Pith. sign in

REVIEW 4 major objections 6 minor 23 references

A self-play game with a known 'odd listener' lets audio models learn fine-grained temporal reasoning without any human labels.

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

2026-08-01 10:35 UTC pith:ARBETKZD

load-bearing objection Clever and probably useful self-play for audio, but the paper never shows that the contrast pairs actually differ in the way the method assumes. the 4 major comments →

arxiv 2607.20166 v1 pith:ARBETKZD submitted 2026-07-22 cs.SD cs.AI

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

Siqian Tong , Xuan Li , Chaozhuo Li , Baolong Bi , Yiwei Wang , Yujun Cai , Shenghua Liu , Chengpeng Hao This is my paper
classification cs.SD cs.AI
keywords audio reasoningself-evolutionself-playlabel-free trainingfine-grained perceptiontemporal reasoningGRPOlarge 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 claims that large audio-language models can improve fine-grained auditory reasoning without human-written captions, question-answer labels, or reasoning traces. The method, Audio-Zero, converts unlabeled audio contrast pairs into a multi-player 'odd listener' game: most players hear a reference clip, one hears a subtle variant, and the model must describe what it hears and then deduce which player heard the variant. Because the odd listener is assigned by the training procedure, the correct answer is known, so the game gives verifiable rewards. Training with group-relative policy optimization on these rewards improves temporal reasoning (event order, counting, duration) and general audio understanding on two 7B audio-language backbones. If true, this is a scalable way to push fine-grained audio perception without expensive annotation.

Core claim

Audio-Zero establishes that a procedurally assigned hidden identity in an auditory guessing game provides a label-free, verifiable training signal. Given a contrast pair, one player receives the variant audio while the rest receive the reference; the model generates a clue for each player and then, as judge, identifies the odd listener. The listening reward combines a rule-based content utility score with a vote-aware penalty that discourages identity-revealing or vague clues, while the attribution reward is simply whether the judge guesses correctly. Alternating GRPO updates between listening and attribution creates a feedback loop: stronger attribution raises the bar for clue quality, and

What carries the argument

The central mechanism is the auditory self-play game with a procedurally assigned odd-listener identity. The game couples two roles: players generate natural-language clues about their private audio, and a judge compares all clues against the reference audio to find the odd listener. Since the odd listener is known by construction, the game outcome is a verifiable reward; the listening reward adds a content-utility term and a vote-aware penalty (λ·v_i/max(1,V−1)) that discourages identity leakage. Alternating GRPO updates between the listening and attribution stages lets the two sides co-evolve, with stronger attribution demanding more discriminative clues and better clues training stronger

Load-bearing premise

The reward is only meaningful if the contrast pairs differ in the fine-grained properties the benchmarks measure and not in easier-to-detect properties like loudness, alignment, or overall quality.

What would settle it

Build contrast pairs by applying only non-temporal acoustic changes (loudness scaling, reverb, filtering) to the same reference clip; if the method still lifts TREA order/count/duration scores under these identical-content variants, the temporal-reasoning mechanism is not what drives the gains. Alternatively, if gains vanish when only event order is changed, the central claim fails.

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

If this is right

  • Fine-grained temporal audio reasoning can be improved without any annotated answers, using only unlabeled paired audio.
  • The alternating game structure is necessary; ablating either stage, dropping vote-aware feedback, or switching to sequential training reduces benchmark gains.
  • During training, generated clues become more event-rich, temporally structured, and less vague, indicating that the game pressure itself drives perceptual specificity.
  • The gains transfer to external multiple-choice benchmarks, not just to the self-play game, for both tested backbones.

Where Pith is reading between the lines

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

  • If the variant audio often differs in loudness, alignment, or overall quality rather than event order, the model could earn the reward by detecting any acoustic mismatch, meaning the TREA gains might reflect general discriminative listening instead of fine-grained temporal reasoning.
  • The same 'known mutant' reward could be applied to other paired sensory data (e.g., video clips or audio-visual pairs) whenever one version of a stimulus differs subtly from a reference.
  • Scaling the number of players and listening rounds, or initializing contrast pairs from different generators, would test whether the game's pressure forces ever-finer perceptual distinctions rather than saturating.
  • The rule-based content scorer is the last hand-crafted component; a learned judge or reward model could make the pipeline fully label-free and potentially sharper.

