Interpretability-Guided Layer Selection over Subspace Projection: SAEs as Stethoscopes, Not Scalpels, for Raw Task Vector Model Editing

Li Lei; Madalina Ciobanu; Qingqing Mao; Ritankar Das

arxiv: 2605.28649 · v1 · pith:VYEFBAQ6new · submitted 2026-05-27 · 💻 cs.LG · cs.CL

Interpretability-Guided Layer Selection over Subspace Projection: SAEs as Stethoscopes, Not Scalpels, for Raw Task Vector Model Editing

Li Lei , Madalina Ciobanu , Qingqing Mao , Ritankar Das This is my paper

Pith reviewed 2026-06-29 13:57 UTC · model grok-4.3

classification 💻 cs.LG cs.CL

keywords sparse autoencodersmodel editingtask vectorsinterpretabilitymathematical reasoninglayer selectionGemma model

0 comments

The pith

Using SAEs only to select layers for raw task vector injection improves math reasoning accuracy.

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

The paper shows that projecting task vectors onto SAE feature subspaces creates an information bottleneck by discarding roughly 97 percent of the modification energy due to misalignment between activation-space directions and weight-space vectors, yielding no gains. It shifts to treating SAEs as diagnostic tools that compute a layer specificity score and then applies the complete unfiltered task vectors only to the chosen layers. This produces a statistically significant rise in number theory accuracy from 29.6 percent to 39.4 percent on the Minerva Math benchmark, with five of seven math subjects improving and none degrading.

Core claim

SAEs function as stethoscopes that identify the right layers rather than scalpels that filter task vectors; injecting the raw vectors into those layers alone produces net gains in mathematical reasoning without the energy loss of subspace projection.

What carries the argument

The SAE-derived specificity score, which ranks layers by how well they align with the task for unfiltered vector injection.

If this is right

Five of seven math subjects show statistically significant gains.
No subject experiences significant degradation.
The procedure adds no inference cost and remains fully deterministic.

Where Pith is reading between the lines

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

The same layer-selection logic could be tested on non-mathematical editing tasks to check generality.
The reported geometric misalignment between SAE directions and task vectors may limit other subspace-based editing approaches.
Combining the specificity score with other layer-ranking methods might further refine the selection.

Load-bearing premise

The specificity score from the SAE correctly flags layers where the raw task vector will improve the target capability without causing degradation or forgetting elsewhere.

What would settle it

Re-running the Minerva Math evaluation on Gemma-3-4B-IT and finding no accuracy change or a drop when task vectors are injected only into the SAE-selected layers would disprove the central result.

Figures

Figures reproduced from arXiv: 2605.28649 by Li Lei, Madalina Ciobanu, Qingqing Mao, Ritankar Das.

**Figure 2.** Figure 2: Per-layer Number Theory specificity (Gemma Scope 2, 16K features). [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Per-subject accuracy across the seven Minerva Math subjects ( [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Method evolution: NT z-score across six approaches in order of development. PPL+CMAES with uniform α (orange) and both SAE Projection variants (green) remain below the p < 0.05 threshold (z = 1.96, red dashed line) and p < 0.01 threshold (z = 2.58, red dotted line). All three Raw Task Vector configurations (blue) cross into significance, with our final SP4 14L α=0.80 reaching z = +3.41. The decisive trans… view at source ↗

**Figure 5.** Figure 5: Energy retention vs. Number Theory significance ( [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

**Figure 6.** Figure 6: Number Theory z-score as a function of α for the SP4 14L configuration (n = 540 NT problems). The green shaded region marks the robust plateau (z ≥ 2.84, α ∈ [0.70, 1.10]). The optimal α ∗ = 0.80 (z = +3.41) is annotated in red. Two dotted reference lines indicate the p < 0.05 (z = 1.96) and p < 0.005 (z = 2.58) thresholds. Performance drops sharply only at α = 1.20, marking the onset of over-modification.… view at source ↗

**Figure 7.** Figure 7: Alpha response surfaces for SP4 14L (blue circles, 14 layers) and SP4 noDeep (red squares, [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗

**Figure 8.** Figure 8: shows per-subject z-scores for the six representative layer-selection configurations referenced in Section 5.3 [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗

**Figure 9.** Figure 9: PPL fitness vs. accuracy z-score for the top-3 CMA-ES configurations in the CP search (left) and NT search (right). Each point is one configuration ranked by PPL fitness (lower PPL = “better” by the fitness signal). Red dashed lines show the Pearson regression fit. CP: r = −0.06 (essentially uncorrelated); NT: r = +0.14 (slightly positive, opposite to the anti-correlation the fitness signal assumes). In bo… view at source ↗

read the original abstract

LLMs increasingly require surgical model editing to enhance domain-specific capabilities without incurring the computational cost or catastrophic forgetting associated with full fine-tuning. Sparse Autoencoders (SAEs) have emerged as a promising tool in this setting, in principle allowing for feature-level identification of where to intervene. In this work, we rigorously evaluate an SAE-guided editing pipeline for mathematical reasoning on Gemma-3-4B-IT and uncover a fundamental failure mode: the intuitively appealing approach of projecting task vectors onto SAE feature subspaces acts as an information bottleneck that discards approximately 97% of the modification energy, yielding no statistically significant improvements across seven math subjects. We show that this failure stems from a geometric misalignment between activation-space SAE directions and weight-space task vectors. We then propose a shift in perspective: SAE as a Stethoscope, Not a Scalpel, where SAEs are used for layer-level diagnosis rather than intervention-level filtering. By injecting unfiltered raw task vectors only into layers identified by an SAE-derived specificity score, we improve Number Theory accuracy from 29.6% to 39.4% (z=+3.41, p=0.0007) on the Minerva Math benchmark; 5 of 7 math subjects significantly improved and none significantly degraded. Our method is fully deterministic, requires no additional inference cost, and provides a principled framework for interpretability-guided model editing.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SAE layer selection for raw task vectors gives measurable math gains, but the paper skips the obvious control of whether any selective editing would do the same.

read the letter

The main result is that projecting task vectors onto SAE subspaces discards 97% of the energy due to misalignment and produces no gains, while using the SAE only to pick layers for the full raw vector lifts Number Theory from 29.6% to 39.4% on Minerva Math with p=0.0007 and improves five of seven subjects without degradation.

The geometric diagnosis is the clearest new piece. It explains why the subspace approach fails in a way that prior editing papers did not flag. The stethoscope framing follows directly and is a clean way to describe the shift from feature filtering to layer diagnosis. The reported numbers are specific, include a z-score, and the method adds no inference cost.

The soft spot is exactly the one in the stress test. The paper shows gains from editing the SAE-chosen layers but does not test the same raw vectors on the same number of layers chosen at random, by depth, or by activation size. Without that comparison, the improvement cannot be credited to the specificity score rather than to the general practice of editing only a subset of layers. The abstract also leaves the exact task-vector construction and run count implicit, though the held-out benchmark setup itself is reasonable.

This is for groups working on efficient LLM adaptation and mechanistic editing. It has enough concrete evidence and a clear negative result to justify referee time, even if the layer-selection claim needs tightening.

Referee Report

1 major / 2 minor

Summary. The paper evaluates SAE-based model editing for mathematical reasoning on Gemma-3-4B-IT. It reports that projecting task vectors onto SAE feature subspaces discards ~97% of modification energy and produces no significant gains across seven math subjects. It then shows that using an SAE-derived specificity score solely to select layers for injecting unfiltered raw task vectors improves Number Theory accuracy from 29.6% to 39.4% (z=+3.41, p=0.0007) on Minerva Math, with significant gains in 5 of 7 subjects and none degraded. The method is presented as deterministic and inference-free.

