Pith. sign in

REVIEW 2 major objections 2 minor 27 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

MANSU achieves unlearning that survives 4-bit quantization by isolating causal circuits and enforcing update magnitudes above bin width.

2026-06-30 20:59 UTC pith:TJ7E7UDU

load-bearing objection The paper flags a real quantization failure in unlearning and offers a circuit-based fix, but its root-cause numbers lack the needed derivation. the 2 major comments →

arxiv 2605.15138 v1 pith:TJ7E7UDU submitted 2026-05-14 cs.LG cs.CLcs.ET

Forgetting That Sticks: Quantization-Permanent Unlearning via Circuit Attribution

Saisab Sadhu , Pratinav Seth , Vinay Kumar Sankarapu This is my paper
classification cs.LG cs.CLcs.ET
keywords machine unlearningpost-training quantizationcircuit attributionnull-space projectionstructural erasurequantization permanenceCAD metric
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that gradient-based unlearning methods produce updates too small to survive post-training quantization, while methods that preserve behavior under compression do not actually forget. It traces both failures to per-parameter changes falling 47-828 times below NF4 bin widths, formalizing this as a sparsity-permanence tradeoff. MANSU counters the problem by using circuit attribution to find the minimal forget subgraph, restricting null-space projection to that subgraph with a Fisher-based retain bound, and applying a magnitude floor so updates clear quantization thresholds by design. It adds Circuit Attribution Divergence to confirm structural change rather than mere behavioral suppression. A reader would care because every deployed model is quantized, making standard unlearning evaluations incomplete if they ignore compression.

Core claim

MANSU resolves both failure modes of prior unlearning by combining causal circuit attribution to isolate the minimal forget-set subgraph, circuit-restricted null-space projection with a diagonal-Fisher retain bound, and a per-parameter magnitude floor guaranteeing quantization survival by construction, while the new CAD metric distinguishes structural erasure from behavioral suppression; across model families and benchmarks it is the first method to satisfy meaningful forgetting, retain preservation, non-positive PTQ gap, and structural erasure with margin on each.

What carries the argument

MANSU, which uses causal circuit attribution to isolate a minimal forget-set subgraph, followed by circuit-restricted null-space projection and a magnitude floor that forces updates to exceed quantization bin width.

Load-bearing premise

The per-parameter updates across baselines lie 47-828 times below the NF4 quantization bin width, and this size mismatch is the root cause of both the loss of forgetting under compression and the lack of actual change in methods that survive compression.

What would settle it

A direct measurement on the same models showing that MANSU updates exceed the NF4 bin width while baseline updates do not, together with an ablation in which removing the magnitude floor makes MANSU lose its PTQ advantage.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Gradient-based baselines recover up to +0.05 accuracy under compression while MANSU maintains the forgetting effect.
  • MANSU is the first method shown to jointly meet all four required properties with margin: meaningful forgetting, retain preservation, non-positive PTQ gap, and structural erasure.
  • The sparsity-permanence tradeoff explains why updates diffused across billions of parameters cannot clear quantization boundaries.
  • CAD provides a verification signal that existing behavioral metrics cannot supply because it measures mechanistic change rather than output suppression.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Unlearning benchmarks that omit post-quantization testing will systematically overstate the effectiveness of gradient-based methods.
  • The magnitude floor may be portable to other editing techniques that need to survive compression, such as model editing or safety fine-tuning.
  • If circuit attribution scales reliably, the same isolation step could reduce the computational cost of null-space methods on larger models.
  • Models quantized to even lower bit widths would likely require proportionally larger magnitude floors to maintain permanence.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 2 minor

