Pith. sign in

REVIEW 2 minor 53 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · grok-4.3

Text rationales from reasoning MLLMs harm dementia classification from speech, while an adaptor on internal representations improves it.

2026-07-03 21:48 UTC pith:FMW2Z3KW

load-bearing objection Text rationales from MLLMs hurt dementia classification from speech, but pulling internal representations through an adaptor plus RL beats baselines on two datasets.

arxiv 2607.00260 v2 pith:FMW2Z3KW submitted 2026-06-30 eess.AS

Do Multimodal Large Language Models Need Reasoning to Classify Dementia from Speech?

Liming Wang , Neguine Rezaii , Bradford C. Dickerson , James Glass This is my paper
classification eess.AS
keywords dementia classificationspeech analysismultimodal large language modelsadaptor frameworkinternal representationsautomatic diagnosisreasoning 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.

The paper tests whether reasoning capabilities in multimodal large language models aid automatic dementia classification from voice recordings. It demonstrates that relying on the models' generated text explanations often produces hallucinations and inconsistent outputs, resulting in performance below standard LLM-free baselines. The authors introduce DeTAiL, an adaptor framework with nonlinear layers and reinforcement learning that accesses the models' internal representations instead. This method delivers higher accuracy than both baselines and rationale-based approaches on two datasets that use different test formats and label types. A reader would care because it identifies a workable way to apply these models to medical speech analysis while sidestepping their explanation failures.

Core claim

Naive use of text-based rationales from reasoning MLLMs for automatic dementia classification from voice produces hallucinated and inconsistent diagnoses that underperform LLM-free baselines. The DeTAiL framework overcomes this by using a nonlinear adaptor and reinforcement learning to exploit internal representations of the same models, yielding consistent gains across two dementia datasets with distinct test formats and label granularities.

What carries the argument

DeTAiL, an adaptor-based framework with nonlinear adaptor and reinforcement learning that extracts dementia signals from internal representations of reasoning MLLMs rather than their text outputs.

Load-bearing premise

The internal representations of reasoning MLLMs contain dementia-relevant information that the adaptor can access and that is more useful than either text rationales or standard baselines.

What would settle it

If DeTAiL shows no accuracy advantage over strong baselines when evaluated on a third dementia speech dataset with comparable characteristics, the superiority claim would not hold.

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

If this is right

  • DeTAiL outperforms strong baselines and text-rationale methods on two datasets with different formats and label granularities.
  • Text rationales from MLLMs lead to hallucinated and inconsistent diagnostic outputs.
  • Internal representations hold accessible dementia-relevant information that text outputs do not reliably convey.
  • The adaptor approach improves classification without depending on explicit reasoning steps from the MLLM.

Where Pith is reading between the lines

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

  • For classification tasks, raw internal states of MLLMs may prove more reliable than their generated reasoning text.
  • The same adaptor strategy could be applied to other speech-based medical detection problems such as Parkinson's or aphasia.
  • Reinforcement learning in the framework likely helps tune the adaptor to task-specific patterns within the representations.

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

0 major / 2 minor

Summary. The manuscript evaluates whether reasoning capabilities in multimodal large language models (MLLMs) benefit automatic dementia classification (ADC) from speech. It finds that text-based rationales from reasoning MLLMs can produce hallucinations and inconsistent outputs, leading to inferior performance versus LLM-free baselines. The authors introduce DeTAiL, an adaptor framework that extracts and processes internal representations of reasoning MLLMs via a nonlinear adaptor and reinforcement learning; across two dementia datasets differing in test format and label granularity, DeTAiL yields consistent gains over both LLM-free baselines and text-rationale methods.

Significance. If the reported ordering holds, the work supplies concrete evidence that internal MLLM representations can be more reliable than generated rationales for speech-based medical classification. The inclusion of ablations on adaptor design, the RL component, and rationale baselines, together with cross-dataset evaluation, strengthens the empirical case and offers a practical path for improving accuracy and reducing hallucination risk in ADC systems.

