pith. sign in

arxiv: 2606.02615 · v1 · pith:ASKJWCLWnew · submitted 2026-05-26 · 📡 eess.AS · cs.AI· cs.SD

FSA-GRPO: Teaching Auditory LLMs to Use Few-shot Demonstrations

Haolong Zheng , Siyin Wang , Xulin Fan , Zengrui Jin , Mark Hasegawa-Johnson This is my paper

Pith reviewed 2026-07-01 15:58 UTC · model grok-4.3

classification 📡 eess.AS cs.AIcs.SD
keywords few-shot promptingauditory large language modelschildren's speech recognitionreinforcement learningASRspeech translationpost-training
0
0 comments X

The pith

Training auditory LLMs only on adult speech data strengthens their few-shot adaptation across tasks.

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

Auditory large language models often do not fully exploit few-shot prompting because they lack explicit training for it. FSA-GRPO uses reinforcement learning with a reward that encourages the model to base answers on provided demonstrations. When this training uses only high-resource adult ASR data, the model improves on children's speech recognition as well as speech translation and audio understanding. The approach works better than direct tuning on related out-of-domain data when in-domain data cannot be used for training.

Core claim

FSA-GRPO is an RL-based post-training recipe that applies a specially designed reward to encourage auditory LLMs to leverage few-shot demonstrations. Training this method exclusively on high-resource adult ASR data improves the model's general few-shot adaptation ability, producing gains in children's speech recognition, speech translation, and audio understanding. When in-domain data are unavailable, FSA-GRPO outperforms direct tuning on related out-of-domain data.

What carries the argument

FSA-GRPO, an RL post-training method that adds a reward encouraging models to condition responses on few-shot demonstrations during inference.

If this is right

  • Performance on children's speech recognition improves without any child data in the training set.
  • Gains transfer to speech translation and audio understanding tasks.
  • FSA-GRPO beats direct tuning on out-of-domain data when in-domain data cannot be used.
  • Data selection and auxiliary reward weighting affect the final effectiveness of the recipe.

Where Pith is reading between the lines

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

  • The method may teach a general meta-skill for using demonstrations that applies beyond the specific training tasks.
  • Similar reward designs could reduce the need to collect domain-specific data for adapting audio models.
  • The approach might extend to other low-resource audio scenarios where demonstrations are the only available adaptation signal.

Load-bearing premise

The specially designed reward in FSA-GRPO will cause the model to leverage few-shot demonstrations during inference rather than other response patterns.

What would settle it

Compare few-shot task performance of models trained with the full FSA-GRPO reward against identical models trained with the same adult ASR data but without the demonstration-leveraging reward component; lack of difference would falsify the claim that the reward drives the gains.

Figures

Figures reproduced from arXiv: 2606.02615 by Haolong Zheng, Mark Hasegawa-Johnson, Siyin Wang, Xulin Fan, Zengrui Jin.

Figure 1
Figure 1. Figure 1: Overview of FSA-GRPO. We construct few-shot ASR training instance by pairing a query utterance with retrieved speech–text in-context examples (ICEs), then optimize a LoRA-augmented auditory LLM with GRPO. Each sampled response receives a reward combining ASR accuracy, measured by WER against the reference transcript, and semantic alignment, which encourages the prediction to remain close to relevant ICE la… view at source ↗
Figure 2
Figure 2. Figure 2: Main results across child ASR, audio understanding/reasoning, multilingual ASR, and speech translation. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Auxiliary semantic-alignment reward weight [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
read the original abstract

