MobiEdit: Resource-efficient Knowledge Editing for Personalized On-device LLMs

Daliang Xu; Dongqi Cai; Fangming Liu; Mengwei Xu; Shangguang Wang; Wei Liu; Zexi Li; Zhenyan Lu

REVIEW 4 major objections 8 minor 22 references

MobiEdit: Resource-efficient Knowledge Editing for Personalized On-device LLMs

T0 review · 4 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read MobiEdit claims the first practical on-device knowledge-editing framework for 3B-parameter LLMs, replacing backpropagation with quantized forward-only gradient estimation.

desk verdict MobiEdit is a credible engineering attempt at on-device knowledge editing, but its headline efficiency numbers are inflated by a CPU-versus-NPU baseline confound and an unsupported latency ratio. read the letter →

arxiv 2506.13772 v1 pith:PKQ5ZMGZ submitted 2025-06-05 cs.LG cs.AI

Zhenyan Lu , Daliang Xu , Dongqi Cai , Zexi Li , Wei Liu , Fangming Liu , Shangguang Wang , Mengwei Xu This is my paper

classification cs.LGcs.AI

keywords knowledgeeditingon-deviceLLMpersonalizationforward-onlygradientestimationzeroth-orderoptimizationmobileNPUquantizedlocate-and-editresourceefficiency

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that personalizing an LLM to a user's facts can be done entirely on a commercial phone if the editing update no longer needs backpropagation. The authors replace the training-style gradient with a central-difference estimate taken from forward passes, keep only a tiny fraction of weights in floating point, and add early stopping plus a prefix cache to trim the extra steps that forward-only optimization requires. On a 3B-parameter Qwen2.5 model they report 7.6x less memory, 14.7x less energy, and 3.6x lower latency than prior locate-and-edit methods, at the cost of a roughly 14-point drop in edit success. If correct, this makes local, private knowledge editing feasible for mobile assistants rather than a cloud-only operation.

What carries the argument

The central mechanism is the central-difference directional gradient estimator: with a perturbation direction $u \sim \mathcal{N}(0, I)$, the gradient is estimated as $(L(v + \mu u) - L(v - \mu u))/(2\mu) \cdot u$, averaged over $N$ random directions. This replaces backpropagation with forward passes, so activations can be discarded immediately and mobile NPUs, which only accelerate inference, can execute the update. The ROME-style rank-one update $W + \Lambda(C^{-1}k^*)^\top$ then inserts the edited association, while the early-stopping controller and the prefix cache reduce the large number of forward passes that zeroth-order optimization would otherwise require.

What would settle it

A controlled benchmark on the same three phones, running ROME, MEMIT, AlphaEdit, and WISE with a memory-optimized runtime and the same W8A16 quantization and stopping setup as MobiEdit; if any baseline matches or beats 6.2GB memory, roughly 0.02J energy, and roughly 1200 seconds per edit while preserving its own edit quality, the paper's central resource-efficiency claim fails.

Watch

Extended reading notes

Core claim

The paper's discovery is that backpropagation is not a necessary ingredient for effective locate-and-edit knowledge editing on resource-constrained hardware. MobiEdit keeps the ROME-style key-value memory view of an MLP layer and the closed-form rank-one update, but obtains the value vector by minimizing the editing loss with a forward-only gradient estimator, quantizes all layers except the edited one and its predecessor, and demonstrates a 3B model being edited on three commercial phones with 6.2GB memory, roughly 25 minutes of latency, and under 0.03J of energy per edit. The paper also argues that this estimator's gradient noise stays bounded with model depth under quantization, in contrast to backpropagation's multiplicative noise amplification along the chain rule.

Load-bearing premise

The paper's Section 3.2 efficiency numbers treat the CPU-based baseline runs, which it notes lack training-side memory optimization, as representative prior-method costs; if those methods were ported to the same NPU stack and quantization, the 7.6x, 14.7x, and 3.6x gaps could shrink substantially.

