Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Efficient Knowledge Transfer in Multi-Task Learning through Task-Adaptive Low-Rank Representation

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read TA-LoRA: a low-rank prompt-tuning method that beats full fine-tuning on unseen NLP tasks while tuning only 1.3M parameters per task.

desk verdict A plausible PEFT recipe whose main cross-task generalization claim is undercut by evaluating on target training sets. read the letter →

arxiv 2505.00009 v1 pith:GAAYE2VN submitted 2025-04-20 cs.CL

Xiao Zhang , Kangsheng Wang , Tianyu Hu , Huimin Ma This is my paper · ORCID
classification cs.CL
keywords multi-tasklearningprompttuninglow-rankrepresentationparameter-efficientfine-tuningfast-slowweightszero-initializedattentionfew-shotknowledgetransfer
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

The paper claims that prompt tuning's limited capacity to represent task differences, not the prompt form itself, is what holds multi-task transfer back, and that a low-rank task-adaptive representation fixes it. TA-LoRA keeps a shared prompt vector (slow weight) that encodes common knowledge and per-task rank-1 vectors (fast weights) that encode heterogeneity, with a zero-initialized attention gate so immature adapters do not disturb the base prompt during warm-up. On 16 Chinese NLP tasks with Qwen2.5-7B, the method is reported to reach 83.4% average on unseen data and 74.3% on unseen tasks, beating full fine-tuning (81.6% and 69.7%) while training roughly 1.3M parameters per task instead of 7B. If true, this means a single frozen backbone plus small per-task adapters can replace per-task full fine-tuning and generalize better to new tasks with only a few dozen examples.

What carries the argument

The central object is the task-adaptive low-rank representation: each task's prompt-level knowledge is written as $\theta_0 + s B (u_i \otimes v_i)$, with $\theta_0$ the shared (slow) prompt, $B$ a shared low-rank matrix, and $(u_i, v_i)$ fast rank-1 task vectors. The shared $B$ is trained with a slower learning rate than the fast vectors, and an orthogonality regularizer $\sum_{j\neq i} \|A_i^\top A_j - I\|_2^2$ pushes different tasks' subspaces apart. A zero-initialized attention mechanism with a learnable gating factor $\tanh(g_l)$ scales the adaptable prompt's attention scores so random low-rank components do not disrupt original tokens during warm-up. This machinery is what carries the argument: it turns task heterogeneity into a small set of directions in parameter space that can be learned and then reused for new tasks by initializing a fresh rank-1 $A_i$.

What would settle it

Re-run the unseen-task evaluation with genuine held-out splits (the official validation/test sets of ChnSent, TNews, OCNLI, LCQMC, DRCD, C3, COTE-BD, and FinRE) and compare TA-LoRA against MPT and full fine-tuning; if the average margin over full fine-tuning shrinks to near zero, the core transfer claim is mostly pretraining overlap. A second check: finetune Qwen2.5-7B with the same 1.3M-parameter budget but random source-task labels; if the few-shot gains persist, the shared-prompt mechanism is not what transfers knowledge.

Watch

Extended reading notes

Core claim

On its own terms, the paper discovers that the difference between a shared prompt and each task's prompt can be approximated by a rank-1 correction $B(u_i \otimes v_i)$, where the shared low-rank matrix $B$ is trained slowly and the task vectors $u_i, v_i$ quickly, and that this decomposition plus an orthogonality penalty and a zero-initialized attention gate cleanly separates shared from task-specific knowledge. The authors argue this is the first multi-task prompt-tuning framework to use low-rank representations for task heterogeneity, and they support it with comparisons against full fine-tuning, Adapter, BitFit, vanilla prompt tuning, SPoT, ATTEMPT, and MPT. Their headline numbers: on Qwen2.5-7B with 1.3M trainable parameters per task, TA-LoRA averages 83.4 on held-out data of seen tasks and 74.3 on eight unseen tasks, versus 81.6 and 69.7 for full fine-tuning; in few-shot evaluation it stays competitive at 16, 32, and 64 shots. The claim is not that low-rank is always better, but that the fast-slow decomposition avoids the entanglement of shared and task-specific knowledge that arises when low-rank adapters are trained from scratch.

Load-bearing premise

The reported transfer to 'unseen tasks' is measured on the full training sets of the target datasets, with no held-out test split; the method assumes Qwen2.5 was never trained on those exact texts, so the gains could partly be memorization rather than the TA-LoRA mechanism.

Editorial extensions

If this is right

  • A 7B model can serve many tasks from one frozen checkpoint; adding a task costs about 1.3M trainable parameters rather than a full fine-tune.
  • Unseen-task generalization improves with multi-source training: the transferred shared prompt beats per-task fine-tuning from scratch on average.
  • Few-shot adaptation to a new task can be done by training only a rank-1 task vector on 16–64 examples, with no update to the shared base.
  • The orthogonality penalty gives a concrete way to keep task subspaces distinct, pointing to a regularizer that other parameter-efficient fine-tuning methods could adopt.
  • The gap between TA-LoRA and MPT/ATTEMPT suggests prompt-vector capacity, not prompt-vector length alone, limits existing multi-task prompt methods.

Reading between the lines

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

  • The paper's 'unseen task' numbers are computed on the full training sets of the target datasets; a fair test of transfer would need held-out test splits, so the reported 74.3% may shrink under a cleaner protocol.
  • If the mechanism is right, the same decomposition should transfer to English and multilingual backbones, since neither the low-rank split nor the attention gate is language-specific.
  • A direct test of the entanglement story: train the same low-rank adapters with one learning rate (no fast-slow split) and measure how much the slow weight $B$ drifts across tasks; the paper's ablation suggests performance drops by roughly 1.7–1.8 points, which can be checked on more tasks.
  • The rank-1 constraint sets an upper bound on task-specific capacity; pushing to rank 2 or 3 might help on harder tasks such as C3 or hurt via overfitting, a trade-off the paper does not explore.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes TA-LoRA, a parameter-efficient multi-task learning method built on prompt tuning with a shared low-rank matrix B and task-specific rank-1 matrices A_i, together with a fast-slow learning-rate mechanism and a zero-initialized attention gate. The authors claim state-of-the-art results on 16 Chinese NLP tasks with Qwen2.5-7B, including gains over full fine-tuning in both full-data and few-shot settings while using only ~1.3M trainable parameters per task. The main evidence is in Tables I and II, with ablations in Table III and an inter-layer similarity analysis in Appendix A.

Significance. If the empirical results hold, the method is a simple and parameter-efficient recipe for multi-task prompt tuning, and the fast-slow weight decomposition plus zero-initialized gating are reasonable, well-motivated components. The paper is clearly written and the experiments cover a diverse task set. However, the significance is currently conditional: the central generalization claims rest on an evaluation protocol that appears to score target tasks on their own training sets, and one of the method's core equations (Eq. 10) does not implement the orthogonality penalty that the text describes.

major comments (3)
  1. [Appendix B.3 / Section III.E / Table I / Table II] The 'Unseen Task' evaluation uses the training sets of ChnSent, TNews, OCNLI, LCQMC, DRCD, C3, COTE-BD, and FinRE, and Section III.E explains that the target task's A_i is initialized and optimized through LPLM on the target task. No train/test split is reported for these target tasks, so the 'Unseen Task' rows of Table I and all of Table II appear to measure performance on the very instances used for adaptation. This makes the headline generalization claim (e.g., 74.3 versus 69.7 for full fine-tuning in Table I) unsupported. The authors must either use genuine held-out test sets, clearly report the split, and re-run the experiments, or substantially weaken the claim to in-sample adaptation performance.
  2. [Section III.E, Eq. (10)] Equation (10) writes the regularization term as \|A_i^T A_j - I\|_2^2 for j ≠ i. If the goal is to enforce orthogonality among task-specific low-rank matrices, the term should target zero (for example \|A_i^T A_j\|_2^2), not the identity. As written, the regularizer drives A_i^T A_j toward I, which is the opposite of orthogonality and also impossible for rank-1 matrices A_i = u_i ⊗ v_i when the inner dimension exceeds one. This discrepancy should be corrected, and the ablation in Table III that depends on this regularizer should be re-examined.
  3. [Section III.B, Eqs. (1)-(3)] Equation (1) is an identity: after substituting θ0 = (1/t)\sum_i θ_i, the bracketed term equals θ_i, so the equation merely restates θ = ⊕ θ_i. Equation (2) is therefore not a derived consequence but a modeling assumption that the residual θ_i − θ0 is well approximated by B_i A_i. This is acceptable as a design choice, but the paper should present it honestly as a low-rank modeling assumption rather than as a derivation. In addition, the left side of Eq. (2) should be written explicitly as (θ_i − θ0), since for a fixed i the average (1/t)\sum_j (θ_i − θ_j) equals exactly that.
minor comments (5)
  1. [Table II] The entry '72.816' in the k=16 row appears to contain an accidental superscript '16' attached to the value, and the table reports no standard deviations or number of seeds; please add variance information or clarify that single runs are shown.
  2. [Appendix A] The text says the similarity of the 1st, 14th, and 28th layers was calculated, but the reported values correspond to layers 14, 21, and 27, while Figure 3 labels layers 14, 21, and 28; these inconsistencies should be resolved.
  3. [Appendix B.3 / Table IV] For the target tasks, Table IV lists only training-set sizes; please also state the sizes of the test or validation sets used for the 'Unseen Task' evaluation so that readers can judge the reliability of the reported averages.
  4. [Section III.C] The notation for A_i is ambiguous: it is described as a rank-1 matrix from the outer product u_i ⊗ v_i, but Eq. (10) treats A_i as a matrix whose transpose can be multiplied by another A_j. Please define the exact dimensions of u_i, v_i, and A_i, and state whether the regularization in Eq. (10) is applied to the factored vectors or to the matrix product.
  5. [Section I / Section II] The claim of being 'the first multi-task prompt tuning framework to leverage low-rank representations' should be softened or more carefully scoped, since low-rank approximations are widely used in PEFT and MTL; a brief comparison with existing low-rank multi-task methods would strengthen the positioning.

Circularity Check

1 steps flagged · score 6.0 of 10

Unseen-task evaluation is performed on the same target training sets used to optimize the target adapter, so the headline cross-task generalization results reduce to in-sample fit.

  1. fitted input called prediction [Appendix B.3 and Section III.E (Unseen Task evaluation protocol)]
    "To assess the model’s capability to generalize across tasks, we employ the training sets of eight downstream tasks: ChnSent, TNews, OCNLI, LCQMC, DRCD, C3, COTE-BD, and FinRE. ... we initialize a task-specific rank-1 Ai = (ui⊗vi) for each target task, and optimize it through LPLM."

    The headline claim is that TA-LoRA generalizes to unseen tasks (Abstract, Table II). But the 'Unseen Task' evaluation uses the full training sets of the target tasks, and Section III.E says the target adapter Ai is initialized and optimized through LPLM on the target task. No held-out split is described. Therefore the reported unseen-task accuracies measure fit to the same target-task examples used for optimization, not prediction on unseen target data. The 'generalization' result reduces by construction to in-sample training performance.

full rationale

The algebraic decomposition in Eq. (1) is an identity and the low-rank approximation in Eq. (2) is a modeling assumption; neither constitutes a circular derivation because the paper does not present them as predictions. The fast-slow weights and zero-initialized attention are design choices evaluated empirically, and the self-citations in the reference list are not load-bearing for the central claim. The serious circularity is in the evaluation protocol: the paper's strongest cross-task generalization evidence is obtained by optimizing the target adapter on the target task's training set and then evaluating on that same training set, with no train/test split described for the Unseen Task setting. The few-shot table has the same ambiguity because no evaluation split is stated for the k-shot sampling. Thus the central claim of superior generalization to unseen tasks is partially circular: it rests on in-sample fit rather than on prediction of held-out target examples.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical or architectural entity beyond standard low-rank matrices and a gating factor; the 'fast-slow weights' phrasing is a learning-rate scheme, not a new object. The main load-bearing choices are the low-rank approximation of task differences, the specific layer range, and the regularization term, all captured above.

free parameters (8)
  • Regularization coefficient lambda
    In Eq (10), lambda controls the orthogonality penalty; its value is not reported, and this term as written appears mis-specified (A_i^T A_j - I should likely be A_i^T A_j). The choice affects how task-specific low-rank factors are shaped.
  • Learning rates for slow weight B and fast weights A_i
    Section III.C assigns different learning rates but gives no values; the ablation (Table III) shows this choice matters (1.8%/1.7% average gain).
  • Rank of shared low-rank matrix B
    Section III.B says 'low-rank matrices' and defines A_i as rank-1 outer products, but the rank of B is never specified; it is a tunable hyperparameter.
  • Scaling factor s
    In Eq (3), s controls the magnitude of the task-specific term; no value or schedule is given.
  • Prompt length = 20
    Section IV.A sets adaptable prompt length 20; affects capacity and parameter count.
  • Number of layers L = 14
    Low-rank representation and zero-initialized attention are applied to the final 14 of 28 Qwen2.5 layers, justified by Appendix A similarities.
  • Initialization scale sigma^2 for B and A
    B and A are initialized from N(0, sigma^2) but the chosen sigma is not reported.
  • Data balancing ratios (up-sampling/down-sampling)
    Appendix B.1 applies up-sampling and down-sampling to balance dataset sizes, but the target sizes are unspecified.
assumptions (5)
  • domain assumption The task-specific parameter difference theta_i - theta_0 can be approximated by a low-rank product B_i A_i (Eq 2).
    This is a modeling assumption, not derived; it is the core of the method.
  • domain assumption Later layers of a PLM (last L=14 of 28) are sufficient to capture task heterogeneity, so low-rank updates are applied only there.
    Justified by the paper's own similarity measurements in Appendix A, not by an independent principle.
  • ad hoc to paper Separate learning rates (slow B, fast A) prevent entanglement of shared and task-specific knowledge.
    Stated in Section III.C without proof or theoretical analysis; the ablation supports it empirically.
  • domain assumption Zero-initialized attention with a tanh-gated factor preserves original prompt behavior during warm-up (from LLaMA-Adapter, ref [15]).
    Borrowed from prior work; the paper relies on [15]'s result and does not re-derive it.
  • domain assumption The orthogonality penalty on A_i (Eq 10) improves generalization to unseen tasks.
    Empirical regularizer; no theoretical guarantee. Additionally, the formula as printed (A_i^T A_j - I) seems inconsistent with the intended orthogonality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Knowledge Transfer in Multi-Task Learning through Task-Adaptive Low-Rank Representation." pith.science (2026). https://pith.science/paper/GAAYE2VN

@misc{pith2026250500009,
  author       = {Pith},
  title        = {Pith review of: Efficient Knowledge Transfer in Multi-Task Learning through Task-Adaptive Low-Rank Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GAAYE2VN}},
  note         = {Machine review of arXiv:2505.00009}
}
read the original abstract

Pre-trained language models (PLMs) demonstrate remarkable intelligence but struggle with emerging tasks unseen during training in real-world applications. Training separate models for each new task is usually impractical. Multi-task learning (MTL) addresses this challenge by transferring shared knowledge from source tasks to target tasks. As an dominant parameter-efficient fine-tuning method, prompt tuning (PT) enhances MTL by introducing an adaptable vector that captures task-specific knowledge, which acts as a prefix to the original prompt that preserves shared knowledge, while keeping PLM parameters frozen. However, PT struggles to effectively capture the heterogeneity of task-specific knowledge due to its limited representational capacity. To address this challenge, we propose Task-Adaptive Low-Rank Representation (TA-LoRA), an MTL method built on PT, employing the low-rank representation to model task heterogeneity and a fast-slow weights mechanism where the slow weight encodes shared knowledge, while the fast weight captures task-specific nuances, avoiding the mixing of shared and task-specific knowledge, caused by training low-rank representations from scratch. Moreover, a zero-initialized attention mechanism is introduced to minimize the disruption of immature low-rank components on original prompts during warm-up epochs. Experiments on 16 tasks demonstrate that TA-LoRA achieves state-of-the-art performance in full-data and few-shot settings while maintaining superior parameter efficiency.

Figures

Figures reproduced from arXiv: 2505.00009 by the authors.

Figure 1
Figure 1. The framework of TA-LoRA serves as a plugin in the final [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Decomposition of the slow weight B and the fast weight Ai for T i, with further decomposition of fast weight Ai into ui and vi. “lr” stands for the learning rate. layers. To demonstrate the suitability of Qwen (PLM with 28 decoder layers) [27] used in the proposed framework, we calculated the similarity of the 1st, 14th, and 28th layers, as shown in Appendix A 1 , verifying the above hypothesis, which led us to deco… view at source ↗
Figure 3
Figure 3. The similarity between base models in different layers obtained by [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 27 canonical work pages

  1. [1]

    Language models are few-shot learners,

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, and et al., “Language models are few-shot learners,” in Neural Information Processing Systems , 2020, pp. 1877–1901

    work page 2020

  2. [2]

    Pre-trained language models and their applications,

    H. Wang, J. Li, H. Wu, E. Hovy, and Y . Sun, “Pre-trained language models and their applications,” Engineering, pp. 51–65, 2023

    work page 2023

  3. [3]

    Enhancing autonomous driving through dual-process learning with behavior and reflection integration,

    Xiao Zhang, Kangsheng Wang, Tianyu Hu, and Huimin Ma, “Enhancing autonomous driving through dual-process learning with behavior and reflection integration,” in ICASSP 2025-2025 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2025, pp. 1–5

    work page 2025

  4. [4]

    Autonomous driving system based on dual process theory and deliberate practice theory,

    Tianyu Hu, Xiao Zhang, Huimin Ma, and et al., “Autonomous driving system based on dual process theory and deliberate practice theory,” PREPRINT (Version 1), mar 2025, available at Research Square

    work page 2025

  5. [5]

    CreDes: Causal Reasoning Enhancement and Dual-End Searching for Solving Long-Range Reasoning Problems using LLMs

    Kangsheng Wang, Xiao Zhang, Hao Liu, Songde Han, Huimin Ma, and Tianyu Hu, “Credes: Causal reasoning enhancement and dual-end searching for solving long-range reasoning problems using llms,” arXiv preprint arXiv:2410.01696, 2024

    work page Pith review arXiv

  6. [6]

    Introducing openai o1-preview: A new reasoning model,

    OpenAI, “Introducing openai o1-preview: A new reasoning model,” Online. Available: https://openai.com/introducing-o1-preview/, 2024

    work page 2024

  7. [7]

    Introducing canvas: A new way to write and code with chat- gpt,

    OpenAI, “Introducing canvas: A new way to write and code with chat- gpt,” Online. Available: https://openai.com/introducing-canvas/, 2024

    work page 2024

  8. [8]

    The power of scale for parameter- efficient prompt tuning,

    L. Brian, A. Rami, and C. Noah, “The power of scale for parameter- efficient prompt tuning,” in Empirical Methods in Natural Language Processing, 2021, pp. 3045–3059

    work page 2021

Show all 29 references
  1. [9]

    Spot: Better frozen model adaptation through soft prompt transfer,

    T. Vu, B. Lester, N. Constant, R. Al-Rfou’, and D. Cer, “Spot: Better frozen model adaptation through soft prompt transfer,” in Association for Computational Linguistics , 2022, pp. 5039–5059

    2022

  2. [10]

    Multitask prompt tuning enables parameter-efficient transfer learning,

    Zhen Wang, Rameswar Panda, Leonid Karlinsky, and et al., “Multitask prompt tuning enables parameter-efficient transfer learning,” in Interna- tional Conference on Learning Representations , 2023

    2023

  3. [11]

    Efficient pareto manifold learning with low-rank structure,

    W. Chen and J.Kwok, “Efficient pareto manifold learning with low-rank structure,” in International Conference on Machine Learning , 2024

    2024

  4. [12]

    Universality and limitations of prompt tuning,

    Y . Wang, J. Chauhan, W. Wang, and C. Hsieh, “Universality and limitations of prompt tuning,” in Neural Information Processing Systems, 2024

    2024

  5. [13]

    When do prompting and prefix- tuning work? a theory of capabilities and limitations,

    P. Aleksandar, T. P. HS, and B. Adel, “When do prompting and prefix- tuning work? a theory of capabilities and limitations,” in arXiv preprint arXiv:2310.19698, 2023, unpublished

    2023 arXiv

  6. [14]

    Towards a unified view of parameter-efficient transfer learning,

    J. He, C. Zhou, X. Ma, T. Berg-Kirkpatrick, and G. Neubig, “Towards a unified view of parameter-efficient transfer learning,” in International Conference on Learning Representations , 2022

    2022

  7. [15]

    Llama- adapter: Efficient fine-tuning of large language models with zero- initialized attention,

    R. Zhang, J. Han, C. Liuand, P. Gao, A. Zhou, and Z. Hu, “Llama- adapter: Efficient fine-tuning of large language models with zero- initialized attention,” in International Conference on Learning Rep- resentations, 2024

    2024

  8. [16]

    Multitask pre-training of modular prompt for Chinese few-shot learning,

    T. Sun, Z. He, Q. Zhu, X. Qiu, and X. Huang, “Multitask pre-training of modular prompt for Chinese few-shot learning,” in Association for Computational Linguistics , 2023, pp. 11156–11172

    2023

  9. [17]

    Effective and efficient few-shot fine-tuning for vision transformers,

    J. Yang, H. Wu, J. Zhang, L. Gao, and J. Song, “Effective and efficient few-shot fine-tuning for vision transformers,” in IEEE International Conference on Multimedia and Expo , 2024, pp. 1–6

    2024

  10. [18]

    Bitfit: Simple parameter- efficient fine-tuning for transformer-based masked language-models,

    B. Z. Elad, G. Yoav, and R. Shauli, “Bitfit: Simple parameter- efficient fine-tuning for transformer-based masked language-models,” in Association for Computational Linguistics , 2022, pp. 1–9

    2022

  11. [19]

    Lst: ladder side-tuning for parameter and memory efficient transfer learning,

    Y . Sung, J. Cho, and M. Bansal, “Lst: ladder side-tuning for parameter and memory efficient transfer learning,” in Neural Information Process- ing Systems , 2024

    2024

  12. [20]

    Parameter-efficient transfer learning for nlp,

    N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. D. Laroussilhe, and A. Gesmundo, “Parameter-efficient transfer learning for nlp,” in International Conference on Machine Learning , 2019, pp. 2790–2799

    2019

  13. [21]

    Parameter- efficient multi-task fine-tuning for transformers via shared hypernet- works,

    K. M. Rabeeh, R. Sebastian, D. Mostafa, and H. James, “Parameter- efficient multi-task fine-tuning for transformers via shared hypernet- works,” in Association for Computational Linguistics , 2021, pp. 565– 576

    2021

  14. [22]

    Adapterfusion: Non-destructive task composition for transfer learning,

    P. Jonas, K. Aishwarya, Andreas R., C. Kyunghyun, and G. Iryna, “Adapterfusion: Non-destructive task composition for transfer learning,” in Association for Computational Linguistics , 2020, pp. 5068–5079

    2020

  15. [23]

    P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks,

    X. Liu, K. Ji, Y . Fu, W. Tam, Z. Du, Z. Yang, and J. Tang, “P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks,” in Association for Computational Linguistics , 2022, pp. 61–68

    2022

  16. [24]

    Bartender: A simple baseline model for task-level heterogeneous federated learning,

    Y . Yang, Y . Lu, S. Huang, S. Sirejiding, C. Liu, M. Yi, Z. Xie, Y . Ding, and H. Lu, “Bartender: A simple baseline model for task-level heterogeneous federated learning,” in IEEE International Conference on Multimedia and Expo , 2024, pp. 1–6

    2024

  17. [25]

    Enhancing multimodal sentiment analysis via learning from large language model,

    N. Pang, W. Wu, Y . Hu, K. Xu, Q. Yin, and L. Qin, “Enhancing multimodal sentiment analysis via learning from large language model,” in IEEE International Conference on Multimedia and Expo , 2024, pp. 1–6

    2024

  18. [26]

    Attempt: Parameter- efficient multi-task tuning via attentional mixtures of soft prompts,

    A. Asai, M. Salehi, M. Peters, and H. Hajishirzi, “Attempt: Parameter- efficient multi-task tuning via attentional mixtures of soft prompts,” in Empirical Methods in Natural Language Processing , 2022, pp. 6655– 6672

    2022

  19. [27]

    Qwen2 technical report,

    A. Yang, B. Yang, B. Hui, B. Zheng, B. Yu, C. Zhou, and et al., “Qwen2 technical report,” arXiv preprint arXiv:2412.15115 , 2024, unpublished

    2024 arXiv

  20. [28]

    Omni-granularity embedding network for text-to-image person retrieval,

    C. Wang, Z. Luo, and S. Li, “Omni-granularity embedding network for text-to-image person retrieval,” in IEEE International Conference on Multimedia and Expo , 2024, pp. 1–6

    2024

  21. [29]

    Csce: Boosting llm reasoning by simultaneous enhancing of causal significance and consistency,

    Kangsheng Wang, Xiao Zhang, Juntao Lyu, Tianyu Hu, and Huimin Ma, “Csce: Boosting llm reasoning by simultaneous enhancing of causal significance and consistency,” 2025. APPENDIX A INTERLAYER SIMILARITY Inspired by the findings in [11], where the similarity be- tween base model...

    2020

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.