minor comments (2)
  1. [Abstract] Abstract: the claim of 'consistent outperformance' is stated without any numerical deltas, dataset sizes, or statistical significance; while the body supplies these details, a one-sentence summary of key metrics would improve the abstract's standalone value.
  2. [Experimental setup] The manuscript states that code and demo will be released upon acceptance; consider adding a brief reproducibility statement (e.g., random seeds, exact hyper-parameter ranges) in the experimental section to facilitate immediate verification.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of the work, including the significance of the empirical findings on internal MLLM representations versus text rationales, the ablations, and the cross-dataset evaluation. The recommendation for minor revision is noted. No major comments were provided in the report, so we have no specific points requiring point-by-point rebuttal or revision at this stage. Any minor issues will be addressed in the revised version.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The manuscript contains no equations, derivations, or mathematical claims that could reduce to inputs by construction. Its central contribution is an empirical comparison of the proposed DeTAiL adaptor framework against LLM-free baselines and text-rationale methods on two external dementia datasets; all reported gains are measured directly from held-out test performance rather than from any fitted parameter renamed as a prediction or from self-citation chains. No load-bearing premise relies on prior work by the same authors in a way that would create circular justification. The evaluation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only abstract available; no explicit free parameters, axioms, or invented entities are stated.

pith-pipeline@v0.9.1-grok · 5730 in / 997 out tokens · 23244 ms · 2026-07-03T21:48:29.854545+00:00 · methodology

0 comments
read the original abstract

Multimodal large language models (MLLMs) have emerged as a promising approach for improving the accuracy, transferability, and explainability of automatic dementia classification (ADC) systems from voice recordings. Yet it remains unclear whether their reasoning capabilities are beneficial for ADC, and how such capabilities should be leveraged. In this paper, we conduct a careful evaluation of reasoning MLLMs for ADC and show that naive strategies, such as relying on text-based rationales, can lead to hallucinated and inconsistent rationales for diagnosis and yield inferior ADC performance compared with LLM-free baselines. To overcome this limitation, we propose \textbf{De}mentia \textbf{T}hinker with Nonlinear \textbf{A}daptor and Re\textbf{i}nforcement \textbf{L}earning (DeTAiL), an adaptor-based framework that exploits the internal representations of reasoning MLLMs for improved dementia classification. Across two dementia datasets with distinct test formats and label granularities, DeTAiL consistently outperforms strong baselines and methods that rely on text-based rationales. Code and demo will be released upon acceptance.

Figures

Figures reproduced from arXiv: 2607.00260 by Bradford C. Dickerson, James Glass, Liming Wang, Neguine Rezaii.

Figure 1
Figure 1. Figure 1: Overall Architecture of DeTAiL. (a) In the distillation and GRPO stages, the MLLM learns to generate both the cognitive label and the textual rationale that explains its prediction; (b) in the MLP adaptor stage, a small MLP classifier is trained on the hidden representation of the MLLM given the prompt and the generated rationale. Although a pretrained LLM can be prompted as an ADC, it often underperforms … view at source ↗
Figure 1
Figure 1. Figure 1: Overall Architecture of DeTAiL. (a) In the distillation and GRPO stages, the MLLM learns to generate both the cognitive label and the textual rationale that explains its prediction; (b) in the MLP adaptor stage, a small MLP classifier is trained on the hidden representation of the MLLM given the prompt and the generated rationale. techniques, including SFT, GRPO, GRPO with ra￾tionale distillation, and a no… view at source ↗
Figure 3
Figure 3. Figure 3: Reliability of the most frequent evidence types in the rationale for DeTAiL on ADReSS. Reliability is estimated by computing the percentage of correct predictions using a given evidence. For the two-class setting, earlier layers tend to work better, suggesting that relatively low-level representations are suf￾ficient for distinguishing controls from cognitively impaired participants. The trend differs for … view at source ↗

Review history (2 revisions) →

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

53 extracted references · 53 canonical work pages · 8 internal anchors

  1. [1]

    Alzheimer’s dementia recognition through spontaneous speech: The ADReSS challenge,

    S. Luz, F. Haider, S. de la Fuente, D. Fromm, and B. MacWhinney, “Alzheimer’s dementia recognition through spontaneous speech: The ADReSS challenge,” inProceedings of INTERSPEECH 2020, 2020

    work page 2020

  2. [2]

    Early dementia detection using multiple spontaneous speech prompts: The PROCESS challenge,

    F. Taoet al., “Early dementia detection using multiple spontaneous speech prompts: The PROCESS challenge,” inICASSP, 2025, signal Processing Grand Challenge

    work page 2025

  3. [3]

    Speech- based cognitive screening: A systematic evaluation of LLM adaptation strategies,

