LoReUn: Adaptive Loss-based Unlearning

• LoReUn is a machine unlearning technique that dynamically reweights individual samples based on loss proxies to target and remove unwanted knowledge.
• It employs adaptive weighting strategies across classification, generative, and language models to balance the trade-off between forgetting and retaining important data.
• Empirical evaluations on datasets like CIFAR-10 demonstrate that LoReUn achieves near-retraining performance with minimal computational cost and leakage.

Loss-based Reweighting Unlearning (LoReUn) refers to a class of machine unlearning algorithms that dynamically and adaptively reweight individual samples or regions in the objective function based on loss- or difficulty-related proxies, thus enabling more precise, efficient, and utility-preserving removal of unwanted knowledge from trained machine learning models. LoReUn encompasses multiple instantiations—ranging from classification and generative models to LLMs—each leveraging per-sample or per-region loss signals (or surrogates such as energy or embedding-based difficulty scores) to guide the unlearning dynamics in a plug-and-play, computationally economical manner (Li et al., 30 Jul 2025, Simone et al., 23 Jun 2025, Yang et al., 17 May 2025, Yu et al., 26 Jan 2026, Liu et al., 28 Aug 2025).

1. Motivation and Foundational Principles

The primary goal of machine unlearning is to transform a model $f(x;\theta_0)$ trained on a dataset $\mathcal{D} = \mathcal{D}_r \cup \mathcal{D}_f$ into an “unlearned” model $f(x;\theta^*)$ that mimics the behavior of retraining from scratch on $\mathcal{D}_r$ alone, thereby erasing the influence of the forget set $\mathcal{D}_f$ at far less computational cost. Traditional approximate MU mechanisms assign uniform weights to each forgetting example, but empirical analyses demonstrate heterogeneity in forgetting difficulty: well-memorized, low-loss examples are harder to erase, while high-loss (poorly fit) ones are easy to forget (Li et al., 30 Jul 2025, Yang et al., 17 May 2025). Uniform weighting thus wastes gradient signal on already-forgotten points and under-forgets hard cases, leading to increased residual leakage or excessive utility degradation.

The LoReUn paradigm is motivated by the observation that per-sample loss or related proxies (cross-entropy, negative log-likelihood, energy, or embedding distance) strongly correlate with unlearning hardness. By adaptively accentuating “difficult” points in the unlearning objective, LoReUn methods efficiently allocate optimization resources, close the gap to ideal retrain-based unlearning, and minimize collateral model damage.

2. Mathematical Formulations and Core Algorithms

Across MU settings, LoReUn operationalizes the following principle: the effective sample weight $w_i$ in the unlearning loss is set to depend monotonically (often inversely) on a measure of loss, memorization, or related difficulty proxy for each point $x_i$.

Canonical Classification/Generative Case

For a mini-batch $B_f \subset \mathcal{D}_f$ and corresponding per-example evaluation loss $\ell_{\mathrm{eval}}(x_i, y_i)$,

$$w_i = \exp\left(-\,\ell_{\mathrm{eval}}(x_i,y_i)/\tau\right), \quad \tau > 0$$

with normalized weights $\mathcal{D} = \mathcal{D}_r \cup \mathcal{D}_f$0 (Li et al., 30 Jul 2025). The overall unlearning objective becomes: $\mathcal{D} = \mathcal{D}_r \cup \mathcal{D}_f$1 where $\mathcal{D} = \mathcal{D}_r \cup \mathcal{D}_f$2 and $\mathcal{D} = \mathcal{D}_r \cup \mathcal{D}_f$3 trades off forgetting/retaining.

Generative Flow/EBM Setting (ContinualFlow)

Given a base distribution $\mathcal{D} = \mathcal{D}_r \cup \mathcal{D}_f$4 and an energy proxy $\mathcal{D} = \mathcal{D}_r \cup \mathcal{D}_f$5 for the forget region, the soft mass-subtracted target is

$\mathcal{D} = \mathcal{D}_r \cup \mathcal{D}_f$6

where $\mathcal{D} = \mathcal{D}_r \cup \mathcal{D}_f$7 and $\mathcal{D} = \mathcal{D}_r \cup \mathcal{D}_f$8 controls suppression. The associated Energy-Reweighted Flow Matching (ERFM) loss is

$\mathcal{D} = \mathcal{D}_r \cup \mathcal{D}_f$9

This approach ensures gradients that match flow to a target where the forget region’s mass is softly removed (Simone et al., 23 Jun 2025).

Long-Tail and Class-Imbalanced Forgetting

FaLW extends LoReUn to long-tailed settings by computing the forget sample’s confidence deviation from unseen data in its class: $f(x;\theta^*)$0 where $f(x;\theta^*)$1 is a class-balancing factor controlled by class frequency, and $f(x;\theta^*)$2 are mean and std of “unseen” confidences for class $f(x;\theta^*)$3 (Yu et al., 26 Jan 2026).

LLM and Token-wise Reweighting

For LLMs, LoReUn manifests at the token level. Saturation-based weighting assigns high weights to tokens with large likelihood $f(x;\theta^*)$4 (still memorized), using

$f(x;\theta^*)$5

while importance-based weighting emphasizes manually-defined or learned important tokens. The SatImp method combines both: $f(x;\theta^*)$6 with $f(x;\theta^*)$7 adjusted for desired trade-off (Yang et al., 17 May 2025).

