Pith. sign in

REVIEW 4 major objections 6 minor 7 cited by

Fine-tuned vision-language models and a reasoning LLM can classify H-reflex waveforms from images, and the paper presents this as the first such clinical decision-support system.

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 ↗

A consortium of fine-tuned VLMs plus a reasoning LLM is proposed for H-reflex image analysis, but the claimed high accuracy is backed only by anecdotal examples.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Plausible architecture, but the central accuracy claim is unsupported: no held-out metrics, no baselines, just loss curves and a few selected examples. the 4 major comments →

arxiv 2508.12473 v1 pith:YP2UHOGV submitted 2025-08-17 cs.CV cs.AI

Standardization of Neuromuscular Reflex Analysis -- Role of Fine-Tuned Vision-Language Model Consortium and OpenAI gpt-oss Reasoning LLM Enabled Decision Support System

Eranga Bandara , Ross Gore , Sachin Shetty , Ravi Mukkamala , Christopher Rhea , Atmaram Yarlagadda , Shaifali Kaushik , L.H.M.P.De Silva
show 4 more authors
Andriy Maznychenko Inna Sokolowska Amin Hass Kasun De Zoysa
This is my paper
classification cs.CV cs.AI
keywords H-reflexEMG waveform analysisvision-language modelreasoning LLMLoRA fine-tuningQLoRAneuromuscular diagnosticsclinical decision support
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 reading

The paper tries to show that H-reflex waveform interpretation—normally done by hand on EMG traces and prone to rater variability—can be automated by a consortium of fine-tuned vision-language models whose outputs are reconciled by a reasoning LLM. Each VLM, trained on about 1,200 expert-annotated waveform images plus athlete metadata, is claimed to read amplitude, latency, and recovery status directly from the plotted image, and the LLM consensus step is claimed to make the final assessment more consistent and explainable. If true, this would give clinicians and sports scientists a standardized, scalable way to monitor fatigue, injury, and recovery from the same images they already look at. The paper supports the claim with training-loss curves and side-by-side before/after model outputs, not with held-out accuracy numbers.

Core claim

The central claim is that a hybrid architecture—three fine-tuned vision-language models (Llama-Vision, Pixtral-Vision, Qwen2-VL) coordinated by a reasoning LLM (gpt-oss)—can produce clinically interpretable H-reflex assessments from EMG waveform images. The authors contend that fine-tuning shifts each model from generic, verbose commentary to concise, expert-aligned statements about waveform features, injury type, and recovery phase, and that the reasoning LLM then synthesizes the model outputs into one consensus diagnosis. The paper presents this as the first integration of a fine-tuned VLM consortium with a reasoning LLM for image-based H-reflex analysis, and positions the platform as a fo

What carries the argument

The load-bearing machinery is the VLM consortium plus reasoning-LLM loop: LoRA/QLoRA fine-tuning adapts general vision-language models to the H-reflex image domain on about 1,200 annotated records; the LLM agent layer builds prompts; each fine-tuned VLM extracts features and predicts neuromuscular state; and the reasoning LLM aggregates the individual predictions into a final consensus assessment. The consensus step is what the paper credits for reducing single-model variability and adding explainability.

Load-bearing premise

The whole validation rests on the assumption that a VLM fine-tuned on roughly 1,200 annotated waveform images can extract accurate amplitude, latency, and recovery-state features straight from a plotted image and generalize to athletes it has never seen; the paper reports no held-out test-set accuracy and no comparison with conventional signal-processing baselines.

What would settle it

Take an independent set of, say, 200 H-reflex EMG waveforms with expert labels, render them as images, run the three fine-tuned VLMs and the gpt-oss consensus step, and compare the final diagnosis against both the expert label and a numeric measurement of amplitude and latency. If the consensus accuracy falls below expert inter-rater agreement, or if the VLM-read amplitude and latency deviate from the numeric values beyond clinical tolerance, the central claim is refuted.

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

If this is right

  • If the hybrid system works as claimed, H-reflex analysis no longer requires manual waveform measurement or numeric EMG signal processing; reading a screenshot of the waveform suffices for a standardized assessment.
  • The same architecture can be retrained for other evoked-potential waveforms (e.g., F-waves, M-waves) by swapping the curated image dataset and fine-tuning targets.
  • The consensus step should make final diagnoses less sensitive to any one VLM's blind spot, provided the models make independent errors.
  • Deployment on consumer-grade hardware via 4-bit quantization means the platform could run in clinics or at the field side, not just in a data center.
  • Athlete metadata (sport, training phase, injury history) becomes part of the diagnostic input, enabling recovery timelines tailored to the individual.