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. Audio-Zero proposes a label-free post-training framework for large audio-language models (LALMs). Given an unlabeled audio contrast pair, the method assigns one player to hear a variant audio while the other players hear a reference, asks the model to generate natural-language clues for each player, and then has the model act as a judge to identify the odd listener. The odd-listener identity is known to the training environment, yielding a verifiable attribution reward; the listening-stage reward combines a rule-based content-utility score with a vote-aware penalty. Optimization alternates between the two stages via GRPO. Experiments on Qwen2-Audio-7B-Instruct and Qwen2.5-Omni-7B report improvements on TREA, MMAU Test-mini, and MMAR, with ablations, a diagnostic analysis of emerging fine-grained descriptions, and a qualitative case study.

Significance. If the central mechanism is validated, the paper would make a notable contribution: it offers a genuinely label-free way to improve fine-grained audio reasoning, converting unlabeled contrast pairs into a verifiable self-play game. The design has several strengths: the attribution reward is verifiable by construction, the main evaluation uses external benchmarks rather than the training distribution, and the improvements are reported on two backbones with ablations supporting the two-stage interaction. However, the paper's central claim is contingent on an unverified property of the training pairs: that they differ predominantly in event order, repetition, or duration rather than in global acoustic quality or artifacts. Because the odd listener always receives the rejected audio from Audio-Alpaca, the verifiable reward may be earned by generic mismatch detection. This issue, together with an underspecified vote-counting reward and an unvalidated rule-based content scorer, means that the paper is currently not ready for acceptance.

major comments (4)
  1. [Sec. 2.1, Sec. 3.1, App. A.1.1] The central assumption is that the contrast pairs 'share high-level semantics but exhibit subtle acoustic differences' such as event order, repetitions, and duration. The actual pairs are chosen/rejected Tango-2 generations from Audio-Alpaca, and the paper itself says they 'may differ in event order, missing events, repetitions, or acoustic quality' (Sec. 3.1) and 'missing or extra events, repetitions, acoustic quality, or background scene' (App. A.1.1). Since the odd listener always receives the rejected audio, a judge can identify it from global quality/alignment/loudness differences without fine-grained temporal reasoning. No experiment characterizes the actual difference distribution of the 2k training pairs, and no control rules out clue-style leakage (length, vocabulary, vagueness) as the judge's cue. This is load-bearing for the claim that TREA gains and the emergence of fine-grai
  2. [Eq. (5), App. A.1.3] The vote-aware penalty is underspecified and internally inconsistent. In Eq. (5), vi is the number of attribution votes received by player i and V is the total number of votes. App. A.1.3 states V = G × N, but the attribution stage samples G completions per prompt, so there are G votes (one per sampled judge prediction), not G × N. The denominator max(1, V−1) therefore has the wrong normalization if V = G × N. It is also unclear whether vi is counted over the G sampled judge outputs, and how players with zero votes are treated relative to players with one. Since Table 2 shows that removing vote-aware feedback causes a sizable TREA drop, this component must be specified precisely and reproducibly.
  3. [Sec. 2.3, App. A.1.3, App. C] The content utility reward R_utility is a rule-based scorer, but the manuscript provides only high-level descriptions and example keyword lists; the exact matching rules, thresholds, term lists, and handling of partial matches are not specified. This scorer is a direct component of the listening reward, so the policy update depends on an unvalidated ad-hoc function. The paper implicitly assumes that this scorer faithfully measures grounded, fine-grained auditory content, but no human correlation, inter-annotator agreement, or alternative-scorer ablation is reported. Please release the exact scorer and validate it, e.g., by correlation with human judgments or by showing that clues with higher R_utility actually improve judge accuracy.
  4. [Sec. 3.1, Table 1, App. E] All results are single-seed point estimates; App. E states that 'all experiments are run with the same random seed.' Many reported differences are small on individual subsets (e.g., Qwen2-Audio TREA Count 30.0 vs. 29.5; Qwen2.5-Omni MMAR Speech 60.54 for both Audio-Thinker and Audio-Zero). Without multiple seeds, variance estimates, or significance tests, the claim of consistent gains, especially over label-dependent baselines, is not fully supported. This is important because the headline improvements on TREA and MMAR are averages over subsets with heterogeneous gaps.
minor comments (6)
  1. [Reproducibility] No code, data, or model checkpoints are released. Given the paper's reliance on a custom rule-based scorer, precise prompt templates, and game construction, releasing artifacts is important for reproducibility and would also allow external verification of the pair-difference concern.
  2. [Sec. 1 / Sec. 3.1] The term 'label-free' is used broadly. The Audio-Alpaca pairs originate from text-to-audio prompts with chosen/rejected preference labels; although these labels are not used as supervision, the pairs are not fully unlabeled. Please clarify what information is and is not used.
  3. [Sec. 2.2] Typo: 'acoustic preperties' should be 'acoustic properties'.
  4. [Fig. 3 caption] The caption for Fig. 3(c) reads 'mathematical transfer performance on downstream audio benchmarks.' The curves show audio reasoning benchmarks, so 'mathematical' should be removed or corrected.
  5. [Table 2 header] The header says 'Qwen2.5-Omni-7B-Instruct' while the rest of the paper uses 'Qwen2.5-Omni-7B.' Make the naming consistent.
  6. [Sec. 2.2] The model sequentially plays all N player roles but is described as not knowing the odd-listener identity. Please specify whether clues for different players are generated with a shared public history that would allow the model to compare its own private audio across roles, and how the prompt prevents identity leakage beyond the instruction 'Do not state whether you are the odd listener.'

