Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

An audio question-answering model can improve its own accuracy at test time using only unlabeled questions.

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

2026-08-04 11:18 UTC pith:WLFRQAGV

load-bearing objection Transductive evaluation undermines the headline; the method and ablations are plausible but the central claim needs a seen/unseen split. the 4 major comments →

arxiv 2510.05478 v3 pith:WLFRQAGV submitted 2025-10-07 eess.AS

AQA-TTRL: Self-Adaptation in Audio Question Answering with Test-Time Reinforcement Learning

Haoyu Zhang , Jiaxian Guo , Dong Yang , Yusuke Iwasawa , Yutaka Matsuo This is my paper
classification eess.AS
keywords audio question answeringtest-time adaptationreinforcement learningpseudo-labelingmajority votingGRPOself-improvementlarge audio language models
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 tries to establish that a large audio-language model can improve its own accuracy on audio question-answering at test time, using only unlabeled test questions—no human annotations or labeled training data. The method, AQA-TTRL, has the model generate multiple answers per question, take a majority vote as a pseudo-label, and then fine-tune its own policy with a reinforcement-learning step (GRPO) that rewards agreement with those pseudo-labels. Two additions make the loop stable: weighting the learning signal by the model's self-consistency (confidence) and resampling when all answers in a group collapse to one. Across three benchmarks the paper reports average gains of 4.42% for a 7B model and 11.04% for a 3B model, with the adapted 3B model even beating the static 7B model. If this holds, it suggests deployed audio models can keep improving in the field without annotation costs.

Core claim

The central claim is that a frozen large audio-language model can bootstrap its own performance on audio question answering by treating its own majority-vote consensus as a reward signal. Specifically, AQA-TTRL generates multiple sampled answers per question, takes the majority answer as a pseudo-label, estimates a confidence score from the vote share, and then runs Group Relative Policy Optimization (GRPO) with a confidence-weighted advantage so that high-consensus labels dominate the update. To prevent the training signal from vanishing when a sampled group of rollouts is unanimous, the method falls back to additional sampled groups. Across the MMAU (test-mini and test), MMAR, and MMSU ben

What carries the argument

The load-bearing mechanism is the self-supervised reward loop: majority-voting pseudo-labels combined with Group Relative Policy Optimization (GRPO). The confidence-weighted advantage scales each rollout's advantage by f(Conf), a monotone function of the vote share of the pseudo-label, so consensus answers contribute larger gradients. Multiple-attempt sampling provides fallback response groups when an entire group is identical, avoiding advantage collapse. The paper also relies on the observed positive correlation between vote confidence and answer correctness, which justifies the weighting.

Load-bearing premise

The method is evaluated on the same test set it adapts to, so the reported accuracy gains could be overfitting to those specific questions rather than genuine capability improvement; the paper does not test on a held-out set.

What would settle it

Adapt the model on one set of unlabeled audio questions, then evaluate it on a disjoint set of similar questions; if accuracy on the held-out set does not improve (or drops), the self-adaptation is mostly overfitting to the adaptation set.

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

If this is right

  • If correct, deployed audio assistants can adapt to new acoustic conditions or question distributions without any labeled data collection.
  • Smaller models can close some of the gap to larger ones through targeted test-time adaptation; the adapted 3B model surpassed the static 7B model's direct inference.
  • Reinforcement learning is more noise-robust than supervised fine-tuning for pseudo-label training, since the adapted model outperforms its own pseudo-labels (the DIMV baseline).
  • The confidence-weighting scheme is justified by a strong empirical correlation between majority-vote confidence and answer accuracy across benchmarks.

Where Pith is reading between the lines

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

  • The transductive evaluation (reporting accuracy on the same test set used for adaptation) means the reported gains may partly reflect memorization of test questions; a held-out evaluation would clarify whether the method yields genuine generalization.
  • A natural extension is to combine the confidence-weighting with external uncertainty or calibration estimates to flag when the model's consensus is untrustworthy, potentially making the self-adaptation loop more conservative on novel domains.
  • The fixed-step selection noted in the paper could be replaced by a label-free early-stopping heuristic, such as monitoring the stability of the model's votes on the unlabeled test set, to avoid overfitting and reduce compute.