Where Pith is reading between the lines

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

  • A testable extension the paper leaves implicit is a direct comparison of the VLM consortium's amplitude and latency readings against gold-standard numeric EMG analysis on the same waveforms; the method stands or falls on whether image-derived feature readings match signal-derived numbers.
  • The before/after examples suggest that fine-tuning mostly improves output format and terminology; a sharper test would ask whether the consensus diagnosis changes the clinical decision (e.g., return-to-play) compared with expert judgment.
  • If the system generalizes, the 1,200-image scale is unusually small for fine-tuning a VLM, which hints that these waveform images are highly self-similar and the classification may be easier than natural-image benchmarks; this also means the system may be brittle to unusual waveform morphologies or acquisition settings.
  • The reasoning LLM step could introduce a subtle circularity: it is prompted with the same expert-annotated outcomes the VLMs were trained on, so its 'consensus' may reinforce the training distribution rather than independently validate it.
Share X Bluesky LinkedIn Reddit HN

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 presents NeuroLens, an end-to-end decision-support platform that interprets H-reflex EMG waveform images using a consortium of three fine-tuned vision-language models (Llama-Vision, Pixtral-Vision, Qwen2-VL) and a reasoning LLM (OpenAI-gpt-oss). The pipeline combines a data lake of ~1,200 expert-annotated records (waveform image, athlete metadata, clinical observations, recovery timelines), QLoRA-based fine-tuning on a 2/3-1/6-1/6 train/validation/test split, per-model prediction, and a consensus/refinement step by the reasoning LLM. The authors claim that the system delivers 'highly accurate, consistent, and interpretable H-reflex assessments' and that it is the first integration of a fine-tuned VLM consortium with a reasoning LLM for image-based H-reflex analysis. The reported evaluation, however, consists exclusively of training/validation loss curves for one of the three models and qualitative before/after text outputs for the three models plus one consensus example. No quantitative accuracy, held-out test-set metrics, baselines, or external validation are reported.

Significance. If the accuracy claims were substantiated, this would be a genuinely useful contribution to automated neuromuscular diagnostics: standardization of H-reflex interpretation is a real clinical need, and the architecture—multiple fine-tuned VLMs plus a reasoning LLM for consensus and explainability—is well motivated. The system design is described in enough detail to reimplement, the explicit data split is good practice, and the reporting of training time and memory use is useful engineering information. The decisive weakness is that the paper's headline claims (accuracy, consistency, clinical readiness) are not backed by any quantitative evaluation: there are no test-set metrics, no error analysis, no baseline comparison, and no external validation. As it stands, the manuscript provides an architecture description and anecdotal examples, not a validated system.

major comments (4)
  1. [§5.2, §5.3, Abstract] The abstract's claim that the system 'delivers highly accurate, consistent, and interpretable H-reflex assessments' is unsupported by the evidence reported. Section 5 defines a held-out test split (1/6 of ~1,200 records), but no metric on that split is ever reported. The evaluation in §5.2 and §5.3 is qualitative—before/after text outputs for three models (Figs. 11-13) and a single consensus example (Fig. 14)—with no statement of how the examples were selected, no counts, no confusion matrix, no sensitivity/specificity, no error bars, and no comparison against any baseline (e.g., peak-detection amplitude/latency extraction, or the EMGLLM/INSPIRE approaches cited in §6). Without test-set accuracy figures, the central claim is not established.
  2. [§5.2, §7] The evaluation is circular in a way the manuscript does not address: the 'expert-validated observations' used as ground truth in Figs. 11-13 are drawn from the same annotated Data Lake used for fine-tuning (§4.1, §5), and the paper does not state that the displayed examples come from the held-out test split or that the ground-truth annotations were produced independently of the training annotations. The close agreement of post-fine-tuning outputs with expert observations may therefore reflect fitting to the training distribution rather than generalization to new athletes. Section 7 further concedes that 'validating the system in large-scale clinical and sports environments' is future work, which directly contradicts the abstract's assertion of demonstrated accuracy.
  3. [§5.1] The fine-tuning evaluation is limited to a single model. Loss curves are reported only for Llama-3.2-11B-Vision-Instruct (Figs. 7-10); no equivalent convergence or validation evidence is given for Pixtral-Vision or Qwen2-VL, so the consortium-level claim is not individually verified. Moreover, training/validation loss and the 'area between curves' characterize optimization behavior, not diagnostic accuracy; the conclusion in §5.1 that a shrinking gap 'confirms the model's ability to generalize well' conflates a low generalization gap in loss with task performance—precisely the relationship that the missing test-set metrics should establish.
  4. [§4.2, §5.2] The central design choice—interpreting H-reflex amplitude, latency, and recovery status from rendered waveform images rather than from the numeric EMG time series—carries an unquantified precision cost that the paper never examines. A conventional baseline that detects M-wave and H-wave peaks and measures amplitude/latency directly would provide a natural comparison, but no such baseline is evaluated. The paper should demonstrate what the VLM-based approach adds over direct signal processing, e.g., by reporting agreement with a reference extraction algorithm on the same test split.
minor comments (6)
  1. [References, §2.4] Ref. [12], cited for Low-Rank Adaptation (LoRA), is actually a study of LoRa (long-range radio) for IoT; the standard LoRA reference (Hu et al.) is missing. Ref. [28], cited as 'Unsloth,' is a LlamaFactory paper. Ref. [4], cited as 'INSPIRE' in §6.4, is actually the EMGLLM paper. These citation mismatches should be corrected.
  2. [§5, §3.2, §4.4] The reasoning model is named inconsistently: 'OpenAI gpt-oss' in the title/abstract/keywords, 'OpenAI-o3' in §3.2 and §4.4, and 'OpenAI-gpt-oss' in §5. The VLM consortium is also described as Llama-Vision/Mistral-Vision/Qwen2-VL in §5 but as Llama-Vision/Pixtral-Vision/Qwen2 in §4.2, and the caption of Fig. 12 says 'Pixtral-12B-2409.' Please harmonize model names.
  3. [References] Ref. [5] is an exact duplicate of Ref. [1]; Ref. [17] ('Sallma'), cited for Llama-Vision, is a multi-agent software-architecture paper unrelated to vision-language models; and several 'Ollama'-related citations (e.g., Ref. [36]) point to unrelated papers. The bibliography needs a careful audit.
  4. [§5] The memory-utilization sentence ('14.605 GB reserved, 5.853 GB actual, equivalent to 39.69% of reserved memory and 99.03% of peak allocation') is numerically inconsistent (5.853/14.605 is ≈ 40.1%, not 39.69%). Clarify the definition of 'peak allocation.'
  5. [Fig. 10] The caption ('Ratio of training to validation loss') does not match the text, which also describes loss difference and loss derivatives; the multi-panel figure should be captioned panel by panel.
  6. [§6, Table 1] Table 1 is a self-reported qualitative comparison (✓/✗) with no supporting evidence for the entries; the claim that prior systems lack 'fine-tuning support' is asserted without citing their documentation or code. Please provide evidence for the table's entries or soften the presentation.

