Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Under equal trainable capacity, splitting tasks into separate QLoRA adapters beats one shared adapter.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 03:46 UTC pith:2XM5GHVW

load-bearing objection Plausible capacity-matched multi-policy PEFT result, but the abstract overstates consistency and the grouping mechanism is not validated; worth reviewing with conditions. the 4 major comments →

arxiv 2607.29601 v1 pith:2XM5GHVW submitted 2026-07-31 cs.LG

The Parts Are Greater Than the Sum: Automated Task Sequencing for Efficient Training of Multi-Policy LLMs

Jiajia Tang , Sizhe Yuen , Francisco Gomez Medina , Yali Du , Adam Sobey This is my paper
classification cs.LG
keywords parameter-efficient fine-tuningLoRAQLoRAtask groupingtask sequencingcatastrophic forgettingcontinual learningoptimization interference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that the main obstacle in parameter-efficient fine-tuning of large language models is not adapter capacity but how heterogeneous tasks are organized before training. It proposes a multi-policy framework that automatically groups tasks by optimization compatibility and sequences them within each group, implemented as independent QLoRA adapters under a fixed total rank. On the eight-task TRACE benchmark, the method reaches an overall performance of 44.78 on LLaMA-2-7B-Chat, above the 42.12 of a single shared rank-128 adapter, with the same trainable budget. The same pattern holds on Vicuna-7B-V1.5 (41.14 vs 39.50). The takeaway is that where tasks share an optimization path matters as much as how much capacity that path has.

Core claim

On the paper's own terms, the central discovery is that the organization of adaptation paths, not adapter expressiveness, is the decisive factor in heterogeneous sequential fine-tuning. By partitioning eight TRACE tasks into two groups via a fused distance (normalized gradient distance computed on a temporary shared model plus a hand-crafted behavior distance), then ordering each group with a composite path objective, the method trains two independent rank-64 QLoRA adapters whose total rank equals the single rank-128 baseline. This yields OP 44.78 (BWT +0.013) on LLaMA-2-7B-Chat versus 42.12 (BWT −0.041) for the shared adapter, and 41.14 vs 39.50 on Vicuna. The automated grouping and sequenc

What carries the argument

The load-bearing mechanism is the optimization-path organization performed in two stages before training. Stage 1 computes a pairwise fused task distance D_ij = (1−λ)D_grad + λD_beh, where gradient distance comes from PCA-compressed gradients of a temporary shared model and behavior distance from normalized dataset statistics; balanced clustering then partitions tasks into K groups, each assigned an independent QLoRA adapter. Stage 2 scores candidate permutations within each group with a composite objective C = C_trans + C_tail + C_dir + C_prog − R_head − R_transfer, capturing adjacent transition cost, head-task protection, tail-task safety, directional interference, positive transfer, and g

Load-bearing premise

The paper assumes that the gradient distances collected on a temporary shared model, combined with hand-picked behavior statistics, correctly predict which tasks will actually interfere during sequential QLoRA fine-tuning; if that mapping is wrong, the automatic groups and sequences lose their claimed advantage.

What would settle it

Train the same two rank-64 QLoRA adapters with random grouping but the automatic per-group sequences, and with automatic grouping but random sequences; if either variant holds OP close to 44.78 on LLaMA-2-7B-Chat, the contribution attributed to that stage is not load-bearing. A sharper test: swap the fused distance for a plain behavior-only distance while keeping all other settings fixed — if OP does not drop, the gradient term is superfluous.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A fixed parameter budget for PEFT is better spent as several smaller, decoupled adapters than as one larger shared adapter when tasks are heterogeneous.
  • Automatic organization can replace manual task-ordering search in continual fine-tuning, since the automated pipeline matched or slightly exceeded an expert-designed grouping and sequence.
  • Balanced rank allocation between policies matters: the 64+64 split outperformed 32+96 and 96+32 under the same total rank, suggesting symmetric budgets are a safe default.
  • Measured backward transfer can be positive (BWT +0.013 on LLaMA-2), meaning later tasks can help rather than harm earlier ones when paths are organized compatibly.

