acceptodds
Under review as a conference paper at ICLR 2027

Fine-tuning Disrupts Edit Readout: A Mechanistic Perspective on Model Editing Robustness

Abstract

Model editing provides a precise way to update the parametric knowledge of Large Language Models (LLMs). However, it has been shown that fine-tuning of edited models for downstream adaptation will erase the intended edits in existing works. This paper studies this problem through the lens of layer-wise information flow in Transformers. Our systematic layer-wise ablation study reveals that edited memory is mainly anchored in shallow editing layers, while fine-tuning primarily disrupts the deep editing layers and adjacent post-edit layers, jointly forming the readout circuitry that decodes this memory into the final prediction. As a result, reverting this readout circuitry alone is often sufficient to recover editing success. We further verify this layer-wise information-flow perspective through null-space analysis and attention-based visualization. Motivated by this key finding, we propose Layer-wise Model Interpolation (LMI), which selectively restores the disrupted readout circuitry to preserve edited knowledge while retaining downstream adaptation. LMI interpolates between the edited and fine-tuned model weights with layer-aware coefficients, up-weighting the edited model in the readout circuitry and the fine-tuned model elsewhere. Across four public datasets, three language models, and two editing methods, LMI achieves the best balance between edit preservation and downstream utility.

Primary area:
interpretability and explainable AI
Keywords:
Knowledge Editing, Edit Robustness under Fine-tuning, Mechanistic Interpretability, Model Merging, Layer-wise Interpolation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_QRK7OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates why knowledge edits become ineffective after subsequent downstream fine-tuning. Through layer-wise restoration, subspace analysis, and attention-based attribution, the authors argue that fine-tuning primarily disrupts the readout circuitry of edited knowledge rather than erasing the edit signal itself. Motivated by these findings, the paper proposes Layer-wise Model Interpolation (LMI), which selectively interpolates between edited and fine-tuned model weights to recover edited knowledge while preserving downstream adaptation. Experiments across multiple models, editing methods, and fine-tuning settings demonstrate substantial improvements in edit retention.

Critical strengths
  1. The paper presents a clear and coherent research story. It starts from the practical problem of edit degradation after fine-tuning, investigates the underlying mechanism from an information-flow perspective, and proposes a corresponding recovery method.

  2. Some of the mechanistic findings are insightful. In particular, the observation that fine-tuning primarily disrupts the deep editing and adjacent post-editing layers, rather than simply erasing the original edit signals, provides an interesting perspective on model-editing robustness.

Critical weaknesses
  1. The generality of the readout-circuitry hypothesis and LMI remains insufficiently validated. The paper only studies MEMIT and AlphaEdit, which both edit similar early MLP layers (L3–8/L4–8). However, other effective editing methods, such as RLEdit, UltraEdit, and LocFT-BF, employ different editing mechanisms and can operate on different modules or layer locations. It remains unclear whether the identified readout circuitry would still lie in the deep-editing and adjacent post-editing regions under these methods, and whether LMI would remain effective.

  2. The fine-tuning setup remains insufficiently representative of realistic downstream adaptation. The main experiments still rely on 10-epoch, MLP-only fine-tuning over 1,000 ZsRE examples, often achieving nearly 100% training accuracy but relatively low test accuracy. Although the authors have added more realistic settings, including NQ-open and full-parameter fine-tuning, these evaluations are less comprehensive. In particular, Table 2 reports edit accuracy without downstream task performance, making it difficult to assess whether LMI truly preserves both edited knowledge and downstream capabilities under these settings.

Questions for authors

Question

Could the authors provide a more comprehensive evaluation of general capability preservation? Table 3 only evaluates full-scale MEMIT editing followed by ZsRE fine-tuning, reporting MMLU on Llama-3 and LAMBADA on GPT-J. Extending this evaluation to other editing methods, models, and downstream settings would provide stronger evidence that LMI preserves general capabilities.

Suggestion

I recognize that a comprehensive evaluation of this pipeline is computationally expensive, as it involves model editing, downstream fine-tuning, recovery, and general task evaluations. Nevertheless, I believe that at least one carefully controlled end-to-end experiment with a more representative downstream adaptation setting and a different editing paradigm would substantially strengthen the paper's conclusions.