Editorial extensions

If this is right

A 3B-parameter LLM can be edited on a 16GB phone instead of needing over 46GB of memory, so on-device personalization no longer requires shipping user data to the cloud.
Knowledge editing becomes an NPU-compatible workload: the update runs as forward passes, which are the only operations mobile NPUs accelerate efficiently.
Per-edit energy below 0.03J means edits could run in the background without thermal throttling or making the phone unusable for the 1.5-to-3-hour stretches reported for baselines.
The paper's own limitations section restricts the method to simple subject-object factual pairs and reports lower accuracy on ambiguous or multi-hop facts, so the resource gains do not yet cover the full range of personalization tasks.
The theoretical variance bound predicts that forward-only quantized editing remains stable in deeper or lower-bit networks, which is the paper's basis for scaling the approach beyond 3B models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper does not pursue this, but the same forward-only, quantized update loop could be applied to other on-device adaptation tasks, such as continual fine-tuning or federated personalization, where backpropagation is equally blocked by NPU hardware.
The paper's comparison uses CPU-based baseline implementations; if ROME, MEMIT, AlphaEdit, and WISE were reimplemented on the same NPU stack, the reported 7.6x, 14.7x, and 3.6x resource reductions would likely shrink, though the core feasibility result might still stand.
The prefix-cache experiment suggests activations stay stable under small parameter updates, implying that a more aggressive cache-invalidation policy than the fixed loss-drop heuristic could be safe; the authors did not test that variant.
A missing comparison a reader might want is the same editing task performed with retrieval-augmented prompting on the same phone, to see whether parameter editing is worth its complexity versus simply injecting user context into the prompt.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Punchline: this is a real systems contribution—forward-only, quantized knowledge editing on mobile NPUs—but the headline numbers don't survive a close read. The baseline comparison is apples-to-oranges, and the paper's own Table 2 doesn't support the advertised 3.6x latency. Still, it deserves a serious referee.

What's genuinely useful: the authors combine MeZO-style zeroth-order gradients with ROME's editing objective, run it on three commercial phones with Qwen2.5-3B, and show that early stopping and prefix cache each help. The mixed-precision design—keeping the editing layer and its preceding layer in float while quantizing the rest—is pragmatic and well-motivated. The ablation in Figure 6 is honest about the edit-success/time trade-off. They also state limitations directly. The related work and self-citations are fine; FwdLLM is genuinely prior work on forward-only fine-tuning.

The soft spots are real. First, the baseline comparison is confounded: ROME/MEMIT/AlphaEdit/WISE run on CPU via llm.c, which the paper itself says 'lacks memory optimization on training part of parameters.' That explains the 46GB and multi-hour per-edit costs. Porting those methods to the same NPU/quantization stack would likely shrink the claimed 7.6x memory and 14.7x energy margins substantially. Second, Table 2 doesn't support the advertised 3.6x latency: the largest ratio I see is about 2.9x (K70, ZsRE), and most are between 2.2x and 2.7x. The abstract and conclusion overstate the result. Third, the quantization-noise variance argument in Section 2.2 is internally muddled: they define sigma_L^2 as per-pass output noise, then claim the centered difference variance is independent of depth. But sigma_L^2 is the accumulated output noise, which itself grows with L. The depth-independence claim doesn't follow. Fourth, key hyperparameters—step size mu, number of directions N, early stopping threshold m—are omitted, which hurts reproducibility. Fifth, calling 20-30 minutes per edit 'real-time' is a stretch; it's fine as a background task, but not interactive.

Who it's for: systems/ML folks working on on-device LLM adaptation. The idea is sensible and the engineering is non-trivial. It deserves peer review, but I'd want to see the baseline comparison redone on equal footing, hyperparameters disclosed, and the latency claim corrected before acceptance. A conditional accept with those requirements seems right.

Referee Report

4 major / 8 minor

Summary. The paper presents MobiEdit, a knowledge-editing framework for on-device LLMs that replaces backpropagation with forward-only zeroth-order gradient estimation, quantizes most weights to W8A16 while keeping the edited layer and its predecessor in floating point, and adds an early-stopping controller and a prefix cache. The proposed system is evaluated on Qwen2.5-3B-Instruct on three smartphones (Xiaomi K60 Pro, K70, OnePlus 13) against ROME, MEMIT, AlphaEdit, and WISE, reporting roughly 7.6x lower memory, 14.7x lower energy, and 3.6x lower latency while retaining competitive editing quality on ZsRE and CounterFact. The main algorithmic components are clearly grounded in ROME's locate-and-edit formulation and MeZO-style zeroth-order optimization, so the method itself is coherent.