    F. Taherinezhad, M. J. M. Nezhad, S. Karimi, S. Rashidi, A. Zolnour, M. Dadkhah, Y . Haghbin, H. AzadMaleki, and M. Zolnoori, “Speech- based cognitive screening: A systematic evaluation of LLM adaptation strategies,” 2025

    work page 2025

  4. [4]

    Unifying acoustic features and text with multimodal LLMs for neurodegenerative screening,

    Q. Zhang, Y . Guo, and Y . Gong, “Unifying acoustic features and text with multimodal LLMs for neurodegenerative screening,” 2026

    work page 2026

  5. [5]

    De- tecting cognitive decline using speech only: The ADReSSo challenge,

    S. Luz, F. Haider, S. de la Fuente, D. Fromm, and B. MacWhinney, “De- tecting cognitive decline using speech only: The ADReSSo challenge,” inProceedings of INTERSPEECH 2021, 2021

    work page 2021

  6. [6]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. V . Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” inNeural Information Processing System, vol. 35, 2022, pp. 24 824–24 837

    work page 2022

  7. [7]

    Capabilities of gemini models in medicine,

    K. Saabet al., “Capabilities of gemini models in medicine,” 2024

    work page 2024

  8. [8]

    GPT-4 Technical Report

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Ale- man, D. Almeidaet al., “GPT-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  9. [9]

    ChatGPT-4o for few-shot pathological speech detec- tion from spectrograms,

    H. Amiriet al., “ChatGPT-4o for few-shot pathological speech detec- tion from spectrograms,” 2025, preprint/proof-of-concept study on UA- Speech; verify exact title and venue before submission

    work page 2025

  10. [10]

    Language models are few-shot learners,

    T. B. Brownet al., “Language models are few-shot learners,” inNeural Information Processing System, vol. 33, 2020, pp. 1877–1901

    work page 2020

  11. [11]

    Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting,

    M. Turpin, J. Michael, E. Perez, and S. R. Bowman, “Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting,” inNeural Information Processing System, 2023, pp. 74 952–74 965. [Online]. Available: https://proceedings.neurips.cc/paper files/paper/2023/hash/ ed3fea9033a80fea1376299fa7863f4a-Abstract-Conference.html

    work page 2023

  12. [12]

    Walk the talk? measuring the faithfulness of large language model explanations,

    K. Matton, R. O. Ness, J. Guttag, and E. Kıcıman, “Walk the talk? measuring the faithfulness of large language model explanations,” inProceedings of the Thirteenth International Conference on Learning Representations, 2025. [Online]. Available: https: //openreview.net/forum?id=4ub9gpx9xw

    work page 2025

  13. [13]

    Faithful or just plausible? evaluating the faithfulness of closed-source LLMs in medical reasoning,

    H. Afolabiet al., “Faithful or just plausible? evaluating the faithfulness of closed-source LLMs in medical reasoning,” inProceedings of the Fifth Machine Learning for Health Symposium, ser. Proceedings of Machine Learning Research, vol. 297. PMLR, 2026, pp. 1562–1591. [Online]. Available: https://proceedings.mlr.press/v297/afolabi26a.html

    work page 2026

  14. [14]

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

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . K. Li, Y . Wu, and D. Guo, “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,”arXiv preprint arXiv:2402.03300, 2024. [Online]. Available: https://arxiv.org/abs/2402.03300

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  15. [15]

    CALM: Class-conditional sparse attention vectors for large audio- language models,

    V . Mehta, L. Wang, H. Kuehne, R. Feris, J. R. Glass, and M. J. Mirza, “CALM: Class-conditional sparse attention vectors for large audio- language models,”arXiv preprint arXiv:2602.07077, 2026

    work page arXiv 2026

  16. [16]

    Chain-of-thought reasoning with large language models for clinical alzheimer’s disease assessment and diagnosis,

    T. Zhang, J.-E. Ding, M. Ozolcer, F.-M. Hung, A. C.-C. Yang, F. Liu, Y .-R. Ji, and S. W. Bae, “Chain-of-thought reasoning with large language models for clinical alzheimer’s disease assessment and diagnosis,” 2026. [Online]. Available: https://arxiv.org/abs/2602.13979

    work page arXiv 2026

  17. [17]

    Qwen2.5-Omni Technical Report

    J. Xuet al., “Qwen2.5-Omni technical report,” 2025. [Online]. Available: https://arxiv.org/abs/2503.20215

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    Qwen3 Technical Report

    A. Yanget al., “Qwen3 technical report,” 2025. [Online]. Available: https://arxiv.org/abs/2505.09388

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [19]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inInternational Conference on Learning Representations, 2022

    work page 2022

  20. [20]