Despite these remaining concerns, I appreciate the clear research motivation, coherent analysis, and potentially insightful mechanistic findings. I therefore lean toward Weak Accept, while encouraging the authors to further strengthen the experimental evidence and clarify the scope and generality of their conclusions.

3 comments

We appreciate the reviewer's comments on the coherence of the research narrative and are pleased that they found our mechanistic perspective on model-editing robustness insightful. We also thank the reviewer for the constructive comment on more comprehensive experiments. In response, we add a complete end-to-end experiment (LocFT-BF edit → Dolly instruction tuning → LMI → evaluation of edit retention, downstream performance and general capability), covering a new downstream task and a new editing method.

Setup. Llama-3-8B, Full scale (about 20.9K CounterFact edits). LocFT-BF is a different paradigm from MEMIT: fine-tuning-based editing that updates only the MLP output matrix of L22. The downstream stage is full-parameter instruction tuning on Dolly-15K (lr 3e-5, 3 epochs). The LMI (μ, σ) are selected by EA × Tr on a random 10% of the edits, and ES / EA are reported on the remaining 90%. Tr / Te are response-token accuracies on the Dolly training / held-out split. ROUGE-L is computed on the held-out split and measures downstream generation quality. MMLU is 5-shot.

LocFT-BF and Dolly fine-tuning

ModelESEATrTeTest lossROUGE-LMMLU
Original27.465.3
Edited100.099.750.150.62.40019.565.3
+ Dolly fine-tuning79.858.689.458.02.09935.461.3
LMI (μ=24, σ=3)99.296.381.658.41.93936.562.5

Instruction tuning improves held-out performance (Te 50.6 → 58.0, ROUGE-L 19.5 → 35.4), so the task is learned rather than memorised. The edits nevertheless collapse (EA 99.7 → 58.6), and LMI recovers 92% of the drop. Its ROUGE-L (36.5 vs. 35.4) and MMLU (62.5 vs. 61.3) are not below those of the fine-tuned model. LocFT-BF edits L22 and the selected LMI centre is L24, consistent with our finding that the readout band lies just downstream of the edit layers.

Thank you for your response. The additional complete end-to-end experiment on LocFT-BF addresses most of my concerns. I encourage the authors to further strengthen the experimental evaluation in the revised version, for example, by considering a broader range of downstream tasks beyond MMLU.

Thank you very much for your positive feedback and for taking the time to review our response. We are glad that the additional end-to-end experiment on LocFT-BF addresses most of your concerns.

We fully agree with your suggestion. In the revised version, we will strengthen the experimental evaluation by including a broader range of downstream tasks beyond MMLU, to further verify the generality of our findings.

Thank you again for your constructive comments, which have helped us improve the paper.

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

This paper investigates the phenomenon that fine-tuning edited models for downstream adaptation can erase the intended edits. To address this issue, the authors propose a Layer-wise Model Interpolation (LMI) method and conduct extensive experiments from multiple perspectives to validate its effectiveness.

Critical strengths

The evaluation methods are diverse, and the experimental comparisons are comprehensive and detailed.

The paper is logically well organized, with a clear structure and smooth transitions between sections.

The finding that “restoring downstream layers is more effective than restoring the editing layers” is an interesting empirical observation.

Critical weaknesses

Please also use hyphens consistently and correctly between words, and carefully distinguish whether a single hyphen (-) or a double hyphen (--) is appropriate.

It may be worth reconsidering whether the choices of learning rate and number of epochs should be discussed in the main text or instead specified in the Experimental Setup section.

Could Table 2 also be formatted consistently using the three-line table style?

Please consider shortening the caption of Figure 1 and moving some of the detailed explanations to the main text.

Please provide the specific value used for the small tolerance ϵ\epsilon.

Although LMI is only one component of the paper, its weighted interpolation-based modification offers limited methodological novelty.

Questions for authors

Refer to the above text.