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 / 6 minor

Summary. The paper proposes AQA-TTRL, a test-time reinforcement-learning method for audio question answering. The model generates pseudo-labels by majority voting over 64 samples, then updates itself with GRPO using a confidence-weighted advantage (Eq. 5) and a multiple-attempt sampling scheme (Algorithm 1) intended to prevent advantage collapse. Experiments on MMAU (test-mini/test), MMAR, and MMSU with Qwen2.5-Omni 7B and 3B report average accuracy gains of 4.42 and 11.04 percentage points, and the adapted 3B model is claimed to outperform direct inference of the unadapted 7B model. The method is clearly described and the two technical additions are plausible, but the evaluation protocol is transductive and uses the test set for adaptation, model selection, and final reporting, which is the central weakness.

Significance. If the empirical claims survive a stricter protocol, this would be a useful demonstration that large audio language models can improve on unlabeled test data without manual annotation, a relatively unexplored direction for audio QA. The confidence-weighted advantage and multiple-attempt sampling are sensible and clearly motivated, and the paper includes ablations isolating their contributions. However, the current experiments do not separate genuine model improvement from memorization of pseudo-labels on the exact test instances, and several hyperparameters (reporting step, confidence-weight function) are selected using the same test benchmarks. The headline numbers are therefore not yet convincing evidence for the 'self-adaptation' claim. The contribution is conditional on a stronger, held-out evaluation.

major comments (4)
  1. [Sec. 4.1/4.2, Table 1, Eq. (1)] The evaluation is transductive in a way that undermines the central claim. The model is adapted on each benchmark's test set and accuracy is reported on that same set. With a global batch of 8 and 500 steps (Sec. 4.1), smaller test sets can be traversed multiple times, while larger sets (e.g., MMSU with 5,000 samples) are only partially seen. The reported gains may reflect memorization of majority-vote pseudo-labels for the exact test instances rather than a better model. Please split the test data into an adaptation set and a held-out evaluation set, or report accuracy separately for examples seen vs. unseen during adaptation. If the intended claim is purely transductive, it should be stated explicitly, but the current wording ('self-improvement', 'surpasses 7B') implies a general capability improvement.
  2. [Sec. 4.1, Fig. 4, Table 2(a)] Model selection is performed on the test benchmarks. The reporting step is fixed per dataset scale (100th step for small sets, 500th for large sets), but Fig. 4's caption explicitly says 'better performance may occur at intermediate steps,' meaning the choice of step is informed by test accuracy even though the authors say they 'preserve the label-free setting.' Similarly, the confidence-weight function f is selected by comparing Sqrt/Linear/Exp on the same test benchmarks (Table 2a). This selection on the test set optimistically biases the reported gains. A label-free validation set or a pre-specified selection rule is needed; at minimum, report results for all steps and all weight functions, and show that the conclusion is not an artifact of the chosen checkpoint.
  3. [Table 1, Tables 2(a)-2(b)] All results are from a single run with no error bars or significance tests. Differences such as 73.74 vs. 73.79 in Table 2(b) or the 0.8-1.3 point gaps in several benchmark cells may be within run-to-run variance. With 4 H100 GPUs and 500 steps, running multiple seeds (at least 2-3) is feasible and should be reported as mean ± std. Without this, the headline improvements and the 3B-vs-7B comparison are not statistically grounded.
  4. [Sec. 3.3, Eq. (5), Fig. 3] The confidence-weighting mechanism relies on a confidence-accuracy correlation measured on the same test sets used for adaptation and final evaluation (Fig. 3). This does not establish that the correlation holds for unseen data or that the confidence signal is sufficiently reliable for reward weighting. Please report the pseudo-label accuracy and confidence distribution, and ideally a separate validation set, to support the claim that weighting by confidence improves label quality rather than merely amplifying the model's existing biases.
minor comments (6)
  1. [Abstract / Sec. 4.2] 'Average improvements of 4.42%' and '11.04%' should be phrased as 'absolute accuracy percentage points' or 'relative improvement' to avoid ambiguity. In Table 1, 64.39 to 68.81 is a 4.42-point gain, not a 4.42% relative improvement.
  2. [Sec. 3.2, Eq. (3)] Equation (3) is hard to read because of the formatting of the min/clip expression. Define c_{i,t} before the equation and use standard GRPO notation with clear parentheses to distinguish the clipped and unclipped surrogate terms.
  3. [References] References [8] and [23] appear to be the same Qwen2.5-Omni technical report with different author order. Please consolidate or explain the distinction.
  4. [Sec. 4.1] The number of samples used for Direct Inference with Majority Voting (DIMV) is not stated in the caption of Table 1. Please add it for completeness.
  5. [Algorithm 1] The sampling notation in lines 3-5 is informal. Please specify the sampling distribution and temperature, and consider adding a sentence on the computational overhead of generating three groups.
  6. [Fig. 4] The caption says 'To preserve the label-free setting, we report performance at the fixed 500th step.' This sentence conflicts with the earlier statement that the reporting step is fixed per dataset scale (100th for smaller sets). Please clarify the exact selection rule used for each benchmark.

