pith. sign in

arxiv: 2606.05016 · v1 · pith:YGOCINHPnew · submitted 2026-06-03 · 💻 cs.CL

TaDA: Calibrated Probe Gating for Task-Domain LoRA Merging

Huy Quoc To , Fuyi Li , Guangyan Huang , Ming Liu This is my paper

Pith reviewed 2026-06-28 05:46 UTC · model grok-4.3

classification 💻 cs.CL
keywords LoRA mergingtask-domain adaptationadapter fusiondepth-dependent asymmetryprobe gatingscientific QAvision transformersparameter-efficient fine-tuning
0
0 comments X

The pith

Calibrated per-layer gating that accounts for depth-dependent task-domain asymmetry outperforms uniform LoRA merging on QA and classification benchmarks.

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

The paper establishes that task and domain LoRA adapters display a consistent depth-dependent asymmetry in transformers, with shallower layers carrying stronger task signals and deeper layers showing greater domain dominance. It introduces TaDA as a training-free method that uses a magnitude-invariant probe to set individual weights per layer and projection type, then merges remaining components after discarding conflicting singular directions. This produces a standard rank-r LoRA with no inference cost. Experiments on Llama-2-7B across six scientific QA benchmarks yield 0.452 average accuracy, exceeding DARE-TIES by 3.6 points, with similar gains shown on ViT-L/16 image classification tasks.

Core claim

By identifying that domain dominance increases with layer depth while shallower layers retain stronger task-relevant signals, TaDA applies calibrated probe-guided per-layer gating using a magnitude-invariant probe signal and discards conflicting singular directions in subspace-aware merging to create an effective unified rank-r LoRA adapter.

What carries the argument

Calibrated probe-guided per-layer gating combined with per-component subspace-aware merging that exploits the observed depth-dependent task-domain asymmetry.

If this is right

  • Merged adapters reach higher accuracy than uniform-weight baselines on scientific QA tasks with Llama-2-7B.
  • The same merging procedure improves results on image classification with ViT-L/16 while leading on three of six benchmarks.
  • The output remains a standard rank-r LoRA adapter requiring no extra computation at inference time.
  • Gains hold when merging task and domain adapters trained on separate objectives without joint retraining.

Where Pith is reading between the lines

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

  • The depth pattern could inform weight assignment when merging three or more distinct adapters.
  • Probe-based gating might transfer to merging strategies for other parameter-efficient tuning methods.
  • The invariance of the probe signal to adapter magnitude suggests it could stabilize merging under different training regimes.

Load-bearing premise

Task and domain adapters exhibit a consistent depth-dependent asymmetry across transformer architectures.

What would settle it

Repeating the per-layer signal strength measurements on additional transformer models and finding either uniform task-domain contributions across depths or reversed patterns would falsify the asymmetry premise.

Figures

Figures reproduced from arXiv: 2606.05016 by Fuyi Li, Guangyan Huang, Huy Quoc To, Ming Liu.

Figure 1
Figure 1. Figure 1: TaDA overview. 3.1 Calibrated Probe-Guided Per-Layer Gating Probe construction. We construct two probe sets of N=32 inputs each: a domain probe Pd (target-domain sentences or images) and a general probe Pg (Wikipedia text or ImageNet images). Both are passed through the frozen base model to extract mean hidden states h (ℓ) (P) ∈ R din at each layer ℓ. Calibrated domain relevance score. The raw activation r… view at source ↗
Figure 2
Figure 2. Figure 2: Per-layer task weight α (ℓ,m˜ ) as a function of normalised depth ˜ℓ = ℓ/L for Llama-2-7b (left) and ViT-L/16 (right). Blue: task-dominant (α → 1); green: domain￾dominant (α → 0); white: balanced. Rows (top to bottom): q, k, v, o (τm=1.5). Domain dominance prevails in both models but weakens in shallower layers (˜ℓ < 0.25). 7 Analysis and Ablation Study 7.1 Cross-Modal Layer-Wise Gating Pattern [PITH_FULL… view at source ↗
read the original abstract

Combining a task LoRA adapter with a domain LoRA adapter into a single unified model is a practical yet largely unexplored challenge. Existing methods treat both adapters as symmetric peers, applying uniform weights across all layers. We argue that task and domain adapters exhibit a consistent depth-dependent asymmetry across transformer architectures. Domain dominance increases with layer depth, while shallower layers retain stronger task-relevant signals. Motivated by this observation, we propose $\textbf{TaDA}$ ($\textbf{Ta}$sk-$\textbf{D}$omain LoR$\textbf{A}$ Merging), a training-free algorithm that exploits this structure through calibrated probe-guided per-layer gating and per-component subspace-aware merging. The gating assigns individual weights per layer and projection type using a probe signal proved invariant to adapter weight magnitude. The merging discards conflicting singular directions before combining the remaining components. $\textbf{TaDA}$ produces a standard rank-$r$ LoRA adapter with zero inference overhead. On six scientific QA benchmarks with Llama-2-7B, TaDA achieves an average accuracy of 0.452, outperforming DARE-TIES by +3.6 percentage points and obtaining the best result on all six benchmarks. On six image classification benchmarks with ViT-L/16, TaDA reaches 85.9\% average accuracy, improving over the strongest merging baseline while leading in three of the six individual benchmarks.