Where Pith is reading between the lines

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

  • The same grouping-and-sequencing principle should transfer to other PEFT mechanisms besides QLoRA — prompt tuning, prefix tuning, or additive adapters — whenever tasks are trained sequentially; the paper's machinery is agnostic to the update rule beyond the gradient signal.
  • If the compatibility measure is the right one, then a cheap proxy for D_ij could allow online grouping for streaming task arrivals, relaxing the paper's pre-deployment assumption that all tasks are known in advance.
  • A natural stress test: replace the proposed gradient distance with a purely behavioral distance or with final-adapter overlap, and measure whether grouping quality (and final OP) degrades; the paper does not compare compatibility measures, so this is an open question.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes an optimization-path organization framework for PEFT of LLMs, implemented as a multi-policy QLoRA architecture. In Stage 1, tasks are automatically grouped into independent adapters using a fused gradient/behavior distance matrix and balanced clustering. In Stage 2, each group is trained according to an automatically constructed task sequence obtained by minimizing a composite path objective. Experiments on TRACE with LLaMA-2-7B-Chat and Vicuna-7B-V1.5 compare the automatic method against single-policy, random, manual, and O-LoRA baselines under a matched total LoRA rank. The authors report best OP of 44.78 on LLaMA-2 and 41.14 on Vicuna, and argue that the gain comes from organizing optimization paths rather than increasing adapter capacity.

Significance. If the central mechanism is validated, the paper offers a useful and relatively simple design principle: task organization, not adapter capacity, is the key lever for heterogeneous sequential PEFT. The capacity-matched experimental design is a strength, as is the inclusion of both automatic and expert/manual organization conditions. However, the proposed compatibility measure is never directly validated against the interference it is supposed to predict, and several load-bearing hyperparameters are undisclosed. The current evidence is suggestive but not yet conclusive enough to support the paper's general claim.

major comments (4)
  1. [§3.2, Eq. (6)-(9)] The compatibility measure Dij is computed from PCA-projected gradients on a temporary shared model plus hand-crafted behavior statistics. The paper does not validate that this distance predicts the actual interference that occurs during sequential QLoRA fine-tuning. The ablations in Tables 3 and 5 show that changing grouping/sequencing matters, but they do not isolate the mechanism. A direct validation would measure pairwise transfer/interference for the task pairs under the target fine-tuning protocol and correlate it with Dij; also compare Eq. (8) against simpler alternatives (gradient-only, behavior-only, random). Without this, the claim that the observed gains arise from 'optimization compatibility' is not established.
  2. [§3.2, Eq. (8), (13)] Several free parameters of the method are never reported: λ in Eq. (8), the PCA dimension, the number of mini-batches used for gradient collection, and the 'fixed weights' combining the five terms in Eq. (13). The number of policies K and the maximum group-size constraint (two balanced four-task groups) are also imposed ad hoc. Without these values, the method is not reproducible, and it is not possible to rule out selection on the TRACE test set during the design of these hyperparameters. Please report all hyperparameters and provide sensitivity analyses, at least for λ, PCA dimension, K, and the group-size constraint.
  3. [§4.2, Table 1] The abstract claims performance 'consistently improves' from single-policy to multi-policy PEFT. Table 1 contradicts this for the uninformed multi-policy baseline: Random/Random is worse than Single-policy on both backbones (36.94 vs 42.12 on LLaMA-2; 36.49 vs 39.50 on Vicuna). The improvement appears only when grouping/sequencing is automatic or manual. The abstract and conclusions should be qualified to reflect that multi-policy PEFT is beneficial when the optimization paths are organized.
  4. [§4.2, Tables 1-5] The paper does not report the number of seeds, standard deviations, or significance tests; the described setting appears to be a single run per configuration. Several head-to-head differences are small (44.78 vs 44.53 for LLaMA in Table 1; 44.78 vs 42.58 in Table 4; 41.14 vs 39.45 for Vicuna). At 7B scale with 5,000 samples per task, these gaps could easily be within run-to-run variance. Please report at least three seeds with error bars on OP and BWT, and significance tests for the key comparisons.
minor comments (5)
  1. [§3.1, Eq. (2)] Eq. (2) presents a joint multi-task objective, but the paper is about sequential fine-tuning. The shared-adapter baseline is trained sequentially, not jointly. Consider writing the sequential objective explicitly to avoid a mismatch between formulation and experiments.
  2. [§3.2] The text describes PCA as a 'parameter-free projection' and then uses a PCA dimension; PCA dimension is a hyperparameter, so the description should be revised.
  3. [§4.2] The 'Manual' grouping/sequence is not described or tabulated. Please provide the manual groups and the manual sequence used for each backbone, or a reference, so that the expert baseline is reproducible.
  4. [§4.3, Figure 2] Figure 2 is referenced as visualizing the fused task-compatibility structure, but the text does not explain how the visualization is produced or what the distance/color scale means.
  5. [General] No code or artifact link is provided. Given the number of undisclosed hyperparameters, a public implementation would substantially improve reproducibility.

