Skip to content

Maintaining utility in continual unlearning of large language models via orthogonal projection

Preserving Utility in Continual LLM Unlearning via Orthogonal Projections

arXiv: Jailbreaks, Prompt Injection, Poisoning & Defenses· arXiv:2610.40483· Juan Belieni, Ana Carolina Erthal, Eliezer de Souza da Silva, Diego MesquitaSchool of Applied Mathematics, Getulio Vargas Foundation; University of Coimbra, CISUC/LASI, DEIIndexed by this site · 20 days ago· Originally publishedDefenses

Quick read

Defenses

Based on the paper PDF (AI-generated); refer to the original

The authors propose COP, which projects each unlearning update onto the orthogonal complement of the activation subspace of the historical forget set; the authors state that on TOFU and RWKU it preserves utility without weakening forgetting.

Questions
In continual unlearning, deletion requests arrive in sequence; the authors state that existing LLM unlearning methods degrade in both unlearning and utility under sequential requests, because later updates interfere with prior deletions.
Methods
The authors propose Continual Orthogonal Projection (COP): it caches a low-rank basis of the input activations of past unlearning sets at each layer and projects each gradient update onto its orthogonal complement, claiming no first-order interference with prior deletions and no modification of the original unlearning objective.
Experiments and results
On TOFU (1B/3B) and RWKU (3B), the authors report that from the 3rd request onward the COP variants have higher utility than the base objectives on TOFU, and prevent the utility collapse of GA and IDK; O3 forgets weakly on TOFU, and ALKN's utility collapses.
Limitations and future work
The authors note that the protection is approximate: compression, AdamW, and unprojected embedding and normalization parameters all relax the theorem's conditions; the evaluation uses only a single seed, 5 requests, and models up to 3B, and does not include relearning or extraction attacks.
Experimental setupLlama-3.2-1B-Instruct、Llama-3.2-3B-Instruct · TOFU、RWKU · FR、FP et al.
Evaluated models
Llama-3.2-1B-InstructLlama-3.2-3B-Instruct
Benchmarks
TOFURWKU
Metrics
FRFPESEMMU
AI summaryResearchers propose continuous orthogonal projection (COP), a gradient projection scheme that enables large language models to maintain utility under successive unlearning requests without weakening the unlearning effect. Most existing unlearning methods are not suitable for iterative deletion; successive deletions interfere with each other through the updates, causing the unlearning effect and model utility to decline sharply. COP restricts each update to the orthogonal complement of the activation subspace cached from previous forget sets, with a provable guarantee of first-order non-interference with previous deletions; each layer only needs to store a low-rank basis, without requiring past gradients or previous forget sets, and it can be inserted into unlearning methods that update the model via differentiable loss gradient steps without modifying their objective. Experiments on the standard unlearning benchmarks TOFU and RWKU show that COP maintains utility under successive requests, avoids the collapse of the standard objective, and outperforms continual-unlearning-specific baselines in unlearning effect.

Researchers propose continuous orthogonal projection (COP), a gradient projection scheme that enables large language models to maintain utility under successive unlearning requests without weakening the unlearning effect. Most existing unlearning methods are not suitable for iterative deletion; successive deletions interfere with each other through the updates, causing the unlearning effect and model utility to decline sharply. COP restricts each update to the orthogonal complement of the activation subspace cached from previous forget sets, with a provable guarantee of first-order non-interference with previous deletions; each layer only needs to store a low-rank basis, without requiring past gradients or previous forget sets, and it can be inserted into unlearning methods that update the model via differentiable loss gradient steps without modifying their objective. Experiments on the standard unlearning benchmarks TOFU and RWKU show that COP maintains utility under successive requests, avoids the collapse of the standard objective, and outperforms continual-unlearning-specific baselines in unlearning effect.

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    In continual unlearning, subsequent updates interfere with prior deletions; the authors propose COP, which preserves utility while not weakening unlearning.

    Core problem: in continual LLM unlearning, deletion requests arriving sequentially interfere with each other, causing unlearning effectiveness and model utility to degrade as requests accumulate.

