AdaMix, a parameter-efficient fine-tuning method, outperforms full model fine-tuning in few-shot NLU tasks across benchmarks like GLUE. Using prompt-based strategies without extra validation or unlabeled data, AdaMix consistently boosts performance with both BERT and RoBERTa encoders, demonstrating stability and efficiency in few-shot scenarios.AdaMix, a parameter-efficient fine-tuning method, outperforms full model fine-tuning in few-shot NLU tasks across benchmarks like GLUE. Using prompt-based strategies without extra validation or unlabeled data, AdaMix consistently boosts performance with both BERT and RoBERTa encoders, demonstrating stability and efficiency in few-shot scenarios.

Smarter AI Training with Few-Shot Natural Language Tasks

2025/10/02 17:00

Abstract and 1. Introduction

  1. Background

    2.1 Mixture-of-Experts

    2.2 Adapters

  2. Mixture-of-Adaptations

    3.1 Routing Policy

    3.2 Consistency regularization

    3.3 Adaptation module merging and 3.4 Adaptation module sharing

    3.5 Connection to Bayesian Neural Networks and Model Ensembling

  3. Experiments

    4.1 Experimental Setup

    4.2 Key Results

    4.3 Ablation Study

  4. Related Work

  5. Conclusions

  6. Limitations

  7. Acknowledgment and References

Appendix

A. Few-shot NLU Datasets B. Ablation Study C. Detailed Results on NLU Tasks D. Hyper-parameter

A Few-shot NLU Datasets

Data. In contrast to the fully supervised setting in the above experiments, we also perform fewshot experiments following the prior study (Wang et al., 2021) on six tasks including MNLI (Williams et al., 2018), RTE (Dagan et al., 2005; Bar Haim et al., 2006; Giampiccolo et al., 2007; Bentivogli et al., 2009), QQP[1] and SST-2 (Socher et al.). The results are reported on their development set following (Zhang et al., 2021). MPQA (Wiebe et al., 2005) and Subj (Pang and Lee, 2004) are used for polarity and subjectivity detection, where we follow (Gao et al., 2021) to keep 2, 000 examples for testing. The few-shot model only has access to |K| labeled samples for any task. Following true few-shot learning setting (Perez et al., 2021; Wang et al., 2021), we do not use any additional validation set for any hyper-parameter tuning or early stopping. The performance of each model is reported after fixed number of training epochs. For a fair comparison, we use the same set of few-shot labeled instances for training as in (Wang et al., 2021). We train each model with 5 different seeds and report average performance with standard deviation across the runs. In the few-shot experiments, we follow (Wang et al., 2021) to train AdaMix via the prompt-based fine-tuning strategy. In contrast to (Wang et al., 2021), we do not use any unlabeled data.

\

B Ablation Study

\ Table 11: Ablation study demonstrating the impact of parameter sharing in AdaMix adapter framework.

\

C Detailed Results on NLU Tasks

The results on NLU tasks are included in Table 1 and Table 13. The performance AdaMix with RoBERTa-large encoder achieves the best performance in terms of different task metrics in the GLUE benchmark. AdaMix with adapters is the

\ \ Table 12: Varying the bottleneck dimension of adapters in AdaMix with BERT-base and RoBERTa-large encoder. * denotes the bottleneck dimension used in AdaMix with adapters.

\ \ only PEFT method which outperforms full model fine-tuning on all the tasks and on average score. Additionally, the improvement brought by AdaMix is more significant with BERT-base as the encoder, demonstrating 2.2% and 1.2% improvement over the performance of full model fine-tuning and the best performing baseline UNIPELT with BERTbase. The improvement is observed to be consistent as that with RoBERTa-large on every task. The NLG results are included in Table 4 and 5.

D Hyper-parameter

Detailed hyper-parameter configuration for different tasks presented in Table 15 and Table 16.

\

:::info Authors:

(1) Yaqing Wang, Purdue University ([email protected]);

(2) Sahaj Agarwal, Microsoft ([email protected]);

(3) Subhabrata Mukherjee, Microsoft Research ([email protected]);

(4) Xiaodong Liu, Microsoft Research ([email protected]);

(5) Jing Gao, Purdue University ([email protected]);

(6) Ahmed Hassan Awadallah, Microsoft Research ([email protected]);

(7) Jianfeng Gao, Microsoft Research ([email protected]).

:::


:::info This paper is available on arxiv under CC BY 4.0 DEED license.

:::

[1] https://www.quora.com/q/quoradata/

Sorumluluk Reddi: Bu sitede yeniden yayınlanan makaleler, halka açık platformlardan alınmıştır ve yalnızca bilgilendirme amaçlıdır. MEXC'nin görüşlerini yansıtmayabilir. Tüm hakları telif sahiplerine aittir. Herhangi bir içeriğin üçüncü taraf haklarını ihlal ettiğini düşünüyorsanız, kaldırılması için lütfen [email protected] ile iletişime geçin. MEXC, içeriğin doğruluğu, eksiksizliği veya güncelliği konusunda hiçbir garanti vermez ve sağlanan bilgilere dayalı olarak alınan herhangi bir eylemden sorumlu değildir. İçerik, finansal, yasal veya diğer profesyonel tavsiye niteliğinde değildir ve MEXC tarafından bir tavsiye veya onay olarak değerlendirilmemelidir.

Ayrıca Şunları da Beğenebilirsiniz

Polygon Tops RWA Rankings With $1.1B in Tokenized Assets

Polygon Tops RWA Rankings With $1.1B in Tokenized Assets

The post Polygon Tops RWA Rankings With $1.1B in Tokenized Assets appeared on BitcoinEthereumNews.com. Key Notes A new report from Dune and RWA.xyz highlights Polygon’s role in the growing RWA sector. Polygon PoS currently holds $1.13 billion in RWA Total Value Locked (TVL) across 269 assets. The network holds a 62% market share of tokenized global bonds, driven by European money market funds. The Polygon POL $0.25 24h volatility: 1.4% Market cap: $2.64 B Vol. 24h: $106.17 M network is securing a significant position in the rapidly growing tokenization space, now holding over $1.13 billion in total value locked (TVL) from Real World Assets (RWAs). This development comes as the network continues to evolve, recently deploying its major “Rio” upgrade on the Amoy testnet to enhance future scaling capabilities. This information comes from a new joint report on the state of the RWA market published on Sept. 17 by blockchain analytics firm Dune and data platform RWA.xyz. The focus on RWAs is intensifying across the industry, coinciding with events like the ongoing Real-World Asset Summit in New York. Sandeep Nailwal, CEO of the Polygon Foundation, highlighted the findings via a post on X, noting that the TVL is spread across 269 assets and 2,900 holders on the Polygon PoS chain. The Dune and https://t.co/W6WSFlHoQF report on RWA is out and it shows that RWA is happening on Polygon. Here are a few highlights: – Leading in Global Bonds: Polygon holds 62% share of tokenized global bonds (driven by Spiko’s euro MMF and Cashlink euro issues) – Spiko U.S.… — Sandeep | CEO, Polygon Foundation (※,※) (@sandeepnailwal) September 17, 2025 Key Trends From the 2025 RWA Report The joint publication, titled “RWA REPORT 2025,” offers a comprehensive look into the tokenized asset landscape, which it states has grown 224% since the start of 2024. The report identifies several key trends driving this expansion. According to…
Paylaş
BitcoinEthereumNews2025/09/18 00:40