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.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes TaDA, a training-free algorithm for merging task and domain LoRA adapters. It is motivated by an observed depth-dependent asymmetry in which domain adapters dominate deeper layers while task signals are stronger in shallower layers. The method uses a calibrated probe for per-layer and per-projection gating (claimed invariant to adapter magnitude) and subspace-aware merging to discard conflicting directions. It reports an average accuracy of 0.452 on six scientific QA benchmarks using Llama-2-7B (outperforming DARE-TIES by 3.6 pp and best on all), and 85.9% on six image classification tasks with ViT-L/16.

Significance. If the asymmetry holds and the gating exploits it causally rather than as generic weighting, this could offer an effective, zero-overhead way to combine specialized adapters, which is practically valuable for efficient model customization. The consistent outperformance on all benchmarks in one setting is a notable result if the experimental protocol is fully specified and reproducible.

major comments (2)
  1. [Abstract] Abstract: The depth-dependent asymmetry ('domain dominance increases with layer depth, while shallower layers retain stronger task-relevant signals') is asserted as motivation and the basis for the per-layer gating, but the text supplies no quantitative measurements (e.g., per-layer cosine similarities or activation norms) or ablations removing the depth-dependent component; without this, it is unclear whether the reported +3.6 pp margin over DARE-TIES is driven by the claimed structure rather than other implementation choices.
  2. [Abstract] Abstract: The gating mechanism relies on a 'probe signal proved invariant to adapter weight magnitude,' yet no derivation, equation, or proof sketch is provided to establish this invariance, which is central to the calibrated per-layer weighting claim.
minor comments (1)
  1. The abstract reports results on ViT-L/16 but does not clarify whether the depth-dependent asymmetry was verified for vision transformers or if the method required adaptation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for these comments. Both points identify genuine gaps in the current manuscript: the absence of quantitative support for the claimed depth-dependent asymmetry and the lack of a derivation for the probe-signal invariance. We will address both by adding the requested material in the revision.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The depth-dependent asymmetry ('domain dominance increases with layer depth, while shallower layers retain stronger task-relevant signals') is asserted as motivation and the basis for the per-layer gating, but the text supplies no quantitative measurements (e.g., per-layer cosine similarities or activation norms) or ablations removing the depth-dependent component; without this, it is unclear whether the reported +3.6 pp margin over DARE-TIES is driven by the claimed structure rather than other implementation choices.

    Authors: We agree that the manuscript currently asserts the asymmetry without presenting supporting measurements or an ablation that isolates its contribution. In the revised version we will add a new subsection containing (i) per-layer cosine-similarity and activation-norm statistics between task and domain adapters on the Llama-2-7B and ViT backbones, and (ii) an ablation that replaces the depth-dependent gating with uniform layer weights while keeping all other components fixed. These additions will allow readers to assess whether the reported gains are attributable to the depth-dependent structure. revision: yes

  2. Referee: [Abstract] Abstract: The gating mechanism relies on a 'probe signal proved invariant to adapter weight magnitude,' yet no derivation, equation, or proof sketch is provided to establish this invariance, which is central to the calibrated per-layer weighting claim.

    Authors: We accept that the manuscript states the invariance without supplying the supporting derivation. In the revision we will insert a dedicated paragraph (with equations) in the Methods section that derives the probe signal, shows the normalization step that removes magnitude dependence, and includes a short proof sketch establishing invariance under scaling of the adapter weights. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical results on external benchmarks with independent motivation

full rationale

The paper reports measured accuracies on six scientific QA benchmarks and six image classification benchmarks, outperforming baselines like DARE-TIES. The method is motivated by an observed depth-dependent asymmetry in task/domain adapters, implemented via probe-guided gating claimed to be invariant to weight magnitude. No equations, fitted parameters, or self-citations are exhibited that reduce the reported gains to quantities defined from the same data by construction. The derivation chain is self-contained against the external benchmarks and does not rely on renaming, self-definition, or load-bearing self-citations.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The method rests on the stated depth-asymmetry observation and the claim that the probe signal is invariant to adapter magnitude; both are introduced without external verification in the abstract.