Significance. MobiEdit attacks a real deployment problem: existing locate-and-edit methods require backpropagation and are therefore hard to run on mobile NPUs. The combination of forward-only gradient estimation with static W8A16 quantization and mixed-precision execution is a sensible and nontrivial systems contribution, and the paper includes real-device measurements and ablations showing that the early-stopping controller and prefix cache each reduce editing time. If the efficiency numbers survive a controlled comparison, this would be a practical step toward private, on-device personalization of 3B-class LLMs. However, because the headline efficiency ratios are currently based on a confounded baseline (CPU/llm.c vs NPU/mllm-npu) and are inconsistent with the numbers in Table 2, the quantitative contribution is not yet established. The paper is not circular; it builds on external work (ROME, MeZO, FwdLLM) and does not assume its conclusions.

major comments (4)

[Section 3.2, Table 2] The central efficiency claim is not supported by a controlled comparison. In Section 3.2 and Table 2, ROME, MEMIT, AlphaEdit, and WISE are run on CPUs with llm.c, which the paper itself notes 'lacks memory optimization on training part of parameters,' while MobiEdit runs on NPUs with mllm-npu under W8A16 quantization. The reported 46GB memory and 3,200-11,359s latencies therefore conflate the editing algorithm with the choice of an unoptimized CPU training stack, and the 7.6x/14.7x/3.6x ratios cannot be attributed to MobiEdit's algorithmic design. To establish the headline claim, the authors should port the baselines to the same NPU/mixed-precision stack (or at least to a memory-optimized CPU training implementation), and report time, memory, and energy with and without quantization separately.
[Abstract and Table 2] The advertised 3.6x latency reduction is contradicted by the paper's own data. From Table 2, ZsRE on K60 gives 4543.78/1902.88 = 2.39x, K70 gives 4276.49/1477.67 = 2.89x, OnePlus 13 gives 3252.81/1211.83 = 2.68x, and the CounterFact ratios are similar; no device/dataset combination approaches 3.6x, and the conclusion's 'latency by 72%' is also not derivable from the table. The abstract, introduction, and conclusion should either be corrected to match the measurements or the calculation behind the 3.6x factor should be reported explicitly.
[Section 2.2, Eq. (12)] The quantization-noise robustness argument is internally inconsistent. Eq. (8) states that the output noise accumulates over layers, with total noise growing linearly or even exponentially with L; Eq. (12) then denotes the per-pass output noise variance by sigma_L^2 and concludes that the centered-difference estimator's variance 'does not grow with network depth L.' But if sigma_L^2 is the variance of the accumulated output noise from Eq. (8), it is not depth-independent, and the conclusion does not follow. The authors should either define sigma_L^2 as a depth-independent per-perturbation noise and justify that, or remove the 'independent of depth' claim and replace it with a more careful (likely empirical) comparison of the two estimators.
[Sections 2.3 and 3.1] Key hyperparameters are missing, which prevents reproducibility and obscures the efficiency trade-offs: the step size mu and number of sampled directions N in Eqs. (4)-(5), the early-stopping interval M and confidence threshold m in Section 2.3, and the prefix-cache staleness criterion ('loss does not decrease by 0.001 over 3 steps') are never given concrete values in the evaluation setup. Report the exact values used, and ideally a sensitivity analysis for N and mu, since the claimed speedups depend on how many forward passes per step are amortized.

minor comments (8)

