Forgetting That Sticks: Quantization-Permanent Unlearning via Circuit Attribution

• URL: http://arxiv.org/abs/2605.15138v1
• Date: Thu, 14 May 2026 17:44:10 GMT
• Title: Forgetting That Sticks: Quantization-Permanent Unlearning via Circuit Attribution
• Abstract summary: We show gradient-based methods that achieve meaningful forgetting lose it under compression, while methods that survive quantization barely change the model.<n>We present MANSU, which resolves both modes by combining causal circuit attribution to isolate the minimal forget-set subgraph.<n>We additionally introduce Circuit Attribution Divergence (CAD), a mechanistic verification metric distinguishing structural erasure from behavioral suppression.
• Score: 3.6704226968275253
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• 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.

Related papers

• QuaSAR: Quantization Compensation via Stable Activation-Aware Rank Truncation [0.0]
  Recent training-free post-training quantization methods restore model accuracy through closed-form residual compensation.<n>We show that under the low-bit W4A4 setting, this gating mechanism fails to distinguish poorly predictable quantization error from numerical solver failure.<n>We propose a parameter-free truncated pseudoinverse solver which removes collapsed directions prior to inversion.
  arXiv  Detail & Related papers  (2026-08-14T10:00:36Z)
• Detecting Safety Training Modification in Language Models via Activation Analysis [0.0]
  We introduce AMS, a tool that detects modifications to safety training in language models by measuring the geometric structure of safety-relevant concepts in activation space.<n>We validate AMS across 14 model configurations spanning 4 architecture families.<n>Mechanistic analysis identifies a four-class taxonomy of safety-training modifications distinguished by activation-space signature.
  arXiv  Detail & Related papers  (2026-08-06T03:57:38Z)
• Closed-Form Spectral Regularization for Multi-Task Model Merging [96.82449201305234]
  Model merging combines several independently fine-tuned experts into a single multi-task model without any training data.<n>State-of-the-art merging methods formulate merging as a layer-wise interference problem.<n>We propose SWUDI, a closed-form method that combines a soft exponential filter, which matches the gradient-flow trajectory of iterative descent.
  arXiv  Detail & Related papers  (2026-06-05T14:00:47Z)
• The Deterministic Horizon: Impossibility Results as Design Specifications for Trustworthy AI Systems [0.0]
  This thesis turns impossibility results from curiosities into design rules.<n>Its flagship result proves an accuracy ceiling set by architecture alone.<n>The same argument recasts across subfields.
  arXiv  Detail & Related papers  (2026-05-21T20:48:35Z)
• Quantization Dominates Rank Reduction for KV-Cache Compression [0.0]
  quantization consistently outperforms rank reduction by 4-364 PPL depending on model and compression level.<n>We formalize this via a result showing projection damage exceeds quantization damage by 3 x 2 (2b) per direction under the softmax Fisher metric.
  arXiv  Detail & Related papers  (2026-04-13T14:06:18Z)
• BASIS: Balanced Activation Sketching with Invariant Scalars for "Ghost Backpropagation" [0.0]
  Activation memory required for exact backpropagation scales linearly with network depth, context length, and feature dimensionality.<n>In this paper, we introduce an efficient backpropagation algorithm that fully decouples activation memory from the batch and sequence dimensions.
  arXiv  Detail & Related papers  (2026-03-05T20:38:25Z)
• Overconfident Errors Need Stronger Correction: Asymmetric Confidence Penalties for Reinforcement Learning [17.384089089363382]
  We identify a root cause that existing methods overlook: the uniform penalization of errors.<n>Current approaches treat all incorrect rollouts within a group identically.<n>We propose the Asymmetric Confidence-aware Error Penalty (ACE)
  arXiv  Detail & Related papers  (2026-02-24T22:46:43Z)
• BPDQ: Bit-Plane Decomposition Quantization on a Variable Grid for Large Language Models [56.504879072674015]
  We propose Bit-Plane Decomposition Quantization (BPDQ), which constructs a variable quantization grid via bit-planes and scalar coefficients.<n>BPDQ enables serving Qwen2.5-72B on a single GTX 3090 with 83.85% GSM8K accuracy (vs. 90.83% at 16-bit)
  arXiv  Detail & Related papers  (2026-02-04T02:54:37Z)
• Phase-space entropy at acquisition reflects downstream learnability [54.4100065023873]
  We propose an acquisition-level scalar $S_mathcal B$ based on instrument-resolved phase space.<n>We show theoretically that (S_mathcal B) correctly identifies the phase-space coherence of periodic sampling.<n>$|S_mathcal B|$ consistently ranks sampling geometries and predicts downstream reconstruction/recognition difficulty emphwithout training.
  arXiv  Detail & Related papers  (2025-12-22T10:03:51Z)
• Geometric-Disentangelment Unlearning [106.99160454669902]
  gradient ascent on forget samples often harms retained knowledge.<n>We propose the Geometric-disment Unlearning (GU) that decomposes any candidate forget gradient update into tangential and normal components to retain space and executes only the normal component.<n>Our method is plug-and-play and can be attached to existing gradient-based unlearning procedures to mitigate side effects.
  arXiv  Detail & Related papers  (2025-11-21T09:58:25Z)
• CAGE: Curvature-Aware Gradient Estimation For Accurate Quantization-Aware Training [73.46600457802693]
  We introduce a new method that counteracts the loss induced by quantization.<n>CAGE significantly improves upon the state-of-theart methods in terms of accuracy, for similar computational cost.<n>For QAT pre-training of Llama models, CAGE matches the accuracy achieved at 4-bits (W4A4) with the prior best method.
  arXiv  Detail & Related papers  (2025-10-21T16:33:57Z)
• Post-Hoc Split-Point Self-Consistency Verification for Efficient, Unified Quantification of Aleatoric and Epistemic Uncertainty in Deep Learning [5.996056764788456]
  Uncertainty quantification (UQ) is vital for trustworthy deep learning, yet existing methods are either computationally intensive or provide only partial, task-specific estimates.<n>We propose a post-hoc single-forward-pass framework that jointly captures aleatoric and epistemic uncertainty without modifying or retraining pretrained models.<n>Our method applies emphSplit-Point Analysis (SPA) to decompose predictive residuals into upper and lower subsets, computing emphMean Absolute Residuals (MARs) on each side.
  arXiv  Detail & Related papers  (2025-09-16T17:16:01Z)
• Rethinking Classifier Re-Training in Long-Tailed Recognition: A Simple Logits Retargeting Approach [102.0769560460338]
  We develop a simple logits approach (LORT) without the requirement of prior knowledge of the number of samples per class. Our method achieves state-of-the-art performance on various imbalanced datasets, including CIFAR100-LT, ImageNet-LT, and iNaturalist 2018.
  arXiv  Detail & Related papers  (2024-03-01T03:27:08Z)
• Towards Continual Learning Desiderata via HSIC-Bottleneck Orthogonalization and Equiangular Embedding [55.107555305760954]
  We propose a conceptually simple yet effective method that attributes forgetting to layer-wise parameter overwriting and the resulting decision boundary distortion. Our method achieves competitive accuracy performance, even with absolute superiority of zero exemplar buffer and 1.02x the base model.
  arXiv  Detail & Related papers  (2024-01-17T09:01:29Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.