axioms (2)
  • domain assumption Task and domain adapters exhibit consistent depth-dependent asymmetry across transformer architectures
    Invoked as the motivation for per-layer gating.
  • domain assumption Probe signal is invariant to adapter weight magnitude
    Used to justify calibrated per-layer weights.

pith-pipeline@v0.9.1-grok · 5782 in / 1220 out tokens · 12077 ms · 2026-06-28T05:46:40.936379+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 5 canonical work pages

  1. [1]

    Understanding the User Experience of Customer Service Chatbots: What Can We Learn from Customer Satisfaction Surveys?Chatbot Re- search and Design (CONVERSATIONS 2020)

    Chen, Q., Hong, Y.: Medblip: Bootstrapping language-image pretraining from 3d medical images and texts. In: Computer Vision – ACCV 2024: 17th Asian Conference on Computer Vision, Hanoi, Vietnam, December 8–12, 2024, Pro- ceedings, Part III. p. 98–113. Springer-Verlag, Berlin, Heidelberg (2024).https: //doi.org/10.1007/978- 981- 96- 0908- 6_6,https://doi.o...

    work page doi:10.1007/978- 2024

  2. [2]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2014) 7 Task-Domain LoRA Merging 15

    Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2014) 7 Task-Domain LoRA Merging 15

    2014

  3. [3]

    arXiv:1803.05457v1 (2018) 6

    Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., Tafjord, O.: Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv:1803.05457v1 (2018) 6

    Pith/arXiv arXiv 2018

  4. [4]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248–255 (2009).https://doi.org/10.1109/CVPR.2009. 52068486, 7

    work page doi:10.1109/cvpr.2009 2009

  5. [5]

    In: In- ternational Conference on Learning Representations (2021),https://openreview

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: In- ternational Conference on Learning Representations (2021),https://openreview. net/forum?id=YicbFdNTTy3, 6

    2021

  6. [6]

    In: Proceedings of the Asian Con- ference on Computer Vision

    He, H., Liu, W., Xing, W.: Biefficient: Bidirectionally prompting vision-language models for parameter-efficient video recognition. In: Proceedings of the Asian Con- ference on Computer Vision. pp. 108–125 (2024) 3

    2024

  7. [7]

    Helber, P., Bischke, B., Dengel, A., Borth, D.: Introducing eurosat: A novel dataset anddeeplearningbenchmarkforlanduseandlandcoverclassification.In:IGARSS 2018 - 2018 IEEE International Geoscience and Remote Sensing Symposium. pp. 204–207 (2018).https://doi.org/10.1109/IGARSS.2018.85192487

    work page doi:10.1109/igarss.2018.85192487 2018

  8. [8]

    Proceedings of the International Conference on Learning Representations (ICLR) (2021) 6

    Hendrycks, D., Burns, C., Basart, S., Critch, A., Li, J., Song, D., Steinhardt, J.: Aligning ai with shared human values. Proceedings of the International Conference on Learning Representations (ICLR) (2021) 6

    2021

  9. [9]

    Proceedings of the In- ternational Conference on Learning Representations (ICLR) (2021) 6

    Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., Steinhardt, J.: Measuring massive multitask language understanding. Proceedings of the In- ternational Conference on Learning Representations (ICLR) (2021) 6

    2021

  10. [10]

    In: International Con- ference on Learning Representations (2022),https://openreview.net/forum?id= nZeVKeeFYf91, 3

    Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Con- ference on Learning Representations (2022),https://openreview.net/forum?id= nZeVKeeFYf91, 3

    2022

  11. [11]

    In: The Eleventh International Conference on Learning Representations (2023),https://openreview.net/forum?id=6t0Kwf8- jrj3, 7, 9, 10

    Ilharco, G., Ribeiro, M.T., Wortsman, M., Schmidt, L., Hajishirzi, H., Farhadi, A.: Editing models with task arithmetic. In: The Eleventh International Conference on Learning Representations (2023),https://openreview.net/forum?id=6t0Kwf8- jrj3, 7, 9, 10

    2023

  12. [12]

    Applied Sciences11(14) (2021).https://doi.org/10.3390/ app11146421,https://www.mdpi.com/2076-3417/11/14/64216

    Jin, D., Pan, E., Oufattole, N., Weng, W.H., Fang, H., Szolovits, P.: What dis- ease does this patient have? a large-scale open domain question answering dataset from medical exams. Applied Sciences11(14) (2021).https://doi.org/10.3390/ app11146421,https://www.mdpi.com/2076-3417/11/14/64216