Few-shot prompting provides an effective way to adapt auditory large language models to low-resource tasks such as children's speech recognition. However, most auditory large language models are not explicitly trained to perform inference in this demonstration-conditioned format, limiting the extent to which they can benefit from few-shot prompting. To address this limitation, we introduce Few-Shot Aware GRPO (FSA-GRPO), an RL-based post-training recipe that uses a specially designed reward to encourage the model to leverage few-shot demonstrations, thereby strengthening its few-shot adaptation ability. Notably, training with only high-resource adult ASR data improves the model's general few-shot adaptation ability, yielding gains not only in children's speech recognition but also in speech translation and audio understanding. We further study data selection and auxiliary reward weighting to identify an effective training recipe. Our experiments show that when in-domain data are unavailable or cannot be used for training, FSA-GRPO is more effective than direct tuning on related out-of-domain data.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces Few-Shot Aware GRPO (FSA-GRPO), an RL-based post-training recipe for auditory LLMs. It employs a specially designed reward to encourage models to leverage few-shot demonstrations during inference. Training occurs exclusively on high-resource adult ASR data, yet the method is claimed to improve general few-shot adaptation, producing gains on children's speech recognition, speech translation, and audio understanding tasks. The approach is further claimed to outperform direct tuning on related out-of-domain data when in-domain data cannot be used for training.

Significance. If the reward mechanism is shown to specifically induce demonstration-conditioned behavior rather than generic task improvements, the result would be significant for low-resource auditory LLM adaptation, as it would allow effective few-shot gains without access to in-domain training data.

major comments (2)
  1. [Reward function definition (method section)] The central claim that FSA-GRPO strengthens few-shot adaptation (rather than yielding generic RL gains on ASR) rests on the reward function specifically rewarding demonstration-conditioned outputs. The manuscript must provide the exact reward formulation (including any auxiliary terms and weighting) and evidence that it enforces conditioning on provided demonstrations, for example via ablations that remove demonstrations at inference time or compare against a reward that ignores demonstration content.
  2. [Experiments and results] The superiority over direct tuning on out-of-domain data is a key result, but without reported metrics, baseline comparisons, or tables showing effect sizes on the downstream tasks (children's ASR, speech translation, audio understanding), it is impossible to assess whether the gains are attributable to the few-shot mechanism or other factors.
minor comments (2)
  1. The abstract states that data selection and auxiliary reward weighting were studied to identify an effective recipe; the corresponding analysis and any hyperparameter sensitivity results should be presented with concrete values and ablation tables.
  2. Notation for the reward components and GRPO variant should be defined clearly before use, and any new symbols introduced in the method should be listed in a table or appendix.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the reward formulation and experimental reporting. We address each major comment below and will revise the manuscript to strengthen these aspects.

read point-by-point responses

  1. Referee: [Reward function definition (method section)] The central claim that FSA-GRPO strengthens few-shot adaptation (rather than yielding generic RL gains on ASR) rests on the reward function specifically rewarding demonstration-conditioned outputs. The manuscript must provide the exact reward formulation (including any auxiliary terms and weighting) and evidence that it enforces conditioning on provided demonstrations, for example via ablations that remove demonstrations at inference time or compare against a reward that ignores demonstration content.

    Authors: We agree that the precise reward formulation is essential to substantiate the central claim. The manuscript describes the reward at a high level but does not include the full mathematical expression with auxiliary terms and weights. In the revision we will add the exact FSA-GRPO reward definition. We will also report new ablations that evaluate performance when demonstrations are removed at inference and compare against a reward variant that ignores demonstration content, to isolate the conditioning effect. revision: yes

  2. Referee: [Experiments and results] The superiority over direct tuning on out-of-domain data is a key result, but without reported metrics, baseline comparisons, or tables showing effect sizes on the downstream tasks (children's ASR, speech translation, audio understanding), it is impossible to assess whether the gains are attributable to the few-shot mechanism or other factors.

    Authors: We concur that detailed quantitative results are required for proper evaluation. While the manuscript states gains on the listed tasks and superiority over out-of-domain tuning, it does not present comprehensive tables with all metrics and effect sizes. In the revision we will add tables reporting absolute and relative performance on children's ASR, speech translation, and audio understanding, including all relevant baselines and direct comparisons to out-of-domain tuning. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical training recipe evaluated on held-out tasks

full rationale

The paper introduces FSA-GRPO as an RL post-training method using a designed reward on adult ASR data, then reports measured gains on children's ASR, speech translation, and audio understanding. No equations, parameters, or claims reduce by construction to the inputs; results are presented as experimental outcomes rather than tautological re-statements of the reward definition or prior self-citations. The derivation chain consists of standard RL training followed by independent downstream evaluation.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Abstract-only review limits visibility; the reward function is implied to contain tunable components, and the core premise is an RL assumption about behavior shaping.