Circularity Check

0 steps flagged

No significant circularity: the training signal is self-generated, but the reported target (ground-truth accuracy) is external and not defined by the method's equations.

full rationale

The paper's derivation chain is empirical and self-contained. Pseudo-labels (Eq. 1) and confidence (Eq. 4) are computed from the model's own sampled outputs; the reward (Eq. 2) and confidence-weighted advantage (Eq. 5) are functions of those pseudo-labels; the GRPO objective (Eq. 3) updates the policy. The reported accuracy is measured against ground-truth labels in the MMAU/MMAR/MMSU benchmarks, so the target metric is not defined by the training signal. No equation reduces to another by construction, and no parameter is renamed as a prediction. The paper does not rely on load-bearing self-citations; TTRL [11], GRPO [12,13], and R1-AQA [9] are external prior work. The transductive protocol (adapting on the test set and reporting accuracy on the same set) and the test-set-based ablation choice of the confidence-weighting function (Table 2a) are methodological/validity limitations, not circularity: they do not make the reported improvement an algebraic consequence of the inputs. The paper also explicitly discloses the fixed-step reporting limitation in Fig. 4 and the Conclusion, noting that intermediate steps may be better. Thus, while the evaluation may overstate generalization to unseen audio questions, the derivation itself is not circular. Score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim depends on several free hyperparameters, many chosen on the same test benchmarks, plus strong domain assumptions about self-generated labels. No new physical entities are introduced.

free parameters (6)
  • Majority-vote sample count M = 64
    Number of model generations used to form each pseudo-label; hand-chosen, not swept.
  • Rollout group size G = 4
    Number of GRPO rollouts per question during policy update.
  • Number of sampling attempts = 3
    Multiple-attempt sampling uses 3 groups of rollouts to mitigate advantage collapse.
  • Confidence weight function f = exponential
    Chosen by comparing exponential, linear, and sqrt on the same test benchmarks (Table 2a); this is test-set-based model selection.
  • Reporting step = 100 (small datasets) / 500 (large datasets)
    Checkpoint selection is fixed by dataset scale to avoid overfitting, but the choice is made after observing training curves; no label-free selection.
  • GRPO hyperparameters epsilon, beta = 0.2, 0
    Taken from prior work [21], not tuned here.
axioms (5)
  • domain assumption Majority voting over multiple model outputs yields pseudo-labels more reliable than individual outputs.
    Basis of Eq. 1; the paper cites self-consistency [22] but does not prove it for audio.
  • domain assumption Confidence (fraction of votes agreeing) correlates with label correctness.
    Used to justify Eq. 5; Fig. 3 shows a trend on 3 datasets without reporting correlation strength or error bars.
  • domain assumption Optimizing the policy to match pseudo-labels with GRPO improves accuracy on ground-truth labels.
    The entire method rests on this transfer; the paper provides empirical evidence but no mechanism or theoretical guarantee.
  • domain assumption Adapting on the test set and evaluating on the same test set is a valid measure of self-improvement.
    No held-out split is used; the paper assumes transductive improvements generalize or are the target of interest.
  • domain assumption The reproduced DI/DIMV baselines are faithful to the original models' inference behavior.
    Results are 'reproduced results for a consistent inference setting', not official numbers; prompting and decoding choices could shift baselines.