Circularity Check

1 steps flagged

Emergence analysis partly reduces to the reward's own rubric; central benchmark claims are independent.

specific steps
  1. self definitional [Sec. 2.3 (Listening Reward) with Sec. 3.5 and App. C / Table 5 (diagnostic metrics)]
    "Content utility rewards clues that are (i) grounded in specific auditory evidence (sound events, temporal relations, acoustic attributes), (ii) non-repetitive across rounds and (iii) free from identity leakage... computed via a rule-based scorer that checks the presence of audio-descriptive terms and penalizes verbatim repetitions and banned phrases. ... As shown in Figure 4, all detail-related metrics generally increase during training, while Vagueness consistently decreases."

    Table 5 defines 'fine-grained auditory descriptions' via Distinct Event Count, Temporal Relation Count, Quantity Expression Frequency, Acoustic Attribute Density, and Vagueness Rate -- the same lexical categories that the R_utility rule-based scorer rewards. Optimizing R_utility therefore directly forces these metrics up and vagueness down, so Figure 4's 'emergence' is a restatement of the reward design rather than independent evidence. This does not contaminate the external TREA/MMAU/MMAR results.

full rationale

The central derivation is not circular: odd-listener identity is procedurally sampled by the environment (Sec. 2.2), the reward is a verifiable game outcome rather than a fitted target, and the paper's main evidence is held-out accuracy on TREA, MMAU Test-mini, and MMAR. No load-bearing self-citation or imported uniqueness theorem is present. The one substantive circularity is secondary: the claim that fine-grained descriptions 'emerge' is measured by Figure 4 with features that are literally the content of the hand-coded listening reward. The unverified assumption that Audio-Alpaca contrast pairs differ in event order/count/duration rather than overall generation quality is a correctness/validity threat, not a circularity, because it is an input premise rather than an output restatement. Overall, the paper's core comparison against label-dependent and label-free baselines stands on independent benchmarks, so the circularity score is low.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The method's verifiable reward is genuinely constructed from game mechanics, but the listening-stage reward imports hand-crafted semantic priors (content utility lists) and unverified assumptions about the data pairs. The main burden is the domain assumption that chosen/rejected Tango-2 audios differ in fine-grained acoustic structure rather than overall quality.