Significance. If the central empirical result holds after controls, the work supplies concrete evidence that interpretability tools can guide layer selection for task-vector editing without the energy-loss bottleneck of feature projection, yielding measurable accuracy gains on held-out math benchmarks with reported p-values and an energy-loss quantification.

major comments (1)

[Results on Minerva Math benchmark (as described in abstract and experimental evaluation)] The central claim attributes performance gains to the SAE specificity score for layer selection, yet the manuscript reports results only for the SAE-chosen layers and does not include ablations that apply the identical raw task vectors to the same number of layers chosen at random, by depth, or by activation magnitude. Without these controls, the improvement cannot be distinguished from the general benefit of selective (rather than full-model) editing.

minor comments (2)

[Abstract] The abstract states concrete accuracy gains and p-values but provides no details on the number of runs, exact task-vector construction procedure, or baseline controls, which limits immediate verification.
Notation for the specificity score and its computation from SAE activations should be defined explicitly with an equation or pseudocode to allow reproduction.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on the need for controls to isolate the contribution of the SAE specificity score. We address the major comment below.

read point-by-point responses

Referee: [Results on Minerva Math benchmark (as described in abstract and experimental evaluation)] The central claim attributes performance gains to the SAE specificity score for layer selection, yet the manuscript reports results only for the SAE-chosen layers and does not include ablations that apply the identical raw task vectors to the same number of layers chosen at random, by depth, or by activation magnitude. Without these controls, the improvement cannot be distinguished from the general benefit of selective (rather than full-model) editing.

Authors: We agree that the manuscript currently lacks these ablations, which are required to establish that gains arise specifically from the SAE-derived layer selection rather than from selective editing in general. In the revised manuscript we will add the requested controls using the identical raw task vectors: random selection of the same number of layers, depth-based selection (e.g., earliest or latest layers), and selection by activation magnitude. Results will be reported on the Minerva Math benchmark with the same statistical tests to allow direct comparison. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The manuscript reports an empirical pipeline in which an SAE-derived layer specificity score is used to select injection sites for raw task vectors, with performance gains measured on held-out Minerva Math benchmarks (e.g., Number Theory 29.6% → 39.4%). No equations, fitted parameters, or self-citations are shown to define the target metric by construction; the central result is an external benchmark comparison rather than a quantity that reduces to its own inputs. The derivation chain therefore remains self-contained against external measurements.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the specificity score and task-vector construction are not detailed enough to classify.

pith-pipeline@v0.9.1-grok · 5801 in / 1057 out tokens · 35284 ms · 2026-06-29T13:57:30.894416+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

26 extracted references · 11 canonical work pages · 4 internal anchors

[1]

LoRA: Low-rank adaptation of large language models

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shengyu Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022

2022
[2]

Editing models with task arithmetic

Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. InInternational Conference on Learning Representations (ICLR), 2023

2023
[3]

Towards monosemanticity: Decomposing language models with dictionary learning.Transformer Circuits Thread, 2023

Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, et al. Towards monosemanticity: Decomposing language models with dictionary learning.Transformer Circuits Thread, 2023

2023
[4]

Sparse autoencoders find highly interpretable directions in language models.International Conference on Learning Representations (ICLR), 2024

Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoencoders find highly interpretable directions in language models.International Conference on Learning Representations (ICLR), 2024

2024
[5]

Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2

Tom Lieberum, Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Nicolas Sonnerat, Vikrant Varma, János Kramár, Amos Drori, and Neel Nanda. Gemma scope: Open sparse autoencoders everywhere all at once on gemma 2.arXiv preprint arXiv:2408.05147, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[6]

Gemma scope 2: Sparse autoencoders and transcoders for gemma 3

Google DeepMind. Gemma scope 2: Sparse autoencoders and transcoders for gemma 3. Technical report, 2025.https://deepmind.google/models/gemma/gemma-scope/

2025
[7]

Gemma 3 Technical Report

Gemma Team. Gemma 3 technical report.arXiv preprint arXiv:2503.19786, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[8]

Open Problems in Mechanistic Interpretability

Lee Sharkey, Bilal Chughtai, Dan Braun, Beren Millidge, et al. Open problems in mechanistic interpretability.Transactions on Machine Learning Research, 2025. arXiv:2501.16496

work page internal anchor Pith review Pith/arXiv arXiv 2025
[9]

Are sparse autoencoders useful? a case study in sparse probing

Subhash Kantamneni, Joshua Engels, Senthooran Rajamanoharan, Max Tegmark, and Neel Nanda. Are sparse autoencoders useful? a case study in sparse probing. InProceedings of the 42nd International Conference on Machine Learning (ICML), volume 267, pages 29018–29049. PMLR, 2025