Circularity Check

0 steps flagged

No demonstrated circularity: the pipeline is ordinary supervised fine-tuning plus aggregation; the serious weakness is the absence of held-out quantitative evaluation, not a self-referential derivation.

full rationale

The derivation chain is: expert-annotated records are used to fine-tune VLMs (supervised learning); fine-tuned VLMs produce predictions from image+metadata inputs; OpenAI-gpt-oss aggregates those predictions into a consensus. Each step is a forward function of its inputs; there is no equation-level identity between the prediction and the training labels at inference time, and no fitted parameter is renamed as a prediction. The strongest candidate for circularity would be Section 5.2's comparison of fine-tuned predictions to 'expert-validated H-reflex observations,' because the training targets are likewise expert-annotated observations. However, Section 5 states a 2/3, 1/6, 1/6 train/validation/test split, and the paper does not state that the examples in Figures 11-14 are training samples. Claiming circularity would require showing that the evaluation set is the training set, which the text does not establish. What the paper actually omits is any reported accuracy/confusion-matrix result on the held-out test split: Section 5.1 reports only training/validation loss curves, and Sections 5.2-5.3 give qualitative before/after examples. Section 7 itself defers validation ('validating the system in large-scale clinical and sports environments'), confirming that the clinical-accuracy claim is unsupported rather than circular. The only citation concern is [23], a same-author preprint using the same VLM-consortium-plus-reasoning-LLM architecture; it is cited for the reasoning step, but the present paper's conclusions rest on its own fine-tuning curves and examples, so this self-citation is not load-bearing. Under the hard rules, unsupported evidence is a correctness/validity problem, not a demonstrated circular loop.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

This is a supervised machine learning pipeline rather than a derivation. The central accuracy claim depends on trained VLM weights that are fitted to a private expert-annotated dataset, on the choice of fine-tuning hyperparameters that are never disclosed, and on a reasoning-LLM aggregation step whose added value is not measured. No artifact to reproduce the fit is released.

free parameters (3)
  • LoRA adapter weights for the three fine-tuned VLMs = not released
    The models are fine-tuned on about 1,200 expert-annotated images; the resulting weights are the fitted values that the central 'accurate predictions' claim relies on, but neither weights nor training checkpoints are provided.
  • Fine-tuning hyperparameters (learning rate, LoRA rank, batch size, epochs) = not reported
    Training used Unsloth and QLoRA, but no hyperparameter values are given, hindering replication and leaving the fit underspecified.
  • Expert annotation and labeling criteria for fatigue, injury, and recovery = not specified
    The ground truth used for fine-tuning and evaluation is described only generically; no inter-rater reliability or protocol details are given.
axioms (4)
  • domain assumption H-reflex EMG waveform images retain sufficient electrophysiological detail for a vision-language model to infer amplitude, latency, and neuromuscular state.
    The whole approach converts a physiological time series to an image and assumes the VLM can read it; no ablation or comparison with numeric signal processing is provided (Sections 4.2-4.3).
  • domain assumption Expert annotations on about 1,200 records are a reliable ground truth for fatigue, injury, and recovery.
    The evaluation compares model outputs to these same 'expert-validated observations' without measuring inter-rater reliability or clinical validity (Section 5.2).
  • domain assumption Supervised fine-tuning on this dataset generalizes to new athletes and settings.
    No external validation set or prospective study is presented; only a 1/6 hold-out split is mentioned, but no hold-out metrics are reported (Section 5).
  • domain assumption A reasoning LLM improves consensus accuracy by aggregating VLM outputs.
    No comparison of consortium-only versus consortium-plus-reasoning-LLM accuracy is provided (Section 5.3).