    • Problem setting: a deployed model continuously receives deletion requests; each deletion must preserve the effect of prior deletions and must not access the data targeted by prior requests, while preventing utility loss from accumulating (per the authors).
    • Shortcomings of existing methods: the authors state that gradient-based unlearning objectives such as GA, NPO, and SimNPO are designed for isolated requests, and when applied sequentially, later updates undo prior deletions or worsen utility erosion; the continual-unlearning-specific method O3 relies on an inference-time OOD detector and target information can still be recovered from the base weights, while ALKN separates only parameters and not effects; in the authors' experiments, neither preserved both unlearning and utility.
    • Core observation: the authors formalize continual unlearning as a constrained sequential ERM problem, requiring each update not to push prior unlearning objectives away from the values they have reached; using layer-wise gradient decomposition, this constraint can be enforced as an orthogonality condition in a low-dimensional input activation space, without storing historical gradients.
    • The paper proposes: Continual Orthogonal Projection (COP), a gradient projection plug-in that projects each update onto the orthogonal complement of the cached activation subspace of past unlearning sets; it stores only one low-rank basis per layer, does not modify the unlearning objective, and can be plugged into any unlearning method based on a differentiable loss.
    • Setting assumptions: at each step k an unlearning set Fk is received, and step k cannot access data from prior unlearning sets; a fixed retain set R is accessible as a proxy for knowledge that should be preserved.
  2. What related research exists?

    Related work is grouped into four categories: LLM unlearning objectives, continual unlearning, gradient projection in continual learning, and projection methods in unlearning.

    LLM unlearning objectives

    • GA (Jang et al., 2023)—maximizes the next-token loss on the unlearning set; the paper notes that its loss is unbounded and erodes general utility.
    • NPO (Zhang et al., 2024)—adapts DPO to the setting with only refusal responses; SimNPO (Fan et al., 2025) removes the reference model and uses length-normalized log-likelihood; IDK (Maini et al., 2024)—trains the model to respond to unlearning queries with a refusal.
    • Other routes: mapping to a generic alternative answer (Eldan and Russinovich, 2023), perturbing internal representations (Li et al., 2024), subtracting task vectors (Ilharco et al., 2023).

    Continual unlearning

    • O3 (Gao et al., 2025)—freezes the base model and trains all deletions in a shared LoRA adapter, with an OOD detector deciding the adapter strength at inference; the paper notes that its behavior depends on detection error and that target information can still be recovered from the base weights.
    • ALKN (Wuerkaixi et al., 2025)—negates task vectors plus gradient masking, so that each request modifies a different subset of parameters; the paper notes that separating parameters is not the same as separating effects.
    • FIT (Xu et al., 2026)—filters redundant requests and selects algorithms and target layers by importance; CURaTE (Bae et al., 2026)—does not change weights and has a sentence embedder refuse similar queries, with target information still remaining in the parameters.

    Gradient projection in continual learning

    • OGD (Farajtabar et al., 2020)—projects new gradients onto directions orthogonal to stored gradients of past tasks.
    • GPM (Saha et al., 2021)—stores a low-rank basis of past-task input activations per layer and projects orthogonally; the authors state that COP follows GPM's activation-space projection, but the subspace comes from past unlearning sets and protects the values reached by prior unlearning objectives.

    Projection methods in unlearning

    • PGU (Hoang et al., 2024) and UNSC (Chen et al., 2024)—on image classifiers, project the unlearning update onto the null space of retain-data activations, protecting only the retain data within a single request.
    • GU (Zhou et al., 2025)—in LLMs, projects the update onto directions orthogonal to the retain gradient, protecting the retain set rather than prior deletions; FG-OrIU (Feng et al., 2025) and Lee et al. (2026)—project across requests, but on vision models and text-to-image diffusion models, respectively.

    Baselines and benchmarks: the compared baselines are O3 and ALKN, the base objectives are GA, NPO, SimNPO, IDK; the benchmarks are TOFU and RWKU. The authors position COP's distinction as: targeting continual unlearning in LLMs, protecting the prior unlearning sets themselves via cached activation subspaces, and not modifying the unlearning objective.

  3. How does the paper address this problem?

    COP projects each update onto the orthogonal complement of the activation subspace of the historical forget set, keeping the prior forget objective unchanged to first order.

    Overall approach: continual unlearning is formalized as constrained sequential ERM; a first-order approximation converts "do not interfere with prior deletions" into a linear constraint on the update direction; then, using layer-wise gradient decomposition, the constraint is imposed on the cached input activation subspace, yielding Continual Orthogonal Projection (COP).