pith-pipeline@v1.3.0-alltime-deepseek · 8054 in / 14558 out tokens · 111513 ms · 2026-08-04T11:18:43.694665+00:00 · methodology

0 comments
read the original abstract

Large Audio Language Models (LALMs) exhibit strong capabilities in general audio understanding but remain static after deployment, limiting their adaptability to real-world data. Since supervised fine-tuning is costly, we propose AQA-TTRL, a novel framework for audio understanding that enables on-the-fly evolution via test-time reinforcement learning using only unlabeled test data. It generates pseudo-labels via majority voting and optimizes the model through reinforcement learning. To address the noise in self-generated labels, we introduce confidence weighting to adjust training signals. Furthermore, multiple-attempt sampling mitigates advantage collapse and stabilizes training. Across MMAU, MMAR, and MMSU, AQA-TTRL achieves significant average improvements of 4.42% for Qwen2.5-Omni 7B and 11.04% for the 3B model. Notably, the adapted 3B model outperforms direct inference of the unadapted 7B model, highlighting the effectiveness of test-time adaptation in audio understanding.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Audio-Zero: Label-Free Self-Evolution for Fine-Grained Audio Reasoning

    cs.SD 2026-07 conditional novelty 6.0

    A self-play game with a known 'odd listener' converts unlabeled audio contrast pairs into a verifiable reward, improving fine-grained audio reasoning on TREA, MMAU, and MMAR.

Reference graph

Works this paper leans on

27 extracted references · 16 linked inside Pith · cited by 1 Pith paper

  1. [1]

    advantage collapse

    INTRODUCTION The application of Large Audio Language Models (LALMs) has led to notable improvements in audio understanding [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]. However, their capabilities are typically frozen after de- ployment; they cannot improve themselves when faced with novel, real-world test data. The traditional approach to this problem— collecting, fi...

  2. [2]

    Large Audio Language Models (LALMs) demonstrate strong abilities in understanding and reason- ing across a variety of tasks involving auditory input such as speech, © 2025 IEEE

    RELATED WORKS Large Audio Language Models. Large Audio Language Models (LALMs) demonstrate strong abilities in understanding and reason- ing across a variety of tasks involving auditory input such as speech, © 2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, in...

    Pith/arXiv arXiv 2025

  3. [3]

    𝑟#…RewardsAdvantage 𝐴! 𝐴

    METHOD In this section, we first present the problem setting of adapting the model at test time using unlabeled test data. Next, we present the overview of our method, covering pseudo-label generation and model update, followed by its two key innovations: Confidence- Weighted Advantage and Multiple-Attempt Sampling. 3.1. Problem Setting Audio understandin...

  4. [4]

    {question}Please choose the answer from the follow- ing options:{choice string}. Output the final answer in<answer> </answer>

    EXPERIMENTS In this section, we introduce our experiment setting and the perfor- mance on the following AQA benchmarks: MMAU-v05.15.25 (test- mini/test), MMAR, and MMSU. 4.1. Experiment Setting We conducted the training with 4 H100 GPUs. We use Qwen2.5- Omni 7B and Qwen2.5-Omni 3B[23] as the base model for adapta- tion training. The batch size is 1 per GP...

    arXiv

  5. [5]

    Our framework establishes a closed-loop learning process by generating confidence-weighted pseudo-labels from majority voting to guide policy optimization with GRPO

    CONCLUSION We propose AQA-TTRL, a self-adaptation framework that enables model to learn and evolve on-the-fly using only the unlabeled data on audio question answering tasks (AQA). Our framework establishes a closed-loop learning process by generating confidence-weighted pseudo-labels from majority voting to guide policy optimization with GRPO. We further...

