Google released a framework to audit 'machine unlearning.' The composition is strongly academic—one official Google Research source plus four arXiv papers—research-led rather than a product, anchored on a proposed evaluation method. It's a framework for verifying whether a specific data's influence has truly been removed from a trained model, measuring the effectiveness of 'forgetting' required for privacy and copyright. The through-line is research broadening from how to train models toward accountability and governance after training. It's still at the research stage; whether it becomes standard in production or regulatory practice is to confirm.
Algorithms & Theory × Developer Tools
Google audits machine unlearning
Google audits machine unlearning
✎ Story body
▲ Official & Press
Official
New framework for auditing machine unlearning
Google introduces a new framework for auditing machine unlearning
Academic (arxiv etc.) 10 ▾
Academic
SPEA2$^+$: Improved Density Estimation in SPEA2 with Provable Runtime Guarantees
SPEA2+ improves density estimation with provable runtime guarantees
Academic
CCKS: Consensus-based Communication and Knowledge Sharing
CCKS improves cooperative MARL via consensus-based knowledge sharing
Academic
Mathematical perspective on genetic algorithms with optimization guided operators
A mathematical model of genetic algorithms with optimization-guided operators
Academic
Latent-diffusion data assimilation for subsurface flow: realism vs. uncertainty
Academic
Test-Time Gradient Guidance of Flow Policies in Reinforcement Learning
QGF: improving flow policies in RL entirely at test time
Academic
Unifying Local Communications and Local Updates for LLM Pretraining
GASLoC: gossip-based communication-efficient LLM pretraining
Academic
Bellman-Taylor score decoding for MDPs with state-dependent action sets
Academic
Learning Doubly Sparse Explicitly Conditioned Transforms
Learning data-adaptive doubly sparse, condition-number-controlled transforms
Academic
Speaker Group Encoding in Self-supervised Speech Recognition Models
How self-supervised speech models encode speaker group traits
Academic
Causal Ensemble Agent: Hierarchical Causal Discovery with LLM-guided Expert Reweighting
LLM referee reweights expert ensemble for causal discovery