acceptodds
Under review as a conference paper at ICLR 2027

Certificates Expire: Durable Unlearning for Large Language Models

Abstract

Machine unlearning removes designated knowledge from a trained language model, with release-time checks assessing whether the requested knowledge erasure has been achieved. Durable unlearning faces a distinctive obstacle: subsequent adversarial fine-tuning moves the released parameters and can recover knowledge that passed those checks. Identical release-time outputs need not fix recovery cost, while resistance to sampled attacks does not certify every trajectory within a budget. To address these, we first establish a separation between release-time behavior and durability, locating the missing information in the geometry of the landing point. We then design PLATEAU, a framework that converts this geometry into a certified budget of fine-tuning updates. Within this framework, Directional Measure Estimation reads the forget-loss margin and directional gradient norm at release together with a uniform negative-curvature bound over a prescribed neighborhood. Barrier Certification then converts valid bounds into a closed-form safety radius and an integer budget of complete updates, covering every admissible trajectory and evaluation objective in the declared class without running attacks. In addition, Certified Durability Training uses the radius dependencies to regularize existing unlearning objectives, followed by re-evaluation of the updated release. Under a common declaration on TOFU, MUSE and WMDP-bio at matched forgetting and utility, PLATEAU attains the largest certified budget and the longest recovery on TOFU and MUSE, and every one of the 51 compared method–benchmark cells stays unrecovered within its budget. Together, the separation and certificate make durability a property of the released parameters: any unlearning procedure can be assessed through the same geometric quantities, with a guarantee expressed in the complete updates an adversary can take.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
machine unlearning, large language models, relearning attacks, fine-tuning robustness, certified unlearning, durability, barrier certificates

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GXMPOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Fine-tuning a previously unlearned model can cause unlearned knowledge to resurface in the model. In this work, they study the geometry and curvature of the loss space to determine a bound on the number of fine-tuning steps an unlearned model can tolerate before the loss on the forget set falls below a designated threshold. This becomes a certified durable unlearning budget. Using the insights from the geometry of the loss space, the proposed method regularizes standard unlearning objectives to find unlearning solutions with favorable geometries and curvatures. To validate their approach, they compare their approach to other unlearning methods on standard unlearning benchmarks such as MUSE and TOFU.

Critical strengths

The emergence of unlearned information after subsequent fine-tuning is an important problem in the unlearning literature. The formalization of a durable unlearning budget is an interesting contribution. Studying how the loss curvature affects the "durability/persistence" of unlearning is an interesting perspective.

Critical weaknesses

The main table in the paper (Table 1) uses CB as the primary metric for comparing unlearning methods. CB estimates the budget using the curvature proxy, but it is unclear whether CB is a reliable estimate. In Table 6 in the appendix, the estimate B^\hat{B} (which I believe corresponds to CB) overestimates the certified budget B⋆B^\star by 1000s (ex. CB=B^=4379CB = \hat{B} = 4379, B⋆=2396B^\star = 2396). Why not use the certified budget B⋆B^\star in the main table to allow for a proper comparison? I am not convinced I can trust CB as a metric. CB uses the curvature proxy as part of its calculation. In the ablations in Table 4, the curvature proxy of PLATEAU is 186, but in Table 6, the certified curvature Mˉs⋆\bar{M}_s^\star is 802. Thus, CB differs greatly from the actual budget B⋆B^\star. Therefore, I do not know how to interpret the results in Table 1, Table 2, and Table 4, which all rely on the CB estimate and curvature proxy as the primary comparison metric. How is the curvature proxy computed or estimated? How do we know if it is reliable? Why is CB a reliable metric? Why aren't the ablations done with the certified curvature Mˉs⋆\bar{M}_s^\star to properly understand how each component affects curvature and budget?

Computing MsM_s (Equation 5) is extremely expensive, and this costly computation is not addressed in the paper. The appendix proposes an upper bound using a local mˉs\bar{m}_s and Hessian-Lipschitz KK, but these computations are expensive as well. How are these values being computed/estimated? What is the computational cost?

The Certified Durability Training (CDT) only minimizes a proxy for the curvature and radius. The solution from CDT is then checked to see if the budget is satisfied. Otherwise, CDT has to be run again. It is unclear how long this back-and-forth would take. Furthermore, certifying the budget can be expensive. How expensive is the certification procedure for the budget and the curvature?

In general, the computations of both the certified values and the estimated values are very unclear. Can you specify the procedures to compute each of the relevant certified and estimated values?

Questions for authors

See the questions in the Weaknesses section.

Reviewer_HX2nOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the durability of unlearning in large language models: how resistant forgotten knowledge is to recovery during subsequent fine-tuning. The authors show that models with identical outputs at release can have substantially different recovery costs, motivating an analysis of parameter-space geometry. They propose PLATEAU, which measures three quantities at release, the margin, directional gradient, and negative curvature within a neighborhood, and uses a second-order Taylor lower bound to derive a lower bound on the number of update steps required for recovery. These quantities also serve as regularization terms during unlearning. Across comparisons with 16 methods on TOFU, MUSE, and WMDP-bio, PLATEAU achieves the largest estimated budget and longest recovery time on TOFU and MUSE.

Critical strengths

Novel perspective and deep understanding. The paper clearly distinguishes forgetting at release from its durability under subsequent fine-tuning. Its attempt to provide a guarantee over all admissible trajectories using release-point geometry complements existing durability defenses based on simulated attacks.