  6. [6]

    Listen, think, and understand,

    Yuan Gong, Hongyin Luo, Alexander H Liu, Leonid Karlin- sky, and James Glass, “Listen, think, and understand,”arXiv preprint arXiv:2305.10790, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    Salmonn: Towards generic hearing abilities for large language models,

    Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun MA, and Chao Zhang, “Salmonn: Towards generic hearing abilities for large language models,” inThe Twelfth International Conference on Learning Repre- sentations, 2024

    2024

  8. [8]

    Audio flamingo: A novel audio language model with few-shot learning and dialogue abilities,

    Zhifeng Kong, Arushi Goel, Rohan Badlani, Wei Ping, Rafael Valle, and Bryan Catanzaro, “Audio flamingo: A novel audio language model with few-shot learning and dialogue abilities,” inInternational Conference on Machine Learning. PMLR, 2024, pp. 25125–25148

    2024

  9. [9]

    Audio flamingo 2: An audio-language model with long-audio understanding and expert reasoning abilities,

    Sreyan Ghosh, Zhifeng Kong, Sonal Kumar, S Sakshi, Jae- hyeon Kim, Wei Ping, Rafael Valle, Dinesh Manocha, and Bryan Catanzaro, “Audio flamingo 2: An audio-language model with long-audio understanding and expert reasoning abilities,” inForty-second International Conference on Ma- chine Learning, 2025

    2025

  10. [10]

    Audio flamingo 3: Advancing audio intelligence with fully open large audio language models,

    Arushi Goel, Sreyan Ghosh, Jaehyeon Kim, Sonal Kumar, Zhifeng Kong, Sang-gil Lee, Chao-Han Huck Yang, Ramani Duraiswami, Dinesh Manocha, Rafael Valle, and Bryan Catan- zaro, “Audio flamingo 3: Advancing audio intelligence with fully open large audio language models,”arXiv preprint arXiv:2507.08128, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models,

    Yunfei Chu, Jin Xu, Xiaohuan Zhou, Qian Yang, Shil- iang Zhang, Zhijie Yan, Chang Zhou, and Jingren Zhou, “Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models,”arXiv preprint arXiv:2311.07919, 2023

    Pith/arXiv arXiv 2023

  12. [12]

    Qwen2-audio technical report,

    Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhi- fang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, Chang Zhou, and Jingren Zhou, “Qwen2-audio technical report,”arXiv preprint arXiv:2407.10759, 2024

    Pith/arXiv arXiv 2024

  13. [14]

    Reinforcement learning outperforms supervised fine-tuning: A case study on audio question answer- ing,

    Gang Li, Jizhong Liu, Heinrich Dinkel, Yadong Niu, Junbo Zhang, and Jian Luan, “Reinforcement learning outperforms supervised fine-tuning: A case study on audio question answer- ing,”arXiv preprint arXiv:2503.11197, 2025

    Pith/arXiv arXiv 2025

  14. [15]

    Omni-r1: Do you really need audio to fine-tune your audio llm?,

    Andrew Rouditchenko, Saurabhchand Bhati, Edson Araujo, Samuel Thomas, Hilde Kuehne, Rogerio Feris, and James Glass, “Omni-r1: Do you really need audio to fine-tune your audio llm?,”arXiv preprint arXiv:2505.09439, 2025

    arXiv 2025

  15. [16]

    Ttrl: Test-time reinforcement learning,

    Yuxin Zuo, Kaiyan Zhang, Li Sheng, Shang Qu, Ganqu Cui, Xuekai Zhu, Haozhan Li, Yuchen Zhang, Xinwei Long, Ermo Hua, et al., “Ttrl: Test-time reinforcement learning,”arXiv preprint arXiv:2504.16084, 2025

    Pith/arXiv arXiv 2025

  16. [17]

    Deepseekmath: Pushing the limits of math- ematical reasoning in open language models,

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al., “Deepseekmath: Pushing the limits of math- ematical reasoning in open language models,”arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  17. [18]

    Deepseek-r1: Incentivizing reasoning ca- pability in llms via reinforcement learning,

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al., “Deepseek-r1: Incentivizing reasoning ca- pability in llms via reinforcement learning,”arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  18. [19]