free parameters (1)
  • auxiliary reward weights
    The abstract states that auxiliary reward weighting was studied to identify an effective recipe, indicating these are adjustable parameters.
axioms (1)
  • domain assumption Reinforcement learning with the specially designed reward will increase the model's use of few-shot demonstrations at inference time
    This premise underpins the entire FSA-GRPO method and the claimed generalization benefits.

pith-pipeline@v0.9.1-grok · 5715 in / 1319 out tokens · 57462 ms · 2026-07-01T15:58:49.626630+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

50 extracted references · 21 canonical work pages · 5 internal anchors

  1. [1]

    2024 , eprint=

    COSMIC: Data Efficient Instruction-tuning For Speech In-Context Learning , author=. 2024 , eprint=

    2024

  2. [2]

    Proceedings of the 2022 conference of the North American chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    Metaicl: Learning to learn in context , author=. Proceedings of the 2022 conference of the North American chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2022

  3. [3]

    2509.13395 , archivePrefix=

    Haolong Zheng and Yekaterina Yegorova and Mark Hasegawa-Johnson , year=. 2509.13395 , archivePrefix=

    work page arXiv

  4. [4]

    2025 , organization=

    Zhou, Jiaming and Zhao, Shiwan and He, Jiabei and Wang, Hui and Zeng, Wenjia and Chen, Yong and Sun, Haoqin and Kong, Aobo and Qin, Yong , booktitle=. 2025 , organization=

    2025

  5. [5]

    2024 , organization=

    Wang, Siyin and Yang, Chao-Han and Wu, Ji and Zhang, Chao , booktitle=. 2024 , organization=

    2024

  6. [6]

    2512.18263 , archivePrefix=

    Haolong Zheng and Yekaterina Yegorova and Mark Hasegawa-Johnson , year=. 2512.18263 , archivePrefix=

    work page arXiv

  7. [7]

    LLM-Core-Team Xiaomi , year=

  8. [8]

    Brown, Tom and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared D and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and others , booktitle =

  9. [9]

    Huang, Shaohan and Dong, Li and Wang, Wenhui and Hao, Yaru and Singhal, Saksham and Ma, Shuming and Lv, Tengchao and Cui, Lei and Mohammed, Owais Khan and Patra, Barun and others , booktitle=

  10. [10]

    Keren, Gil and Kozhevnikov, Artyom and Meng, Yen and Ropers, Christophe and Setzler, Matthew and Wang, Skyler and Adebara, Ife and Auli, Michael and Balioglu, Can and others , journal=

  11. [11]

    2024 , organization=

    Chen, Zhehuai and Huang, He and Andrusenko, Andrei and Hrinchuk, Oleksii and Puvvada, Krishna C and Li, Jason and Ghosh, Subhankar and Balam, Jagadeesh and Ginsburg, Boris , booktitle=. 2024 , organization=

    2024

  12. [12]

    Kong, Zhifeng and Goel, Arushi and Badlani, Rohan and Ping, Wei and Valle, Rafael and Catanzaro, Bryan , booktitle=

  13. [13]

    2025 , volume=

    Ihori, Mana and Yamane, Taiga and Kawata, Naotaka and Makishima, Naoki and Tanaka, Tomohiro and Suzuki, Satoshi and Orihashi, Shota and Masumura, Ryo , booktitle=. 2025 , volume=

    2025

  14. [14]

    arXiv preprint arXiv:2409.10429 , year=

    SMILE: speech meta in-context learning for low-resource language automatic speech recognition , author=. arXiv preprint arXiv:2409.10429 , year=

    work page arXiv

  15. [15]