    Spoken Language Biomarkers for Detecting Cognitive Impairment

    T. Alhanai, R. Au, and J. Glass, “Spoken language biomarkers for detecting cognitive impairment,” in2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2017, pp. 409–416. [Online]. Available: https://arxiv.org/abs/1710.07551

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  21. [21]

    Temporal integration of acoustic and textual features for alzheimer’s dementia recognition,

    M. Martinc, S. Pollak, and M. Robnik- ˇSikonja, “Temporal integration of acoustic and textual features for alzheimer’s dementia recognition,” in Proceedings of INTERSPEECH 2021, 2021, venue inferred from the ADReSS/ADReSSo challenge literature; verify page numbers before camera-ready use

    work page 2021

  22. [22]

    Detecting dementia from long neuropsychological interviews,

    N. Dawalatabad, Y . Gong, S. Khurana, R. Au, and J. Glass, “Detecting dementia from long neuropsychological interviews,” inFindings of the Association for Computational Linguistics: EMNLP 2022. Abu Dhabi, United Arab Emirates: Association for Computational Linguistics, 2022, pp. 5270–5283. [Online]. Available: https://aclanthology.org/ 2022.findings-emnlp.386/

    work page 2022

  23. [23]

    Alzheimer’s dementia recogni- tion using acoustic, lexical, disfluency and speech pause features robust to noisy inputs,

    M. Rohanian, J. Hough, and M. Purver, “Alzheimer’s dementia recogni- tion using acoustic, lexical, disfluency and speech pause features robust to noisy inputs,” inProceedings of INTERSPEECH 2021, 2021

    work page 2021

  24. [24]

    Detection of dementia on voice recordings using deep learning: a framingham heart study,

    C. Xue, C. Karjadi, I. C. Paschalidis, R. Au, and V . B. Kolachalama, “Detection of dementia on voice recordings using deep learning: a framingham heart study,”Alzheimer’s Research & Therapy, vol. 13, p. 146, Aug. 2021

    work page 2021

  25. [25]

    Recognizing dementia from neuropsychological tests with state space models,

    L. Wang, S. Bhati, C. Karjadi, R. Au, and J. Glass, “Recognizing dementia from neuropsychological tests with state space models,” in 2025 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). Honolulu, HI, USA: IEEE, 2025, pp. 1–7

    work page 2025

  26. [26]

    Comparing acoustic-based approaches for alzheimer’s disease detection,

    A. Balagopalan and J. Novikova, “Comparing acoustic-based approaches for alzheimer’s disease detection,” inInterspeech, 2021, pp. 3800–3804

    work page 2021

  27. [27]

    Classifying alzheimer’s disease using audio and text-based representations of speech,

    R. Haulcy and J. Glass, “Classifying alzheimer’s disease using audio and text-based representations of speech,”Frontiers in Psychology, vol. 11, p. 624137, 2021

    work page 2021

  28. [28]

    Leveraging pretrained representations with task-related keywords for alzheimer’s disease detection,

    J. Li, K. Song, J. Li, B. Zheng, D. Li, X. Wu, X. Liu, and H. Meng, “Leveraging pretrained representations with task-related keywords for alzheimer’s disease detection,” inArXiv, 2023. [Online]. Available: https://arxiv.org/pdf/2303.08019.pdf

    work page arXiv 2023

  29. [29]

    A unified approach to interpreting model predictions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” inAdvances in Neural Information Processing Systems, vol. 30. Curran Associates, Inc., 2017

    work page 2017

  30. [30]

    Explainable alzheimer’s disease detection using linguistic features from automatic speech recog- nition,

    L. Tang, Z. Zhang, F. Feng, L. Yang, and H. Li, “Explainable alzheimer’s disease detection using linguistic features from automatic speech recog- nition,”Dementia and Geriatric Cognitive Disorders, vol. 52, no. 4, pp. 240–248, 2023

    work page 2023

  31. [31]

    Optimizing and evaluating pre-trained large language models for alzheimer’s disease detection,

    F. Casu, E. Grosso, A. Lagorio, and G. A. Trunfio, “Optimizing and evaluating pre-trained large language models for alzheimer’s disease detection,” inProceedings of the 32nd Euromicro International Confer- ence on Parallel, Distributed and Network-Based Processing. Dublin, Ireland: IEEE, 2024, pp. 277–284

    work page 2024

  32. [32]