Summary. The manuscript claims that gradient-based unlearning methods exhibit a dual failure under 4-bit post-training quantization (PTQ): they either lose forgetting or achieve only minimal behavioral change. Both modes are traced to per-parameter updates lying 47-828x below the NF4 quantization bin width, formalized as a sparsity-permanence tradeoff. The proposed MANSU method combines causal circuit attribution to isolate minimal forget-set subgraphs, circuit-restricted null-space projection with a diagonal-Fisher retain bound, and a per-parameter magnitude floor to guarantee quantization survival by construction. It also introduces Circuit Attribution Divergence (CAD) to distinguish structural erasure from behavioral suppression. Experiments across model families and hazard benchmarks position MANSU as the first method to jointly achieve meaningful forgetting, retain preservation, non-positive PTQ gap, and structural erasure, while baselines recover up to +0.05 accuracy post-compression.

Significance. If the empirical margins and mechanistic justification hold, the work identifies a practically relevant limitation in current unlearning pipelines given that deployed LLMs are routinely quantized. The constructive use of a magnitude floor and the CAD metric for verifying structural changes (rather than relying solely on behavioral metrics) are positive contributions. The approach of restricting interventions to causally attributed circuits also offers a more targeted alternative to full-model gradient updates.

major comments (2)
  1. [Abstract] Abstract: the central root-cause claim that 'per-parameter updates lie 47-828x below the NF4 quantization bin width' is presented without derivation, error bars, dataset sizes, or explicit computation (which parameters measured, L2 vs. per-element norm, how bin width is obtained per model family, or scaling checks). This ratio is load-bearing for both the sparsity-permanence tradeoff and the claim that MANSU evades the failure modes by construction.
  2. [Abstract] Abstract and tradeoff formalization: the sparsity-permanence tradeoff is asserted as explaining the dual failure, yet no equations or step-by-step derivation link the update-to-bin-width mismatch to the observed recovery of up to +0.05 accuracy under compression; if this appears later in the manuscript it must be cross-referenced and shown to be independent of the evaluation models.
minor comments (2)
  1. [Abstract] Abstract: dataset sizes, number of runs, and error bars are omitted from the central measurements and four-way success claim.
  2. [Abstract] Abstract: the potential circularity risk (if the 47-828x ratios were measured on the same models later used to evaluate MANSU) should be explicitly addressed or ruled out.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for highlighting the need for greater transparency in the presentation of our root-cause analysis. We address each comment below and will revise the manuscript to incorporate explicit cross-references, additional measurement details, and an expanded validation of the tradeoff.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central root-cause claim that 'per-parameter updates lie 47-828x below the NF4 quantization bin width' is presented without derivation, error bars, dataset sizes, or explicit computation (which parameters measured, L2 vs. per-element norm, how bin width is obtained per model family, or scaling checks). This ratio is load-bearing for both the sparsity-permanence tradeoff and the claim that MANSU evades the failure modes by construction.

    Authors: We agree the abstract states the ratio without supporting details. The per-element norm computations (not L2), exact parameter subsets (forget-set updates only), NF4 bin-width derivation (per-tensor scaling), dataset sizes, error bars across seeds, and model-family scaling are provided in Section 4.1 and Appendix B. We will revise the abstract to add a concise cross-reference to these sections plus a footnote summarizing the protocol, ensuring the claim is traceable while respecting length limits. revision: yes

  2. Referee: [Abstract] Abstract and tradeoff formalization: the sparsity-permanence tradeoff is asserted as explaining the dual failure, yet no equations or step-by-step derivation link the update-to-bin-width mismatch to the observed recovery of up to +0.05 accuracy under compression; if this appears later in the manuscript it must be cross-referenced and shown to be independent of the evaluation models.

