Pith. sign in

REVIEW 3 major objections 5 minor 4 references

The paper claims that a single linear projection, trained jointly with an LLM, suffices to map raw ECG signals into language-model space and match far more complex encoder-based ECG-language models—and that current benchmarks reward languag

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 single linear projection from raw ECG to LLM embeddings matches complex encoder-based ECG-language models, while perturbation tests show such models largely ignore the ECG signal.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A useful diagnostic for whether ECG-language models actually use ECGs, packaged inside an architecture paper whose central claim the data don't support. the 3 major comments →

arxiv 2601.18798 v3 pith:HQXHW5HP submitted 2026-01-05 cs.MM cs.AI

ELF: A Family of Encoder-Free ECG-Language Models

William Han , Tony Chen , Chaojing Duan , Xiaoyu Song , Yihang Yao , Yuzhe Yang , Michael A. Rosenberg , Emerson Liu
show 1 more author
Ding Zhao
This is my paper
classification cs.MM cs.AI
keywords ECG-language modelsencoder-freelinear projectionmultimodal LLMECG-QAperturbation analysislanguage priorsbenchmark evaluation
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 sets out to test whether the architectural machinery dominant in current ECG-language models—pretrained ECG encoders, vision encoders, and multi-stage training—is genuinely necessary for strong text-generation performance. It introduces ELF, which flattens a raw 12-lead ECG signal and maps it into an LLM's embedding space through a single trainable linear projection, trained end-to-end with the LLM on an autoregressive objective. Across five ECG-text datasets and three LLM backbones, ELF matches or slightly beats or trails within noise of far more complex encoder-based baselines, while two slightly augmented variants (patching, convolutions) give no meaningful gain. The paper's sharper, second claim is diagnostic: replacing the ECG with zeros or omitting it entirely leaves ELF's accuracy largely intact, from which the authors conclude that current ELMs often answer from language priors and benchmark artifacts rather than from the physiological signal. If both claims hold, the field can safely discard encoder complexity and must instead invest in benchmarks that genuinely require ECG understanding.

Core claim

On its own terms, the paper establishes two main findings. The first is a sufficiency result: a single projection z = Wx + b applied to the flattened 12-lead ECG signal and inserted as a single embedding token into the LLM input is enough to keep pace with ELMs using separately pretrained ECG/vision encoders or learned symbolic tokenizers, across ECG-QA, instruction-following, and report-generation datasets. Ablations that add inductive biases—patching the signal into subsequences (Patch ELF) or adding two convolutional layers before pooling (Conv. ELF)—yield only marginal differences, and the relative ordering holds across three LLM backbones. The second finding is a negative result about e

What carries the argument

Two mechanisms do the work. The first is the linear projection layer: for a 12-lead ECG X ∈ R^{C×L}, ELF flattens it to x ∈ R^{CL} and computes z = Wx + b with W ∈ R^{d×CL}, producing a single d-dimensional embedding that replaces a '<signal>' placeholder token among the text-query embeddings, then the whole sequence is processed autoregressively by the LLM with only the response tokens masked. The second is the perturbation probe: running the same model with the ECG replaced by a zeros tensor or dropped altogether isolates how much of the response depends on the ECG path versus on text and learned priors. Together they let the paper attribute performance differences between ELM architecture

Load-bearing premise

The load-bearing premise is that the evaluated benchmarks and exact-string-match accuracy actually measure ECG-grounded understanding; if, as the paper concedes, many questions can be answered from majority-class priors and clinically irrelevant text, then ELF's parity with encoder models does not by itself show that a linear projection preserves ECG information.

What would settle it

A benchmark with class-balanced answers and ECG-specific questions that text-only models cannot solve: if, on such a benchmark, an ELF trained with ECG inputs performs no better than an ELF trained with text-only inputs (or a majority-class predictor), while a model with a dedicated ECG encoder does substantially better, then the encoder-free sufficiency claim fails; conversely, if ELF matches the encoder model, the claim is supported.

