Confess What You Know: Forget-Set Misalignment with Model Knowledge in LLM Unlearning
Miso Kim, Georu Lee, Seungwon Jeong, Woojin Lee
cs.LG · cs.AI · cs.CL
Quick verdict
Proposes CONFS, a data-blind pipeline eliciting model-aligned forget sets for LLM unlearning; careful gradient diagnostics, yet gold-fact recall stays 0.378.
Summary
This paper studies forget-set misalignment in LLM unlearning: the gap between a user-specified forget set and what the model actually memorized. The authors define two failure modes — Under Unlearning, where omitted memorized data keeps leaking, and Out-of-Knowledge Unlearning, where requests covering never-learned data degrade utility. Using a first-order gradient inner-product diagnostic on controlled TOFU partitions (leaked/requested splits), they show forgetting stays target-localized while gradients on never-seen data dominate retain-set damage. They propose CONFS, a data-blind forget-set constructor: the target model 'confesses' memorized knowledge about an entity, claims are formalized into atomic SRO triplets, reconfession recursively probes attribute-level details, GPT-4o generates competency questions, and DeBERTa-NLI self-consistency filters unreliable pairs. Across TOFU, CLEAR, and RWKU with GA/GD/NPO/RT objectives, CONFS attains the best forget-set F1 among data-blind constructions (0.402 vs 0.195 RWKU-style and 0.123 FreeRecall-QA; gold self-judged 0.983), the lowest out-of-knowledge share (54% vs 68%/91%), and the smallest retain-loss increase (+0.11 vs +0.31/+3.53). Under GA on TOFU it trades forgetting strength (forget Prob 0.372 vs gold 0.303) for much better utility preservation than baselines. Limitations: gold-fact recall is only 0.378, GPT-4o acts as both structurer and judge, results use three seeds without significance tests, and scope is limited to entity-centric factual knowledge.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Forget-set misalignment causes two distinct unlearning failures: Under Unlearning leaves memorized data leaking, and Out-of-Knowledge Unlearning degrades general utility (Sec. 3, Table 1).
- These failures arise from misaligned targets rather than optimization choices: forgetting stays confined to requested facts (Eeg near zero) and never-seen-data gradients dominate retain damage (Eq. 2-5, Fig. 2, Table 5).
- CONFS builds model-aligned forget sets fully data-blind via confession, SRO triplet structuring, reconfession, and consistency filtering, achieving the best F1 (0.402) among data-blind constructions (Table 8).
- CONFS's better alignment translates to outcomes: the lowest out-of-knowledge share (54% vs 68%/91%) and smallest retain-loss increase (+0.11 vs +0.31/+3.53) (Table 5).
- Across TOFU/CLEAR/RWKU with GA/GD/NPO/RT, CONFS approaches gold-standard on several metrics and preserves utility better than other data-blind forget-set constructions (Tables 3, 4, 6).
Strengths
Well-motivated, precisely defined problem
Forget-set misalignment is formalized with a clean two-axis partition (leaked/requested, Table 2) yielding DL+, DL-, DN+ subsets, making both failure modes directly measurable rather than anecdotal.
Careful gradient-level diagnostics
Etf/Eeg metrics (Eq. 3-4) subtract global utility decay from targeted forgetting; Table 1 shows DL- probability decaying at the retain-set rate (0.624 vs 0.658 at step 60, from 0.996 pretrained), and Eq. 5 decomposition shows never-seen gradients dominate retain damage.
Stage-by-stage ablation of the pipeline
Honest evaluation of the gold gap
Unusually complete reproducibility detail
Weaknesses
Residual misalignment is still substantial
'Approaches Gold-standard' is metric-dependent
Thin statistical rigor
Judge-structurer circularity
Lax verification threshold semantics
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper addresses LLM machine unlearning under a realistic assumption gap: the user-supplied forget set rarely coincides with what the model actually memorized. It claims (1) that this forget-set misalignment produces two distinct failure modes — Under Unlearning (unrequested but memorized data keeps leaking) and Out-of-Knowledge Unlearning (requested but never-learned data damages utility); (2) that gradient analysis attributes both to misaligned targets rather than the optimizer; and (3) that CONFS, a data-blind forget-set constructor, approaches gold-standard forgetting-utility balance across TOFU, CLEAR, and RWKU. Claims 1-2 are well supported by the controlled experiments; claim 3 is well supported relative to data-blind baselines, but the 'approaches gold' phrasing needs a careful reading (see below).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is forget-set misalignment in LLM unlearning?
It is the mismatch between the data a user asks a model to forget and what the model actually memorized during training. The paper shows this causes two failures: omitted memorized data keeps leaking (Under Unlearning), and requests about data the model never learned damage general utility (Out-of-Knowledge Unlearning).
How does CONFS build a forget set without access to training data?
CONFS prompts the target model to reveal ('confess') what it knows about an entity, converts the answers into atomic subject-relation-object triplets, recursively probes attribute details (reconfession), turns the resulting facts into competency questions using GPT-4o, and keeps only QA pairs the model reproduces consistently according to a DeBERTa-based self-check.
How close is CONFS to using the true forget set?
Its forget-set F1 against the gold TOFU data is 0.402 (gold self-consistency is 0.983) and it recovers 37.8% of gold facts, yet downstream unlearning results approach the gold baseline on several metrics while preserving more utility than other data-blind constructions such as RWKU-style prompting and FreeRecall-QA.
Which models and benchmarks were used in the paper?
TOFU and RWKU with LLaMA-2-7B-Chat, and the multimodal CLEAR benchmark with LLaVA-1.5-7B; unlearning used GA, GD, NPO, and RT objectives. Forget-set construction takes about 4 minutes per entity on an H200 GPU plus GPT-4o API calls.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt: segments marked [...sampling gap...] skip parts of Secs. 3.1-3.2, 4.3-4.4, Appendix B.5, and Appendix F; judgments about those spans rely on surrounding text only.
- Figure 2's plotted curves are described in captions and text but their numeric values are not extractable from the provided text.
- The excerpt ends inside the Appendix F/G prompt templates; any content beyond that point in the PDF was not visible.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.