Circularity Check

0 steps flagged

No significant circularity: grouping/sequencing are computed from gradient and behavior statistics, distinct from the performance metrics being predicted.

full rationale

The paper's central claim is that optimization-path organization via task grouping and sequencing improves multi-policy PEFT. The grouping distance Dij (Eq. 8) is computed from PCA-projected gradients (Eq. 6) and hand-crafted behavior statistics (Eq. 7), neither of which is defined in terms of the reported performance (OP, BWT) or fitted to it. The clustering (Eq. 9) and sequencing (Eq. 12) are then solved on these inputs, and the resulting policies are trained on the same task data. The improvements over random and manual baselines are empirical ablations, not tautological predictions. No parameter is fitted to the target metric and then reported as a prediction; the rank allocation ablation is a model-selection exercise, not a circularity. The citations to the authors' prior RL work (Bossens & Sobey 2024; Birkbeck et al. 2024) are motivational, not load-bearing to the derivation. Undisclosed hyperparameters (λ, PCA dimension, sequencing weights) and the lack of direct validation of Dij against true interference are substantive limitations, but they concern whether the heuristic truly captures optimization compatibility, not whether the derivation reduces to its inputs. Under the specified circularity criteria, no step meets the standard of exhibiting a constructional equivalence or a fitted-value-as-prediction.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The framework rests on several manually selected hyperparameters and domain assumptions about what makes tasks compatible. No new physical or conceptual entities are introduced; 'optimization path' is a framing term. The most honest measure is that the method's 'automatic' behavior depends critically on free parameters (K, group size, lambda, objective weights) that are not disclosed.

free parameters (6)
  • lambda (fusion weight in Eq. 8) = not reported
    Controls the trade-off between gradient and behavior distances; no ablation or value given.
  • Number of policies K = 2
    Fixed to two policies in all experiments; the 'automatic' grouping actually requires K as an input.
  • Maximum group size = 4
    Balanced clustering constraint forcing two four-task groups on TRACE; not derived from data.
  • Weights of composite path objective (Eq. 13) = not reported
    C_trans, C_tail, C_dir, C_prog, R_head, R_transfer are combined with 'fixed weights' that are never specified; likely chosen after seeing TRACE results.
  • PCA dimension for gradient representations = not reported
    Stage 1 uses PCA to compress gradients; the dimensionality is not stated.
  • Number of mini-batches for gradient collection = not reported
    Gradients are collected from a temporary shared model over 'multiple mini-batches', but the count is unspecified.
axioms (4)
  • domain assumption Gradient distance computed on a temporary shared model is a valid proxy for optimization interference during later QLoRA fine-tuning.
    Stage 1 uses this to partition tasks, but it is a heuristic with no theoretical justification or validation against other conflict metrics.
  • domain assumption Hand-crafted behavior statistics (prompt/answer lengths, answer tendencies) capture capability and output-structure compatibility.
    These features are simple dataset-level aggregates; no evidence that they reflect the true capability requirements of the tasks.
  • ad hoc to paper Two balanced groups of four tasks is the appropriate organization for the TRACE task set.
    The paper imposes K=2 and a max group size of 4, so the grouping reduces to choosing which four tasks go together; these constraints are not derived from a principle.
  • ad hoc to paper The composite sequencing objective with unspecified fixed weights is a sufficient curriculum.
    Five heuristic terms are combined into a scalar objective; the weights are fixed but not stated, making the curriculum an ad-hoc scoring function.

pith-pipeline@v1.3.0-daily-deepseek · 11593 in / 8494 out tokens · 79116 ms · 2026-08-03T03:46:36.658335+00:00 · methodology

0 comments
read the original abstract

Parameter-Efficient Fine-Tuning (PEFT) commonly adapts large language models using a single shared Low-Rank Adapter (LoRA). This shared optimization space often suffers from interference when adapting heterogeneous task sequences, leading to poor transfer and catastrophic forgetting. Existing approaches mainly improve adapter expressiveness by increasing parameter capacity or composing multiple adapters, yet they still rely on a shared optimization path. In this paper, we propose an optimization-path organization framework for parameter-efficient fine-tuning of large language models, implemented as an automatic multi-policy PEFT architecture. Specifically, optimization-compatible adaptation paths are automatically organized through task grouping and task sequencing under a fixed parameter budget. The organized optimization paths are implemented as independent Quantized Low-Rank Adapters (QLoRA), enabling heterogeneous tasks to be optimized in decoupled adaptation spaces while preserving positive transfer among compatible tasks. Experiments on the TRACE benchmark demonstrate that performance consistently improves from conventional single-policy PEFT to multi-policy PEFT, with the proposed automatic multi-policy framework achieving the best performance of 44.78 under the same trainable capacity. This suggests that optimization-path organization is more effective than simply increasing adapter capacity for heterogeneous parameter-efficient fine-tuning.

