Unlearning Without Losing Focus: Leverage-Targeted Attention Edits

One promising way to make a language model forget a fact is to blur its attention. Attention Smoothing Is All You Need For Unlearning (ASU) observes that recalling a memorized fact depends on attention that is sharply concentrated on a few tokens. ASU raises the softmax temperature in middle layers to build a "forget-teacher" with flattened attention, then distills the model toward it. Work on contextual noise points the other way. The main documented weakness of Transformers in long or noisy prompts is attention spread across irrelevant tokens. Improving Zero-shot Reader by Reducing Distractions from Irrelevant Documents in Open-Domain Question Answering shows readers derailed by irrelevant retrieved passages. UDA reports GPT-4 falling from 45.9 with retrieval to 0.4 when given the full document on FinHybrid. If unlearning works by flattening attention everywhere in a layer, it may quietly make the model worse at ignoring distractors. Current unlearning benchmarks would not catch this, because they test utility only on clean inputs such as MMLU, TOFU's retain set, and GSM8K.

This note proposes to test that suspicion directly. It also proposes a fix borrowed from matrix completion. There, "coherence" measures whether a matrix's energy sits in a few spiky coordinates, and "leverage scores," the per-coordinate share of a subspace's energy, tell you where those spikes are. The proposal is to locate, inside each attention head, the few directions in key space that carry unusually high leverage for the forget set relative to benign data, and to shrink only those directions. Attention elsewhere stays sharp. The claim is that this removes facts about as well as global smoothing while leaving the model's ability to focus in noisy contexts intact.

Core Claim

The work makes two claims, and the first gates the second. The diagnostic claim is that unlearning methods which act by delocalizing attention, ASU in particular, increase benign-task sensitivity to distractor documents compared with the original model and with methods that do not touch attention. The increase should show up as a measurable rise in the uniformity of attention rows on benign noisy prompts. The constructive claim is that a leverage-targeted edit can match ASU's forgetting while avoiding this cost. The edit confines unlearning to the head-level key directions on which forget-set attention is spiky and retain-set attention is not.

If the diagnostic claim fails, the constructive half is unnecessary, and the paper becomes a negative result that clears attention-smoothing unlearning of this concern. That outcome is still worth publishing, because no one has run these evaluations.

To make "spikiness" of attention concrete, we use the direct analogue of the incoherence parameter from Exact Matrix Completion via Convex Optimization. That paper bounds $\max_i \|U^T e_i\|^2$ by $\mu_0 r/n$. For an attention row $a \in \Delta^N$ over $N$ context tokens, we define

$$\mu(a) = N \cdot \max_j a_j,$$

which equals 1 for perfectly uniform attention and $N$ for attention placed on a single token. Distraction literally means low $\mu$ on the rows that should be attending to the answer-bearing passage. ASU-style forgetting means low $\mu$ on the rows that retrieve a forgotten fact. The tension is whether you can obtain the second without the first.

Research Program

The first component is measurement. For each unlearned model, we record per-head $\mu$ distributions on two kinds of prompts. The first kind is forget-set prompts. The second kind is benign open-domain questions padded with $k$ retrieved distractor passages, for $k$ from 0 to 50. For each head we also compute the attention mass placed on the gold passage. This tells us where and how much each method changed attention focus. Existing work does not provide it: ASU reports forgetting and utility but not attention behavior under distraction, and RMU from The WMDP Benchmark is evaluated only on clean multiple-choice tests.

The second component is the method, Leverage-Targeted Unlearning. For head $h$ with query-key bilinear form $M_h = W_Q W_K^T$, we collect two covariance matrices of key vectors in the head's subspace. The first, $C_F$, comes from answer-bearing positions in forget prompts and is weighted by the attention those positions receive. The second, $C_R$, comes from the same kind of positions in retain prompts. The directions to edit are the top generalized eigenvectors

$$C_F v = \lambda C_R v,$$

meaning directions where forget-set keys have much higher leverage than retain-set keys. This plays the role of local coherence in Coherent Matrix Completion, where sampling effort is directed by row and column leverage rather than spread uniformly. The forget-teacher is the original model with $M_h$ shrunk along these few directions, by a factor $\gamma$, in heads where $\lambda$ is large. It replaces ASU's uniform temperature increase. The student is distilled toward this teacher exactly as in ASU, using a low-rank adapter on attention projections only.

We add a row-norm incoherence penalty on the adapter factors. This follows the regularizer that Matrix Completion has No Spurious Local Minimum needs for its benign-landscape result, and that A Unified Computational and Statistical Framework for Nonconvex Low-rank Matrix Estimation also relies on. Its purpose is to keep the update from concentrating its own energy into new spikes. We treat this penalty as an engineering stabilizer and do not claim a guarantee: those landscape results are for symmetric matrix factorization, not for distillation through a Transformer.

The third component tests the opposite failure. Noise might bring a forgotten fact back. If a retrieved passage mentions related entities, the context might supply the associative cue that sharp attention used to provide internally. When Machine Unlearning Meets RAG shows that its purely behavioral, retrieval-based unlearning drops to about 20% success under high exposure. No one has run the corresponding test on parametrically unlearned models. We will.