    Problem formulation

    • The unlearning objective at step k is LFk(θ) = 1/(|Fk|) ∑(p,c)∈ Fk ℓ(θ; p, c), where ℓ is any differentiable unlearning loss; the total loss is Lk = LFk + λLRk, with λ ≥ 0 controlling the unlearning-retention trade-off.
    • Anchoring constraint: each prior unlearning objective LFi is required to stay within a tolerance ϵi of its value at θi (no drift in either direction), i.e., the constraint |LFi(θ) − LFi(θi)| ≤ ϵi; increasing LFi partially undoes the deletion, while decreasing it over-unlearns and harms utility.
    • A first-order Taylor expansion linearizes the constraint as −ϵi − ai ≤ ⟨giF, ∆θ⟩ ≤ ϵi − ai; the homogeneous condition ⟨giF, ∆θ⟩ = 0 keeps drift first-order invariant but cannot correct already accumulated drift.

    From parameter space to activation space

    • The gradient of a linear layer decomposes as ∇W L = δx⊤, and the unlearning-set gradient is the mean of δi(z)xi(z)⊤ over samples (Eq. 1); ⟨GFi, ∆W⟩ depends only on the effect of ∆W on the input xi(z).
    • Sufficient condition: if ∆W x = 0 for all x ∈ Xi (activations zeroed), then ⟨GFi, ∆W⟩ = 0; this condition is stronger than the original constraint—it holds for every sample and any differentiable loss—so it can be encoded by a low-rank basis in the input activation space without parameter-space gradients.

    Projection rule and basis construction

    • Let Vk−1 be an orthogonal basis spanning the protected subspace Sk−1, with projector Pk−1 = Vk−1Vk−1⊤; the projected gradient is G̃k = Gk(I − Pk−1), and the update is W ← W − ηG̃k; the projection includes the full gradient of both the forget and retain parts, so the retain loss repairs utility only in directions orthogonal to prior deletions.
    • Basis construction: after training at step k, collect the layer input activations Ak of Fk under θk, first pre-project against the current basis to obtain (I − Pk−1)Ak, take the top left singular vectors of the squared Frobenius norm at retention ratio τ ∈ (0,1], at most rmax per layer per request, concatenate with Vk−1, and QR-orthogonalize to obtain the nested subspace Sk−1 ⊆ Sk.
    • Theorem 1: if all trainable parameters are linear weights, the subspaces are nested, and each update satisfies ∆WVn−1 = 0, then the outputs for covered samples remain unchanged for any step size (fθk(z) = fθi(z)); the authors also note three premises: under Adam-type optimizers the projection must act on the final step, embedding/normalization/biases must be frozen, and the per-layer inputs of protected samples must lie in the subspace—the third holds only approximately in practice.

    Complexity and capacity

    • Each layer stores only the basis Vk (r × din entries), so memory grows linearly with rank rather than with the number of requests; the projection is computed as Gk − (GkVk−1)Vk−1⊤, adding two matrix multiplications per update, plus one forward pass, one SVD and one QR per request.
    • Capacity: after k requests, r ≤ min(din, k·rmax); once the subspace spans the entire layer input space, that layer can no longer be updated; if the current unlearning-set activations already lie in Sk−1, the projected gradient vanishes in that layer. τ and rmax control growth, at the cost that discarded directions are no longer protected.

    Evaluation

    • Unlearning metrics: FR, FP, EM, ES (TOFU reports FR/ES/EM, RWKU reports FR/FP/EM), computed after the k-th request over the union of all requests so far; the utility metric MU is the harmonic mean of benchmark-specific utility quantities; definitions are in Appendix B.
  4. What experiments did the paper conduct?

    On TOFU (1B/3B) and RWKU (3B), the COP variants preserve utility and unlearn comparably to the base objectives; O3 and ALKN each have shortcomings.