free parameters (3)
  • lambda (vote penalty coefficient) = 0.6
    Eq. (5), Appendix A.1.3; hand-set, no sensitivity analysis, directly changes listening reward.
  • Content-utility rule set = hand-crafted keyword lists and penalties
    R_utility in Eq. (5); encodes what counts as fine-grained auditory evidence; not validated against human labels.
  • Game hyperparameters N, R, G = N=4, R=2, G=6
    Table 4; chosen by hand, no sweep; controls how many clues and votes shape the reward.
axioms (3)
  • domain assumption Contrast pairs from Audio-Alpaca share high-level semantics but differ in subtle acoustic structure (event order, repetitions, durations, missing sounds).
    Stated in Sec. 2.1 and App. A.1.1; not verified with distribution analysis. Chosen/rejected Tango-2 generations may differ in quality.
  • domain assumption A judge that hears the reference audio and reads clues can only identify the odd listener if clues capture fine-grained perceptual detail.
    Core coupling of listening and attribution; no control rules out stylistic or shortcut solutions.
  • ad hoc to paper Rule-based content utility scorer faithfully measures grounded, fine-grained auditory content.
    Appendix A.1.3 and C; hand-crafted lists determine rewards; validity not demonstrated.

pith-pipeline@v1.3.0-alltime-deepseek · 14157 in / 13996 out tokens · 142258 ms · 2026-08-01T10:35:16.652649+00:00 · methodology

0 comments
read the original abstract

Large Audio Language models (LALMs) have made rapid progress on acoustic understanding, yet they still struggle with fine-grained audio reasoning (e.g., recognizing event order, repetitions and duration). Existing post-training methods heavily rely on expensive external labels or provide only coarse semantic signals. To bridge this gap, we introduce Audio-Zero, the first label-free self-evolution framework in the field of LALMs that improves fine-grained auditory perception and reasoning. Audio-Zero constructs an auditory self-play game from unlabeled audio contrast pairs: most players hear a reference audio, while one odd listener hears a subtle variant. The model first generates clues describing what it hears and then identifies the odd listener by reasoning over inconsistencies among clues. Since the odd listener is known by construction, the game provides verifiable rewards without any annotated answers. Experiments with Qwen2-Audio-7B-Instruct and Qwen2.5-Omni-7B on TREA, MMAU Test-mini and MMAR show that Audio-Zero improves fine-grained audio reasoning while preserving broad audio understanding. Evolutionary and diagnostic analyses further reveal that increasingly fine-grained auditory descriptions emerge naturally from game pressure.

Figures

Figures reproduced from arXiv: 2607.20166 by Baolong Bi, Chaozhuo Li, Chengpeng Hao, Shenghua Liu, Siqian Tong, Xuan Li, Yiwei Wang, Yujun Cai.

Figure 1
Figure 1. Figure 1: From external supervision to self-evolution. Existing audio-language learning paradigms [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Audio-Zero. Audio-Zero formulates label-free auditory learning as a multi [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Analysis of self-evolution dynamics across training iterations. Subfigures depict (a) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Diagnostic analysis of generated listening-stage descriptions across training iter￾ations. All metrics are independently min-max scaled to [0, 1] for visualization after length normalization where applicable. Higher val￾ues indicate stronger fine-grained auditory evi￾dence, except for Vagueness, where lower val￾ues indicate better specificity. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_4.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

23 extracted references · 13 linked inside Pith

  1. [1]

    Benchmarking and confidence evaluation of lalms for temporal reasoning.arXiv preprint arXiv:2505.13115,

    Debarpan Bhattacharya, Apoorva Kulkarni, and Sriram Ganapathy. Benchmarking and confidence evaluation of lalms for temporal reasoning.arXiv preprint arXiv:2505.13115,

    Pith/arXiv arXiv

  2. [4]

    Minicpm-o 4.5: Towards real-time full-duplex omni-modal interaction.arXiv preprint arXiv:2604.27393,

    Junbo Cui, Bokai Xu, Chongyi Wang, Tianyu Yu, Weiyue Sun, Yingjing Xu, Tianran Wang, Zhihui He, Wenshuo Ma, Tianchi Cai, et al. Minicpm-o 4.5: Towards real-time full-duplex omni-modal interaction.arXiv preprint arXiv:2604.27393,

    Pith/arXiv arXiv

  3. [6]

    Visplay: Self-evolving vision-language models from images.arXiv preprint arXiv:2511.15661,

    Yicheng He, Chengsong Huang, Zongxia Li, Jiaxin Huang, and Yonghui Yang. Visplay: Self-evolving vision-language models from images.arXiv preprint arXiv:2511.15661,

    arXiv

  4. [7]