    Mmau: A massive multi-task audio understanding and reasoning benchmark,

    S Sakshi, Utkarsh Tyagi, Sonal Kumar, Ashish Seth, Ra- maneswaran Selvakumar, Oriol Nieto, Ramani Duraiswami, Sreyan Ghosh, and Dinesh Manocha, “Mmau: A massive multi-task audio understanding and reasoning benchmark,” arXiv preprint arXiv:2410.19168, 2024

    Pith/arXiv arXiv 2024

  19. [20]

    Mmar: A challenging benchmark for deep reasoning in speech, audio, music, and their mix,

    Ziyang Ma, Yinghao Ma, Yanqiao Zhu, Chen Yang, Yi-Wen Chao, Ruiyang Xu, Wenxi Chen, Yuanzhe Chen, Zhuo Chen, Jian Cong, et al., “Mmar: A challenging benchmark for deep reasoning in speech, audio, music, and their mix,”arXiv preprint arXiv:2505.13032, 2025

    Pith/arXiv arXiv 2025

  20. [21]

    Mmsu: A massive multi-task spoken language understanding and rea- soning benchmark,

    Dingdong Wang, Jincenzi Wu, Junan Li, Dongchao Yang, Xueyuan Chen, Tianhua Zhang, and Helen Meng, “Mmsu: A massive multi-task spoken language understanding and rea- soning benchmark,”arXiv preprint arXiv:2506.04779, 2025

    Pith/arXiv arXiv 2025

  21. [22]

    Desta: Enhancing speech language models through descrip- tive speech-text alignment,

    Ke-Han Lu, Zhehuai Chen, Szu-Wei Fu, He Huang, Boris Ginsburg, Yu-Chiang Frank Wang, and Hung yi Lee, “Desta: Enhancing speech language models through descrip- tive speech-text alignment,” inInterspeech 2024, 2024, pp. 4159–4163

    2024

  22. [23]

    Developing instruction-following speech language model without speech instruction-tuning data,

    Ke-Han Lu, Zhehuai Chen, Szu-Wei Fu, Chao-Han Huck Yang, Jagadeesh Balam, Boris Ginsburg, Yu-Chiang Frank Wang, and Hung-yi Lee, “Developing instruction-following speech language model without speech instruction-tuning data,” inICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  23. [24]

    Desta2.5- audio: Toward general-purpose large audio language model with self-generated cross-modal alignment,

    Ke-Han Lu, Zhehuai Chen, Szu-Wei Fu, Chao-Han Huck Yang, Sung-Feng Huang, Chih-Kai Yang, Chee-En Yu, Chun- Wei Chen, Wei-Chih Chen, Chien-yu Huang, et al., “Desta2.5- audio: Toward general-purpose large audio language model with self-generated cross-modal alignment,”arXiv preprint arXiv:2507.02768, 2025

    arXiv 2025

  24. [25]

    Seed-grpo: Semantic entropy enhanced grpo for uncertainty- aware policy optimization,

    Minghan Chen, Guikun Chen, Wenguan Wang, and Yi Yang, “Seed-grpo: Semantic entropy enhanced grpo for uncertainty- aware policy optimization,”arXiv preprint arXiv:2505.12346, 2025

    Pith/arXiv arXiv 2025

  25. [26]

    Spurious rewards: Rethink- ing training signals in rlvr,

    Rulin Shao, Shuyue Stella Li, Rui Xin, Scott Geng, Yiping Wang, Sewoong Oh, Simon Shaolei Du, Nathan Lambert, Se- won Min, Ranjay Krishna, et al., “Spurious rewards: Rethink- ing training signals in rlvr,”arXiv preprint arXiv:2506.10947, 2025

    Pith/arXiv arXiv 2025

  26. [27]

    Self-consistency improves chain of thought reasoning in lan- guage models,

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou, “Self-consistency improves chain of thought reasoning in lan- guage models,”arXiv preprint arXiv:2203.11171, 2022

    Pith/arXiv arXiv 2022

  27. [28]

    Qwen2. 5-omni technical report,

    Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, et al., “Qwen2. 5-omni technical report,”arXiv preprint arXiv:2503.20215, 2025. 5

    Pith/arXiv arXiv 2025