    Reasoning-based approach with chain-of-thought for alzheimer’s detection using speech and large language models,

    C. Park, A. S. G. Choi, S. Cho, and C. Kim, “Reasoning-based approach with chain-of-thought for alzheimer’s detection using speech and large language models,” inProceedings of INTERSPEECH 2025, 2025, arXiv:2506.01683; page numbers/proceedings metadata should be added from the official proceedings when available

    work page arXiv 2025

  33. [33]

    LLMCARE: Alzheimer’s detection via transformer models enhanced by LLM-generated synthetic data,

    A. Zolnour, H. Azadmaleki, Y . Haghbin, F. Taherinezhad, M. J. M. Nezhad, S. Rashidi, M. Khani, A. Taleban, S. M. Sani, M. Dad- khah, J. M. Noble, S. Bakken, Y . Yaghoobzadeh, A.-H. Vahabie, M. Rouhizadeh, and M. Zolnoori, “LLMCARE: Alzheimer’s detection via transformer models enhanced by LLM-generated synthetic data,” Front. Artif. Intell., 2025

    work page 2025

  34. [34]

    What do LLMs know about alzheimer’s disease? fine-tuning, probing, and data synthesis for AD detection,

    L. Jiang, Y . Zhou, and N. Parde, “What do LLMs know about alzheimer’s disease? fine-tuning, probing, and data synthesis for AD detection,” 2026

    work page 2026

  35. [35]

    Cross- lingual alzheimer’s disease detection with multimodal llms via speech cue-augmented prompting and instruction tuning,

    Y .-L. Liu, Y . Li, Y . Chen, L. He, R. Feng, J. Chen, and J. Yuan, “Cross- lingual alzheimer’s disease detection with multimodal llms via speech cue-augmented prompting and instruction tuning,” inICASSP, 2026, pp. 16 557–16 561

    work page 2026

  36. [36]

    Alzheimer’s dementia detection using perplexity from paired large language models,

    Y . Xiao, H. Christensen, and S. Goetze, “Alzheimer’s dementia detection using perplexity from paired large language models,” inInterspeech, Rotterdam, The Netherlands, 2025

    work page 2025

  37. [37]

    CARE-AD: A multi-agent large language model framework for alzheimer’s disease prediction using longitudinal clinical notes,

    R. Li, X. Wang, D. Berlowitz, J. Mez, H. Lin, and H. Yu, “CARE-AD: A multi-agent large language model framework for alzheimer’s disease prediction using longitudinal clinical notes,”npj Digital Medicine, vol. 8, no. 1, p. 541, 2025

    work page 2025

  38. [38]

    Agentic cognitive profiling: Realigning automated alzheimer’s disease detection with clinical construct validity,

    J. Kang, K. Li, D. Han, J. Li, J. Li, L. Meng, X. Wu, and H. Meng, “Agentic cognitive profiling: Realigning automated alzheimer’s disease detection with clinical construct validity,” 2026

    work page 2026

  39. [39]

    ADAgent: LLM agent for alzheimer’s disease analysis with collaborative coordinator,

    W. Hou, G. Yang, Y . Du, Y . Lau, L. Liu, J. He, L. Long, and S. Wang, “ADAgent: LLM agent for alzheimer’s disease analysis with collaborative coordinator,” inAI for Clinical Applications, J. Qiu, J. Wu, C. Langlotz, B. Huang, Z. Lei, H. Wu, H. Liu, and W. Xie, Eds. Springer Nature Switzerland, 2026, pp. 23–32. [Online]. Available: https://doi.org/10.1007...

    work page doi:10.1007/978-3-032-06004-4 2026

  40. [40]

    Linguistic changes in spontaneous speech for detecting parkinson’s disease using large language models,

    J. L. Crawford, “Linguistic changes in spontaneous speech for detecting parkinson’s disease using large language models,”PLOS Digital Health, vol. 4, no. 2, p. e0000757, 2025

    work page 2025

  41. [41]

    Clinical efficacy of pre-trained large language models through the lens of aphasia,

    Y . Cong, A. N. LaCroix, and J. Lee, “Clinical efficacy of pre-trained large language models through the lens of aphasia,”Scientific Reports, vol. 14, p. 15573, 2024

    work page 2024

  42. [42]