    2021 , booktitle =

    CoVoST 2 and Massively Multilingual Speech Translation , author =. 2021 , booktitle =. doi:10.21437/Interspeech.2021-2027 , issn =

    work page doi:10.21437/interspeech.2021-2027 2021

  16. [16]

    and Branson, M

    Ardila, R. and Branson, M. and Davis, K. and Henretty, M. and Kohler, M. and Meyer, J. and Morais, R. and Saunders, L. and Tyers, F. M. and Weber, G. , title =. Proceedings of the 12th Conference on Language Resources and Evaluation , pages =

  17. [17]

    Wang, Dingdong and Wu, Jincenzi and Li, Junan and Yang, Dongchao and Chen, Xueyuan and Zhang, Tianhua and Meng, Helen , journal=

  18. [18]

    Jin, Xu and Zhifang, Guo and Jinzheng, He and Hangrui, Hu and Ting, He and Shuai, Bai and Keqin, Chen and Jialin, Wang and Yang, Fan and Kai, Dang and Bin, Zhang and Xiong, Wang and Yunfei, Chu and Junyang, Lin , journal=

  19. [19]

    Xiuwen Zheng and Bornali Phukon and Jonghwan Na and Ed Cutrell and Kyu J. Han and Mark Hasegawa-Johnson and Pan-Pan Jiang and Aadhrik Kuila and Colin Lea and Bob MacDonald and Gautam Mantena and Venkatesh Ravichandran and Leda Sari and Katrin Tomanek and Chang D. Yoo and Chris Zwilling , year =. doi:10.21437/Interspeech.2025-566 , issn =

    work page doi:10.21437/interspeech.2025-566 2025

  20. [20]

    Pradhan, Sameer and Cole, Ronald and Ward, Wayne , booktitle=

  21. [21]

    2019 , publisher=

    Redmond, Sean M and Ash, Andrea C and Christopulos, Tyler T and Pfaff, Theresa , journal=. 2019 , publisher=

    2019

  22. [22]

    NeurIPS , pages=

    Uniaudio 1.5: Large language model-driven audio codec is a few-shot audio task learner , author=. NeurIPS , pages=

  23. [23]

    arXiv preprint arXiv:2311.02248 , year=

    Cosmic: Data efficient instruction-tuning for speech in-context learning , author=. arXiv preprint arXiv:2311.02248 , year=

    work page arXiv

  24. [24]

    EMNLP , pages=

    Bayesian Example Selection Improves In-Context Learning for Speech, Text and Visual Modalities , author=. EMNLP , pages=

  25. [25]

    2026 , eprint=

    SICL-AT: Another way to adapt Auditory LLM to low-resource task , author=. 2026 , eprint=

    2026

  26. [26]

    2026 , eprint =

    MetaSICL: Adapting Auditory LLM via Meta Speech In-Context Learning , author =. 2026 , eprint =

    2026

  27. [27]

    Audio Flamingo Next: Next-Generation Open Audio-Language Models for Speech, Sound, and Music

    Audio Flamingo Next: Next-Generation Open Audio-Language Models for Speech, Sound, and Music , author=. arXiv preprint arXiv:2604.10905 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    International Conference on Learning Representations , volume=

    Mmau: A massive multi-task audio understanding and reasoning benchmark , author=. International Conference on Learning Representations , volume=

  29. [29]

    Advances in Neural Information Processing Systems , volume=

    Mmar: A challenging benchmark for deep reasoning in speech, audio, music, and their mix , author=. Advances in Neural Information Processing Systems , volume=

  30. [30]

    Proceedings of the 40th International Conference on Machine Learning , pages=

    Robust Speech Recognition via Large-Scale Weak Supervision , author=. Proceedings of the 40th International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  31. [31]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    Proximal Policy Optimization Algorithms

    Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  34. [34]

    International Conference on Learning Representations , year=

    LoRA: Low-Rank Adaptation of Large Language Models , author=. International Conference on Learning Representations , year=

  35. [35]

    arXiv preprint arXiv:2511.01299 , year=

    Towards General Auditory Intelligence: Large Multimodal Models for Machine Listening and Speaking , author=. arXiv preprint arXiv:2511.01299 , year=

    work page arXiv

  36. [36]