[Abstract] The sentence 'MobiEdit replaces full-precision backpropagation with quantized forward-only gradient estimation...' appears twice in the abstract; remove the duplicate.
[Throughout] There are several typos: 'stoping' -> 'stopping', 'COST' -> 'COTS' in Section 3.1, 'AhphaEdit' -> 'AlphaEdit' in Table 2, 'eﬁcieny' -> 'efficiency' in Section 3.2, 'papaer' -> 'paper' in Section 3.1, and 'zhof' in Section 3.3.
[Figure 5 and Section 3.1] The description of the 'system efficiency' metric is too vague: it says values are normalized to [40, 100] using min-max normalization and then inverted, but it does not state the direction of each raw metric or how the six dimensions are aggregated; clarify this so the reader can interpret the radar plot.
[Section 3.2] The phrase 'real-time editing' is inconsistent with the reported 1,200-2,000 seconds per edit; replace it with 'feasible on-device' or 'within a few tens of minutes' to avoid overclaiming.
[Section 3.2] The sentence 'Although slightly behind in quality (13.9 reduction in edit success loss compared to MEMIT and AlphaEdit)' is unclear: specify whether 13.9 is a percentage-point drop in edit success and report the exact numeric values for the three quality metrics.
[Table 2 and Figure 5] No variance estimates or repeated-run statistics are reported; given that the efficiency ratios are the main empirical contribution, please report standard deviations across at least three runs or seeds.
[References] Reference [9] lists llm.c with year 2013; the URL indicates a GitHub repository by karpathy that is actively maintained, so the year and citation should be updated.
[Section 2.3, Figure 4] The prefix-cache justification relies on cosine similarity values, but the actual editing quality with and without the cache is only shown indirectly in Figure 6; report the final edit success/locality/portability numbers for both configurations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MobiEdit composes external building blocks (ROME-style editing and MeZO-style forward-only gradients) and evaluates them on-device; the headline efficiency ratios are confounded by an unequal hardware/implementation comparison, but no derived quantity is equivalent to its inputs by construction.

full rationale

I walked the claimed derivation chain. The core editing objective (Eq. 3) and the rank-one update (Eq. 6) are taken from the external ROME framework, not derived from MobiEdit's own conclusions. The forward-only gradient estimator (Eqs. 4-5) is the standard central-difference zeroth-order estimator, explicitly cited to Baydin et al. and to MeZO; MobiEdit does not claim to derive it, nor does it define it in terms of its own efficiency targets. The quantization-robustness argument (Eqs. 7-12) is a self-contained variance comparison: it assumes a quantized network noise model and then compares the multiplicative noise growth of backpropagation with the depth-independent variance of a two-sided forward difference. Whether or not that model is accurate, the conclusion follows from the stated equations rather than being assumed. The prefix cache and early-stopping optimizations are presented as empirical system improvements, validated by ablations (Figure 6) and by the cosine-similarity check (Figure 4); they are not fitted parameters renamed as predictions. The self-citations to FwdLLM, mllm-npu, and the authors' small-language-model survey support the implementation stack and related work, but none is invoked as a uniqueness theorem or as the sole justification for a central claim. The most serious concern is the efficiency comparison in Section 3.2 / Table 2: baseline ROME/MEMIT/AlphaEdit/WISE run on CPU through llm.c, which the paper itself says 'lacks memory optimization on training part of parameters,' while MobiEdit runs on NPU with W8A16 quantization via mllm-npu. This is a fairness and validity problem for the 46 GB / multi-hour baseline numbers, and the advertised 3.6x latency is not supported by Table 2's raw ratios. However, that is an experimental-control issue, not circularity: no equation or fitted parameter makes MobiEdit's resource numbers equal to its inputs by construction. Therefore the circularity score is 0, with the caveat that the headline efficiency gain is not yet established by a controlled comparison.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim depends on several free hyperparameters that are not disclosed, and on assumptions inherited from ROME and MeZO. The quantization noise model is internally inconsistent, so the theoretical support for robustness is not reliable.

free parameters (5)

step size mu
The perturbation size for central-difference gradient estimation in Eq. (4), not specified in the paper.
number of sampled directions N
Number of perturbation directions averaged in Eq. (5), not specified.
early stopping interval M
Editing is evaluated every M steps (Section 2.3); no concrete value given.
success confidence threshold m
The early stopping success criterion references a threshold m, but the paper leaves an unfinished note ('and explicitly describe the threshold we used in the eval setup') without providing it.
prefix cache staleness threshold = loss decrease < 0.001 over 3 steps
The rule to recompute the prefix cache is given in Section 2.3, but the 0.001 and 3-step values are hand-chosen without sensitivity analysis.