    Detecting primary progressive aphasia (PPA) from text: A benchmarking study,

    G. Merhbene, F. Lecron, P. Fortemps, B. C. Dickerson, M. Kurpicz- Briki, and N. Rezaii, “Detecting primary progressive aphasia (PPA) from text: A benchmarking study,” inFindings of the Association for Computational Linguistics: EACL 2026, V . Demberg, K. Inui, and L. Marquez, Eds. Rabat, Morocco: Association for Computational Linguistics, Mar. 2026, pp. 3...

    work page 2026

  43. [43]

    AP-GRPO: Anchor-gated phonetic alignment with policy optimization for pathological speech reconstruction,

    P. Zhang, H. H. Nguyen, Y . Song, W. Huang, T. I. Imu, H. P. Zou, J. Wu, H. Xu, and A. M. Rahmani, “AP-GRPO: Anchor-gated phonetic alignment with policy optimization for pathological speech reconstruction,” 2026. [Online]. Available: https://arxiv.org/abs/2606. 15540

    work page 2026

  44. [44]

    Towards understanding chain-of-thought prompting: An empirical study of what matters,

    B. Wang, S. Min, X. Deng, J. Shen, Y . Wu, L. Zettlemoyer, and H. Sun, “Towards understanding chain-of-thought prompting: An empirical study of what matters,” inProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Toronto, Canada: Association for Computational Linguistics, 2023, pp. 2717–2739. [O...

    work page 2023

  45. [45]

    Position: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!

    S. Kambhampati, K. Valmeekam, S. Bhambri, V . Palod, L. Saldyt, K. Stechly, S. R. Samineni, D. Kalwar, and U. Biswas, “Position: Stop anthropomorphizing intermediate tokens as reasoning/thinking traces!” inInternational Conference on Machine Learning, 2026. [Online]. Available: https://arxiv.org/abs/2504.09762

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  46. [46]

    TTRL: Test-Time Reinforcement Learning

    Y . Zuo, K. Zhang, S. Qu, L. Sheng, X. Zhu, B. Qi, Y . Sun, G. Cui, N. Ding, and B. Zhou, “TTRL: Test-time reinforcement learning,” inNeural Information Processing System, 2025. [Online]. Available: https://arxiv.org/abs/2504.16084

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  47. [47]

    V oiceprints of cognitive impair- ment: analyzing digital voice for early detection of alzheimer’s and related dementias,

    N. Rezaii, B. Wong, P. Aisenet al., “V oiceprints of cognitive impair- ment: analyzing digital voice for early detection of alzheimer’s and related dementias,”npj Dementia, vol. 1, no. 35, 2025

    work page 2025

  48. [48]

    Qwen2.5 Technical Report

    A. Yanget al., “Qwen2.5 technical report,” 2024. [Online]. Available: https://arxiv.org/abs/2412.15115

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  49. [49]

    SWIFT: A scalable lightweight infrastructure for fine-tuning,

    Y . Zhao, J. Huang, J. Hu, X. Wang, Y . Mao, D. Zhang, Z. Jiang, Z. Wu, B. Ai, A. Wang, W. Zhou, and Y . Chen, “SWIFT: A scalable lightweight infrastructure for fine-tuning,” 2024. [Online]. Available: https://arxiv.org/abs/2408.05517

    work page arXiv 2024

  50. [50]

    Robust speech recognition via large-scale weak super- vision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak super- vision,” inInternational Conference on Machine Learning, 2023

    work page 2023

  51. [51]

    BERT: Pre- training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre- training of deep bidirectional transformers for language understanding,” inProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2019, pp. 4171–4186

    work page 2019

  52. [52]

    Publicly available clinical BERT embed- dings,

    E. Alsentzer, J. R. Murphy, W. Boag, W.-H. Weng, D. Jindi, T. Naumann, and M. B. A. McDermott, “Publicly available clinical BERT embed- dings,” inProceedings of the 2nd Clinical Natural Language Processing Workshop, 2019, pp. 72–78

    work page 2019

  53. [53]

    Exploring deep transfer learning techniques for alzheimer’s dementia detection,

    Y . Zhu, X. Liang, J. A. Batsis, and R. M. Roth, “Exploring deep transfer learning techniques for alzheimer’s dementia detection,”Frontiers in Computer Science, vol. 3, p. 624683, 2021

    work page 2021