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Standardization of Neuromuscular Reflex Analysis -- Role of Fine-Tuned Vision-Language Model Consortium and OpenAI gpt-oss Reasoning LLM Enabled Decision Support System." pith.science (2026). https://pith.science/paper/YP2UHOGV

@misc{pith2026250812473,
  author       = {Pith},
  title        = {Pith review of: Standardization of Neuromuscular Reflex Analysis -- Role of Fine-Tuned Vision-Language Model Consortium and OpenAI gpt-oss Reasoning LLM Enabled Decision Support System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YP2UHOGV}},
  note         = {Machine review of arXiv:2508.12473}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Accurate assessment of neuromuscular reflexes, such as the H-reflex, plays a critical role in sports science, rehabilitation, and clinical neurology. Traditional analysis of H-reflex EMG waveforms is subject to variability and interpretation bias among clinicians and researchers, limiting reliability and standardization. To address these challenges, we propose a Fine-Tuned Vision-Language Model (VLM) Consortium and a reasoning Large-Language Model (LLM)-enabled Decision Support System for automated H-reflex waveform interpretation and diagnosis. Our approach leverages multiple VLMs, each fine-tuned on curated datasets of H-reflex EMG waveform images annotated with clinical observations, recovery timelines, and athlete metadata. These models are capable of extracting key electrophysiological features and predicting neuromuscular states, including fatigue, injury, and recovery, directly from EMG images and contextual metadata. Diagnostic outputs from the VLM consortium are aggregated using a consensus-based method and refined by a specialized reasoning LLM, which ensures robust, transparent, and explainable decision support for clinicians and sports scientists. The end-to-end platform orchestrates seamless communication between the VLM ensemble and the reasoning LLM, integrating prompt engineering strategies and automated reasoning workflows using LLM Agents. Experimental results demonstrate that this hybrid system delivers highly accurate, consistent, and interpretable H-reflex assessments, significantly advancing the automation and standardization of neuromuscular diagnostics. To our knowledge, this work represents the first integration of a fine-tuned VLM consortium with a reasoning LLM for image-based H-reflex analysis, laying the foundation for next-generation AI-assisted neuromuscular assessment and athlete monitoring platforms.

Figures

Figures reproduced from arXiv: 2508.12473 by Amin Hass, Andriy Maznychenko, Atmaram Yarlagadda, Christopher Rhea, Eranga Bandara, Inna Sokolowska, Kasun De Zoysa, L.H.M.P.De Silva, Ravi Mukkamala, Ross Gore, Sachin Shetty, Shaifali Kaushik.

Figure 1
Figure 1. Figure 1: Platform architecture. 2 Background This section provides a foundational overview of the core scientific and technological concepts underpinning the proposed AI-assisted neuromuscular reflex analysis platform. In particular, we highlight the basis and clinical importance of the H-reflex in neuromuscular analysis, recent advancements in Large Language Models (LLMs), reasoning-capable LLMs, fine-tuning techn… view at source ↗
Figure 2
Figure 2. Figure 2: LLM integration flow with Ollama LLM-API [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative samples from the H-reflex neuromuscular dataset used to fine-tune the VLMs. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Fine-tune VLM with Qlora and deploy with Ollama. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Prompt for OpenAI-gpt-oss reasoning LLM for final prediction reasoning. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The required data format of the unsloth library to fine-tune the VLMs. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 3
Figure 3. Figure 3: The Unsloth framework requires input data to be structured in an instruction-based format [28]. To meet this requirement, the dataset was preprocessed and transformed into the required schema, shown in [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 7
Figure 7. Figure 7: Training loss during fine-tuning of the Llama-3.2-11B-Vision-Instruct LLM [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Validation loss during fine-tuning of the Llama-3.2-11B-Vision-Instruct LLM [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Training vs. Validation Loss and Area Between Curves during Fine-Tuning of the Llama-3.2-11B-Vision [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Ratio of training to validation loss during the fine-tuning of the Llama-3.2-11B-Vision-Instruct LLM. [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The prediction results of Llama-3.2-11B-Vision-Instruct vision language model. [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The prediction results of Pixtral-12B-2409 vision language model. [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The prediction results of Qwen2-VL-7B-Instruct vision language model. [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Final reasoning prediction done by the OpenAI-gpt-oss LLM through multi-model consensus. [PITH_FULL_IMAGE:figures/full_fig_p017_14.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 7 Pith papers

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

  1. MindGap: A Conversational AI Framework for Upstream Neuroplastic Intervention in Post-Traumatic Stress Disorder

    cs.AI 2026-05 unverdicted novelty 5.0

    The paper proposes MindGap, an on-device AI framework using dependent origination to guide PTSD patients through progressive observation layers for upstream neuroplastic intervention rather than downstream symptom management.