2 comments

We thank the reviewer for highlighting the breadth of our experiments, the clarity of the presentation and our key finding. We adopt the first four points in full and have revised the paper accordingly; we respond to the fifth and sixth below.

1. Hyphens. We went through the whole paper and made the usage consistent. We also unified the spelling of recurring terms (e.g., fine-tuning / fine-tuned throughout, no longer "finetune").

2. Learning rate and number of epochs. Agreed. The learning rates, numbers of epochs, batch sizes and other training settings are moved from the running text of Section 4 to the Experimental Setup (Sec. 6.1). Section 4 keeps a single sentence pointing to Sec. 6.1.

3. Format of Table 2. Table 2 now uses the same three-line style as the other tables (booktabs: \toprule / \midrule / \bottomrule, no vertical rules).

4. Caption of Figure 1. Shortened. The caption now gives one sentence for each of the panels (A), (B) and (C); the mechanistic explanation of the readout band and the five-region partition are moved to the Introduction and to Sec. 4.3.

5. The value of ε. We thank the reviewer for pointing out this unclear statement. In Eq. (1), ε formalises the requirement that the recovery procedure should not noticeably degrade downstream performance. It is not a hyperparameter that we tune, and no fixed threshold is used in the experiments. The criterion used in practice is to maximise EA × Tr over the candidate configurations, which is a threshold-free surrogate of the constrained objective in Eq. (1).

6. Methodological novelty of LMI. We agree that LMI is, in form, a simple weighted interpolation, and we do not claim it as a new model-merging algorithm. The main contribution of the paper is the mechanistic finding.

Thank you for your detailed point-by-point responses. After thoroughly consideration, I have decided to maintain my original rating.

Reviewer_2kZ4OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper asks why knowledge edits are erased when an edited model is fine-tuned for a downstream task. Using MEMIT/AlphaEdit edits (COUNTERFACT, ZsRE) on GPT-J-6B, Llama-3-8B and Qwen2.5-7B, it reports three mechanistic findings. (i) Edits collapse after fine-tuning, but the usual Edit Success metric hides this: ES stays ~96–97% while Edit Accuracy falls from 100% to ~25–30%. (ii) The edit signal is not erased from parameters: key drift into the edited MLPs is ~66% but the part that couples with the fine-tuning update (signal error) is only ~17%, with readout direction cosine similarity > 0.99 — i.e., the drift lies mostly in the null space of the update. (iii) Sliding-window MLP restoration (activation patching) shows that reverting the layers downstream of the edit region (L9–L14) recovers far more EA (89%/95% for MEMIT/AlphaEdit) than reverting the editing layers themselves (L3–L8, 76%), and the peak also appears under a uniform-residual control; Grad-ELLM attribution and an SVD null-space analysis are offered as corroboration. The authors conclude that fine-tuning disrupts a "readout circuitry" (deep editing + adjacent post-edit layers) rather than the stored edit, and propose LMI: per-layer convex interpolation between the edited and fine-tuned checkpoints with coefficients from a Gaussian kernel centered on the readout band (α ∈ [0.1, 0.9], µ and σ selected by maximizing EA × train accuracy). LMI is evaluated against restoring crucial layers (RCL), uniform interpolation (UMI), Task Arithmetic and TIES-Merging across 4 datasets, 3 models, 2 editors, edit scales 10²–21K, and MLP-only / attention-only / full-parameter / LoRA fine-tuning, plus PS/NS/GE/RS behavioral profiles and MMLU/LAMBADA. LMI attains the best edit-accuracy/downstream trade-off in most settings (e.g., 5.9% → 73.0% EA on GPT-J MEMIT@Full; 2.0% → 83.0% on Llama-3 MEMIT@10²), sometimes improving downstream test accuracy.