    Authors: The tradeoff is formalized in Section 3.3 (Equation 5) via a probabilistic survival model P(survive) = 1 - CDF(bin_width/2; observed magnitude distribution), which is shown to predict the measured post-PTQ recovery. We will insert an explicit abstract cross-reference to Section 3.3. To confirm independence from the primary models, the revision will report the same magnitude-to-bin mismatch and recovery pattern on a held-out model family. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical observation does not reduce to self-definition

full rationale

The provided abstract and description present the 47-828x update-to-bin-width ratio as an empirical measurement across baselines that motivates the sparsity-permanence tradeoff and the design of MANSU. No equations, self-citations, or derivations are shown that reduce a claimed result to its own inputs by construction (e.g., no fitted parameter renamed as prediction, no uniqueness theorem imported from self-citation, no ansatz smuggled via prior work). MANSU is described as combining circuit attribution, null-space projection with Fisher bound, and magnitude floor, with CAD as a new metric; these are presented as independent contributions evaluated on external benchmarks. The measurement is framed as observation, not as a load-bearing derivation that collapses. Per rules, absent explicit quotes exhibiting reduction, score remains 0.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the unstated premise that circuit attribution can reliably isolate a minimal forget-set subgraph whose null-space projection will not harm retain performance, plus the assumption that enforcing a magnitude floor does not degrade model utility outside the forget set.

free parameters (1)
  • per-parameter magnitude floor
    Introduced to guarantee updates exceed NF4 bin width; value not specified in abstract.
axioms (1)
  • domain assumption Causal circuit attribution isolates the minimal forget-set subgraph without affecting retain circuits
    Invoked when describing MANSU construction; no justification given in abstract.
invented entities (1)
  • Circuit Attribution Divergence (CAD) no independent evidence
    purpose: Mechanistic verification metric to distinguish structural erasure from behavioral suppression
    Newly introduced; no independent evidence provided in abstract.

pith-pipeline@v0.9.1-grok · 5792 in / 1531 out tokens · 23449 ms · 2026-06-30T20:59:25.014616+00:00 · methodology

0 comments
read the original abstract

Standard unlearning evaluations measure behavioral suppression in full precision, immediately after training, despite every deployed language model being quantized first. Recent work has shown that 4-bit post-training quantization can reverse machine unlearning; we show this is not a tuning artefact but a systematic dual failure: gradient-based methods that achieve meaningful forgetting lose it under compression, while methods that survive quantization barely change the model. Both failures trace to the same root cause: across all baselines, per-parameter updates lie 47-828x below the NF4 quantization bin width; updates diffused across billions of parameters cannot clear quantization bin boundaries, a consequence we formalize as a sparsity-permanence tradeoff. We present MANSU (Mechanistic-Aligned Null-Space Unlearning), which resolves both modes by combining causal circuit attribution to isolate the minimal forget-set subgraph, circuit-restricted null-space projection with a diagonal-Fisher retain bound, and a per-parameter magnitude floor guaranteeing quantization survival by construction. We additionally introduce Circuit Attribution Divergence (CAD), a mechanistic verification metric distinguishing structural erasure from behavioral suppression, a distinction existing metrics cannot make. Across multiple model families and hazard benchmarks, MANSU is the first method to jointly satisfy all four properties with margin on each (meaningful forgetting, retain preservation, non-positive PTQ gap, and structural erasure), while gradient-based baselines recover up to +0.05 accuracy under compression.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages · 1 internal anchor

  1. [1]

    Nathaniel Li, Alexander Pan, Anjali Gopal, Summer Yue, Daniel Berrios, Alice Gatti, Justin D. Li, Ann-Kathrin Dombrowski, Shashwat Goel, Gabriel Mukobi, Nathan Helm-Burger, Rassin Lababidi, Lennart Justen, Andrew Bo Liu, Michael Chen, Isabelle Barrass, Oliver Zhang, Xiaoyuan Zhu, Rishub Tamirisa, Bhrugu Bharathi, Ariel Herbert-V oss, Cort B Breuer, Andy Z...

    work page 2024

  2. [2]

    Knowledge unlearning for mitigating privacy risks in language models

    Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. Knowledge unlearning for mitigating privacy risks in language models. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors,Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1...

    work page 2023

  3. [3]