    2021

  13. [13]

    URLhttps://doi.org/10.18653/v1/D19-1259

    Jin,Q.,Dhingra,B.,Liu,Z.,Cohen,W.,Lu,X.:PubMedQA:Adatasetforbiomed- ical research question answering. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Pro- cessing(EMNLP-IJCNLP).pp.2567–2577.AssociationforCom...

    work page doi:10.18653/v1/d19-1259 2019

  14. [14]

    Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. Rep. 0, University of Toronto, Toronto, Ontario (2009),https://www.cs. toronto.edu/~kriz/learning-features-2009-TR.pdf7

    2009

  15. [15]

    In: Flores, G., Chen, G.H., Pollard, T., Ho, J.C., Naumann, T

    Pal, A., Umapathi, L.K., Sankarasubbu, M.: Medmcqa: A large-scale multi-subject multi-choice dataset for medical domain question answering. In: Flores, G., Chen, G.H., Pollard, T., Ho, J.C., Naumann, T. (eds.) Proceedings of the Conference 16 Huy To et al. on Health, Inference, and Learning. Proceedings of Machine Learning Research, vol. 174, pp. 248–260....

    2022

  16. [16]

    In: The Thirteenth International Conference on Learning Representations (2025),https://openreview.net/forum?id=67X93aZHII3

    Stoica, G., Ramesh, P., Ecsedi, B., Choshen, L., Hoffman, J.: Model merging with SVD to tie the knots. In: The Thirteenth International Conference on Learning Representations (2025),https://openreview.net/forum?id=67X93aZHII3

    2025

  17. [17]

    Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model.https: //github.com/tatsu-lab/stanford_alpaca(2023) 6

    2023

  18. [18]

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A

    Tenney, I., Das, D., Pavlick, E.: BERT rediscovers the classical NLP pipeline. In: Korhonen, A., Traum, D., Màrquez, L. (eds.) Proceedings of the 57th An- nual Meeting of the Association for Computational Linguistics. pp. 4593–4601. Association for Computational Linguistics, Florence, Italy (Jul 2019).https: //doi.org/10.18653/v1/P19-1452,https://aclantho...

    work page doi:10.18653/v1/p19-1452 2019

  19. [19]

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023), https://arxiv.org/abs/2302.139716

    Pith/arXiv arXiv 2023

  20. [20]

    In: Derczynski, L., Xu, W., Ritter, A., Baldwin, T

    Welbl, J., Liu, N.F., Gardner, M.: Crowdsourcing multiple choice science ques- tions. In: Derczynski, L., Xu, W., Ritter, A., Baldwin, T. (eds.) Proceedings of the 3rd Workshop on Noisy User-generated Text. pp. 94–106. Association for Com- putational Linguistics, Copenhagen, Denmark (Sep 2017).https://doi.org/10. 18653/v1/W17-4413,https://aclanthology.org...

    2017

  21. [21]

    Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., Schmidt, L.: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time (2022),https://arxiv.org/abs/2203.054821, 3, 7, 9, 10

    arXiv 2022

  22. [22]

    In: Thirty-seventh Conference on Neu- ral Information Processing Systems (2023),https://openreview.net/forum?id= xtaX3WyCj11, 3, 7, 9, 10

    Yadav, P., Tam, D., Choshen, L., Raffel, C., Bansal, M.: TIES-merging: Resolv- ing interference when merging models. In: Thirty-seventh Conference on Neu- ral Information Processing Systems (2023),https://openreview.net/forum?id= xtaX3WyCj11, 3, 7, 9, 10

    2023

  23. [23]

    Scientific Data10(1), 41 (2023) 6, 7

    Yang,J.,Shi,R.,Wei,D.,Liu,Z.,Zhao,L.,Ke,B.,Pfister,H.,Ni,B.:Medmnistv2- a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data10(1), 41 (2023) 6, 7

    2023

  24. [24]

    In: International Conference on Learning Representations (2023),https://openreview.net/forum?id=CIoSZ_HKHS73

    Yang, T., Zhu, Y., Xie, Y., Zhang, A., Chen, C., Li, M.: Aim: Adapting image models for efficient video understanding. In: International Conference on Learning Representations (2023),https://openreview.net/forum?id=CIoSZ_HKHS73

    2023

  25. [25]

    In: Proceedings of the 41st International Conference on Machine Learning

    Yu, L., Yu, B., Yu, H., Huang, F., Li, Y.: Language models are super mario: absorbing abilities from homologous models as a free lunch. In: Proceedings of the 41st International Conference on Machine Learning. ICML’24, JMLR.org (2024) 1, 3, 7, 9, 10

    2024