HiddenLens: A Multi-Layer LLM Framework for Detecting and Resolving Hidden Assumptions in Software Requirements
We propose HiddenLens, a novel six-layer BERT-LLM pipeline and, to our knowledge, the first framework engineered to detect, classify, question, and resolve hidden assumptions rather than merely flag ambiguity.
ProofPaper ↗
Key points
- Natural-language software requirements often omit assumptions about performance, operating conditions, user behaviour, and domain constraints, affecting architecture, testing, and deployment.
- We also introduce the HiddenLens Corpus (HLC), comprising 312 requirements annotated with 847 hidden assumptions across six NFR categories.
- HiddenLens outperforms keyword-based and QuARS-Adapted baselines by 71% and 61%, respectively, achieving an end-to-end F1-score of 0.79; its BERT-based NFR classifier achieves an F1-score of 0.847.
- These results indicate that HiddenLens can help analysts resolve hidden assumptions before they propagate through software development.
Sources (1)
- [1]HiddenLens: A Multi-Layer LLM Framework for Detecting and Resolving Hidden Assumptions in Software RequirementsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 04:35 PM
We propose HiddenLens, a novel six-layer BERT-LLM pipeline and, to our knowledge, the first framework engineered to detect, classify, question, and resolve hidden assumptions rather than merely flag ambiguity.
Natural-language software requirements often omit assumptions about performance, operating conditions, user behaviour, and domain constraints, affecting architecture, testing, and deployment.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 21, 2026Conflicted Records: Typed Adversarial Contracts for Auditable LLM Output Quality Enforcement
- Oct 21, 2026OmniSimulator: Aligning Small Language Models for Authentic Heterogeneous Behavior Modeling
- Oct 21, 2026A Large-Scale Evaluation of Role-Enhanced Zero-Shot Chain-of-Thought Prompting
- Oct 20, 2026CovNav: Conformal Prediction-Guided Adaptive Inference for Object Goal Navigation
- Oct 19, 2026Sparse RL Relaxation for LLM Reasoning
- Oct 19, 2026EDMA: Entropy-Driven Multimodal Answering