Pseudocode Summary

Below is a concise pseudocode for static LoReUn (classification/generative):

$\mathcal{D}_r$4 Dynamic LoReUn re-evaluates $f(x;\theta^*)$8 on current $f(x;\theta^*)$9 every step.

3. Empirical Evaluations and Benchmarks

LoReUn methods have been extensively evaluated on image classification (CIFAR-10, CIFAR-100, SVHN, Tiny ImageNet), text-to-image diffusion (CIFAR-10/DDPM, Imagenette/Stable Diffusion, NSFW filtering), and LLM QA/knowledge benchmarks (TOFU, MUSE, WMDP). Key metrics include:

• Unlearning Accuracy (UA): Accuracy on $\mathcal{D}_r$0 (should be minimized).
• Retaining Accuracy (RA): Accuracy on $\mathcal{D}_r$1 (should be maximized).
• Forget Rate / Leakage: Classifier-based evaluation of residual forbidden knowledge.
• FID: Generation quality post-unlearning (diffusion models).
• Membership Inference Attack (MIA): Degree of privacy risk post-unlearning.
• Trade-off (ToW, Avg Gap, MU/FQ): Combined measures of forget-retain balance (Li et al., 30 Jul 2025, Yang et al., 17 May 2025).

Representative empirical findings:

Method	UA	RA	Test Acc	FID	Avg Gap	Overhead (min)
RL+LoReUn	~0	~Retrain	≈Retrain	23.3	0.3%	0.2-0.5
Standard RL	>0	↓	↓	24.0	1.3%	-
SatImp (LLM)	best	best	best	-	-	-

LoReUn consistently matches or surpasses retrain baselines in UA, RA, and trade-off metrics, often with negligible computational cost vs. conventional unlearning.

4. Mechanistic Insights, Design Variants, and Ablations

The central efficacy of LoReUn arises from up-weighting low-loss, hard-to-forget examples, focusing gradients on critical points while suppressing over-forgetting of easy/high-loss cases. Dynamic variants (e.g., LoReUn-d) adapt as unlearning progresses, further improving coverage. Temperature $\mathcal{D}_r$2 and normalization strategy affect sensitivity; excessive sharpness induces instability, excessive uniformity weakens selectivity (Li et al., 30 Jul 2025).

Token-wise reweighting at fine granularity (LLMs) outperforms instance- or batch-wise approaches. Combination schemes (e.g., SatImp) empirically yield the state-of-the-art unlearn-retain balance across model families.

FaLW introduces further instance- and class-awareness, correcting both “heterogeneous deviation” (variability across points) and “skewed deviation” (across long-tail classes). This suggests the generality of loss-based reweighting when aligned with the underlying difficulty or statistical imbalance (Yu et al., 26 Jan 2026).

5. Theoretical Guarantees and Limitations

Theoretical analyses for LoReUn focus on properties of the proxy weights and the objectives they induce. In flow-based generative models, LoReUn’s reweighted flow matching is formally equivalent to pushing the model toward a soft-mass-subtracted target distribution (Simone et al., 23 Jun 2025). In the classification setting, weighting by loss proxies is justified by empirical correlation between loss and the model’s “forgetting state,” and is further supported by analysis of seen/unseen confidence statistics in FaLW (Yu et al., 26 Jan 2026).

Noted limitations include:

• Reliance on proxy metrics (energy, loss): If proxies are poor, the unlearning is mis-targeted.
• Hyperparameter sensitivity: $\mathcal{D}_r$3 and trade-off factors require tuning per scenario.
• Soft forgetting: Small residuals, or “leakage,” may persist unless weights are made extreme.
• Current coverage: Most evaluations are on classification and diffusion models; LLMs require token-wise proxies and additional regularization.
• No formal DP or removal guarantees: LoReUn’s guarantees are pragmatic but non-cryptographic.

6. Extensions, Open Problems, and Future Directions

Extensions of LoReUn include:

• Proxy learning: Jointly optimizing the proxy (e.g., adversarial energy) along with the unlearner (Simone et al., 23 Jun 2025).
• Dynamic temperature/weight annealing for further adaptivity (Li et al., 30 Jul 2025).
• Application to graph and sequence models; domain-specific proxies for time-series, graphs, or highly structured data (Li et al., 30 Jul 2025).
• Certified unlearning: Incorporating DP noise or constraint-based regularization for formal privacy guarantees (Li et al., 30 Jul 2025).
• Optimality analysis: Theoretical determination of weight functional forms (e.g., optimal β parameters in SatImp) remains open (Yang et al., 17 May 2025).
• Continual/multi-task forgetting: Handling sequences of unlearning requests efficiently (Simone et al., 23 Jun 2025, Yang et al., 17 May 2025).

A plausible implication is that continued advances in proxy design, theoretical characterization, and modality-specific adaptations will further generalize the LoReUn approach across machine learning, especially as requests for safe, targeted, and scalable unlearning grow in regulatory and safety-sensitive domains.

7. Connections to Related Work

LoReUn relates to the broader literature on influence estimation, importance sampling, and active learning, but is distinctive in its operationalization for targeted unlearning via loss/energy-based proxies. Early loss-based reweighting appeared in approximate data deletion and model correction. Contemporary frameworks such as FaLW (Yu et al., 26 Jan 2026) and SatImp (Yang et al., 17 May 2025) generalize and clarify the mechanistic design space, establishing LoReUn as an adaptive, pluggable enhancement for a wide array of unlearning workflows.