Critical strengths
  1. "The paper converts an empirical observation ("edits are fragile under fine-tuning") into a specific, testable mechanism, supported by three independent lines of evidence". Layer-wise activation restoration, a null-space subspace analysis of input–output drift, and Grad-ELLM attribution all converge on the same claim — the edit survives in the early editing layers while fine-tuning breaks the later circuit that decodes it — and the key result survives a uniform-residual control and replicates on three model families. This is consequential because it reframes the problem: recovery should target readout, not re-injecting knowledge, and it explains prior findings (e.g., that fine-tuning outside the editing range is especially damaging).
  2. "LMI is simple, training-free, and validated broadly". It is a closed-form per-layer weight merge requiring no data, no re-editing and no optimization, yet it is tested across 3 LMs, 2 editors, edit scales from 100 to ~21K, four downstream datasets, four fine-tuning regimes, and two extra merging baselines. The evaluation goes beyond EA to paraphrase/neighborhood/entropy/reference scores and to general capability (MMLU, LAMBADA), showing merging adds no capability loss. Hyper-parameters, full sweeps and code are documented, so the gains are reproducible.
Critical weaknesses
  1. "The one baseline that would most threaten the method is missing". Since the recovery map is given both checkpoints and the edit set (needed to compute M_ed in Eq. 1), re-applying the same editor to the fine-tuned model — or simply freezing the edited layers during fine-tuning — is available and cheap, and would plausibly dominate LMI on both edit accuracy and downstream accuracy. Without this comparison, the paper does not establish when weight merging is preferable to the obvious alternative, so the practical significance of LMI is unproven even though the mechanistic analysis stands.
  2. "Both the localization claim and the reported gains have unaddressed threats to validity". The sliding-window result lacks a control for the generic fact that deeper layers are more causally influential on the output: no restoration is run on unedited or held-out facts, so the L6–L14 peak may not be edit-specific, and the selected band moves a lot across models (µ* = 6 for GPT-J, 7–10 for Llama-3, 13–14 for Qwen2.5), weakening the claim of a stable five-region partition. Separately, (µ, σ, α) are chosen by maximizing EA × train accuracy on the very edit set used for evaluation (no held-out split), and fine-tuning is run to 100% training accuracy on 1k examples — a memorization regime — so both the collapse and the recovery magnitudes may be optimistic and setting-specific.
Questions for authors
  1. Re-editing baseline. Please report EA / Tr. / Te. for re-applying MEMIT or AlphaEdit to the fine-tuned checkpoint, and for edit-preserving fine-tuning (e.g., L2 penalty toward W_ed, or freezing the edited MLPs). If either dominates LMI, the method's practical claim needs to be reframed.
  2. Is the readout band edit-specific? Run the same sliding-window restoration on unedited facts (or a disjoint edit set) to show the L9–L14 peak is not a generic depth effect, and reconcile µ* = 6 on GPT-J with µ* = 13–14 on Qwen2.5.
  3. Hyper-parameter selection and fine-tuning regime. Report LMI with (µ, σ, α) selected on a held-out subset of edits, and repeat the main comparison under an early-stopped / lower-LR fine-tune that does not reach 100% training accuracy.
  4. EA vs. LR. EA falls to ~25% while LR stays ~83–84%, so much of the "erasure" may be a top-1 / output-format shift rather than knowledge loss. Does LMI's recovery also hold under LR and under format-matched prompts?
7 comments

We thank the reviewer for recognising the readout-disruption mechanism and LMI as a simple, training-free recovery method. We hope the response could address the concerns and are happy to discuss further if there are any follow-up questions.

Q1. Re-edit Baseline and Freezing edited MLP layers

(a) The setting we target.

We appreciate the raised question that helps us further highlight the significance of our contributions.

Scientifically, our research aims to investigate how downstream fine-tuning disrupts the readout of edited knowledge in an LLM. Re-editing supplies all target facts again, so it could succeed even if fine-tuning had erased every original edit. The problem then reduces to the quality of a single edit applied to an arbitrary model. It measures the editing method and not whether an existing edit can be preserved and recovered after downstream fine-tuning. Thus, re-editing is irrelevant to our research question.

Practically, a downstream user can meet its target and still deploy a model that gives incorrect answers on facts the provider already corrected via model editing. Re-editing requires all original requests, which probably won't hold in practice, and repeats target optimisation for each edit on every adapted checkpoint, which is computationally heavy. In contrast, our LMI requires only the two checkpoints and one weight merge after choosing hyperparameters on a tiny validation set (5% of original edits), which takes about 30 seconds on an H200 (App. B.8).