    Catastrophic failure of LLM unlearning via quantization

    Zhiwei Zhang, Fali Wang, Xiaomin Li, Zongyu Wu, Xianfeng Tang, Hui Liu, Qi He, Wenpeng Yin, and Suhang Wang. Catastrophic failure of LLM unlearning via quantization. InThe Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  4. [4]

    Locating and editing factual associations in GPT

    Kevin Meng, David Bau, Alex J Andonian, and Yonatan Belinkov. Locating and editing factual associations in GPT. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors,Advances in Neural Information Processing Systems, 2022

    work page 2022

  5. [5]

    Toy models of superposition, 2022

    Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield- Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy models of superposition, 2022

    work page 2022

  6. [6]

    Attribution patching outperforms automated circuit discovery

    Aaquib Syed, Can Rager, and Arthur Conmy. Attribution patching outperforms automated circuit discovery. In Yonatan Belinkov, Najoung Kim, Jaap Jumelet, Hosein Mohebbi, Aaron Mueller, and Hanjie Chen, editors, Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pages 407–416, Miami, Florida, US, November 2024. A...

    work page 2024

  7. [7]

    Have faith in faithfulness: Going beyond circuit overlap when finding model mechanisms

    Michael Hanna, Sandro Pezzelle, and Yonatan Belinkov. Have faith in faithfulness: Going beyond circuit overlap when finding model mechanisms. InConference on Language Modeling (COLM), 2024. Introduces EAP-IG (Edge Attribution Patching with Integrated Gradients)

    work page 2024

  8. [8]

    Negative preference optimization: From catastrophic collapse to effective unlearning

    Ruiqi Zhang, Licong Lin, Yu Bai, and Song Mei. Negative preference optimization: From catastrophic collapse to effective unlearning. InFirst Conference on Language Modeling, 2024. 10 Quantization-Permanent Unlearning via Circuit Attribution

    work page 2024

  9. [9]

    Simplicity prevails: Rethinking negative preference optimization for LLM unlearning

    Chongyu Fan, Jiancheng Liu, Licong Lin, Jinghan Jia, Ruiqi Zhang, Song Mei, and Sijia Liu. Simplicity prevails: Rethinking negative preference optimization for LLM unlearning. InNeurips Safe Generative AI Workshop 2024, 2024

    work page 2024

  10. [10]

    Unified gradient-based machine unlearning with remain geometry enhancement

    Zhehao Huang, Xinwen Cheng, JingHao Zheng, Haoran Wang, Zhengbao He, Tao Li, and Xiaolin Huang. Unified gradient-based machine unlearning with remain geometry enhancement. InProceedings of the 38th International Conference on Neural Information Processing Systems, NIPS ’24, Red Hook, NY , USA, 2024. Curran Associates Inc

    work page 2024

  11. [11]

    Shen, Xinchi Qiu, Meghdad Kurmanji, Alex Iacob, Lorenzo Sani, Yihong Chen, Nicola Cancedda, and Nicholas D

    William F. Shen, Xinchi Qiu, Meghdad Kurmanji, Alex Iacob, Lorenzo Sani, Yihong Chen, Nicola Cancedda, and Nicholas D. Lane. LLM unlearning via neural activation redirection. InAdvances in Neural Information Processing Systems (NeurIPS), 2025

    work page 2025

  12. [12]

    Continual learning and private unlearning

    Bo Liu, Qiang Liu, and Peter Stone. Continual learning and private unlearning. InProceedings of the First Conference on Lifelong Learning Agents (CoLLAs), 2022

    work page 2022

  13. [13]

    Manning, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. Direct preference optimization: your language model is secretly a reward model. InProceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, Red Hook, NY , USA, 2023. Curran Associates Inc

    work page 2023

  14. [14]

    TOFU: A task of fictitious unlearning for LLMs

    Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary Chase Lipton, and J Zico Kolter. TOFU: A task of fictitious unlearning for LLMs. InFirst Conference on Language Modeling, 2024

    work page 2024

  15. [15]

