mlsanityData to 7 Oct 2026

SAFE-KD: Risk-Controlled Early-Exit Distillation for Vision Backbones

arXiv 2602.03043 ↗First posted 3 Feb 2026cs.LG cs.AI cs.CVcited by 0

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

Abstract (arXiv)

Early-exit networks reduce inference cost by allowing ``easy'' inputs to stop early, but practical deployment hinges on knowing when early exit is safe. We introduce SAFE-KD, a universal multi-exit wrapper for modern vision backbones that couples hierarchical distillation with conformal risk control. SAFE-KD attaches lightweight exit heads at intermediate depths, distills a strong teacher into all exits via Decoupled Knowledge Distillation (DKD), and enforces deep-to-shallow consistency between exits. At inference, we calibrate per-exit stopping thresholds on a held-out set using conformal risk control (CRC) to guarantee a user-specified selective misclassification risk (among the samples that exit early) under exchangeability. Across multiple datasets and architectures, SAFE-KD yields improved accuracy compute trade-offs, stronger calibration, and robust performance under corruption while providing finite-sample risk guarantees.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Early exit and edge inference, one of 12 directions in Efficient LLM inference and serving. Builds on 18 of its 27 references that are in the map. First posted on arXiv 3 Feb 2026.

Builds on

18 of 27 references.

References · 18

  1. 2015Distilling the Knowledge in a Neural Networkcited by 10,076
  2. 2017On Calibration of Modern Neural Networkscited by 4,860
  3. 2016Simple and Scalable Predictive Uncertainty Estimation using Deep Ensemblescited by 3,888
  4. 2019Benchmarking Neural Network Robustness to Common Corruptions and Perturbationscited by 2,857
  5. 2014FitNets: Hints for Thin Deep Netscited by 1,619
  6. 2016Paying More Attention to Attention: Improving the Performance of Convolutional Neural Networks via Attention Transfercited by 1,059
  7. 2019Relational Knowledge Distillationcited by 873
  8. 2021DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsificationcited by 849
  9. 2017Selective Classification for Deep Neural Networkscited by 799
  10. 2021A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantificationcited by 777
  11. 2017BranchyNet: Fast Inference via Early Exiting from Deep Neural Networkscited by 591
  12. 2015Weight Uncertainty in Neural Networkscited by 554
6 more
  1. 2021Dynamic Neural Networks: A Surveycited by 405
  2. 2020DeeBERT: Dynamic Early Exiting for Accelerating BERT Inferencecited by 322
  3. 2022Conformal Risk Controlcited by 290
  4. 2020BERT Loses Patience: Fast and Robust Inference with Early Exitcited by 267
  5. 2025PolygoNet: Leveraging Simplified Polygonal Representation for Effective Image Classificationcited by 6
  6. 2026TopoLoRA-SAM: Topology-Aware Parameter-Efficient Adaptation of Foundation Segmenters for Thin-Structure and Cross-Domain Binary Semantic Segmentationcited by 6

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025CalexNet: Soft Cascade-Aligned Training and Calibration for Lightweight Early-Exit Branches
  2. 2026CA-OPD: Confidence-Aware On-Policy Distillation for Structured Visual Prediction
  3. 2026A Cross-Architecture Audit of Direction-Based Inference-Time Defences in Vision-Language Models
  4. 2024Fast yet Safe: Early-Exiting with Risk Control
  5. 2026IntroConformal: Conformal Factuality Guarantees for Large Vision-Language Models via Introspective Signals
  6. 2026Geometry-Aware Distillation for Prompt Tuning Biomedical Vision-Language Models
  7. 2025Beyond Greedy Exits: Improved Early Exit Decisions for Risk Control and Reliability
  8. 2026BOKBO (Best of K Bad Options): Calibrated Abstention for VLA Policies