In summary, re-editing is irrelevant to our research scientifically and not a cheap and valid alternative practically. We have included the above discussion in the updated paper to make it more clear.

(b) On D_edit in Eq. (1). To formalise the task, the objective in Eq. (1) is written over the full edit set D_edit, which expresses that we want to recover as much of the edit as possible, but the full edit set is not required in our pipeline. Specifically in our experiment, the parameters are swept on 1K edits only, and the resulting (μ, σ) are transferred unchanged to the Full scales with about 21K edits, this means that about 5% of the edits are used for parameter selection. We have updated Table 1 in the main text so that the Full-scale rows use exactly the same (μ, σ) as the other three scales (the table and the corresponding text are marked in blue). The difference is marginal.

(c) On freezing the edited MLPs and L2 penalty toward W_ed . Table R1-1 reports freezing the edited layers on GPT-J(100 MEMIT edits at L3-8, ZsRE fine-tuning). In the frozen condition the MLPs of L3-8 are kept at their edited weights throughout fine-tuning and only the MLPs of L9-27 are trained.

Table R1-1: Performance of freezing the edited MLPs.

MethodEATrTe
Edited100.00.00.0
Standard FT25.0100.021.0
Freeze L3–8, train L9–2762.0100.019.0
LMI (μ=6, σ=5, from 1K)100.094.621.0

Freezing the edited MLPs raises EA only from 25.0 to 62.0, so 38% of the edits are still lost, whereas LMI recovers 100.0 at a cost of 5.4 points of Tr. We also ran the L2 penalty toward W_ed requested by the reviewer: the term (λ/2)·‖W − W_ed‖² on the MLPs of the edited layers L3-8 is added to the fine-tuning loss, with all MLPs trained and everything else unchanged. These runs use a re-edit of the same 100 facts, on which standard fine-tuning gives EA 28.0 and freezing L3-8 gives EA 54.0.

Table R1-2: Performance of L2 penalty toward W_ed.

λ1e-31e-20.1110
EA30.040.035.028.038.0
Tr100.0100.0100.0100.0100.0
Te20.018.015.020.021.0

For every penalty coefficient EA stays between 28.0 and 40.0, so the penalty does not preserve the edits better than freezing does. Keeping the edited layers close to W_ed is not sufficient to preserve the edit.

Q2. Is the readout band edit-specific?

The hypothesis that deeper layers are more causally influential on the output does not hold. Figure 3 of MEMIT (Meng et al.), sweeps the layer at which a fact is edited and shows that editing success is concentrated at the early-to-mid layers identified by causal tracing, and drops at deeper layers rather than increasing with depth. We then run the same sliding-window recovery on facts that were never edited, where there is likewise no layer-wise difference (Table R1-3, R1-4).

(a). Sliding-window recovery on unedited facts

Setup. GPT-J (1K MEMIT edits at L3-8, ZsRE fine-tuning with lr 6e-5 for 10 epochs) and Qwen2.5-7B (1K MEMIT edits at L4-8, ZsRE fine-tuning with lr 4e-5). Each window is scored on two sets of facts: edited facts: the 1,000 edited facts, scored by EA; unedited facts: 964 ZsRE facts that were never edited.

Table R1-3: Sliding-window recovery on unedited facts with GPT-J.

EA (edited)Acc (unedited, 964)
Edited96.50.0
FT8.719.0
L0–525.217.1
L3–837.218.6
L6–1155.719.3
L9–1450.820.0
L12–1724.719.4
L15–2015.619.5
L18–2311.419.7
L22–2710.219.4

Table R1-4: Sliding-window recovery on unedited facts with Qwen2.5. Before fine-tuning, the edited model answers 112 of the unedited facts correctly; the third column is computed on these 112 facts and the fourth on all 964.