Figures

Figures reproduced from arXiv: 2607.29601 by Adam Sobey, Francisco Gomez Medina, Jiajia Tang, Sizhe Yuen, Yali Du.

Figure 1
Figure 1. Figure 1: Our approach automatically organizes optimization paths through task grouping and task sequencing. These organized [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of the fused task-compatibility struc [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

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

36 extracted references · 9 linked inside Pith

  1. [1]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    Orthogonal subspace learning for language model continual learning , author=. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    2023

  2. [2]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    HOP to the Next Tasks and Domains for Continual Learning in NLP , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  3. [3]

    Nature Machine Intelligence , volume=

    Parameter-efficient fine-tuning of large-scale pre-trained language models , author=. Nature Machine Intelligence , volume=. 2023 , month=

    2023

  4. [4]

    Findings of the Association for Computational Linguistics: EACL 2023 , pages =

    Chronopoulou, Alexandra and Peters, Matthew and Fraser, Alexander and Dodge, Jesse , title =. Findings of the Association for Computational Linguistics: EACL 2023 , pages =

    2023

  5. [5]

    Advances in Neural Information Processing Systems , volume =

    Dettmers, Tim and Pagnoni, Artidoro and Holtzman, Ari and Zettlemoyer, Luke , title =. Advances in Neural Information Processing Systems , volume =

  6. [6]

    Proceedings of the 36th International Conference on Machine Learning , series =

    Houlsby, Neil and Giurgiu, Andrei and Jastrzebski, Stanislaw and Morrone, Bruna and de Laroussilhe, Quentin and Gesmundo, Andrea and Attariyan, Mona and Gelly, Sylvain , title =. Proceedings of the 36th International Conference on Machine Learning , series =

  7. [7]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , title =

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , title =. International Conference on Learning Representations , journal=

  8. [8]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages =

    Lester, Brian and Al-Rfou, Rami and Constant, Noah , title =. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages =

    2021

  9. [9]

    Li, Xiang Lisa and Liang, Percy , title =. 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 =

  10. [10]

    , title =

    Liu, Haokun and Tam, Derek and Muqeeth, Mohammed and Mohta, Nikhil and Huang, Tenghao and Bansal, Mohit and Raffel, Colin A. , title =. Advances in Neural Information Processing Systems , volume =

  11. [11]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages =

    Liu, Xiao and Ji, Kaixuan and Fu, Yicheng and Tam, Wing Yin and Du, Zhengxiao and Yang, Zhilin and Tang, Jie , title =. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages =

  12. [12]

    Advances in Neural Information Processing Systems , volume =

    Karimi Mahabadi, Rabeeh and Henderson, James and Ruder, Sebastian , title =. Advances in Neural Information Processing Systems , volume =

  13. [13]

    Karimi Mahabadi, Rabeeh and Ruder, Sebastian and Dehghani, Mostafa and Henderson, James , title =. 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 =

  14. [14]

    UniPELT: A unified framework for parameter-efficient language model tuning , booktitle =

    Mao, Yuning and K. UniPELT: A unified framework for parameter-efficient language model tuning , booktitle =

  15. [15]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Mallya, Arun and Lazebnik, Svetlana , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  16. [16]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    Pfeiffer, Jonas and Vuli. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    2020

  17. [17]

    AdapterFusion: Non-destructive task composition for transfer learning , booktitle =

    Pfeiffer, Jonas and Kamath, Aishwarya and R. AdapterFusion: Non-destructive task composition for transfer learning , booktitle =

  18. [18]

    and Rabinowitz, Neil C

    Rusu, Andrei A. and Rabinowitz, Neil C. and Desjardins, Guillaume and Soyer, Hubert and Kirkpatrick, James and Kavukcuoglu, Koray and Pascanu, Razvan and Hadsell, Raia , title =. arXiv preprint arXiv:1606.04671 , year =