    Experimental Setup

    • Models tested: Llama-3.2-1B-Instruct and Llama-3.2-3B-Instruct (TOFU, checkpoints fine-tuned on the full data by Dorna et al. (2025)); RWKU uses only Llama-3.2-3B-Instruct, with no further fine-tuning.
    • Benchmarks and request splits: TOFU forget05 is split into T = 5 requests, each with 2 authors and 40 QA pairs, with retain95 as R; RWKU uses level 2 probes, randomly sampling 20 real people into 5 groups (4 people each), with probes for the remaining 180 people as R.
    • Baselines and variants: four base objectives GA, NPO, SimNPO, IDK, each paired with a COP variant (main table rmax = 256); continual unlearning baselines O3, ALKN; each method is tuned for the best forget-utility trade-off (Appendix C).
    • Metrics: TOFU reports FR↓, ES↓, EM↓, MU↑; RWKU reports FR↓, FP↓, EM↓, MU↑; the forget metric after the k-th request is computed over the union of all requests so far.
    • Training details: full-model bfloat16 training, 10 epochs per request (about 20 steps for TOFU), AdamW (weight decay 0.01), effective batch 32, retain NLL weight λ = 1; each COP run uses the same seed, learning rate, data and budget as the corresponding base objective, differing only in the projection.
    • Number of repetitions: each grid point is run only once, seed 0 (Appendix C); the 20 forget people in RWKU are sampled with a fixed seed, the same for all methods.

    Main Results

    Results after the 5th request on TOFU (from Table 1, Table 2; higher MU is better, lower FR/ES/EM is better):

    Swipe horizontally to read the table

    Method1B: FR1B: EM1B: MU3B: FR3B: EM3B: MU
    Original0.830.970.5990.930.990.667
    GA0.170.300.1590.070.120.224
    GA+COP0.340.630.4610.380.710.566
    NPO0.320.630.4960.210.520.456
    NPO+COP0.300.560.5320.330.560.571
    SimNPO0.360.650.4630.360.660.554
    SimNPO+COP0.370.690.5170.380.710.600
    IDK0.020.490.0370.030.550.056
    IDK+COP0.100.590.4650.040.670.477
    O30.300.840.5760.340.910.639
    ALKN0.020.060.0040.040.170.040
    • The authors state that from the 3rd request onward, on 1B every COP variant has higher MU than its base objective; on 3B, SimNPO and IDK improve at every request, GA from the 2nd request onward and NPO from the 3rd request onward.
    • The authors state that the low FR of base GA and IDK is a byproduct of utility collapse rather than effective forgetting; O3 has the highest utility but weak forgetting (EM close to the original model, with EM above 0.9 at every step on 3B); ALKN's forget metrics collapse together with utility.
    • Exception: on 3B, NPO+COP has higher FR than base NPO (0.33 vs. 0.21), which the authors state trades forgetting for higher utility; on 1B, NPO+COP's FR is instead lower (0.30 vs. 0.32).

    RWKU Results

    • What was measured: continual unlearning of real-person knowledge, with knowledge from pretraining (Table 3, Figure 3).
    • Results: after the 5th request, GA+COP has MU 0.485 (base GA 0.366, and base GA's MU was once 0.004 at the 4th request); SimNPO+COP 0.720 (base 0.620); IDK+COP 0.569 (base 0.559); NPO+COP 0.605 (base 0.603), but NPO+COP has FR 0.04 and EM 0.20, lower than base NPO's 0.12 and 0.30.
    • Authors' interpretation: COP still stabilizes utility on pretrained knowledge; NPO shows a different trade-off—the COP version forgets more while final utility is the same. O3 forgets effectively on RWKU but MU drops to 0.062, and ALKN retains more target knowledge with low utility (MU 0.156). The authors also note that most runs and all COP variants have slightly higher MU than the original model, because the retain loss is computed on R and MU is partly evaluated on the same distribution.

    Rank Cap Ablation (Appendix D)

    • What was measured: the effect of rmax ∈ {16, 64, 256} on final MU and ES on TOFU 1B (Figure 4).
    • Results: the authors state that final MU for all methods rises as rmax increases, with GA and IDK rising by a large margin and NPO and SimNPO more slowly; ES is nearly unchanged across ranks and far below the original model.
    • Authors' interpretation: within the tested range, increasing rmax improves utility without harming forgetting, with the main cost being memory for storing the basis.