EA (edited)Acc (112 known)Acc (964)
Edited99.7100.011.6
FT30.788.436.1
L0–540.890.235.3
L3–851.190.235.5
L6–1161.292.036.8
L9–1465.689.336.4
L12–1762.784.835.6
L15–2070.588.436.6
L18–2354.090.237.6
L22–2737.687.534.3

On both models, EA on the edited facts peaks with window position (GPT-J 55.7 at L6–11; Qwen2.5 70.5 at L15–20), while accuracy on the unedited facts stays flat across all windows and near the fine-tuned level, so the peak is specific to the edited facts and is not a generic depth effect.

(b). On μ = 6 for GPT-J versus μ = 13–14 for Qwen2.5

The two models differ not only in μ* but also in σ*: (μ*, σ*) = (6, 5) for GPT-J, (13–14, 10) for Qwen2.5 and (7–10, 14) for Llama-3. LMI selects a wider Gaussian centred further downstream on Qwen2.5 and Llama-3, and a narrow Gaussian close to the edit layers on GPT-J. This difference comes from the shape of the readout band of each model, which can be seen directly in the finer-grained sliding-window experiments (Figure 4 in the main text, Figures 10 and 13 in the appendix; stride 1, 100 edits):

Table R1-5: Summary of finer-grained sliding-window experiments.

Model (layers, edit layers)Sliding-window curve (MEMIT, 100 edits)(μ*, σ*)
GPT-J (28 layers, L3-8), Figure 4Single peak right after the edit layers: EA 91 to 95 for windows L5-10 to L8-13, maximum at L7-12 (95), then a monotonic decrease to 34(6, 5)
Llama-3 (32 layers, L4-8), Figure 10Elevated region from L7-11 to L12-16 (EA 47 to 63), maximum at L8-12 (63); 17 to 38 outside it(7–10, 14)
Qwen2.5 (28 layers, L4-8), Figure 13Elevated region from L9-13 to L17-21, wider and further downstream, with maxima at L16-20 (42) and L9-13 (33); 15 to 27 outside it(13–14, 10)

What the three models share is that the EA peak always lies downstream of the edit layers. What differs is the width and extent of this downstream readout band: on GPT-J it is concentrated in the few layers right after the edit layers, whereas on Qwen2.5 and Llama-3 it is wider and reaches deeper layers. The Gaussian of LMI has to cover the whole readout band, so on Qwen2.5 its centre moves to 13–14 and its width grows to 10. With a Gaussian of this variance, μ is also not sensitive: at the Full scale on Qwen2.5, EA stays between 44.7 and 48.8 and Tr between 84.6 and 88.3 for μ from 11 to 16.

Q3. Hyper-parameter selection and fine-tuning regime.

As clarified in Q1(b), α* (UMI) and (μ*, σ) (LMI) are selected on the 1K edits only, by maximising EA × Tr, and transferred unchanged to all other scales, so they are never selected on the edits they are evaluated on. We agree that held-out evaluation is necessary: at the Full scale we report a held-out protocol (a, b), add an instruction-tuning task that requires generalisation (c), and repeat the main comparisons under three lighter fine-tuning regimes (d). In-sample and held-out EA differ by only 0.3 points on average (at most 1.2; Table R1-6), always within one standard deviation, and the UMI/LMI ranking is unchanged.

(a). LMI results with hyperparameters selected on held-out edits

Protocol. For every setting at the Full scale (20,877 edits) we draw 100 random splits of the edit set. In each split, the hyperparameters are selected by maximising EA × Tr on 10% of the edits (2,088), and EA is reported on the remaining 90%. Tr / Te do not depend on the split. EA entries are mean ± std over the 100 splits; the "in-sample" column is the paper's original protocol.

Table R1-6: Held-out selection (In/Held: in-sample/held-out EA).

SettingMethodInHeldTrTe
GPT-J MEMITUMI α=0.668.968.6±1.682.120.0
LMI μ=5,σ=573.272.9±0.787.517.3
GPT-J AlphaEditUMI α=0.641.141.1±0.172.212.0
LMI μ=5,σ=430.229.0±1.582.111.4
Llama-3 MEMITUMI α=0.850.750.7±0.183.143.0
LMI μ=8,σ=1451.051.0±0.294.238.8
Llama-3 AlphaEditUMI α=0.860.860.5±2.075.623.1
LMI μ=8,σ=1459.559.4±0.295.620.9

σ was not swept jointly with μ in the paper; it was fixed before looking at downstream results by the rule of thumb in the appendix (σ ≈ half-width of the readout band: 4–5 for GPT-J, 14 for Llama-3).