assumptions (5)

domain assumption MLP layers in transformers act as key-value memory (ROME)
Section 2.1 relies on this interpretation to justify the rank-one update.
domain assumption The averaged activation k* from sampled prompts is a valid key
Equation (2) estimates the key from a set of random prompts; this comes from ROME, not re-derived.
domain assumption Zeroth-order gradient estimation converges for the editing loss
The whole method depends on central-difference gradients being reliable, which is not proven for this loss and model.
ad hoc to paper Stale prefix cache activations remain valid
Section 2.3 empirically justifies reusing activations from the first step, but it is an assumption that the activation shift is small.
domain assumption Quantization noise is i.i.d. and independent across passes
Section 2.2 assumes zero-mean i.i.d. errors for the variance analysis; real quantization noise is input-dependent.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MobiEdit: Resource-efficient Knowledge Editing for Personalized On-device LLMs." pith.science (2026). https://pith.science/paper/PKQ5ZMGZ

@misc{pith2026250613772,
  author       = {Pith},
  title        = {Pith review of: MobiEdit: Resource-efficient Knowledge Editing for Personalized On-device LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PKQ5ZMGZ}},
  note         = {Machine review of arXiv:2506.13772}
}

abstract

Large language models (LLMs) are deployed on mobile devices to power killer applications such as intelligent assistants. LLMs pre-trained on general corpora often hallucinate when handling personalized or unseen queries, leading to incorrect or outdated responses. Knowledge editing addresses this by identifying and adjusting a small crucial portion of model weights, without compromising the general knowledge. However, prior knowledge editing methods are impractical to run on local devices due to the resource-heavy backpropagation (BP) needed for updates. We present MobiEdit, the first mobile knowledge editing framework that enables efficient LLM personalization on commercial off-the-shelf (COTS) mobile devices. MobiEdit replaces full-precision BP with quantized forward-only gradient estimation, thus compatible with the energy-efficient mobile neural processing units (NPUs). MobiEdit replaces full-precision backpropagation with quantized forward-only gradient estimation, making it compatible with energy-efficient mobile NPUs. To further improve gradient estimation efficiency, we introduce two optimizations: an early stoping mechanism that adaptively terminates editing upon success and a prefix cache that reuses computation across steps. Our approach enables real-time editing of a 3B-parameter model (Qwen2.5-3B-Instruct) on COTS mobile devices with 7.6$\times$ less memory, 14.7 $\times$ less energy and 3.6$\times$ less latency compared to previous knowledge editing methods.

Figures

Figures reproduced from arXiv: 2506.13772 by the authors.

**Figure 2.** MobiEdit quantization workflow and strategy. MobiEdit quantizes all activation functions, with only the editing layer and its preceding layer executed in floating-point format. Quantization workflow and strategy [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** The edit success step number. Early stopping controller To address this, we first analysis the successful editing step distribution of various knowledge. As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: The cosine similarity of QKV representations at ea [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: The comprehensive performance comparison of know [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Edit success vs. time on ZsRE. Time is averaged across all devices [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 13 canonical work pages

[1]

Apple intelligence

Apple. Apple intelligence. https://www.apple.com/ap ple-intelligence/, 2024. URL https://www.apple.com/apple-intelligence/

work page 2024
[2]

Gradi- ents without backpropagation

Atılım Güne¸ s Baydin, Barak A Pearlmutter, Don Syme, Fra nk Wood, and Philip Torr. Gradi- ents without backpropagation. arXiv preprint arXiv:2202.08587, 2022

arXiv 2022
[4]

Knowl- edge neurons in pretrained transformers

Damai Dai, Li Dong, Y aru Hao, Zhifang Sui, Baobao Chang, a nd Furu Wei. Knowl- edge neurons in pretrained transformers. In Smaranda Mures an, Preslav Nakov, and Aline Villavicencio, editors, Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 8493–8502. Associa- tion for Computational Li...

doi:10.18653/v 2022
[5]

Optimal rates for zero-order convex optimization: The power of two function e valuations

John C Duchi, Michael I Jordan, Martin J Wainwright, and A ndre Wibisono. Optimal rates for zero-order convex optimization: The power of two function e valuations. IEEE Transactions on Information Theory, 61(5):2788–2806, 2015

work page 2015
[6]