2025
[10]

Sparse autoencoders can interpret randomly initialized transformers.arXiv preprint arXiv:2501.17727, 2025

Thomas Heap, Tim Lawson, Lucy Farnik, and Laurence Aitchison. Sparse autoencoders can interpret randomly initialized transformers.arXiv preprint arXiv:2501.17727, 2025

work page arXiv 2025
[11]

Where to edit? complementary protein property control from weight and activation spaces

Armaity Katki, Nathan Choi, Son Sophak Otra, George Flint, and Kevin Zhu. Where to edit? complementary protein property control from weight and activation spaces. InNeurIPS 2025 Workshop on Biosecurity Safeguards for Generative AI (BioSafe GenAI), 2025. URL https://openreview.net/forum?id=KiZxvtn3JE

2025
[12]

Measuring mathematical problem solving with the MATH dataset

Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the MATH dataset. NeurIPS Datasets and Benchmarks, 2021

2021
[13]

TIES-merging: Resolving interference when merging models

Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. TIES-merging: Resolving interference when merging models. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

2023
[14]

Language models are super mario: Absorbing abilities from homologous models as a free lunch.arXiv preprint arXiv:2311.03099, 2024

Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch.arXiv preprint arXiv:2311.03099, 2024

work page arXiv 2024
[15]

Localize-and-stitch: Efficient model merging via sparse task arithmetic

Yifei He, Yuzheng Hu, Yong Lin, Tong Zhang, and Han Zhao. Localize-and-stitch: Efficient model merging via sparse task arithmetic. InTransactions on Machine Learning Research (TMLR), 2025. arXiv:2408.13656

work page arXiv 2025
[16]

Subspace-boosted model merging.arXiv preprint arXiv:2506.16506, 2025

Ronald Skorobogat, Karsten Roth, and Mariana-Iuliana Georgescu. Subspace-boosted model merging.arXiv preprint arXiv:2506.16506, 2025

work page arXiv 2025
[17]

Locating and editing factual associations in GPT.Advances in Neural Information Processing Systems (NeurIPS), 2022

Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in GPT.Advances in Neural Information Processing Systems (NeurIPS), 2022

2022
[18]

Mass- editing memory in a transformer

Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass- editing memory in a transformer. InInternational Conference on Learning Representations (ICLR), 2023

2023
[19]

Task arithmetic in the tangent space: Improved editing of pre-trained models

Guillermo Ortiz-Jiménez, Alessandro Favero, and Pascal Frossard. Task arithmetic in the tangent space: Improved editing of pre-trained models. InAdvances in Neural Information Processing Systems (NeurIPS), 2023. 12

2023
[20]

Llama Scope: Extracting millions of features from Llama-3.1-8b with sparse autoencoders.arXiv preprint arXiv:2410.20526,

Zhengfu He, Wentao Shu, Xuyang Ge, Lingjie Chen, Junxuan Wang, Yunhua Zhou, Frances Liu, Qipeng Guo, Xuanjing Huang, Zuxuan Wu, Yu-Gang Jiang, and Xipeng Qiu. Llama scope: Extracting millions of features from Llama-3.1-8B with sparse autoencoders.arXiv preprint arXiv:2410.20526, 2024

work page arXiv 2024
[21]

Into the Rabbit Hull: From Task-Relevant Concepts in DINO to Minkowski Geometry

Thomas Fel, Binxu Wang, et al. Into the rabbit hull: From task-relevant concepts in DINO to minkowski geometry. InInternational Conference on Learning Representations (ICLR), 2026. arXiv:2510.08638

work page internal anchor Pith review Pith/arXiv arXiv 2026
[22]

Priors in time: Missing inductive biases for language model interpretability.arXiv preprint arXiv:2511.01836, 2025

Ekdeep Singh Lubana et al. Priors in time: Missing inductive biases for language model interpretability.arXiv preprint arXiv:2511.01836, 2025. ICLR 2026 poster

work page arXiv 2025
[23]

Joint localization and activation editing for low- resource fine-tuning

Wen Lai, Alexander Fraser, and Ivan Titov. Joint localization and activation editing for low- resource fine-tuning. InProceedings of the 42nd International Conference on Machine Learning (ICML), volume 267, pages 32206–32227. PMLR, 2025