(b). Agreement between held-out and in-sample selection

For each setting we draw 100 random splits, select the hyperparameter on a 5%, 10% or 20% subset of the edits, and count how often the selection equals the in-sample choice.

Table R1-7: Number of splits (out of 100) that select the in-sample hyperparameter.

SettingMethodIn-sample5% (1,044 edits)10% (2,088)20% (4,175)
GPT-J MEMITUMIα=0.6849899
LMIμ=5678698
GPT-J AlphaEditUMIα=0.6100100100
LMIμ=5535256
Llama-3 MEMITUMIα=0.8100100100
LMIμ=8708891
Llama-3 AlphaEditUMIα=0.89098100
LMIμ=8648193

The in-sample choice is the most frequent held-out choice in every setting and at every fraction. The remaining splits pick a nearby grid point (α ± 0.1 or μ ± 1; μ ± 2 in at most 3 splits) whose held-out EA is within 2 points, which is why the standard deviations in Section (a) are small.

(c). Fine-tuning that requires generalisation

We agree that 10-epoch ZsRE fine-tuning is largely memorisation. We therefore add full-parameter instruction tuning on Dolly-15K (lr 3e-5, 3 epochs) for Llama-3 at the Full scale, which raises held-out ROUGE-L from 22.8 to 35.1. (μ, σ) are selected on 10% of the edits; ES / EA are reported on the remaining 90%.

Table R1-8: MEMIT + Dolly fine-tuning. Tr / Te: response-token accuracy on the Dolly training / held-out split; ROUGE-L on the held-out split; MMLU 5-shot. Original model: ROUGE-L 27.4, MMLU 65.3.

ModelESEATrTeROUGE-LMMLU
Edited93.879.959.459.122.858.3
+ Dolly FT64.236.389.158.435.155.8
LMI (μ=10, σ=14)91.075.977.359.933.959.3

Edits collapse under instruction tuning as well (EA 79.9 → 36.3). LMI recovers 91% of the drop while keeping ROUGE-L close to the fine-tuned model (33.9 vs. 35.1) and MMLU above both endpoints.

(d). Early-stopped / lower-learning-rate fine-tuning

As the reviewer requested, we also add three lighter ZsRE regimes at the Full scale: early stopping at epoch 1 and 3, and half the learning rate for 10 epochs, with hyperparameters selected by the held-out protocol of (a). LMI has higher held-out EA than UMI in 8 of the 10 settings (by 1.5 to 18.3 points).

Table R1-9: Results under lighter fine-tuning. FT Tr / FT EA: Tr and EA after fine-tuning alone. UMI and LMI cells: held-out EA / Tr.

SettingFT TrFT EAUMILMIμ*
GPT-J MEMIT, ep 142.717.060.2±3.6/37.677.2±1.6/36.50
GPT-J MEMIT, ep 395.96.167.0±3.5/73.874.0±0.1/78.15
GPT-J MEMIT, half lr98.213.371.6±0.1/82.471.4±1.5/72.614
Llama-3 MEMIT, ep 157.815.963.0±2.0/48.167.4±0.2/58.56
Llama-3 MEMIT, ep 397.62.052.8±0.1/74.754.3±0.2/89.27
Llama-3 MEMIT, half lr100.06.066.0±0.1/93.175.2±0.2/89.08
GPT-J AlphaEdit, ep 133.77.435.1±3.4/23.450.7±0.1/25.30
GPT-J AlphaEdit, ep 394.72.340.4±0.1/59.734.9±1.4/64.44
Llama-3 AlphaEdit, ep 158.523.156.3±1.8/53.774.6±0.2/52.44
Llama-3 AlphaEdit, ep 397.81.748.6±0.1/86.360.5±0.1/89.57

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.