    Smith, and Chiyuan Zhang

    Weijia Shi, Jaechan Lee, Yangsibo Huang, Sadhika Malladi, Jieyu Zhao, Ari Holtzman, Daogao Liu, Luke Zettlemoyer, Noah A. Smith, and Chiyuan Zhang. MUSE: Machine unlearning six-way evaluation for language models. InThe Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  16. [16]

    Mass-editing memory in a transformer

    Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, and David Bau. Mass-editing memory in a transformer. InThe Eleventh International Conference on Learning Representations, 2023

    work page 2023

  17. [17]

    C- ∆Θ: Circuit-restricted weight arithmetic for selective refusal.CoRR, abs/2602.04521, 2026

    Aditya Kasliwal, Pratinav Seth, and Vinay Kumar Sankarapu. C- δθ: Circuit-restricted weight arithmetic for selective refusal.arXiv preprint arXiv:2602.04521, 2026

    work page internal anchor Pith review arXiv 2026

  18. [18]

    Does localization inform unlearning? a rigorous examination of local parameter attribution for knowledge unlearning in language models

    Hwiyeong Lee, Uiji Hwang, Hyelim Lim, and Taeuk Kim. Does localization inform unlearning? a rigorous examination of local parameter attribution for knowledge unlearning in language models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 21868–21880, 2025

    work page 2025

  19. [19]

    Mechanistic unlearning: robust knowledge unlearning and editing via mechanistic localization

    Phillip Guo, Aaquib Syed, Abhay Sheshadri, Aidan Ewart, and Gintare Karolina Dziugaite. Mechanistic unlearning: robust knowledge unlearning and editing via mechanistic localization. InProceedings of the 42nd International Conference on Machine Learning, ICML’25. JMLR.org, 2025

    work page 2025

  20. [20]

    Qlora: efficient finetuning of quantized llms

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: efficient finetuning of quantized llms. InProceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, Red Hook, NY , USA, 2023. Curran Associates Inc

    work page 2023

  21. [21]

    Horn and Charles R

    Roger A. Horn and Charles R. Johnson.Matrix Analysis. Cambridge University Press, USA, 2nd edition, 2012

    work page 2012

  22. [22]

    Unlearning isn’t deletion: Investigating reversibility of machine unlearning in LLMs, 2026

    Xiaoyu Xu, Xiang Yue, Yang Liu, Qingqing Ye, Huadi Zheng, Peizhao Hu, Minxin Du, and Haibo Hu. Unlearning isn’t deletion: Investigating reversibility of machine unlearning in LLMs, 2026

    work page 2026

  23. [23]

    Chandler Davis and W. M. Kahan. The rotation of eigenvectors by a perturbation. iii.SIAM Journal on Numerical Analysis, 7(1):1–46, 1970

    work page 1970

  24. [24]

    New insights and perspectives on the natural gradient method.J

    James Martens. New insights and perspectives on the natural gradient method.J. Mach. Learn. Res., 21(1), January 2020

    work page 2020

  25. [25]

    Curran Associates Inc., Red Hook, NY , USA, 2019

    Frederik Kunstner, Lukas Balles, and Philipp Hennig.Limitations of the empirical fisher approximation for natural gradient descent. Curran Associates Inc., Red Hook, NY , USA, 2019

    work page 2019

  26. [26]

    Transformer feed-forward layers are key-value memories

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. Transformer feed-forward layers are key-value memories. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih, editors,Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5484–5495, Online and Punta Cana, Dominican Republic, November 2...

    work page 2021

  27. [27]

    circuit dismantled

    Minyeong Choe, Haehyun Cho, Changho Seo, and Hyunil Kim. Do all autoregressive transformers remember facts the same way? a cross-architecture analysis of recall mechanisms. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng, editors,Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages...

    work page arXiv 2025