Clear, method-agnostic framework. The margin, directional gradient, and negative curvature have intuitive roles, and the certificate can be applied to models produced by any unlearning method. The directional gradient norm admits a closed-form solution based on truncated SVD. CDT directly regularizes the three components determining the radius, providing a clear motivation for its design. The ablations in Table 4 further support each component’s contribution.

Comprehensive and carefully controlled comparisons. The experiments span three benchmarks, multiple model scales and families, and 16 methods, including recent durability defenses. Comparisons match forget quality and utility and use five attack seeds per model. The evaluation of 60 independently trained models further assesses variation across training runs.

Clearly stated certificate scope. The paper distinguishes local estimates from certificates obtained using validated bounds and reports attacks outside the declared scope, including full-rank updates and unbounded step sizes.

Critical weaknesses

The practical meaning of the budget is unclear under the small step-size bound. The common declaration uses a per-update norm bound of h=2×10−5h = 2 \times 10^{-5} and a learning rate of η=10−7\eta = 10^{-7}, while PLATEAU’s certified radius is approximately 0.048 in L2 norm. The paper does not relate these quantities to update norms under realistic fine-tuning attacks, making the practical protection difficult to assess.

The recovery metric may conflate resistance to relearning with resistance to learning from scratch. The retain reference model, which has never seen the forget set, has an RS of 4137 steps, compared with 9749 for PLATEAU on TOFU. This raises the possibility that PLATEAU broadly impedes learning in the relevant directions, rather than specifically preventing recovery of forgotten knowledge.

Questions for authors
  1. What exact value of the Hessian Lipschitz constant KK is used in Table 6, and how is its validity rigorously established for an LLM with billions of parameters?

  2. How does h=2×10−5h = 2 \times 10^{-5} compare with single-step update norms under realistic fine-tuning attacks, such as full-parameter AdamW fine-tuning with a learning rate of 10−510^{-5}? What is the ratio of the certified radius to the model’s weight norm, and how many realistic fine-tuning steps would exhaust that radius?

This paper is insightful and trying to solve the relearning attack problem in machine unlearning, I am glad to raise my score if the the questions are solved.

Reviewer_2S5NOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PLATEAU, a framework for certifying the durability of machine unlearning against adversarial fine-tuning. It shows that successful knowledge removal at release does not guarantee resistance to subsequent recovery. PLATEAU uses loss geometry, gradient norms, and curvature bounds to certify a safe number of fine-tuning updates without explicitly running attacks. It also proposes durability-aware training to strengthen unlearning. Experiments on TOFU, MUSE, and WMDP-bio demonstrate improved certified durability while maintaining comparable forgetting effectiveness and model utility.

Critical strengths

(1) The paper provides a novel and systematic diagnosis of hidden pipeline couplings in graph-RAG evaluation, demonstrating that observed performance degradation largely results from implementation artifacts rather than graph evidence itself. Its strengths include rigorous controlled experiments, statistically supported findings, comprehensive graph auditing, and transparent analysis of limitations.

(2) The comparison between graph-based and graph-free multi-hop reasoning offers valuable insights into the role of knowledge graphs in small VLMs. Furthermore, the released datasets, graphs, and evaluation tools support reproducibility, while the findings provide practical guidance for improving reliable and efficient RAG systems.

Critical weaknesses

The paper has several limitations that weaken its novelty, experimental rigor, and generalizability.

(1) First, it lacks sufficient comparison with prior cultural-heritage and multimodal KG-RAG studies, as well as established KG-RAG baselines.

(2) Second, its central causal diagnosis is directly validated only for the 2B model, while experiments are restricted to a single VLM family and a small, author-constructed dataset.

(3) Third, the multi-hop comparisons rely on simplified graph-free baselines and oracle hop counts for graph traversal, raising concerns about fairness. Additionally, the limited human evaluation and incomplete reporting reduce confidence in some findings.

(4) Finally, the contribution is primarily diagnostic rather than methodological, and its broader applicability remains insufficiently demonstrated.

Questions for authors

(1) Could the authors clarify the novelty of their diagnostic contribution relative to existing cultural-heritage and multimodal KG-RAG approache? A more comprehensive discussion and comparison with established KG-RAG baselines would better position the work.

(2) Could the authors provide stronger causal evidence for the observed performance degradation in the 7B model, rather than relying mainly on preliminary logs? Additional controlled experiments or statistical analyses would strengthen the central claims.

(3) How well do the identified image-coupling and n-gram blocking effects generalize beyond the Qwen2-VL family? Evaluating at least one additional VLM architecture would help establish the broader applicability of the findings.

(4) Could the authors clarify the fairness of the multi-hop comparisons, particularly given that graph-free baselines are simplified while KG traversal methods are provided with oracle hop counts?

(5) Could the authors expand the human evaluation and provide clearer annotation guidelines and inter-annotator agreement statistics, particularly for hallucination detection and fabricated citations?

(6) Could the authors provide a more consolidated description of the preliminary pipeline configuration and release a minimal reproduction script for the identified image-coupling artifact?

(7) To what extent do the findings generalize to larger datasets, more complex knowledge graphs, and other application domains? Additional sensitivity analyses or a discussion of the boundary conditions would strengthen the practical significance of the conclusions.

(8) Could the authors report strict recall alongside exact match in the main experimental results and provide statistical significance tests for key comparisons?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.