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

If this is right

  • Pretrained ECG encoders become an optional rather than necessary ingredient; a jointly trained linear projection captures what is needed for current ECG-language tasks, cutting training cost and complexity.
  • Accuracy numbers on existing ECG-QA-style benchmarks are not trustworthy signals of ECG understanding, since text-only models reach near-parity; reported scores should be read with this caveat.
  • ELM evaluation should adopt perturbation diagnostics (zeros/omission) as a standard sanity check, and benchmark designers should build class-balanced, ECG-specific questions that text-only models cannot answer.
  • Because ELF uses a single projected token, the LLM must compress a full 5-second, 12-lead signal into one embedding; that it works indicates current tasks are either easy or dominated by language, so harder tasks will stress this bottleneck.
  • The architecture removes the need for separately pretrained encoders and their data-intensive pretraining, lowering the computational barrier to training ELMs.

Where Pith is reading between the lines

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

  • A natural next experiment, not run in the paper, is to test ELF on a purpose-built, class-balanced ECG-QA set selected precisely because text priors fail; if the single projection still holds its own against encoders there, the sufficiency claim moves from 'sufficient for current benchmarks' to 'sufficient for ECG understanding.'
  • The same zeros-versus-text protocol could be transferred to other multimodal biomedical tasks (e.g., chest X-ray or pathology image LLMs) to quantify how often those models rely on textual priors rather than image content.
  • The striking result that training with text-only inputs and testing with ECG inputs slightly improved accuracy suggests the model is using the ECG as a style cue or attending to it only after the text stream establishes an answer; disentangling this would clarify when the signal matters.
  • If single-token compression is the bottleneck, increasing the number of projected tokens (patches) should matter more on harder, ECG-grounded benchmarks than it did here—where conv and patch variants gained at most ~0.9 accuracy points.
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

3 major / 5 minor

Summary. The paper introduces ELF, a family of encoder-free ECG-language models that map a flattened raw ECG into the LLM embedding space with a single linear projection layer and train the projection and LLM jointly. Across several ECG-QA and ECG-text datasets, ELF is compared with ELMs that use pretrained ECG encoders, vision encoders, and symbolic ECG tokenizers. The paper reports that ELF matches or sometimes outperforms these more complex baselines, that adding patch/convolutional inductive biases gives only marginal gains, and that perturbation experiments (replacing the ECG with zeros or omitting it) indicate current ELMs rely heavily on language priors and benchmark artifacts rather than ECG-derived information. The authors also acknowledge in their Limitations that the evaluation benchmarks may be answerable via majority-class shortcuts and clinically irrelevant text.

Significance. If the central claim held, this would be a valuable simplicity result: it would show that a single linear projection from raw ECG is sufficient to match much heavier encoder-based ELMs on current tasks. The paper is also constructive in reporting three-seed means and standard deviations, making the comparison transparent, and in releasing code/data. The perturbation diagnostic is a useful contribution and is likely the most robust part of the paper: it makes a falsifiable negative point about current evaluation practices. However, the positive architectural claim is not actually tested if the benchmarks do not require ECG understanding, as the paper itself suggests. The main value of the paper may therefore lie in its negative result rather than in the sufficiency of a linear projection.

