HAMU: Hardness-Aware Multi-Objective Unlearning

Code for How Hard Can It Be? Hardness-Aware Multi-Objective Unlearning (ICML 2026).

Authors: Jiangwei Chen*, Xinyuan Niu*, Rachael Hwee Ling Sim, Zhengyuan Liu, Nancy F. Chen, Bryan Kian Hsiang Low

ArXiv

HAMU is a machine unlearning method that quantifies the local hardness of unlearning from the alignment between retain and forget gradients. Instead of optimizing a fixed weighted sum of objectives, HAMU solves a per-step constrained update: improve one unlearning objective by a specified amount while minimizing the cost to the other objective.

• HAMU-Q (hamu-q): enforce a minimum improvement in forget quality while minimizing retain utility degradation.
• HAMU-U (hamu-u): reciprocal variant that prioritizes the retain-side constraint.
• Supported tasks: CIFAR-10 vision experiments and WaterDrum-TOFU LLM QA experiments.
• Supported baselines: ft, ga, gdiff, kl, scrub, gru, pcgrad.

Repository Layout

hamu/
  cli/train.py              Unified training and unlearning CLI
  methods/                  HAMU and baseline gradient/update rules
  training/                 Trainer implementations and callbacks
  data/                     Vision and LLM dataset adapters
configs/accelerate/         Default accelerate launch config
scripts/experiments/        Paper experiment launchers

Installation

pip install -e .

Basic Workflow

Unlearning has two phases:

1. Train or fine-tune an original model on retain and forget data.
2. Start from that original model and unlearn with HAMU or a baseline on the retain/forget split.

By default, artifacts are written under artifacts/:

• CIFAR full checkpoints: artifacts/checkpoints/cifar/seed_<seed>/forget_<class>_<rho>/full
• CIFAR retain-only reference checkpoints: artifacts/checkpoints/cifar/seed_<seed>/forget_<class>_<rho>/retain
• WaterDrum-TOFU checkpoints: artifacts/checkpoints/waterdrum_tofu/seed_<seed>/<subset>/<variant>
• Unlearning outputs: artifacts/results/cifar/ and artifacts/results/waterdrum_tofu/
• Dataset cache: artifacts/dataset_cache/

The experiment scripts enable Weights & Biases by default. Use USE_WANDB=0 for local runs without logging.

Quick Start: CIFAR-10

Train the full ResNet-20 model:

scripts/experiments/cifar/train_full.sh 42

Run HAMU-Q unlearning:

METHOD=hamu-q scripts/experiments/cifar/hamu_unlearn.sh 42

Run a baseline:

METHOD=gdiff scripts/experiments/cifar/baseline_unlearn.sh 42

Useful overrides:

USE_WANDB=0
FORGET_CLASS=0
HARD_PROBABILITY=0.5
LR=1e-4
CONSTRAINT_RATIO=0.5
NUM_GPUS=1
GRADIENT_PAIRING=sequential
MODEL_NAME=artifacts/checkpoints/cifar/seed_42/forget_0_0.0/full

For parameter sweeps and ablations, see:

scripts/experiments/cifar/rho_sweep.sh
scripts/experiments/cifar/constraint_sweep.sh
scripts/experiments/cifar/full_grad_ablation.sh
scripts/experiments/cifar/stopping_criterion.sh
scripts/experiments/cifar/batch_size_ablation.sh
scripts/experiments/cifar/optimizer_ablation.sh
scripts/experiments/cifar/timing_all.sh
scripts/experiments/cifar/additional_baselines.sh

Quick Start: WaterDrum-TOFU LLM QA

Train the full LoRA-adapted Llama-2 checkpoint:

DATASET_NAME=Glow-AI/WaterDrum-TOFU  scripts/experiments/llm/train_full_waterdrum_tofu.sh 42

Run HAMU-Q unlearning:

METHOD=hamu-q scripts/experiments/llm/hamu_unlearn_waterdrum_tofu.sh 42

Run a baseline:

METHOD=gdiff scripts/experiments/llm/baseline_unlearn_waterdrum_tofu.sh 42

Useful overrides:

USE_WANDB=0
FORGET_PCT=10
DUPLICATE_SPLIT=semantic_duplicate
ADD_DUPLICATE=1
LR=1e-4
CONSTRAINT_RATIO=0.5
NUM_GPUS=1
GRADIENT_PAIRING=sequential
MODEL_NAME=artifacts/checkpoints/waterdrum_tofu/seed_42/forget_10/full

For parameter sweeps and ablations, see:

scripts/experiments/llm/semantic_similarity_sweep.sh
scripts/experiments/llm/constraint_sweep.sh
scripts/experiments/llm/full_grad_ablation.sh
scripts/experiments/llm/stopping_criterion.sh
scripts/experiments/llm/batch_size_ablation.sh
scripts/experiments/llm/optimizer_ablation.sh
scripts/experiments/llm/timing_all.sh
scripts/experiments/llm/additional_baselines.sh

CLI Usage

All scripts call the same CLI:

python -m hamu.cli.train --help

Important arguments:

• --modality {vision,llm} selects the training stack.
• --unlearning_method none trains a normal model.
• --unlearning_method hamu-q or hamu-u runs HAMU.
• --unlearning_method ft|ga|gdiff|kl|scrub|gru|pcgrad runs a baseline.
• --model_name is either a Hugging Face model name or a saved checkpoint.
• --hamu_constraint sets the per-step HAMU constraint.
• --use_lr_radius sets the HAMU radius from learning rate and gradient norm.
• --stop_on_stopping_criterion stops HAMU once the aggregate stopping criterion is met. Without this flag, the criterion is logged only.

HAMU and paired-gradient baselines support two gradient pairing modes:

• --gradient_pairing sequential: single-process mode. This computes retain and forget gradients sequentially and currently requires --gradient_accumulation_steps 1.
• --gradient_pairing split-gpu: parallel mode for an even number of processes. Retain and forget batches are interleaved so alternate processes compute paired gradients in parallel.

Notes

• split-gpu requires an even process count. For one GPU, set NUM_GPUS=1 GRADIENT_PAIRING=sequential.
• hamu/methods/core.py contains compact HAMU and baseline update formulas to compare the methods without reading the trainer code.

Citation

@inproceedings{chen2026icml,
  title={How hard can it be? Hardness-aware multi-objective unlearning},
  author={Chen, Jiangwei and Niu, Xinyuan and Sim, Rachael Hwee Ling and Liu, Zhengyuan and Chen, Nancy F. and Low, Bryan Kian Hsiang},
  booktitle={Proc. ICML},
  year={2026}
}