AI Paper Atlas

Research map / Optimization / Implicit bias and generalization

Loss is Not Behavior: A Unified Output-Space Analysis of Gradient-Based Machine Unlearning

Xingjian Zhao, Mohammad Mohammadi Amiri, Malik Magdon-Ismail

NeurIPS 2026 · Optimization › Implicit bias and generalization · approach: Edge-of-stability dynamics · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Analyzes gradient-based machine unlearning in output space, arguing that matching forget/retain losses does not imply retraining-like behavior.

Problem

Gradient-based unlearning methods optimize forget and retain losses, but similar losses do not guarantee output-space behavior matching a model retrained on only the retain data.

Approach

A unified output-space analysis of how a gradient-based unlearner deviates from the retrain model, rather than judging unlearning by loss values alone.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use matched forget loss or weights as the evaluation criterion for unlearnersgradient-based unlearners · MNIST class forgettingoutput RMS gap on forget set: SimNPO 0.009 vs ProjGD 0.509 at similar forget CE (4.368-4.404)Hurt
switch from random to class forgettingGradDiff, NPO, SimNPO · MNISTRMS_f gap from retrained model: residual norm 5.28 -> 24.25; RMS_f scaled by 1.8-3.8xHurt
use retained-Hessian leverage geometry instead of Euclidean Gram for gap prediction (ablation)GradDiff gap predictor · MNIST random forget set, retain anchorsPearson correlation: 0.40 -> 0.72Improved
use retained-Hessian leverage geometry instead of Euclidean Gram for gap prediction (ablation)SimNPO gap predictor · MNIST class forget set, forget setPearson correlation: -0.03 -> 0.98Improved
use projected-gradient dynamics (ProjGD) with unprojected predictorProjGD · MNIST and CIFAR-10 headactual output gap vs prediction: direction correlation 0.79-0.97, actual gap about 7-15x smallerMixed
remove retain anchoring (Pure GA, IMU) instead of GradDiff-style retain term (ablation)Pure GA and IMU · MNIST and CIFAR-10output KL from retrained model: 0.7-3 nats vs at most 10^-5Hurt
use SimNPO default beta 2.5 instead of NPO beta 0.1 in class forgetting (ablation)SimNPO vs NPO · MNIST class forgettingRMS_f gap: 0.032 -> 0.009 (saturated weights switch off forget term)Mixed
set per-set rho from leverage rule rho = s n / lambda_max(K_ff) with s=0.3GradDiff/NPO/SimNPO · 330 MNIST forget setsoptimization stabilityImproved
use leverage K-channel scores computed at the original model as example-level diagnosticTOFU Llama-3.2-1B-Instruct LoRA chart · TOFUSpearman rank correlation with disturbed retain examples: 0.26-0.49 (retain), 0.14-0.29 (forget)Improved
use LoRA-restricted curvature leverage scores to predict gap sizeTOFU diagnostics · TOFUgap size predictionNo clear effect
initialize at the original model rather than the retrained reference for the local predictor (ablation)local gap predictor · MNISTprediction validityHurt

Surprising findings

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy