ARIADNE: Agnostic Routing for Inference-time Adapter DyNamic sElection Enrico Cassano1,2 , Michał Brzozowski2 , Zuzanna Dubanowska2 , Paolo Mandica2 , Neo Christopher Chung2 1 University of Turin, 2 Samsung AI Center, Warsaw, Poland Correspondence: [email protected]
arXiv:2606.19079v1 [cs.AI] 17 Jun 2026
The increasing deployment of parameterefficient fine-tuning (PEFT) has led to model ecosystems in which a single backbone is paired with many task-specialized adapters. In this setting, inference-time queries often arrive without task labels, requiring the system to automatically select the most appropriate adapter from a growing and heterogeneous adapter pool. Existing routing methods either depend on access to adapter internals, such as weight decompositions or gradient-based statistics, or require additional router training, which limits scalability and portability as new adapters are added. We introduce ARIADNE, a trainingfree, adapter-agnostic routing framework for dynamic adapter selection at inference time. ARIADNE represents each adapter through a set of centroids computed from embeddings of its training set, capturing the data distribution associated with that adapter. Given an unlabeled input, it selects an adapter by measuring proximity to these centroids in latent space. Because routing is performed entirely in the input embedding space, ARIADNE is compatible with arbitrary PEFT methods and requires no modification to the adapters or training procedures. Primarily evaluated with Llama 3.2 1B Instruct on 23 diverse NLP tasks, ARIADNE recovers 97.44% of the upper bound performance. Scaling to 44 tasks, it achieves 89.7% average selection accuracy, without additional training or access to adapter internals.
1
Introduction
The proliferation of parameter-efficient fine-tuning (PEFT) methods has fundamentally altered the landscape of language model adaptation. Rather than fine-tuning monolithic models end-to-end, practitioners now maintain growing libraries of lightweight adapters (Hu et al., 2022; Houlsby et al., 2019), each specializing a shared backbone for a particular task or domain. This modular paradigm offers compelling advantages in terms of
Accuracy (%)
Abstract 100 75 50 25 0
Arrow
HellaSwag MNLI
SpectR
MRPC
ARIADNE
QQP
SST-2
Figure 1: Adapter SA comparison between ARIADNE and spectral routing methods Arrow and SpectR. ARIADNE consistently outperforms both across all tasks.
storage, compute, and composability. Yet it introduces a critical challenge: given an input without task label and a library of n specialized adapters, how does one select the most appropriate one without the overhead of additional training, labeled data, or privileged access to adapter internals? Existing approaches to this routing problem can be broadly divided into two families. The first employs retrieval-based mechanisms trained on labeled task data. LoraRetriever (Zhao et al., 2024) fine-tunes a sentence embedding model via contrastive learning to align inputs with adapter representations, achieving strong performance but requiring an additional supervised training phase and access to each adapter’s training distribution. A second family of methods uses spectral routing, deriving routing signals directly from adapter weights. Representative examples include Arrow (Ostapenko et al., 2024) and SpectR (Fleshman and Van Durme, 2025), which build prototypes from the SVD of each LoRA weight matrix and assign inputs based on the alignment between these prototypes and the model’s hidden states. Although these approaches operate in a zero-shot setting, they are strictly designed around the LoRA formulation and do not naturally generalize to other PEFT methods (Kopiczko et al., 2023; Liu et al., 2024). Moreover, empirical results show that these
methods struggle when adapters are similar, with Arrow performing close to random chance on several benchmarks (Fleshman and Van Durme, 2025). We propose ARIADNE (Agnostic Routing for Inference-time Adapter DyNamic sElection), a zero-shot routing framework for dynamic adapter selection compatible with any PEFT architecture. The key insight is framing adapter routing as an input classification problem: the latent geometry of a frozen, off-the-shelf text encoder is sufficient to distinguish task distributions, without relying on adapter weights or gradients. Inputs from the same task naturally cluster in this space, making it a reliable signal for routing. For each task, we construct a set of m representative centroids by clustering embeddings of samples drawn from its training set. At inference time, an input is projected into the same space, and the adapter whose centroids set yields the highest cosine similarity is selected. We instantiate ARIADNE on top of Llama 3.2 1B Instruct and Qwen2.5 3B Instruct, and evaluate it end-to-end on 23 diverse NLP tasks, measuring both adapter Selection Accuracy (SA) and Task Performance (TP) against an oracle upper bound. ARIADNE achieves an average TP of 54.74%, recovering 97.44% of the average oracle performance (56.18%). On the 5-task subset shared with Arrow and SpectR, ARIADNE consistently outperforms both baselines in adapter SA. An extended scalability study on 44 tasks confirms that routing performance remains stable as the adapter library grows, reaching an average SA of 89.7%. Our main contributions are: • We reframe adapter routing as an input classification problem and propose a zero-shot centroid-based routing mechanism that operates entirely in the input embedding space. • Thanks to its reliance solely on the input space rather than adapter weight decompositions, ARIADNE is, by construction, compatible with any PEFT architecture. • We show empirically that input geometry alone provides a strong signal for effective adapter selection, outperforming spectral routing methods while scaling robustly. • We conduct a systematic analysis of routing failure modes and show that errors are concentrated within semantically related task clusters, leading to graceful rather than catastrophic degradation.
2
Related Work
Parameter-Efficient Fine-Tuning. The growth of large pretrained language models has made full fine-tuning increasingly impractical. Parameterefficient fine-tuning (PEFT) methods address this by updating only a small subset of parameters while keeping the backbone frozen (Houlsby et al., 2019; Hu et al., 2022; Li and Liang, 2021; Liu et al., 2022). Among these, Low-Rank Adaptation (LoRA) (Hu et al., 2022) has emerged as the dominant paradigm: for each weight matrix W ∈ Rd×k , it introduces a residual update ∆W = BA with B ∈ Rd×r , A ∈ Rr×k , and r ≪ min(d, k), reducing the number of trainable parameters by orders of magnitude while preserving competitive downstream performance. Subsequent variants including VeRA (Kopiczko et al., 2023), DoRA (Liu et al., 2024), AdaLoRA (Zhang et al., 2023), and GPart (Mandica et al., 2026) further improve parameter efficiency and flexibility. Our work does not modify or extend any specific PEFT method; instead, it provides routing that operates independently of the adapter implementation. Adapter Selection. Existing approaches vary considerably in the resources and assumptions they require, ranging from methods that train dedicated routing components to fully zero-shot, data-free alternatives. LoRARetriever (Zhao et al., 2024) frames adapter selection as a retrieval and composition problem, but requires training a dedicated retrieval component on top of each adapter’s training data, a costly overhead that ARIADNE avoids entirely. More closely related to our approach are methods that perform zero-shot routing by exploiting the internal structure of adapter weights. ARROW (Ostapenko et al., 2024) uses the first right singular vector of each adapter’s weight product matrix, obtained via SVD, as a proxy for its training distribution, while SpectR (Fleshman and Van Durme, 2025) extends this idea by leveraging the full covariance spectrum. Both methods require white-box access to adapter internals, and their reliance on the SVD of LoRA weight matrices ties them architecturally to the LoRA family and, implicitly, to the underlying base model. Furthermore, Fleshman and Van Durme (2025) show that these spectral proxies can be unreliable: ARROW degrades to near-random routing accuracy on highly similar task pairs, while SpectR falls even below the random threshold in the same setting. Like ARROW and SpectR, ARIADNE requires
no training of additional components. Unlike these methods, however, it grounds routing entirely in the latent geometry of a frozen, off-the-shelf encoder, decoupling the routing mechanism from both adapter internals and the underlying PEFT architecture. As a result, ARIADNE is agnostic to both adapter type and base model, enabling the same routing method to transfer across model families and scales. This design is further motivated by recent evidence that internal model representations are unreliable as general-purpose routing signals, with spectral approaches often degrading out-of-distribution (Dubanowska et al., 2025).
us to characterize what happens when routing fails. Full details on tasks and training are provided in Appendix A.3. Dynamic Selection. For each task Ti , we represent its input distribution with a set of m taskrepresentative centroids Ci = {ci,j }m j=1 , computed in the embedding space of a frozen auxiliary encoder e(·). To construct these centroids, we sample m subsets Si,j ⊂ Di with different strategies (Appendix A.10), and average the embeddings of their inputs. Formally, for each j ∈ {1, . . . , m}, the j-th centroid for task Ti is defined as ci,j =
3
Methodology
Routing Without Adapter Access. A central design choice in ARIADNE is to ground routing decisions exclusively in the latent geometry of a text encoder, rather than on adapter internals. This decoupling is both principled and practical: adapter weights encode the output of a training process whose data distribution and optimization trajectory are opaque at deployment time, and weightspace signals offer no guarantee of correspondence with task boundaries in input space (Fleshman and Van Durme, 2025; Dubanowska et al., 2025). By operating only on inputs, ARIADNE offers three properties that spectral routing methods cannot provide: compatibility with any PEFT architecture by design, straightforward extension to new tasks by simply computing centroids from training samples, and independence from the underlying backbone, enabling the same routing infrastructure transfers across model families and scales. Problem Formulation. Let T = {T1 , . . . , Tn } denote a set of n tasks. Each task Ti is associi ated with a dataset Di = {(xi,k , yi,k )}N k=1 , where Ni is the number of examples in task Ti , xi,k is the k-th input, and yi,k is its corresponding label. TheSfull multi-task dataset is then defined as D = ni=1 Di . We consider a base language model L and a library of n task-specific adapters Φ = {ϕ1 , . . . , ϕn }, where each ϕi is optimized for Ti . Under a mixed-task scenario, an input x is submitted to L without a task label, and the objective is to select the adapter best suited to process it. Adapter Library. We train LoRA adapters on state-of-the-art tasks that span four semantic categories. These adapters are fundamental to our evaluation: they enable end-to-end measurement of routing quality under realistic conditions, and allow
1 |Si,j |
X
e(x).
(1)
(x,y)∈Si,j
This multi-centroid representation captures intratask variability more effectively than a single global prototype. At inference time, an unlabeled input x is embedded in the same space, and the routing function selects the adapter associated with the most similar task centroid: i∗ = arg max max cos(e(x), c) . (2) i
c∈Ci
The selected adapter is then ϕi∗ .
4
Experiments
Setup. We evaluate ARIADNE on top of Llama 3.2 1B Instruct (Grattafiori et al., 2024) and Qwen2.5 3B Instruct (Team, 2024) (results in Appendix A.2). We use a library of 23 LoRA adapters, each trained independently on a single task in the span of four semantic categories: NLI, QA, Similarity, and Reasoning, covering established benchmarks. The frozen embedder e(·) is intfloat/e5-large-v2 (Wang et al., 2022). We report the motivation for this choice in Appendix A.5. Our main results are computed with up to 500 samples per centroid and centroids number m = 5. A robustness study training samples number is in Appendix A.9, and the selection of the m value is in Appendix A.6. Comparison with Spectral Routing. We compare adapter SA against Arrow (Ostapenko et al., 2024) and SpectR (Fleshman and Van Durme, 2025) on the 5-task intersection shared across evaluations (HellaSwag (Zellers et al., 2019), MNLI (Williams et al., 2018), MRPC (Wang et al., 2018), QQP (Wang et al., 2018), SST-2 (Socher et al., 2013)).
Category
SA
Base Model
Oracle
ARIADNE
Recoup %
NLI QA Similarity Reasoning
81% 83% 96% 100%
23.37% 22.91% 19.42% 23.38%
60.16% 51.55% 66.60% 46.40%
58.06% 49.14% 65.20% 46.40%
96.51% 95.32% 97.92% 100.0%
Avg.
85%
22.27%
56.18%
54.74%
97.44%
5
Results
Comparison with Spectral Routing. Figure 1 compares ARIADNE with Arrow and SpectR on the 5-task benchmark shared across all three methods. The approaches differ in their underlying assumptions: Arrow and SpectR do not need training data but require white-box access to LoRA weight matrices, whereas ARIADNE uses training samples without requiring access to adapter internals. ARIADNE achieves the best performance on every task, with the largest gains on MRPC and QQP, where spectral methods degrade to near-random routing accuracy, consistent with their known failure mode (Fleshman and Van Durme, 2025). Adapter Selection and Performance. Table 1 summarizes results on 23 tasks grouped by semantic category. ARIADNE achieves a zero-shot average SA of 85%, translating to an average TP of 54.74% and recovering 97.44% of Oracle performance. Routing is most reliable in the Similarity and Reasoning categories, where near-perfect selection closes the gap to the Oracle. NLI is the most challenging category, with 81% SA, largely due to the high similarity. Even in this setting, however, ARIADNE recovers 96.51% of oracle performance, indicating that routing errors tend to select semantically related adapters rather than causing catastrophic degradation. Full per-task results are reported in Appendix A.1. In Appendix A.2, are reported the end-to-end performances with an additional backbone, Qwen2.5 3B Instruct. We report the overhead of ARIADNE in Appendix A.7.
Scalability. SA degrades gracefully as the adapter library grows, stabilizing at an average SA of 89.7% across 44 tasks. As shown in Figure 2, the initial drop reflects the increasing difficulty of distinguishing semantically proximate tasks as the candidate pool expands. Crucially, performance plateaus beyond approximately 20 adapters rather than continuing to degrade. This trend is consistent with our failure mode analysis, which shows that routing errors are concentrated within semantic clusters. The SA for each task is reported in Appendix A.8.
Accuracy
Adapter Selection and End-to-End Performance. For each of the 23 selected tasks, we assess the SA on 50 random test samples. We then evaluate the pipeline end-to-end by testing the adapter selected by ARIADNE and measuring the TP against an Oracle. This Oracle is defined as a configuration that always uses the correct adapter. The gap between these two metrics directly quantifies the performance cost attributable to routing errors. Finally, to assess scalability, we report SA across an extended set of 44 tasks.
Table 1: End-to-end performance across 23 tasks grouped by semantic category. SA: adapter Selection Accuracy; TP: Task Performance; Recoup: TP recovered relative to Oracle.
1.0 0.5 0.0
1 6 11 16 21 26 31 36 41 Number of Tasks
Figure 2: SA trend for up to 44 tasks.
Graceful Degradation. Severe routing failure occurs on SQuAD V1 (Rajpurkar et al., 2016), which achieves 0% SA because it is consistently routed to the SQuAD V2 (Rajpurkar et al., 2018) adapter. ARIADNE still recovers 85% of Oracle performance (64% vs 75% TP), since the two tasks are semantically close and their adapters transfer well. This behavior highlights a major advantage of ARIADNE: since routing is performed in a semantically-structured space, errors tend to result in graceful degradation, making routing errors easier to diagnose (Appendix A.4).
6
Conclusions
We introduced ARIADNE, a zero-shot framework for dynamic adapter selection that reframes routing as an input classification problem. By operating in the latent space of a frozen text encoder rather than relying on adapter weight decompositions, ARIADNE is, by construction, compatible with arbitrary PEFT architectures. Evaluated across 23 tasks, it recovers 97.44% of Oracle performance and outperforms spectral routing methods on shared tasks, suggesting that input geometry provides an effective signal for adapter selection.
Limitations The primary limitation of ARIADNE is that it requires access to training data to compute task centroids. As a result, it is not directly applicable to decentralized adapter ecosystems where training data is proprietary or unavailable. A promising direction for future work is to combine ARIADNE with knowledge extraction in order to derive taskdescriptive signals directly from the adapters, enabling the construction of input-space fingerprints without requiring access to training samples. This would make our approach fully data-free while preserving its adapter-agnostic nature. At the same time, in many practical PEFT settings, adapters are released together with at least some information about their training data or task domain.
References
tated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632–642, Lisbon, Portugal. Association for Computational Linguistics. Daniel Cer, Mona Diab, Eneko Agirre, Inigo LopezGazpio, and Lucia Specia. 2017. Semeval-2017 task 1: Semantic textual similarity multilingual and crosslingual focused evaluation. In Proceedings of the 11th international workshop on semantic evaluation (SemEval-2017), pages 1–14. Clark Christopher, Lee Kenton, Chang Ming-Wei, Kwiatkowski Tom, Collins Michael, and Toutanova Kristina. 2019. Boolq: Exploring the surprising difficulty of natural yes/no questions. In NAACL. Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv:1803.05457v1.
Marta Bañón, Pinzhen Chen, Barry Haddow, Kenneth Heafield, Hieu Hoang, Miquel Esplà-Gomis, Mikel L. Forcada, Amir Kamran, Faheem Kirefu, Philipp Koehn, Sergio Ortiz Rojas, Leopoldo Pla Sempere, Gema Ramírez-Sánchez, Elsa Sarrías, Marek Strelec, Brian Thompson, William Waites, Dion Wiggins, and Jaume Zaragoza. 2020. ParaCrawl: Web-scale acquisition of parallel corpora. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4555–4567, Online. Association for Computational Linguistics.
Yiheng Dong, Yi Lin, and Xin Yang. 2025. Copa: Hierarchical concept prompting and aggregating network for explainable diagnosis. In International Conference on Medical Image Computing and ComputerAssisted Intervention, pages 67–76. Springer.
Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. 2020. Piqa: Reasoning about physical commonsense in natural language. In ThirtyFourth AAAI Conference on Artificial Intelligence.
Zuzanna Dubanowska, Maciej Żelaszczyk, Michał Brzozowski, Paolo Mandica, and Michal P. Karpowicz. 2025. Representation-based broad hallucination detectors fail to generalize out of distribution. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 17563–17575, Suzhou, China. Association for Computational Linguistics.
Ond rej Bojar, Rajen Chatterjee, Christian Federmann, Yvette Graham, Barry Haddow, Matthias Huck, Antonio Jimeno Yepes, Philipp Koehn, Varvara Logacheva, Christof Monz, Matteo Negri, Aurelie Neveol, Mariana Neves, Martin Popel, Matt Post, Raphael Rubino, Carolina Scarton, Lucia Specia, Marco Turchi, and 2 others. 2016. Findings of the 2016 conference on machine translation. In Proceedings of the First Conference on Machine Translation, pages 131–198, Berlin, Germany. Association for Computational Linguistics. Ondrej Bojar, Christian Buck, Christian Federmann, Barry Haddow, Philipp Koehn, Johannes Leveling, Christof Monz, Pavel Pecina, Matt Post, Herve SaintAmand, Radu Soricut, Lucia Specia, and Ale s Tamchyna. 2014. Findings of the 2014 workshop on statistical machine translation. In Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 12–58, Baltimore, Maryland, USA. Association for Computational Linguistics. Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large anno-
Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. 2019. DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs. In Proc. of NAACL.
Ondřej Dušek, Jekaterina Novikova, and Verena Rieser. 2020. Evaluating the State-of-the-Art of End-to-End Natural Language Generation: The E2E NLG Challenge. Computer Speech & Language, 59:123–156. William Fleshman and Benjamin Van Durme. 2025. SpectR: Dynamically composing LM experts with spectral routing. arXiv preprint arXiv:2504.03454. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad AlDahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783. Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790–2799. PMLR.
Edward J Hu, yelong shen, Phillip Wallis, Zeyuan AllenZhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations. Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. triviaqa: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension. arXiv e-prints, arXiv:1705.03551. Daniel Khashabi, Snigdha Chaturvedi, Michael Roth, Shyam Upadhyay, and Dan Roth. 2018. Looking beyond the surface:a challenge set for reading comprehension over multiple sentences. In Proceedings of North American Chapter of the Association for Computational Linguistics (NAACL). Dawid J Kopiczko, Tijmen Blankevoort, and Yuki M Asano. 2023. Vera: Vector-based random matrix adaptation. arXiv preprint arXiv:2310.11454. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Matthew Kelcey, Jacob Devlin, Kenton Lee, Kristina N. Toutanova, Llion Jones, Ming-Wei Chang, Andrew Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association of Computational Linguistics. Hector Levesque, Ernest Davis, and Leora Morgenstern. 2012. The winograd schema challenge. In Thirteenth International Conference on the Principles of Knowledge Representation and Reasoning. Citeseer. Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582– 4597. Bill Yuchen Lin, Wangchunshu Zhou, Ming Shen, Pei Zhou, Chandra Bhagavatula, Yejin Choi, and Xiang Ren. 2020. CommonGen: A constrained text generation challenge for generative commonsense reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1823–1840, Online. Association for Computational Linguistics. Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin A Raffel. 2022. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. Advances in Neural Information Processing Systems, 35:1950–1965. Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. 2024. Dora: Weightdecomposed low-rank adaptation. In Forty-first International Conference on Machine Learning.
Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142–150, Portland, Oregon, USA. Association for Computational Linguistics. Paolo Mandica, Michał Brzozowski, Zuzanna Dubanowska, and Neo Christopher Chung. 2026. Gpart: End-to-end isometric fine-tuning via global parameter partitioning. Preprint, arXiv:2605.14841. Linyong Nan, Dragomir Radev, Rui Zhang, Amrit Rau, Abhinand Sivaprasad, Chiachun Hsieh, Xiangru Tang, Aadit Vyas, Neha Verma, Pranav Krishna, Yangxiaokang Liu, Nadia Irwanto, Jessica Pan, Faiaz Rahman, Ahmad Zaidi, Mutethia Mutuma, Yasin Tarabar, Ankit Gupta, Tao Yu, and 5 others. 2021. DART: Open-domain structured data record to text generation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 432–447, Online. Association for Computational Linguistics. Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial nli: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics. Oleksiy Ostapenko, Zhan Su, Edoardo Maria Ponti, Laurent Charlin, Nicolas Le Roux, Matheus Pereira, Lucas Caccia, and Alessandro Sordoni. 2024. Towards modular llms by building and reusing a library of loras. arXiv preprint arXiv:2405.11157. Altaf Rahman and Vincent Ng. 2012. Resolving complex cases of definite pronouns: the winograd schema challenge. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 777–789. Association for Computational Linguistics. Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don’t know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784–789, Melbourne, Australia. Association for Computational Linguistics. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392, Austin, Texas. Association for Computational Linguistics. Rishi Sharma, James Allen, Omid Bakhshandeh, and Nasrin Mostafazadeh. 2018. Tackling the story ending biases in the story cloze test. In Proceedings
of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 752–757, Melbourne, Australia. Association for Computational Linguistics. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631–1642, Seattle, Washington, USA. Association for Computational Linguistics. Qwen Team. 2024. Qwen2.5: A party of foundation models. Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. Glue: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP workshop BlackboxNLP: Analyzing and interpreting neural networks for NLP, pages 353– 355. Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weaklysupervised contrastive pre-training. arXiv preprint arXiv:2212.03533. Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual e5 text embeddings: A technical report. arXiv preprint arXiv:2402.05672. Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). Association for Computational Linguistics. Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. Hellaswag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023. Adalora: Adaptive budget allocation for parameter-efficient finetuning. arXiv preprint arXiv:2303.10512. Sheng Zhang, Xiaodong Liu, Jingjing Liu, Jianfeng Gao, Kevin Duh, and Benjamin Van Durme. 2018. Record: Bridging the gap between human and machine commonsense reading comprehension. arXiv preprint arXiv:1810.12885. Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. Advances in neural information processing systems, 28.
Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. 2025. Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176. Yuan Zhang, Jason Baldridge, and Luheng He. 2019. PAWS: Paraphrase Adversaries from Word Scrambling. In Proc. of NAACL. Ziyu Zhao, Leilei Gan, Guoyin Wang, Wangchunshu Zhou, Hongxia Yang, Kun Kuang, and Fei Wu. 2024. Loraretriever: Input-aware lora retrieval and composition for mixed tasks in the wild. In Findings of the Association for Computational Linguistics: ACL 2024, pages 4447–4462.
A
Appendix
A.1
Full Results for adapters oracle and dynamic selection
In Table 2, we report, for each task, the performance evaluated on Llama 3.2 1B Instruct (Base), the Oracle (the base model with the adapter trained on that specific task) and the base model with the Dynamically selected adapter through ARIADNE. For all the tasks, the performance is computed as Exact Match, with the exception of the CommonGen task, measured with Rouge. Dataset
Base
Oracle
Dynamic
WSC DPR PAWS WIKI Trivia QA QQP MRPC STS-B CB WNLI ANLI R1 ANLI R2 ANLI R3 MNLI Matched MNLI Mismatched SNLI RTE CommonGen SQuAD V1 SQuAD V2 BoolQ NQ ARC-Easy ARC-Challenge
0.0200 0.3365 0.4580 0.1400 0.0000 0.2647 0.0000 0.3214 0.4225 0.1560 0.2100 0.2420 0.1920 0.1980 0.2980 0.2816 0.2470 0.5420 0.1700 0.3700 0.0340 0.1060 0.1137
0.5133 0.6154 0.8980 0.3000 1.0000 0.6961 0.1100 0.4286 0.4366 0.3780 0.4360 0.4360 0.8520 0.8580 0.8840 0.5560 0.4117 0.7480 0.6480 0.7120 0.1140 0.5580 0.3311
0.5133 0.6154 0.8838 0.3000 1.0000 0.6961 0.1100 0.4268 0.4104 0.4240 0.3928 0.4318 0.7232 0.8580 0.8840 0.5560 0.4117 0.6398 0.6221 0.7045 0.1140 0.5418 0.3311
Average
0.2227
0.5618
0.5474
Table 2: Full table for the TP comparison between the 23 selected tasks for Llama3.2 1B Instruct. For all the tasks, the metric employed is Exact Match, with the exception of the CommonGen task, which is measured with Rouge. This table expands the results presented in Table 1
A.2
Performances on other backbones
Even if ARIADNE is completely decoupled from the backbone and the adapters architecture, for experimental completeness, we report the full pipeline performances for Qwen2.5 3B Instruct (Team, 2024) in Table 3. A.3
Adapters Training Details
The adapters were instantiated with a bottleneck rank of r = 64 and a scaling factor α = 128. Training was performed for 3 epochs with a batch size of 4 and a learning rate of 5 × 10−5 , while
Dataset
Base
Oracle
Dynamic
WSC DPR PAWS WIKI Trivia QA QQP MRPC STS-B CB WNLI ANLI R1 ANLI R2 ANLI R3 MNLI Matched MNLI Mismatched SNLI RTE CommonGen SQuAD V1 SQuAD V2 BoolQ NQ ARC-Easy ARC-Challenge
0.6000 0.6200 0.7400 0.1600 0.7000 0.6400 0.0000 0.3400 0.3400 0.5400 0.3400 0.2600 0.7000 0.7600 0.8800 0.5800 0.3130 0.7200 0.2400 0.7000 0.1000 0.8600 0.8400
0.6200 0.6200 0.9000 0.2000 1.0000 0.6400 0.0200 0.2800 0.3400 0.6600 0.4400 0.5800 0.8800 0.9000 0.9200 0.8600 0.4358 0.8000 0.6600 0.8600 0.1600 0.9000 0.9000
0.6200 0.6200 0.8280 0.2000 1.0000 0.6016 0.0200 0.2352 0.3196 0.3036 0.1672 0.5452 0.6336 0.9000 0.9200 0.8600 0.4358 0.7300 0.5940 0.7224 0.1600 0.8280 0.9000
Average
0.5206
0.6337
0.5715
Table 3: Full table for the TP comparison between the 23 selected tasks for Qwen2.5 3B model. For all the tasks, the metric employed is Exact Match, with the exception of the CommonGen task, which is measured with Rouge.
all backbone parameters remained frozen. For the CommonGen task, ROUGE is computed using the HuggingFace evaluate library (version 0.4.6). A.4
Analysis of Routing Failures
Despite high overall precision, we identify three failure modes that inform the system’s limitations: (1) Domain overlap: SQuAD V1 achieves 0% selection accuracy as it is consistently misrouted to the SQuAD V2 adapter. However, the system recovers 85% of Oracle performance (0.64 vs 0.75 TP), demonstrating that semantically proximate adapters can often absorb routing errors. (2) Adversarial variance: ANLI R1 and R2 exhibit lower accuracy (0.46 and 0.38) due to high intra-task embedding variance from adversarial construction. Interestingly, ANLI R1’s downstream TP actually exceeds the target (0.42 vs 0.38), suggesting beneficial cross-task generalization from other NLI adapters. (3) Reasoning ambiguity: Complex tasks like MultiRC are frequently misrouted to general QA centroids (TriviaQA) due to shared linguistic surface features. Across all modes, errors are concentrated within the same semantic cluster, leading to "graceful degradation" rather than catastrophic
t-SNE Visualization of Task Embeddings 0.6 40
0.5 20
0.3
0.2
0.1
t-SNE Dimension 2
0.4
0
20
40
0.0 60
Tasks
Figure 3: Pairwise distances between different tasks centroids. Visual interpretation of the Graceful Degradation. Model
CoLA
RTE
MRPC
SST-2
Avg.
MiniLM† Qwen3-0.6B‡ mE5-small§ E5-large-v2¶
0.111 0.532 0.860 0.910
0.090 0.400 0.848 0.877
0.271 0.415 0.868 0.902
0.433 0.686 0.881 0.891
0.226 0.508 0.864 0.895
Table 4: Similarity scores across GLUE tasks centroids and test samples. Bold = best, justifying our embedder choice. † all-MiniLM-L12-v2; ‡ Qwen/Qwen3-Embedding-0.6B (Zhang et al., 2025); § intfloat/multilingual-e5-small (Wang et al., 2024); ¶ intfloat/e5-large-v2 (Wang et al., 2022).
failure. A visual representation of this is reported in Figure 3, where it’s shown how similar tasks often achieve the highest centroids similarity, meaning that their representations are the closest compared to other tasks. A.5
qnli squad_v2 anli arc copa hellaswag cosmosqa obqa rte boolq triviaqa squad imdb drop cola stsb snli mrpc sst2
60
a arc annli_r1 _ch anli_r l 2 araclleni_r3 _ g de co beoasey fin mm o ite on clq _pr on co c_gebn ou sm op n_r os a eso d qa lut art io h e2 dr n mn mn ellaes_nolp li_mli_m wag ismaticmdgb na h tur atched al_ mm ed q rp p uesulti c paaractionrcs ws raw _w l piqiki qna recqqpli o s rd s q sto uanli ry_ ss d clo t2 wm wm t1 tr stsze w t1 6_t w w ivi -b w mt6_t ran mt1eb_aqa wmmt1166_rtanssllate4_ennlg r a a _ t wm 16 tra ns te__csefr t16_transl latedeen _t ns ate _fi n ye ransllate__roeen lp_ a n po te_rtuen lar r ity wenn _re w li vie sc ws
Tasks
Pairwise Centroid Distances anli_r1 anli_r2 anli_r3 arc_challenge arc_easy boolq cb common_gen copa cosmosqa dart definite_pronoun_resolution drop e2e_nlg hellaswag imdb mnli_matched mnli_mismatched mrpc multirc natural_questions paracrawl paws_wiki piqa qnli qqp record snli squad sst2 story_cloze sts-b triviaqa web_nlg wmt14_enfr wmt16_translate_csen wmt16_translate_deen wmt16_translate_fien wmt16_translate_roen wmt16_translate_ruen wmt16_translate_tren wnli wsc yelp_polarity_reviews
60
40
20
0
t-SNE Dimension 1
20
40
60
Figure 4: T-SNE analysis of the tasks embeddings. Note that this visualization includes tasks from the full 44task pool used in the scalability study.
intra-task embedding variance. Adversarially constructed tasks such as ANLI produce inputs whose embeddings span multiple disjoint regions of the latent space, as illustrated in Figure 4, while tasks with heterogeneous input formats, such as extractive vs. abstractive question answering, may cluster around semantically distinct centroids even within the same task. A single global mean collapses this structure and produces a centroid that may lie in a low-density region, making it a poor representative of any individual input. By partitioning each task’s distribution into local centroids, ARIADNE captures this multimodality explicitly. K-NN routing, by contrast, lacks task-level structure entirely and conflates inter-task proximity with intra-task variance, explaining its intermediate performance in Table 5.
Embedders search
Metric SA
1 Centroid
5 Centroids
K-NN
47.4%
89.4%
65.9%
Due to the large number of existing text embedders, our selection is based upon a search on a subset of tasks. We pick as most suitable encoder the one that yields the highest cosine similarity between the generated task representation, i.e. the centroids, and 20 test set samples. We perform this study on the GLUE tasks (Wang et al., 2018). The results determining the selection are reported in Table 4.
Table 5: The reported results in the table show the selection accuracy achieved over 19 tasks, as a way to empirically choose the best setup for dynamic selection. As reported, the sweetspot is in the middle of having a single centroid for each task, and having no centroid as per K-NN selection strategy.
A.6
A.7
Number of centroids ablation study
To decide the correct number of centroids and selection accuracy, we tested on a subset of 19 tasks, the selection accuracy provided by each method. The multi-centroid formulation is motivated by the observation that many NLP tasks exhibit substantial
Dynamic Selection Inference Cost
Since our method requires choosing an adapter at inference time, it is important to quantify the associated overhead. The embedding step, which uses the intfloat/e5-large-v2 model, takes on average 20.04 ms with a standard deviation of ±3.70
Accuracy Score
1.0 0.8 0.6 0.4 0.2 0.0
Dataset 100
50
Training Samples (%)
10 2
Figure 5: SA trend with less training samples. The best performance is achieved with 500 samples, as reported in the main paper. Here, we show that even with a small fraction (2%) of training samples we can still achieve good SA performances.
ms. The subsequent selection of the most appropriate adapter, based on the pre-computed centroids, adds another 1.98 ms (±0.10 ms). Consequently, the complete selection pipeline incurs a total latency of roughly 22.02 ms (±3.66 ms). After this brief selection phase, inference proceeds as usual by applying the chosen adapter to the base model. A.8
Full Routing results
To evaluate whether ARIADNE’s routing precision degrades as the adapter library grows, we extend the evaluation to the full set of 44 tasks, adding 21 tasks to the primary evaluation. For these additional tasks, Table 6 reports per-task selection accuracy across all 44 tasks. A.9
Training Samples Robustness
We hereby show that drastically reducing the number of training samples used to generate the centroids has a mild effect on the SA, as shown in Figure 5. When using only 2% of the original number of training samples, the SA is 77.1%. This plot shows that ARIADNE is well performing even when a small number of samples is available. A.10
Selection Strategies
When generating the 5 centroids for each task, we employ 3 strategies: beginning, end, and random. They respectively pick the n samples from the beginning of the training set, the end of it, or pick random samples. In our main study, we used n = 500, with one centroid for beginning, one centroid for end, and 3 random samplings. When the number of training samples for a task is smaller than n, we use only one centroid.
Acc.
Primary evaluation (23 tasks) WSC (Levesque et al., 2012) DPR (Rahman and Ng, 2012) PAWS Wiki (Zhang et al., 2019) Trivia QA (Joshi et al., 2017) QQP (Wang et al., 2018) MRPC (Wang et al., 2018) STS-B (Cer et al., 2017) CB (Wang et al., 2018) WNLI (Wang et al., 2018) BoolQ (Christopher et al., 2019) NQ (Kwiatkowski et al., 2019) ARC-Challenge (Clark et al., 2018) ANLI R1 (Nie et al., 2020) ANLI R2 (Nie et al., 2020) ANLI R3 (Nie et al., 2020) MNLI Matched (Williams et al., 2018) MNLI Mismatched (Williams et al., 2018) SNLI (Bowman et al., 2015) RTE (Wang et al., 2018) CommonGen (Lin et al., 2020) SQuAD V1 (Rajpurkar et al., 2016) SQuAD V2 (Rajpurkar et al., 2018) ARC-Easy (Clark et al., 2018)
1.00 1.00 0.92 1.00 1.00 0.94 1.00 0.84 0.94 0.96 1.00 1.00 0.46 0.38 0.94 0.72 1.00 1.00 1.00 1.00 0.00 0.94 0.92
Scalability extension (21 tasks) WMT16 Ro-En (Bojar et al., 2016) DART (Nan et al., 2021) DROP (Dua et al., 2019) ParaCrawl (Bañón et al., 2020) Story Cloze (Sharma et al., 2018) HellaSwag (Zellers et al., 2019) PIQA (Bisk et al., 2020) WMT16 Fi-En (Bojar et al., 2016) WMT16 Tr-En (Bojar et al., 2016) WMT16 Ru-En (Bojar et al., 2016) WMT16 De-En (Bojar et al., 2016) IMDB (Maas et al., 2011) WMT14 En-Fr (Bojar et al., 2014) MultiRC (Khashabi et al., 2018) E2E NLG (Dušek et al., 2020) ReCoRD (Zhang et al., 2018) SST-2 (Socher et al., 2013) WMT16 Cs-En (Bojar et al., 2016) Yelp Reviews (Zhang et al., 2015) COPA (Dong et al., 2025) QNLI (Wang et al., 2018)
0.96 1.00 0.70 0.80 1.00 1.00 0.98 1.00 1.00 1.00 1.00 0.96 0.92 0.56 1.00 0.94 0.98 1.00 0.86 1.00 0.88
Average (44 tasks)
0.897
Table 6: Adapter selection accuracy across all 44 tasks. The upper block corresponds to the 23 primary evaluation tasks; the lower block constitutes the scalability extension. The average is computed over all 44 tasks.