  2. Think Before You Act -- A Neurocognitive Governance Model for Autonomous AI Agents

    cs.AI 2026-04 unverdicted novelty 5.0

    A neurocognitive governance model formalizes a Pre-Action Governance Reasoning Loop that consults global, workflow, agent, and situational rules before each action, yielding 95% compliance accuracy with zero false esc...

  3. Train the Trainers -- An Agentic AI Framework for Peer-Based Mental Health Support in Battlefield Environments

    cs.HC 2026-03 unverdicted novelty 5.0

    The paper introduces an agentic AI platform to train and support recovered soldiers as peer facilitators providing mental health triage and interventions in austere battlefield environments.

  4. Towards an Agent-First Web: Redesigning the Web for AI Agents

    cs.AI 2026-06 unverdicted novelty 4.0

    Proposes ten design principles for an agent-first web with changes to access (agent identification and dual content), economics (intent-based tiers and tokens), and content (ATML and provenance chains) to address bloc...

  5. Toward Zero-Egress Psychiatric AI: On-Device LLM Deployment for Privacy-Preserving Mental Health Decision Support

    cs.AI 2026-04 unverdicted novelty 4.0

    A cross-platform mobile application deploys an ensemble of quantized open-source LLMs for fully local, DSM-5-aligned psychiatric decision support with claimed accuracy comparable to prior cloud versions.

  6. Flowr -- Scaling Up Retail Supply Chain Operations Through Agentic AI in Large Scale Supermarket Chains

    cs.AI 2026-04 unverdicted novelty 3.0

    Flowr is an agentic AI framework that decomposes retail supply chain workflows into coordinated LLM-based agents with human-in-the-loop oversight to automate operations in large supermarket chains.

  7. AI Trust OS -- A Continuous Governance Framework for Autonomous AI Observability and Zero-Trust Compliance in Enterprise Environments

    cs.AI 2026-04 unverdicted novelty 3.0

    AI Trust OS is a proposed always-on operating layer that discovers undocumented AI systems via telemetry and produces continuous zero-trust compliance artifacts for regulations including ISO 42001, EU AI Act, SOC 2, G...

Reference graph

Works this paper leans on

45 extracted references · 27 canonical work pages · cited by 7 Pith papers

  1. [2]

    Changes in h-reflex, v-wave, and contractile properties of the plantar flexors following concurrent exercise sessions—the acute interference effect

    Miguel Gomes, André D Gonçalves, Pedro Pezarat-Correia, and Goncalo V Mendonca. Changes in h-reflex, v-wave, and contractile properties of the plantar flexors following concurrent exercise sessions—the acute interference effect. Journal of Applied Physiology, 138(2):327–341, 2025

    work page 2025

  2. [3]

    Artificial intelligence models using f-wave responses predict amyotrophic lateral sclerosis

    Jennifer M Martinez-Thompson, Kevin A Mazurek, Carolina Parra-Cantu, Elie Naddaf, Venkatsampath Gogineni, Hugo Botha, David T Jones, Ruple S Laughlin, Leland Barnard, and Nathan P Staff. Artificial intelligence models using f-wave responses predict amyotrophic lateral sclerosis. Brain, page awaf014, 2025

    work page 2025

  3. [4]

    Emgllm: Data-to-text alignment for electromyogram diagnosis generation with medical numerical data encoding

    Zefei Long, Zhenbiao Cao, Wei Chen, and Zhongyu Wei. Emgllm: Data-to-text alignment for electromyogram diagnosis generation with medical numerical data encoding. In Findings of the Association for Computational Linguistics: ACL 2025, pages 20470–20480, 2025

    work page 2025

  4. [5]

    Neurophysiological assessment of h-reflex alterations in compressive radiculopathy

    Petr Šádek, Ela Hrušková, Svatopluk Ostr `y, and Jakub Otáhal. Neurophysiological assessment of h-reflex alterations in compressive radiculopathy. Physiological Research, 73(3):427, 2024

    work page 2024

  5. [7]

    Sgva-clip: Semantic-guided visual adapting of vision-language models for few-shot image classification

    Fang Peng, Xiaoshan Yang, Linhui Xiao, Yaowei Wang, and Changsheng Xu. Sgva-clip: Semantic-guided visual adapting of vision-language models for few-shot image classification. IEEE Transactions on Multimedia, 26:3469–3480, 2023

    work page 2023

  6. [8]