Baselines

The naive reference point is the original fine-tuned model, which has not been unlearned. It sets the noise-robustness ceiling and the forgetting floor. Gradient Difference, which applies gradient ascent on the forget set and descent on the retain set, is the standard weak competitor. TOFU reports that it fails to reach statistically significant forget quality at the 1% split (p > 0.05) without losing utility.

The strongest competitor on this specific question is ASU itself. It is the method whose mechanism raises the concern, and it reports better forget-quality/utility trade-offs on TOFU plus stability over ten consecutive unlearning rounds where gradient ascent collapsed. We found no absolute ASU forget-quality number on Lacuna. The file reports only its chosen temperature of 2.3 and its relative superiority. We will therefore rerun ASU ourselves on Llama-2-7B, using layers 6–8 and τ = 2.3 as published, with the same retain set as our method.

RMU is the second strong baseline, because it reaches forgetting without editing attention. The reference number is from The WMDP Benchmark: on Zephyr-7B, WMDP-Bio accuracy falls from 63.7% to 31.2% against a 25% chance level, while MMLU moves only from 58.1% to 57.1%. Stable Forgetting is a third baseline, representing bounded low-rank updates on MLP blocks. It reports forget quality about three orders of magnitude above LoKU on TOFU; the file gives no absolute value.

All methods receive the same forget and retain data, the same adapter rank where one applies, and the same number of hyperparameter trials (twenty) selected on a held-out validation split. Selection uses a fixed rule: the highest forget quality subject to model utility within one point of the original model.

Two ablations isolate the key ingredient. The first replaces the generalized-eigenvector directions with randomly chosen directions of matched rank in the same heads. The second applies ASU's temperature increase only to the heads our method selects, which separates where the edit happens from what it does. The proposal beats ASU if two conditions hold together: forget quality and utility on TOFU are statistically indistinguishable from ASU or better, and the drop in exact match from 0 to 50 distractors is significantly smaller (paired bootstrap over questions). It beats RMU if it matches RMU's WMDP-Bio/MMLU trade-off and is also more robust to resurrection under related-context exposure.

Why not just…?

Why not just run ASU on fewer heads chosen by an existing attribution method? This is the most serious threat. Layer selection by leave-one-out masking as in FIT, or Hessian attribution as in Erase or Hide?, might localize the damage enough on its own. Temperature still flattens a selected head in every direction. Our method flattens a head only along forget-specific key directions, so benign queries routed through that head should keep their focus. The second ablation above (ASU temperature on our selected heads) together with an ASU+FIT-selected-layers variant settles this question. If either matches us on the distractor curve, the leverage directions add nothing, and we will report that.

Why not just fix noise at inference with distraction-aware prompting or passage pruning? Methods such as DAS (Cho et al.), which reports a 64% average exact-match gain for zero-shot readers, or Superposition Prompting, which prunes paths through a document graph, are complementary and cheap. If they fully close any robustness gap that ASU opens, the gap matters little in practice. We will apply DAS on top of every unlearned model. The relevant quantity is the residual gap that remains after DAS. Pruning also does not address the resurrection test, where the dangerous passage is relevant to the query and would survive pruning.

Why not just use RMU, which never touches attention? RMU might show no added distraction sensitivity at all, and the diagnostic experiment may confirm that. The WMDP paper, however, reports worse collateral damage in the chemistry domain, and RMU's resistance to relearning through context has not been tested. If RMU is both robust to noise and resistant to resurrection, it dominates, and our contribution narrows to explaining why attention-level unlearning is the worse design.

Proposed Experiments

The decisive first experiment is the distraction curve. We unlearn on the TOFU 5% and 10% forget splits with each method. Then, on the retain set, Real Authors, and Natural Questions, we measure exact match as 0, 5, 20, and 50 distractor passages are inserted, with the gold passage at the start, middle, or end of the context. We do the same on UDA's FinHybrid and NqText subsets, under both retrieval and full-document settings. The diagnostic claim holds if ASU's curve falls significantly faster than the original model's, and if that faster decline coincides with lower $\mu$ on gold-passage attention in the edited layers. If ASU's curve is parallel to the original model's, we stop and report the negative result.

The second experiment compares the full method against the baselines and ablations on TOFU (forget quality, model utility, truth ratio), on WMDP-Bio, Cyber, and Chem with MMLU using Zephyr-7B, and on the continual setting of the PCH benchmark from FIT. In the continual setting we track the distractor curve after each round, testing whether focus erodes as edits accumulate.

The third experiment is resurrection. For forgotten TOFU authors and WMDP topics, we insert retrieved passages that mention related entities but not the target answer, and we measure how much forgotten-answer accuracy recovers. We also run the fine-tuning relearning attack on FaithUn from Erase or Hide?, where SSIUU's forgetting score of 14.8% compared with 68.4% for gradient ascent provides the reference scale.

A final mechanistic check asks whether the eigenvalue spectrum $\lambda$ predicts which heads matter. If the method's premise is right, ablating the top-$\lambda$ heads in the original model should remove forget-set recall while leaving distractor robustness largely unchanged. Ablating low-$\lambda$ heads in the same layers should do the reverse.