    R-zero: Self-evolving reasoning llm from zero data.arXiv preprint arXiv:2508.05004,

    Chengsong Huang, Wenhao Yu, Xiaoyang Wang, Hongming Zhang, Zongxia Li, Ruosen Li, Jiaxin Huang, Haitao Mi, and Dong Yu. R-zero: Self-evolving reasoning llm from zero data.arXiv preprint arXiv:2508.05004,

    Pith/arXiv arXiv

  5. [8]

    Mmau-pro: A chal- lenging and comprehensive benchmark for holistic evaluation of audio general intelligence.arXiv preprint arXiv:2508.13992,

    Sonal Kumar, Šimon Sedláˇcek, Vaibhavi Lokegaonkar, Fernando López, Wenyi Yu, Nishit Anand, Hyeonggon Ryu, Lichang Chen, Maxim Pli ˇcka, Miroslav Hlaváˇcek, et al. Mmau-pro: A chal- lenging and comprehensive benchmark for holistic evaluation of audio general intelligence.arXiv preprint arXiv:2508.13992,

    Pith/arXiv arXiv

  6. [9]

    Reinforcement learning outperforms supervised fine-tuning: A case study on audio question answering.arXiv preprint arXiv:2503.11197,

    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 answering.arXiv preprint arXiv:2503.11197,

    Pith/arXiv arXiv

  7. [10]

    Spiral: Self-play on zero-sum games incentivizes reasoning via multi-agent multi-turn reinforcement learning.arXiv preprint arXiv:2506.24119, 2025a

    Bo Liu, Leon Guertler, Simon Yu, Zichen Liu, Penghui Qi, Daniel Balcells, Mickel Liu, Cheston Tan, Weiyan Shi, Min Lin, et al. Spiral: Self-play on zero-sum games incentivizes reasoning via multi-agent multi-turn reinforcement learning.arXiv preprint arXiv:2506.24119, 2025a. Bo Liu, Chuanyang Jin, Seungone Kim, Weizhe Yuan, Wenting Zhao, Ilia Kulikov, Xia...

    arXiv

  8. [11]

    Mmau: A massive multi-task audio understanding and reasoning benchmark.arXiv preprint arXiv:2410.19168,

    S Sakshi, Utkarsh Tyagi, Sonal Kumar, Ashish Seth, Ramaneswaran 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,

    Pith/arXiv arXiv

  9. [12]

    Deepseekmath: Pushing the limits of mathemat- ical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    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 mathemat- ical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    Pith/arXiv arXiv

  10. [13]

    Can speech llms think while listening?arXiv preprint arXiv:2510.07497,

    Yi-Jen Shih, Desh Raj, Chunyang Wu, Wei Zhou, SK Bong, Yashesh Gaur, Jay Mahadeokar, Ozlem Kalinli, and Mike Seltzer. Can speech llms think while listening?arXiv preprint arXiv:2510.07497,

    arXiv

  11. [14]

    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. In International Conference on Learning Representations, volume 2024, pp. 16607–16629,

    2024

  12. [15]

    Qwen Team. Qwen3. 5-omni technical report.arXiv preprint arXiv:2604.15804,

    Pith/arXiv arXiv

  13. [17]

    Emotion- thinker: Prosody-aware reinforcement learning for explainable speech emotion reasoning.arXiv preprint arXiv:2601.15668,

    Dingdong Wang, Shujie Liu, Tianhua Zhang, Youjun Chen, Jinyu Li, and Helen Meng. Emotion- thinker: Prosody-aware reinforcement learning for explainable speech emotion reasoning.arXiv preprint arXiv:2601.15668,

    arXiv

  14. [18]

    Vision-zero: Scalable vlm self-improvement via strategic gamified self-play.arXiv preprint arXiv:2509.25541, 2025a

    Qinsi Wang, Bo Liu, Tianyi Zhou, Jing Shi, Yueqian Lin, Yiran Chen, Hai Helen Li, Kun Wan, and Wentian Zhao. Vision-zero: Scalable vlm self-improvement via strategic gamified self-play.arXiv preprint arXiv:2509.25541, 2025a. Shaobo Wang, Zhengbo Jiao, Zifan Zhang, Yilang Peng, Xu Ze, Boyu Yang, Wei Wang, Hu Wei, and Linfeng Zhang. Socratic-zero: Bootstrap...

    arXiv

  15. [19]

    Genius: A generalizable and purely unsupervised self-training framework for advanced reasoning

    Fangzhi Xu, Hang Yan, Chang Ma, Haiteng Zhao, Qiushi Sun, Kanzhi Cheng, Junxian He, Jun Liu, and Zhiyong Wu. Genius: A generalizable and purely unsupervised self-training framework for advanced reasoning. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 13153–13167, 2025a. Jin Xu, Zhifa...