    Llm as a mastermind: A survey of strategic reasoning with large language models

    Yadong Zhang, Shaoguang Mao, Tao Ge, Xun Wang, Adrian de Wynter, Yan Xia, Wenshan Wu, Ting Song, Man Lan, and Furu Wei. Llm as a mastermind: A survey of strategic reasoning with large language models. arXiv preprint arXiv:2404.01230, 2024

    Pith/arXiv arXiv 2024

  7. [9]

    Evaluating ai reasoning models in pediatric medicine: A comparative analysis of o3-mini and o3-mini-high

    Gianluca Mondillo, Mariapia Masino, Simone Colosimo, Alessandra Perrotta, and Vittoria Frattolillo. Evaluating ai reasoning models in pediatric medicine: A comparative analysis of o3-mini and o3-mini-high. medRxiv, pages 2025–02, 2025

    work page 2025

  8. [10]

    Agentic ai: Autonomous intelligence for complex goals–a comprehensive survey

    Deepak Bhaskar Acharya, Karthigeyan Kuppan, and B Divya. Agentic ai: Autonomous intelligence for complex goals–a comprehensive survey. IEEE Access, 2025

    work page 2025

  9. [11]

    Understanding the planning of llm agents: A survey

    Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. Understanding the planning of llm agents: A survey. arXiv preprint arXiv:2402.02716, 2024

    Pith/arXiv arXiv 2024

  10. [12]

    A study of lora: Long range & low power networks for the internet of things

    Aloÿs Augustin, Jiazi Yi, Thomas Clausen, and William Townsley. A study of lora: Long range & low power networks for the internet of things. Sensors, 16(9):1466, 2016

    work page 2016

  11. [13]

    Qlora: Efficient finetuning of quantized llms

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36, 2024

    2024

  12. [14]

    On the benefit of fmg and emg sensor fusion for gesture recognition using cross- subject validation

    Maurice Rohr, Jad Haidamous, Niklas Schäfer, Stephan Schaumann, Bastian Latsch, Mario Kupnik, and Christoph Hoog Antink. On the benefit of fmg and emg sensor fusion for gesture recognition using cross- subject validation. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2025. 19 A PREPRINT - S EPTEMBER 25, 2025

    work page 2025

  13. [15]

    Vision-language models for vision tasks: A survey

    Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  14. [16]

    The llama 3 herd of models

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

  15. [17]

    Sallma: A software architecture for llm-based multi-agent systems

    Marco Becattini, Roberto Verdecchia, and Enrico Vicario. Sallma: A software architecture for llm-based multi-agent systems

    work page

  16. [18]

    Fine-tuning mistral 7b large language model for python query response and code generation: A parameter efficient approach

    Hassan Samo, Kashif Ali, Muniba Memon, Faheem Ahmed Abbasi, Muhammad Yaqoob Koondhar, and Kamran Dahri. Fine-tuning mistral 7b large language model for python query response and code generation: A parameter efficient approach. VAWKUM Transactions on Computer Sciences, 12(1):205–217, 2024

    work page 2024

  17. [19]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024

    Pith/arXiv arXiv 2024

  18. [20]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Nee- lakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020

    1901

  19. [21]

    Google gemini as a next generation ai educational tool: a review of emerging educational technology

    Muhammad Imran and Norah Almusharraf. Google gemini as a next generation ai educational tool: a review of emerging educational technology. Smart Learning Environments, 11(1):22, 2024

    work page 2024

  20. [22]

    Tracing information flow in llama vision: A step toward multimodal understanding

    Alessia Saporita, Vittorio Pipoli, Federico Bolelli, Lorenzo Baraldi, Andrea Acquaviva, Elisa Ficarra, et al. Tracing information flow in llama vision: A step toward multimodal understanding. In Proceedings of the 21st International Conference in Computer Analysis of Images and Patterns, 2025

    work page 2025

  21. [23]

    Proof-of-tbi–fine-tuned vision language model consortium and openai-o3 reasoning llm-based medical diagnosis support system for mild traumatic brain injury (tbi) prediction

    Ross Gore, Eranga Bandara, Sachin Shetty, Alberto E Musto, Pratip Rana, Ambrosio Valencia-Romero, Christo- pher Rhea, Lobat Tayebi, Heather Richter, Atmaram Yarlagadda, et al. Proof-of-tbi–fine-tuned vision language model consortium and openai-o3 reasoning llm-based medical diagnosis support system for mild traumatic brain injury (tbi) prediction. arXiv p...