2025
[24]

A framework for few-shot language model evaluation.Zenodo, 2024

Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, et al. A framework for few-shot language model evaluation.Zenodo, 2024

2024
[25]

The CMA evolution strategy: A comparing review.Towards a New Evolu- tionary Computation, pages 75–102, 2006

Nikolaus Hansen. The CMA evolution strategy: A comparing review.Towards a New Evolu- tionary Computation, pages 75–102, 2006

2006
[26]

surprising

Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. Evolutionary optimization of model merging recipes.arXiv preprint arXiv:2403.13187, 2024. 13 A Reproduction Commands # Step 1: LoRA fine-tuning (produces task vector v2) python experiments/nt_train_lora_v2.py \ --gpu 0 --name lora_v2 --epochs 5 --lora_r 16 # Step 2: Compute task vector and appl...

work page arXiv 2024

[1] [1]

LoRA: Low-rank adaptation of large language models

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shengyu Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022

2022

[2] [2]

Editing models with task arithmetic

Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. InInternational Conference on Learning Representations (ICLR), 2023

2023

[3] [3]

Towards monosemanticity: Decomposing language models with dictionary learning.Transformer Circuits Thread, 2023

Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, et al. Towards monosemanticity: Decomposing language models with dictionary learning.Transformer Circuits Thread, 2023

2023

[4] [4]

Sparse autoencoders find highly interpretable directions in language models.International Conference on Learning Representations (ICLR), 2024

Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoencoders find highly interpretable directions in language models.International Conference on Learning Representations (ICLR), 2024

2024

[5] [5]

Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2

Tom Lieberum, Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Nicolas Sonnerat, Vikrant Varma, János Kramár, Amos Drori, and Neel Nanda. Gemma scope: Open sparse autoencoders everywhere all at once on gemma 2.arXiv preprint arXiv:2408.05147, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024

[6] [6]

Gemma scope 2: Sparse autoencoders and transcoders for gemma 3

Google DeepMind. Gemma scope 2: Sparse autoencoders and transcoders for gemma 3. Technical report, 2025.https://deepmind.google/models/gemma/gemma-scope/

2025

[7] [7]

Gemma 3 Technical Report

Gemma Team. Gemma 3 technical report.arXiv preprint arXiv:2503.19786, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025

[8] [8]

Open Problems in Mechanistic Interpretability

Lee Sharkey, Bilal Chughtai, Dan Braun, Beren Millidge, et al. Open problems in mechanistic interpretability.Transactions on Machine Learning Research, 2025. arXiv:2501.16496

work page internal anchor Pith review Pith/arXiv arXiv 2025

[9] [9]

Are sparse autoencoders useful? a case study in sparse probing

Subhash Kantamneni, Joshua Engels, Senthooran Rajamanoharan, Max Tegmark, and Neel Nanda. Are sparse autoencoders useful? a case study in sparse probing. InProceedings of the 42nd International Conference on Machine Learning (ICML), volume 267, pages 29018–29049. PMLR, 2025

2025

[10] [10]

Sparse autoencoders can interpret randomly initialized transformers.arXiv preprint arXiv:2501.17727, 2025

Thomas Heap, Tim Lawson, Lucy Farnik, and Laurence Aitchison. Sparse autoencoders can interpret randomly initialized transformers.arXiv preprint arXiv:2501.17727, 2025

work page arXiv 2025

[11] [11]

Where to edit? complementary protein property control from weight and activation spaces

Armaity Katki, Nathan Choi, Son Sophak Otra, George Flint, and Kevin Zhu. Where to edit? complementary protein property control from weight and activation spaces. InNeurIPS 2025 Workshop on Biosecurity Safeguards for Generative AI (BioSafe GenAI), 2025. URL https://openreview.net/forum?id=KiZxvtn3JE

2025

[12] [12]

Measuring mathematical problem solving with the MATH dataset

Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the MATH dataset. NeurIPS Datasets and Benchmarks, 2021

2021

[13] [13]

TIES-merging: Resolving interference when merging models

Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. TIES-merging: Resolving interference when merging models. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