A survey on rag meeting llms: Towards retri eval-augmented large language models

Wenqi Fan, Y ujuan Ding, Liangbo Ning, Shijie Wang, Hengy un Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meeting llms: Towards retri eval-augmented large language models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD ’24, page 6491–6501, New Y ork, NY , USA, 2024. As- sociation for Computing Ma...

arXiv 2024
[7]

Alphaedit: Null-space constrained m odel editing for language mod- els

Junfeng Fang, Houcheng Jiang, Kun Wang, Y unshan Ma, Jie S hi, Xiang Wang, Xiangnan He, and Tat-Seng Chua. Alphaedit: Null-space constrained m odel editing for language mod- els. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=HvSytvg3Jh

work page 2025
[8]

Model editing harms general abilities of large language models: Regularization to the rescue

Jia-Chen Gu, Hao-Xiang Xu, Jun-Y u Ma, Pan Lu, Zhen-Hua Li ng, Kai-Wei Chang, and Nanyun Peng. Model editing harms general abilities of large language models: Regularization to the rescue. In Y aser Al-Onaizan, Mohit Bansal, and Y un-Nu ng Chen, editors, Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 16801– 1...

doi:10.18653/v1/2024.emnlp-main 2024
[9]

karpathy. llm.c. https://github.com/karpathy/llm.c , 2013

work page 2013

Show all 22 references

[10]

Zero-shot relation extraction via reading comprehension

Omer Levy, Minjoon Seo, Eunsol Choi, and Luke Zettlemoy er. Zero-shot relation extraction via reading comprehension. In Roger Levy and Lucia Specia, e ditors, Proceedings of the 21st 10 Conference on Computational Natural Language Learning (CoNLL 2017), pages 333–342, V ancouv...

2017
[11]

Personal llm agents: In sights and survey about the capability, efﬁciency and security

Y uanchun Li, Hao Wen, Weijun Wang, Xiangyu Li, Yizhen Y uan, Guohong Liu, Jiacheng Liu, Wenxing Xu, Xiang Wang, Yi Sun, et al. Personal llm agents: In sights and survey about the capability, efﬁciency and security. arXiv preprint arXiv:2401.05459, 2024

2024 arXiv
[12]

Small language models: Survey, measur ements, and insights

Zhenyan Lu, Xiang Li, Dongqi Cai, Rongjie Yi, Fangming L iu, Xiwen Zhang, Nicholas D Lane, and Mengwei Xu. Small language models: Survey, measur ements, and insights. arXiv preprint arXiv:2409.15790, 2024

2024 arXiv
[13]

Lee, Danqi Chen, and Sanjeev Arora

Sadhika Malladi, Tianyu Gao, Eshaan Nichani, Alex Dami an, Jason D. Lee, Danqi Chen, and Sanjeev Arora. Fine-tuning language models with j ust forward passes. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=Vota6rFhBQ

2023
[14]

Locating and editing factual associations in gpt

Kevin Meng, David Bau, Alex Andonian, and Y onatan Belin kov. Locating and editing factual associations in gpt. Advances in neural information processing systems, 35:17359–17372, 2022

2022
[15]

Mass- editing memory in a transformer

Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Y onatan B elinkov, and David Bau. Mass- editing memory in a transformer. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=MkbcAHIYgyS

2023
[16]

Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. Fast model editing at scale. CoRR, 2021. URL https://arxiv.org/pdf/2110.11309.pdf

2021 arXiv
[17]

Qwen2.5: A party of foundation models, Septe mber 2024

Qwen Team. Qwen2.5: A party of foundation models, Septe mber 2024. URL https://qwenlm.github.io/blog/qwen2.5/

2024
[18]

Wise: Rethinking the knowledge memo ry for lifelong model editing of large language models

Peng Wang, Zexi Li, Ningyu Zhang, Ziwen Xu, Y unzhi Y ao, Y ong Jiang, Pengjun Xie, Fei Huang, and Huajun Chen. Wise: Rethinking the knowledge memo ry for lifelong model editing of large language models. Advances in Neural Information Processing Systems, 37:53764– 53797, 2024

2024
[19]