    Pith/arXiv arXiv 2025

  22. [24]

    Estimating worst-case frontier risks of open-weight llms

    Eric Wallace, Olivia Watkins, Miles Wang, Kai Chen, and Chris Koch. Estimating worst-case frontier risks of open-weight llms. arXiv preprint arXiv:2508.03153, 2025

    Pith/arXiv arXiv 2025

  23. [25]

    Stop reasoning! when multimodal llm with chain-of-thought reasoning meets adversarial image

    Zefeng Wang, Zhen Han, Shuo Chen, Fan Xue, Zifeng Ding, Xun Xiao, V olker Tresp, Philip Torr, and Jindong Gu. Stop reasoning! when multimodal llm with chain-of-thought reasoning meets adversarial image. arXiv preprint arXiv:2402.14899, 2024

    Pith/arXiv arXiv 2024

  24. [26]

    Bouk, Sachin Shetty, Ross Gore, Sastry Kompella, Ravi Mukkamala, Abdul Rahman, Peter Foytik, Xueping Liang, Ng Wee Keong, and Kasun De Zoysa

    Eranga Bandara, Safdar H. Bouk, Sachin Shetty, Ross Gore, Sastry Kompella, Ravi Mukkamala, Abdul Rahman, Peter Foytik, Xueping Liang, Ng Wee Keong, and Kasun De Zoysa. Bassa-llama — fine-tuned meta’s llama llm, blockchain and nft enabled real-time network attack detection platform for wind energy power plants. In 2025 International Wireless Communications...

    work page 2025

  25. [27]

    Bouk, Sachin Shetty, Ross Gore, Sastry Kompella, Ravi Mukkamala, Abdul Rahman, Peter Foytik, Xueping Liang, Ng Wee Keong, and Kasun De Zoysa

    Eranga Bandara, Safdar H. Bouk, Sachin Shetty, Ross Gore, Sastry Kompella, Ravi Mukkamala, Abdul Rahman, Peter Foytik, Xueping Liang, Ng Wee Keong, and Kasun De Zoysa. Vindsec-llama — fine-tuned meta’s llama-3 llm, federated learning, blockchain and pbom-enabled data security architecture for wind energy data platforms. In 2025 International Wireless Comm...

    work page 2025

  26. [28]

    Llamafactory: Unified efficient fine-tuning of 100+ language models

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of 100+ language models. arXiv preprint arXiv:2403.13372, 2024

    Pith/arXiv arXiv 2024

  27. [29]

    Performance comparision of tpu, gpu, cpu on google colaboratory over distributed deep learning

    Haklin Kimm, Incheon Paik, and Hanke Kimm. Performance comparision of tpu, gpu, cpu on google colaboratory over distributed deep learning. In 2021 IEEE 14th International Symposium on Embedded Multicore/Many-core Systems-on-Chip (MCSoC), pages 312–319. IEEE, 2021

    work page 2021

  28. [30]

    Lohan: Low- cost high-performance framework to fine-tune 100b model on a consumer gpu

    Changyue Liao, Mo Sun, Zihan Yang, Jun Xie, Kaiqi Chen, Binhang Yuan, Fei Wu, and Zeke Wang. Lohan: Low- cost high-performance framework to fine-tune 100b model on a consumer gpu. arXiv preprint arXiv:2403.06504, 2024

    Pith/arXiv arXiv 2024

  29. [31]

    Llama-recipe—fine-tuned meta’s llama llm, pbom and nft enabled 5g network-slice orchestration and end-to-end supply-chain verification platform

    Eranga Bandara, Safdar H Bouk, Sachin Shetty, Sandip Roy, Ravi Mukkamala, Abdul Rahman, Peter Foytik, Xueping Liang, Ng Wee Keong, and Kasun De Zoysa. Llama-recipe—fine-tuned meta’s llama llm, pbom and nft enabled 5g network-slice orchestration and end-to-end supply-chain verification platform. In 2025 IEEE 22nd Consumer Communications & Networking Confer...

    work page 2025

  30. [32]

    Prompt engineering in large language models

    Ggaliwango Marvin, Nakayiza Hellen, Daudi Jjingo, and Joyce Nakatumba-Nabende. Prompt engineering in large language models. In International Conference on Data Intelligence and Cognitive Informatics, pages 387–402. Springer, 2023. 20 A PREPRINT - S EPTEMBER 25, 2025

    work page 2023

  31. [33]

    Data-efficient fine-tuning for llm-based recommendation

    Xinyu Lin, Wenjie Wang, Yongqi Li, Shuo Yang, Fuli Feng, Yinwei Wei, and Tat-Seng Chua. Data-efficient fine-tuning for llm-based recommendation. In Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval, pages 365–374, 2024

    work page 2024

  32. [34]

    A tutorial on llm reasoning: Relevant methods behind chatgpt o1

    Jun Wang. A tutorial on llm reasoning: Relevant methods behind chatgpt o1. arXiv preprint arXiv:2502.10867, 2025

    Pith/arXiv arXiv 2025

  33. [35]

    Wedagpt—generative-ai (with custom-trained meta’s llama2 llm), blockchain, self sovereign identity, nft and model card enabled indigenous medicine platform