    SALMONN-omni: A Standalone Speech LLM without Codec Injection for Full-duplex Conversation , url =

    Yu, Wenyi and Wang, Siyin and Yang, Xiaoyu and Chen, Xianzhao and Tian, Xiaohai and Zhang, Jun and Sun, Guangzhi and Lu, Lu and Wang, Yuxuan and Zhang, Chao , booktitle =. SALMONN-omni: A Standalone Speech LLM without Codec Injection for Full-duplex Conversation , url =

  37. [37]

    doi:10.21437/Interspeech.2025-764 , issn =

    Seraphina Fong and Marco Matassoni and Alessio Brutti , year =. doi:10.21437/Interspeech.2025-764 , issn =

    work page doi:10.21437/interspeech.2025-764 2025

  38. [38]

    EURASIP Journal on Audio, Speech, and Music Processing , volume=

    Exploring the effect of differences in the acoustic correlates of adults' and children's speech in the context of automatic speech recognition , author=. EURASIP Journal on Audio, Speech, and Music Processing , volume=. 2010 , publisher=

    2010

  39. [39]

    Proceedings of the 2nd Workshop on Child, Computer and Interaction , pages=

    A review of ASR technologies for children's speech , author=. Proceedings of the 2nd Workshop on Child, Computer and Interaction , pages=

  40. [40]

    Computer speech & language , volume=

    Transfer learning from adult to children for speech recognition: Evaluation, analysis and recommendations , author=. Computer speech & language , volume=. 2020 , publisher=

    2020

  41. [41]

    arXiv preprint arXiv:2406.10507 , year=

    Benchmarking Children's ASR with Supervised and Self-supervised Speech Foundation Models , author=. arXiv preprint arXiv:2406.10507 , year=

    work page arXiv

  42. [43]

    arXiv preprint arXiv:2503.11197 , year=

    Reinforcement Learning Outperforms Supervised Fine-Tuning: A Case Study on Audio Question Answering , author=. arXiv preprint arXiv:2503.11197 , year=

    work page arXiv

  43. [44]

    arXiv preprint arXiv:2505.09439 , year=

    Omni-R1: Do You Really Need Audio to Fine-Tune Your Audio LLM? , author=. arXiv preprint arXiv:2505.09439 , year=

    work page arXiv

  44. [45]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    SoundMind: RL-Incentivized Logic Reasoning for Audio-Language Models , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=. 2025 , publisher=

    2025

  45. [46]

    arXiv preprint arXiv:2509.16990 , year=

    Advancing Speech Understanding in Speech-Aware Language Models with GRPO , author=. arXiv preprint arXiv:2509.16990 , year=

    work page arXiv

  46. [47]

    arXiv preprint arXiv:2509.01939 , year=

    Group Relative Policy Optimization for Speech Recognition , author=. arXiv preprint arXiv:2509.01939 , year=

    work page arXiv

  47. [48]

    Sentence- BERT : Sentence Embeddings using Siamese BERT -Networks

    Reimers, Nils and Gurevych, Iryna. Sentence- BERT : Sentence Embeddings using Siamese BERT -Networks. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing. 2019. doi:10.18653/v1/D19-1410

    work page doi:10.18653/v1/d19-1410 2019

  48. [49]

    doi:10.21437/Interspeech.2024-1932 , issn =

    Chang, Kai-Wei and Hsu, Ming-Hao and Li, Shan-Wen and Lee, Hung-yi , year =. doi:10.21437/Interspeech.2024-1932 , issn =

    work page doi:10.21437/interspeech.2024-1932 2024

  49. [50]

    doi:10.21437/Interspeech.2025-1467 , issn =

    Agrawal, Neeraj and Ganapathy, Sriram , year =. doi:10.21437/Interspeech.2025-1467 , issn =

    work page doi:10.21437/interspeech.2025-1467 2025

  50. [51]

    Multimodal In-context Learning for ASR of Low-resource Languages

    Multimodal In-context Learning for ASR of Low-resource Languages , author=. arXiv preprint arXiv:2601.05707 , year=

    work page internal anchor Pith review Pith/arXiv arXiv