    Pith/arXiv arXiv 2026

  16. [20]

    Spell: Self-play reinforcement learning for evolving long-context language models.arXiv preprint arXiv:2509.23863, 2025b

    Ziyi Yang, Weizhou Shen, Chenliang Li, Ruijun Chen, Fanqi Wan, Ming Yan, Xiaojun Quan, and Fei Huang. Spell: Self-play reinforcement learning for evolving long-context language models.arXiv preprint arXiv:2509.23863, 2025b. Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. Star: Bootstrapping reasoning with reasoning.Advances in Neural Information Pro...

    arXiv

  17. [22]

    Aqa-ttrl: Self-adaptation in audio question answering with test-time reinforcement learning.arXiv preprint arXiv:2510.05478,

    Haoyu Zhang, Jiaxian Guo, Yusuke Iwasawa, and Yutaka Matsuo. Aqa-ttrl: Self-adaptation in audio question answering with test-time reinforcement learning.arXiv preprint arXiv:2510.05478,

    Pith/arXiv arXiv

  18. [23]

    We treat the chosen and rejected audios as an audio contrast pair (aref , avar)

    13 A APPENDIX A.1 ADDITIONALIMPLEMENTATIONDETAILS A.1.1 TRAININGDATACONSTRUCTION We instantiate Audio-Zero using Audio-Alpaca (Majumder et al., 2024), a paired audio preference dataset where each example contains a prompt, a chosen audio, and a rejected audio. We treat the chosen and rejected audios as an audio contrast pair (aref , avar). The two clips s...

    2024

  19. [2022]

    Glm-4-voice: Towards intelligent and human-like end-to-end spoken chatbot.arXiv preprint arXiv:2412.02612,

    Aohan Zeng, Zhengxiao Du, Mingdao Liu, Kedong Wang, Shengmin Jiang, Lei Zhao, Yuxiao Dong, and Jie Tang. Glm-4-voice: Towards intelligent and human-like end-to-end spoken chatbot.arXiv preprint arXiv:2412.02612,

    Pith/arXiv arXiv

  20. [2023]

    Incentivizing consistent, effective and scalable reasoning capability in audio llms via reasoning process rewards.arXiv preprint arXiv:2510.20867,

    Jiajun Fan, Roger Ren, Jingyuan Li, Rahul Pandey, Prashanth Gurunath Shivakumar, Ivan Bulyko, Ankur Gandhe, Ge Liu, and Yile Gu. Incentivizing consistent, effective and scalable reasoning capability in audio llms via reasoning process rewards.arXiv preprint arXiv:2510.20867,

    arXiv

  21. [2024]

    Qwen2-audio technical report.arXiv preprint arXiv:2407.10759,

    Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, et al. Qwen2-audio technical report.arXiv preprint arXiv:2407.10759,

    Pith/arXiv arXiv

  22. [2025]

    Self-play fine-tuning converts weak language models to strong language models.arXiv preprint arXiv:2401.01335,

    10 Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models.arXiv preprint arXiv:2401.01335,

    Pith/arXiv arXiv

  23. [2026]

    Game-rl: Synthesizing multimodal verifiable game data to boost vlms’ general reasoning.arXiv preprint arXiv:2505.13886,

    Jingqi Tong, Jixin Tang, Hangcheng Li, Yurong Mou, Ming Zhang, Jun Zhao, Yanbo Wen, Fan Song, Jiahao Zhan, Yuyang Lu, et al. Game-rl: Synthesizing multimodal verifiable game data to boost vlms’ general reasoning.arXiv preprint arXiv:2505.13886,

    arXiv