    Eranga Bandara, Peter Foytik, Sachin Shetty, Ravi Mukkamala, Abdul Rahman, Xueping Liang, Ng Wee Keong, and Kasun De Zoysa. Wedagpt—generative-ai (with custom-trained meta’s llama2 llm), blockchain, self sovereign identity, nft and model card enabled indigenous medicine platform. In 2024 IEEE Symposium on Computers and Communications (ISCC), pages 1–6. IEEE, 2024

    work page 2024

  34. [36]

    Artificial intelligence to automate network meta-analyses: Four case studies to evaluate the potential application of large language models

    Tim Reason, Emma Benbow, Julia Langham, Andy Gimblett, Sven L Klijn, and Bill Malcolm. Artificial intelligence to automate network meta-analyses: Four case studies to evaluate the potential application of large language models. PharmacoEconomics-Open, pages 1–16, 2024

    work page 2024

  35. [37]

    Incorporating dialect understanding into llm using rag and prompt engineering techniques for causal commonsense reasoning

    Benedikt Perak, Slobodan Beliga, and Ana Meštrovi´c. Incorporating dialect understanding into llm using rag and prompt engineering techniques for causal commonsense reasoning. In Proceedings of the Eleventh Workshop on NLP for Similar Languages, Varieties, and Dialects (VarDial 2024), pages 220–229, 2024

    work page 2024

  36. [38]

    On-device qwen2

    Maoyang Xiang, Ramesh Fernando, and Bo Wang. On-device qwen2. 5: Efficient llm inference with model compression and hardware acceleration. arXiv preprint arXiv:2504.17376, 2025

    Pith/arXiv arXiv 2025

  37. [39]

    Vtutor: An open-source sdk for generative ai-powered animated pedagogical agents with multi-media output

    Eason Chen, Chenyu Lin, Xinyi Tang, Aprille Xi, Canwen Wang, Jionghao Lin, and Kenneth R Koedinger. Vtutor: An open-source sdk for generative ai-powered animated pedagogical agents with multi-media output. arXiv preprint arXiv:2502.04103, 2025

    Pith/arXiv arXiv 2025

  38. [40]

    Survey on evaluation of llm-based agents

    Asaf Yehudai, Lilach Eden, Alan Li, Guy Uziel, Yilun Zhao, Roy Bar-Haim, Arman Cohan, and Michal Shmueli- Scheuer. Survey on evaluation of llm-based agents. arXiv preprint arXiv:2503.16416, 2025

    Pith/arXiv arXiv 2025

  39. [41]

    Pitfalls in training and validation of deep learning systems

    Tom Eelbode, Pieter Sinonquel, Frederik Maes, and Raf Bisschops. Pitfalls in training and validation of deep learning systems. Best Practice & Research Clinical Gastroenterology, 52:101712, 2021

    work page 2021

  40. [42]

    Llm unlearning via loss adjustment with only forget data

    Yaxuan Wang, Jiaheng Wei, Chris Yuhao Liu, Jinlong Pang, Quan Liu, Ankit Parag Shah, Yujia Bao, Yang Liu, and Wei Wei. Llm unlearning via loss adjustment with only forget data. arXiv preprint arXiv:2410.11143, 2024

    Pith/arXiv arXiv 2024

  41. [43]

    Probabilistic inference layer integration in mistral llm for accurate information retrieval

    Bing Wang, Shiyu Wang, and Qian Ouyang. Probabilistic inference layer integration in mistral llm for accurate information retrieval. 2024

    work page 2024

  42. [44]

    Vit-hgr: Vision transformer-based hand gesture recognition from high density surface emg signals

    Mansooreh Montazerin, Soheil Zabihi, Elahe Rahimian, Arash Mohammadi, and Farnoosh Naderkhani. Vit-hgr: Vision transformer-based hand gesture recognition from high density surface emg signals. In 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 5115–5119. IEEE, 2022

    work page 2022

  43. [45]

    Interpreting stroke-impaired electromyography patterns through explainable artificial intelligence

    Iqram Hussain and Rafsan Jany. Interpreting stroke-impaired electromyography patterns through explainable artificial intelligence. Sensors, 24(5):1392, 2024

    work page 2024

  44. [46]

    Can llms understand unvoiced speech? exploring emg-to-text conversion with llms

    Payal Mohapatra, Akash Pandey, Xiaoyuan Zhang, and Qi Zhu. Can llms understand unvoiced speech? exploring emg-to-text conversion with llms. arXiv preprint arXiv:2506.00304, 2025

    Pith/arXiv arXiv 2025

  45. [47]

    Advanced multiscale machine learning for nerve conduction velocity analysis.Scientific Reports, 15(1):23399, 2025

    Hossein Sadeghi. Advanced multiscale machine learning for nerve conduction velocity analysis.Scientific Reports, 15(1):23399, 2025. 21

    work page 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.