2023

[14] [14]

Language models are super mario: Absorbing abilities from homologous models as a free lunch.arXiv preprint arXiv:2311.03099, 2024

Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch.arXiv preprint arXiv:2311.03099, 2024

work page arXiv 2024

[15] [15]

Localize-and-stitch: Efficient model merging via sparse task arithmetic

Yifei He, Yuzheng Hu, Yong Lin, Tong Zhang, and Han Zhao. Localize-and-stitch: Efficient model merging via sparse task arithmetic. InTransactions on Machine Learning Research (TMLR), 2025. arXiv:2408.13656

work page arXiv 2025

[16] [16]

Subspace-boosted model merging.arXiv preprint arXiv:2506.16506, 2025

Ronald Skorobogat, Karsten Roth, and Mariana-Iuliana Georgescu. Subspace-boosted model merging.arXiv preprint arXiv:2506.16506, 2025

work page arXiv 2025

[17] [17]

Locating and editing factual associations in GPT.Advances in Neural Information Processing Systems (NeurIPS), 2022

Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in GPT.Advances in Neural Information Processing Systems (NeurIPS), 2022

2022

[18] [18]

Mass- editing memory in a transformer

Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass- editing memory in a transformer. InInternational Conference on Learning Representations (ICLR), 2023

2023

[19] [19]

Task arithmetic in the tangent space: Improved editing of pre-trained models

Guillermo Ortiz-Jiménez, Alessandro Favero, and Pascal Frossard. Task arithmetic in the tangent space: Improved editing of pre-trained models. InAdvances in Neural Information Processing Systems (NeurIPS), 2023. 12

2023

[20] [20]

Llama Scope: Extracting millions of features from Llama-3.1-8b with sparse autoencoders.arXiv preprint arXiv:2410.20526,

Zhengfu He, Wentao Shu, Xuyang Ge, Lingjie Chen, Junxuan Wang, Yunhua Zhou, Frances Liu, Qipeng Guo, Xuanjing Huang, Zuxuan Wu, Yu-Gang Jiang, and Xipeng Qiu. Llama scope: Extracting millions of features from Llama-3.1-8B with sparse autoencoders.arXiv preprint arXiv:2410.20526, 2024

work page arXiv 2024

[21] [21]

Into the Rabbit Hull: From Task-Relevant Concepts in DINO to Minkowski Geometry

Thomas Fel, Binxu Wang, et al. Into the rabbit hull: From task-relevant concepts in DINO to minkowski geometry. InInternational Conference on Learning Representations (ICLR), 2026. arXiv:2510.08638

work page internal anchor Pith review Pith/arXiv arXiv 2026

[22] [22]

Priors in time: Missing inductive biases for language model interpretability.arXiv preprint arXiv:2511.01836, 2025

Ekdeep Singh Lubana et al. Priors in time: Missing inductive biases for language model interpretability.arXiv preprint arXiv:2511.01836, 2025. ICLR 2026 poster

work page arXiv 2025

[23] [23]

Joint localization and activation editing for low- resource fine-tuning

Wen Lai, Alexander Fraser, and Ivan Titov. Joint localization and activation editing for low- resource fine-tuning. InProceedings of the 42nd International Conference on Machine Learning (ICML), volume 267, pages 32206–32227. PMLR, 2025

2025

[24] [24]

A framework for few-shot language model evaluation.Zenodo, 2024

Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, et al. A framework for few-shot language model evaluation.Zenodo, 2024

2024

[25] [25]

The CMA evolution strategy: A comparing review.Towards a New Evolu- tionary Computation, pages 75–102, 2006

Nikolaus Hansen. The CMA evolution strategy: A comparing review.Towards a New Evolu- tionary Computation, pages 75–102, 2006

2006

[26] [26]

surprising

Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. Evolutionary optimization of model merging recipes.arXiv preprint arXiv:2403.13187, 2024. 13 A Reproduction Commands # Step 1: LoRA fine-tuning (produces task vector v2) python experiments/nt_train_lora_v2.py \ --gpu 0 --name lora_v2 --epochs 5 --lora_r 16 # Step 2: Compute task vector and appl...

work page arXiv 2024