before SFT to 82.39% after DPO <ref:2504.05607#pg4>. This demonstrates that techniques like SFT and DPO can serve as effective strategies for bolstering LLMs’ reasoning abilities, increasing the rates of direct refusals and reasoned answers up to 60.14% in the best-performing model <ref:2504.05607#pg4>.

Conclusion

FactGuard enables flexible generation of answerable and unanswerable questions, introducing a paradigm shift in the evaluation and enhancement of long-context machine reading comprehension <ref:2504.05607#pg4>. The development of FactGuard-Bench extends the challenge to significantly longer contexts, pushing the boundaries of current LLMs <ref:2504.05607#pg4>. The empirical results underscore the significant challenges that remain in the domain of machine reading comprehension, especially when dealing with unanswerable questions <ref:2504.05607#pg4>. Future work will focus on enhancing the FactGuard framework and expanding FactGuard-Bench to include a wider range of contexts and question types with lower noise <ref:2504.05607#pg4>.

The paper concludes that handling unanswerable questions in extended textual inputs requires models to both identify gaps in evidence and reject misleading premises, paralleling tasks in recognizing textual entailment (RTE) [16] and relation extraction under uncertain conditions [43] <ref:2504.05607#pg4>. The overall quality of the synthetic data, with 93% for good-quality unanswerable QA pairs, indicates the high value of the method <ref:2504.05607#pg4>.

The paper is a scientific paper on arXiv 2504.05607 titled Towards Reliable Long-Context Reasoning: Detecting Unanswerable Questions via FactGuard and it matters because it proposes a novel multi-agent framework and benchmark to address the critical challenge of reliably recognizing unanswerable queries in long-context LLMs.

The QA Generation Stage leverages QA generation agents to produce tuples in the form of (Fragment, Question, Answer, Evidence) <ref:2504.05607#pg5>. This step ensures that each question is firmly grounded in the provided context <ref:2504.05607#pg5>.

1. textbfThe FactGuard Multi-Agent Framework for Data Augmentation: Generation of Contextually Challenging Examples and Unanswerable Queries; The framework dynamically generates answerable and unanswerable questions by leveraging a multi-agent collaboration process, resulting in contextually difficult examples. This allows the system to autonomously create high-quality training data without costly manual annotation processes like SQuAD 2.0, significantly reducing costs while increasing dataset diversity and complexity for long-context LLMs (8K to 128K tokens).

2. textbfLong-Context Evaluation Benchmark: FactGuard-Bench; This benchmark is specifically curated to evaluate the models’ abilities in addressing answerable and unanswerable questions within extended contexts, providing a necessary evaluation tool where current models show critical shortcomings compared to answerable ones, allowing for targeted performance improvement.

3. textbfEnhanced Reasoning Capability for Unanswerable Queries: Implementation of SFT and DPO Fine-tuning; The combination of SFT and DPO improved the model’s ability to verify the validity of the dataset and resulted in a 61.05% improvement in handling misleading evidence, while also increasing the rates of direct refusals and reasoned answers, suggesting models can now better reject [unanswerable questions] outright or to provide well-reasoned responses.

4. textbfContext-Aware Question Filtering: Retrieval Augmented Generation (RAG) for Negative Example Generation; By using RAG, the system can extract the first N relevant passages from a lengthy article for short-reading comprehension and to filter out data that contain conflicting answers, enhancing efficiency in detecting inconsistencies before generating misleading negative examples.

Despite the success of Large Language Models (LLMs) in long-context tasks, their reliability in recognizing unanswerable questions remains a critical challenge. In information-seeking scenarios, models often suffer from overconfidence and hallucinations when evidence is missing or misleading. In this paper, we propose FACTGUARD, a systematic multi-task workflow for generating evidence-grounded answerable and unanswerable questions. We introduce FACTGUARD-BENCH, a benchmark comprising 25,220 examples with context lengths up to 128K, specifically targeting Lack of Evidence and Misleading Evidence scenarios. Evaluations of eleven representative open-source and proprietary LLMs reveal a significant reliability gap: models struggle to reject unanswerable questions as context scales, with the leading model achieving only 71.62% accuracy. By training with FACTGUARD-BENCH, we improve overall accuracy to 81.17% and significantly enhance models' reasoned rejection capabilities. Our work provides a scalable path toward building more faithful and trustworthy long-context LLMs. Our source is publicly available at https://github.com/biaoliu-kiritsugu/FactGuard.