    Other Ablations and Analyses

    • Figure 2 (TOFU 1B utility trajectory): after the 1st request, the utility of each base objective and COP variant is similar (no prior deletions to protect), diverging from the 2nd request onward, with base objectives declining progressively and COP variants declining slowly.
    • Figure 3 (RWKU 3B utility trajectory): the authors state that under COP utility stays stable as requests accumulate, while base objectives fluctuate or degrade.
    • Hyperparameter search (Appendix C): learning rate {1,2,5}×10⁻⁵; on TOFU, COP additionally searches τ ∈ {0.90, 0.95, 0.99}, while RWKU fixes τ = 0.99; TOFU selects configurations by CRFS (harmonic mean of MU and forget quality), RWKU by the harmonic mean of MU and 1−EM.
    • The paper does not report statistical significance tests; all results are single-seed single runs.
  5. What are some points that can be further explored?

    The authors list limitations including the approximate nature of the protection, capacity growth, and evaluation scale, and state explicitly that it should not be regarded as a certified deletion mechanism.

    Limitations and Future Directions Noted by the Authors

    • The protection is approximate (Limitations): COP exactly preserves the prior forget loss only for inputs whose activations fall within the cached subspace, and requires that updates be projected and that all trainable parameters be linear weights; in actual runs, compression retains at most rmax directions per request, AdamW and weight decay act after the projection, and embeddings and normalization gains are not projected, so the observed protection is approximate.
    • Capacity grows with requests (Limitations): the protected subspace grows with each request, limiting the length of the request stream a fixed rank can support.
    • Evaluation scale (Limitations): the evaluation uses a single seed, 5 requests, and models of at most 3B parameters, and the forget metrics are based on training probes.
    • Attacks not tested (Limitations): the evaluation does not include relearning or extraction attacks.
    • Compliance and misuse (Limitations): the authors state that COP should not be regarded as a certified deletion mechanism and is not sufficient on its own to satisfy legal or regulatory compliance; unlearning methods are dual-use, and the same process may remove safety-relevant knowledge or suppress legitimate information.

    Scope of Experiments

    • The models tested are Llama-3.2-1B-Instruct and Llama-3.2-3B-Instruct, 2 in total; RWKU uses only 3B.
    • The base objectives are GA, NPO, SimNPO, IDK, 4 in total, and the continual unlearning baselines are O3, ALKN, 2 in total.
    • The request stream length is T = 5; TOFU has 40 QA pairs per request, RWKU has 4 people per request.
    • Each hyperparameter grid point is run once (seed 0); the paper does not report multi-seed repetitions or variance.
    • The main table fixes rmax = 256; the rank ablation covers only three levels {16, 64, 256} and is conducted only on TOFU 1B (Appendix D).
  6. Summarize the main content of the paper

    COP prevents mutual interference in continual unlearning through orthogonal projection in activation subspaces; the authors state that it preserves utility on two benchmarks without weakening unlearning.

    A methods paper on continual LLM unlearning: when deletion requests arrive sequentially, later updates undo prior deletions or accumulate utility loss; the authors formalize this problem as constrained sequential ERM and propose Continual Orthogonal Projection (COP).

    • Method: cache a low-rank basis of the input activations of past unlearning sets at each linear layer (at most rmax = 256 new directions per request), and project each full gradient update (including the retain term) onto the orthogonal complement of that subspace; the authors prove that under ideal conditions the outputs for protected samples are exactly unchanged for any step size (Theorem 1), and the method does not modify the unlearning objective and requires neither historical data nor an inference-time classifier.
    • Results (TOFU, forget05, 5 requests): on 1B, after the 5th request, MU of GA+COP is 0.461 (base GA 0.159), and IDK+COP is 0.465 (base IDK 0.037); the authors state that from the 3rd request onward, every COP variant has higher utility than the base objective (Table 1, Table 2).
    • Results (RWKU, 3B): after the 5th request, MU of SimNPO+COP is 0.720 (base 0.620); NPO+COP unlearns more (FR 0.04 vs. base 0.12) with the same final utility (Table 3).
    • Comparison with continual baselines: O3 has the highest utility on TOFU but weak unlearning (on 3B, EM above 0.9 at every step), and on RWKU it unlearns effectively but MU drops to 0.062; ALKN's utility collapses on both benchmarks (TOFU 1B final MU 0.004).
    • Authors' conclusion: COP provides a practical option for deployed models facing continual deletion requests without adding inference cost; the authors also emphasize that the protection is approximate, the evaluation covers only a single seed, 5 requests, and models up to 3B, and relearning or extraction attacks were not tested, so it should not be regarded as a certified deletion mechanism.

Related papers

Show 3 more papers
Read originalarxiv.org