    Pith/arXiv arXiv

  19. [19]

    arXiv preprint arXiv:2010.05874 , year=

    Wang, Zihan and Tsvetkov, Yulia and Firat, Orhan and Cao, Yuan and Mallinson, Jonathan and Li, Xian , title =. arXiv preprint arXiv:2010.05874 , year=

    Pith/arXiv arXiv 2010

  20. [20]

    arXiv preprint arXiv:2310.06762 , year =

    Wang, Xiao and Zhang, Yuansen and Chen, Tianze and Gao, Songyang and Jin, Senjie and Yang, Xianjun and Xi, Zhiheng and Zheng, Rui and Zou, Yicheng and Gui, Tao and Zhang, Qi and Huang, Xuanjing , title =. arXiv preprint arXiv:2310.06762 , year =

    Pith/arXiv arXiv

  21. [21]

    Advances in Neural Information Processing Systems , volume =

    Yu, Tianhe and Kumar, Saurabh and Gupta, Abhishek and Hausman, Karol and Levine, Sergey and Finn, Chelsea , title =. Advances in Neural Information Processing Systems , volume =

  22. [22]

    arXiv preprint arXiv:2303.10512 , year=

    Zhang, Qingru and Chen, Minhao and Bukharin, Alexander and He, Pengcheng and Cheng, Yu and Chen, Weizhu and Zhao, Tuo , title =. arXiv preprint arXiv:2303.10512 , year=

    Pith/arXiv arXiv

  23. [23]

    IEEE Transactions on Knowledge and Data Engineering , volume =

    Zhang, Yu and Yang, Qiang , title =. IEEE Transactions on Knowledge and Data Engineering , volume =

  24. [24]

    arXiv preprint arXiv:2403.14608 , year =

    Parameter-Efficient Fine-Tuning for Large Models: A Comprehensive Survey , author =. arXiv preprint arXiv:2403.14608 , year =

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2410.19878 , year =

    Parameter-Efficient Fine-Tuning in Large Models: A Survey of Methodologies , author =. arXiv preprint arXiv:2410.19878 , year =

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2404.15159 , year =

    MixLoRA: Enhancing Large Language Models Fine-Tuning with LoRA-based Mixture of Experts , author =. arXiv preprint arXiv:2404.15159 , year =

    Pith/arXiv arXiv

  27. [27]

    Han, Jiayi and Du, Liang and Du, Hongwei and Zhou, Xiangguo and Wu, Yiwen and Zhang, Yuanfang and Zheng, Weibo and Han, Donghong , booktitle =

  28. [28]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers) , pages =

    Rehearsal-Free Modular and Compositional Continual Learning for Language Models , author =. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers) , pages =

    2024

  29. [29]

    Chen, Cheng and Zhu, Junchen and Luo, Xu and Shen, Hengtao and Gao, Lianli and Song, Jingkuan , journal =

  30. [30]

    Wang, Yifan and Liu, Yafei and Shi, Chufan and Li, Haoling and Chen, Chen and Lu, Haonan and Yang, Yujiu , booktitle =

  31. [31]

    ACM Computing Surveys , volume =

    Continual Learning of Large Language Models: A Comprehensive Survey , author =. ACM Computing Surveys , volume =

  32. [32]

    arXiv preprint arXiv:2504.13822 , year =

    Parameter-Efficient Continual Fine-Tuning: A Survey , author =. arXiv preprint arXiv:2504.13822 , year =

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2409.03577 , year=

    CHIRPs: Change-Induced Regret Proxy metrics for Lifelong Reinforcement Learning , author=. arXiv preprint arXiv:2409.03577 , year=

    Pith/arXiv arXiv

  34. [34]

    Bossens and Adam J

    David M. Bossens and Adam J. Sobey , title =. AI Communications , volume =

  35. [35]

    2025 , eprint=

    MTL-LoRA: Low-Rank Adaptation for Multi-Task Learning , author=. 2025 , eprint=

    2025

  36. [36]

    M o RE : A Mixture of Low-Rank Experts for Adaptive Multi-Task Learning

    Zhang, Dacao and Zhang, Kun and Chu, Shimao and Wu, Le and Li, Xin and Wei, Si. M o RE : A Mixture of Low-Rank Experts for Adaptive Multi-Task Learning. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.68

    doi:10.18653/v1/2025.findings-acl.68 2025