Knowl- edge editing for large language models: A survey

Song Wang, Y aochen Zhu, Haochen Liu, Zaiyi Zheng, Chen C hen, and Jundong Li. Knowl- edge editing for large language models: A survey. ACM Comput. Surv., 57(3), November 2024. ISSN 0360-0300. doi: 10.1145/3698590. URL https://doi.org/10.1145/3698590

2024 doi
[20]

Fast on-device llm inference with npus

Daliang Xu, Hao Zhang, Liming Y ang, Ruiqi Liu, Gang Huan g, Mengwei Xu, and Xuanzhe Liu. Fast on-device llm inference with npus. In Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, V olume1, pages 4...

2025
[21]

{FwdLLM}: Efﬁcient federated ﬁnetuning of large language models with perturbed inferences

Mengwei Xu, Dongqi Cai, Y aozong Wu, Xiang Li, and Shangg uang Wang. {FwdLLM}: Efﬁcient federated ﬁnetuning of large language models with perturbed inferences. In 2024 USENIX Annual Technical Conference (USENIX A TC24), pages 579–596, 2024

2024
[22]

MQuAKE: Assessing knowledge editing in language models via multi-hop questions

Zexuan Zhong, Zhengxuan Wu, Christopher Manning, Chri stopher Potts, and Danqi Chen. MQuAKE: Assessing knowledge editing in language models via multi-hop questions. In Houda Bouamor, Juan Pino, and Kalika Bali, editors, Proceedings of the 2023 Conference on Empirical Methods i...

2023 doi
[934]

URL https://aclanthology.org/2024.emnlp-main.934/

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Apple intelligence

Apple. Apple intelligence. https://www.apple.com/ap ple-intelligence/, 2024. URL https://www.apple.com/apple-intelligence/

work page 2024

[2] [2]

Gradi- ents without backpropagation

Atılım Güne¸ s Baydin, Barak A Pearlmutter, Don Syme, Fra nk Wood, and Philip Torr. Gradi- ents without backpropagation. arXiv preprint arXiv:2202.08587, 2022

arXiv 2022

[3] [4]

Knowl- edge neurons in pretrained transformers

Damai Dai, Li Dong, Y aru Hao, Zhifang Sui, Baobao Chang, a nd Furu Wei. Knowl- edge neurons in pretrained transformers. In Smaranda Mures an, Preslav Nakov, and Aline Villavicencio, editors, Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 8493–8502. Associa- tion for Computational Li...

doi:10.18653/v 2022

[4] [5]

Optimal rates for zero-order convex optimization: The power of two function e valuations

John C Duchi, Michael I Jordan, Martin J Wainwright, and A ndre Wibisono. Optimal rates for zero-order convex optimization: The power of two function e valuations. IEEE Transactions on Information Theory, 61(5):2788–2806, 2015

work page 2015

[5] [6]

A survey on rag meeting llms: Towards retri eval-augmented large language models

Wenqi Fan, Y ujuan Ding, Liangbo Ning, Shijie Wang, Hengy un Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meeting llms: Towards retri eval-augmented large language models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD ’24, page 6491–6501, New Y ork, NY , USA, 2024. As- sociation for Computing Ma...

arXiv 2024

[6] [7]

Alphaedit: Null-space constrained m odel editing for language mod- els

Junfeng Fang, Houcheng Jiang, Kun Wang, Y unshan Ma, Jie S hi, Xiang Wang, Xiangnan He, and Tat-Seng Chua. Alphaedit: Null-space constrained m odel editing for language mod- els. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=HvSytvg3Jh

work page 2025

[7] [8]

Model editing harms general abilities of large language models: Regularization to the rescue

Jia-Chen Gu, Hao-Xiang Xu, Jun-Y u Ma, Pan Lu, Zhen-Hua Li ng, Kai-Wei Chang, and Nanyun Peng. Model editing harms general abilities of large language models: Regularization to the rescue. In Y aser Al-Onaizan, Mohit Bansal, and Y un-Nu ng Chen, editors, Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 16801– 1...

doi:10.18653/v1/2024.emnlp-main 2024

[8] [9]

karpathy. llm.c. https://github.com/karpathy/llm.c , 2013

work page 2013