mlsanityData to 7 Oct 2026

Bridging Efficiency and Safety: Formal Verification of Neural Networks with Early Exits

arXiv 2512.20755 ↗First posted 23 Dec 2025cs.LG cs.AIcited by 1

On the map: Efficient LLM inference and serving → Early exit and edge inference

Abstract (arXiv)

Ensuring the safety and efficiency of AI systems is a central goal of modern research. Formal verification provides guarantees of neural network robustness, while early exits improve inference efficiency by enabling intermediate predictions. Yet verifying networks with early exits introduces new challenges due to their conditional execution paths. In this work, we define a robustness property tailored to early exit architectures and show how off-the-shelf solvers can be used to assess it. We present a baseline algorithm, enhanced with an early stopping strategy and heuristic optimizations that maintain soundness and completeness. Experiments on multiple benchmarks validate our framework's effectiveness and demonstrate the performance gains of the improved algorithm. Alongside the natural inference acceleration provided by early exits, we show that they also enhance verifiability, enabling more queries to be solved in less time compared to standard networks. Together with a robustness analysis, we show how these metrics can help users navigate the inherent trade-off between accuracy and efficiency.

Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Placed in Early exit and edge inference, one of 12 directions in Efficient LLM inference and serving. Builds on 56 of its 88 references that are in the map. First posted on arXiv 23 Dec 2025.

Builds on

56 of 88 references.

Closest ancestors · 5

Same direction, newest first

  1. 2025A Survey of Early Exit Deep Neural Networks in NLP13 Jan
  2. 2024LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding25 Apr
  3. 2023Training dynamic models using early exits for automatic speech recognition on resource-constrained devices18 Sep
  4. 2022You Need Multiple Exiting: Dynamic Early Exiting for Accelerating Unified Vision Language Model21 Nov
  5. 2022Resource-Constrained Edge AI with Early Exit Prediction15 Jun

Other references · 51

  1. 2015Deep Residual Learning for Image Recognitioncited by 51,809
  2. 2014Very Deep Convolutional Networks for Large-Scale Image Recognitioncited by 15,267
  3. 2017Towards Deep Learning Models Resistant to Adversarial Attackscited by 6,326
  4. 2013Intriguing properties of neural networkscited by 5,114
  5. 2018The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networkscited by 2,250
  6. 2020Conformer: Convolution-augmented Transformer for Speech Recognitioncited by 1,838
  7. 2016Neural Architecture Search with Reinforcement Learningcited by 1,489
  8. 2017Reluplex: An Efficient SMT Solver for Verifying Deep Neural Networkscited by 739
  9. 2017BranchyNet: Fast Inference via Early Exiting from Deep Neural Networkscited by 591
  10. 2017A Survey of Model Compression and Acceleration for Deep Neural Networkscited by 298
  11. 2020BERT Loses Patience: Fast and Robust Inference with Early Exitcited by 267
  12. 2016Safety Verification of Deep Neural Networkscited by 262
38 more
  1. 2017Formal Verification of Piece-Wise Linear Feed-Forward Neural Networkscited by 241
  2. 2021Beta-CROWN: Efficient Bound Propagation with Per-neuron Split Constraints for Complete and Incomplete Neural Network Robustness Verificationcited by 224
  3. 2019Algorithms for Verifying Deep Neural Networkscited by 221
  4. 2018Formal Security Analysis of Neural Networks using Symbolic Intervalscited by 185
  5. 2020Fast and Complete: Enabling Complete Neural Network Verification with Rapid and Massively Parallel Incomplete Verifierscited by 183
  6. 2018Efficient Formal Safety Analysis of Neural Networkscited by 165
  7. 2019Branch and Bound for Piecewise Linear Neural Network Verificationcited by 138
  8. 2018Dynamic Channel Pruning: Feature Boosting and Suppressioncited by 129
  9. 2022General Cutting Planes for Bound-Propagation-Based Neural Network Verificationcited by 102
  10. 2024Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systemscited by 94
  11. 2024Marabou 2.0: A Versatile Formal Analyzer of Neural Networkscited by 81
  12. 2020Robustness Verification for Transformerscited by 76
  13. 2021Adaptive Focus for Efficient Video Recognitioncited by 76
  14. 2020Why should we add early exits to neural networks?cited by 70
  15. 2019An Abstraction-Based Framework for Neural Network Verificationcited by 68
  16. 2015Conditional Deep Learning for Energy-Efficient and Enhanced Pattern Recognitioncited by 67
  17. 2018Training for Faster Adversarial Robustness Verification via Inducing ReLU Stabilitycited by 67
  18. 2023How Deep Learning Sees the World: A Survey on Adversarial Attacks & Defensescited by 65
  19. 2022Certified Training: Small Boxes are All You Needcited by 51
  20. 2018Training verified learners with learned verifierscited by 50
  21. 2017IDK Cascades: Fast Deep Learning by Learning not to Overthinkcited by 47
  22. 2024The Fifth International Verification of Neural Networks Competition (VNN-COMP 2024): Summary and Resultscited by 42
  23. 2024Scalable Neural Network Verification with Branch-and-bound Inferred Cutting Planescited by 33
  24. 2020DeepAbstract: Neural Network Abstraction for Accelerating Verificationcited by 31
  25. 2023Expressive Losses for Verified Robustness via Convex Combinationscited by 23
  26. 2022Scalable Verification of GNN-based Job Schedulerscited by 18
  27. 2022Neural Network Verification using Residual Reasoningcited by 18
  28. 2024CTBENCH: A Library and Benchmark for Certified Trainingcited by 13
  29. 2022Towards Reliable Neural Specificationscited by 12
  30. 2022QVIP: An ILP-based Formal Verification Approach for Quantized Neural Networkscited by 12
  31. 2022Linearity Grafting: Relaxed Neuron Pruning Helps Certifiable Robustnesscited by 10
  32. 2024Verifying message-passing neural networks via topology-based bounds tighteningcited by 10
  33. 2022Fundamental Limits in Formal Verification of Message-Passing Neural Networkscited by 9
  34. 2024Formal Verification of Graph Convolutional Networks with Uncertain Node Features and Uncertain Graph Structurecited by 8
  35. 2023Efficient Certified Training and Robustness Verification of Neural ODEscited by 3
  36. 2024Learning Minimal Neural Specificationscited by 3
  37. 2020Property-Directed Verification of Recurrent Neural Networkscited by 2
  38. 2019Refactoring Neural Networks for Verificationcited by 1

50 most cited of 51 All 51 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026Talking with Verifiers: Automatic Specification Generation for Neural Network Verificationcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2024Fast yet Safe: Early-Exiting with Risk Control
  2. 2026Balancing Stability and Plasticity in Sequentially Trained Early-Exiting Neural Networks
  3. 2021Adaptive Inference through Early-Exit Networks: Design, Challenges and Directions
  4. 2025Rethinking Calibration for Early-Exit Neural Networks
  5. 2024Neural Model Checking
  6. 2025TAO: Tolerance-Aware Optimistic Verification for Floating-Point Neural Networks
  7. 2026Neuro-Formal Verification: Agentic Language-Agnostic Formal Program Reasoning
  8. 2023Towards Anytime Classification in Early-Exit Architectures by Enforcing Conditional Monotonicity