major comments (3)
  1. [§4.5, Figure 4, Tables 6–7] The perturbation experiments undermine the central evaluation. For ELF on PTB-XL, training with ECG and evaluating with only text yields 30.05% accuracy versus 31.43% with the ECG; on ECG-Chat, only-text inference gives 13.74% versus 14.55% with the ECG. The same pattern appears for Llama-3.2-1B-Instruct + CLIP and OpenTSLM (Table 7). This is not a minor caveat: if replacing the ECG with zeros or text-only input changes accuracy by only a few points, the evaluated benchmarks do not require ECG understanding. The paper acknowledges this in Limitations but still uses benchmark parity to support Contribution 1 ('single projection is sufficient'). The parity result is consistent with the null hypothesis that the ECG input is effectively ignored. To make the architectural claim, please add (i) a text-only/majority-class baseline, (ii) per-category accuracy for single-verify/choose/query, and
  2. [Abstract, §4.1, Table 11] The phrase 'often outperform' in the abstract is not supported by the main comparison. On PTB-XL ECG-QA, ELF has accuracy 31.43 ± 0.30, lower than MERL (32.05 ± 0.49) and ST-MEM (32.12 ± 0.31), with differences within one standard deviation. On MIMIC-IV ECG-QA and ECG-Chat, ELF is in the middle of the pack. The clearest ELF wins are the PULSE transfer setting (Table 1) and the Qwen backbone (Table 3). Please replace 'often outperform' with a more precise claim such as 'is competitive with', and if superiority is claimed, support it with paired significance tests or effect sizes.
  3. [§5, Discussion] The paper concludes that 'a single projection is sufficient for current ECG-language tasks', but the perturbation results show that these tasks are substantially language tasks. This conclusion should be reformulated as either 'a single projection is sufficient to match encoder-based ELMs on benchmarks that do not require ECG understanding' or the paper should introduce at least one evaluation setting that forces ECG-grounded reasoning. Without such a setting, the central architectural claim is not demonstrated, regardless of how many standard deviations are reported.
minor comments (5)
  1. [Table 3] The gemma-2-2b-it + CLIP row reports exactly the same values as Qwen2.5-1.5B-Instruct + CLIP (BLEU-4 14.62 ± 0.78, Accuracy 34.89 ± 0.36). This is likely a copy/paste error and should be corrected.
  2. [Abstract vs. Full Text] The abstract gives the code URL as github.com/ELM-Research/ECG-Language-Models, while the full text gives https://github.com/willxxy/ECG-Bench. These should be reconciled.
  3. [§3.1 and §4] The abstract and Contribution 1 say 'across five datasets', but the main comparative table (Table 11) covers three datasets (PTB-XL ECG-QA, MIMIC-IV-ECG ECG-QA, ECG-Chat Instruct). PULSE ECG-Instruct/PULSE ECG-Bench appear in Table 1 with only three models. Please clarify which datasets are used for the headline comparison and which are used in individual experiments.
  4. [§3.2, ECG-Byte] The paper modifies the ECG-Byte tokenizer by using instance normalization instead of the original dataset-wide percentile normalization. Since ECG-Byte serves as a baseline, the impact of this modification on the reported tokenizer performance should be stated explicitly.
  5. [§4.1, Table 2] OpenTSLM is described as evaluated in a zero-shot setting, but it was trained on the ECG-QA-CoT dataset derived from PTB-XL ECG-QA. Calling this 'zero-shot' is misleading; 'out-of-box evaluation with potential training overlap' would be more accurate.

Circularity Check

0 steps flagged

No significant circularity: the paper's central claim is an empirical architecture comparison, not a derivation from self-cited premises.

full rationale

The paper's central claim is empirical: a single jointly-trained linear projection (z = Wx + b) matches encoder-based ELMs on five public ECG-text benchmarks. There is no analytic derivation whose conclusion is assumed in its premises, no fitted parameter relabeled as a prediction, and no uniqueness or ansatz imported from the authors' prior work to force the architecture. The self-citations to ECG-Byte (Han et al., 2024) and the earlier unified representation framework (Han et al., 2025) are used as a baseline input representation and as definitions of ECG representations, not as evidence that ELF's single projection is sufficient; that claim is tested directly against external baselines (CLIP, ViT, SigLIP, ST-MEM, MERL) on held-out test splits. The zero/only-text perturbation experiments are independent probes of ECG usage, and the paper's own Limitations candidly state that current ECG-QA benchmarks can be inflated by majority-choice answers and contain clinically irrelevant text. That admission undermines the external validity of the headline claim, but it is a benchmark-validity/correctness concern, not circular reasoning: ELF's competitive performance is not equivalent by construction to its input representation. No load-bearing step reduces to a self-citation.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No physically new entities are introduced. The only 'invention' is the ELF architecture itself, which requires no independent falsifiable handle beyond the benchmark comparisons. The free parameters listed are modeling choices that the empirical claims depend on.

free parameters (4)
  • Projected ECG embedding length (single token) = 1 token
    Hand-chosen design based on VLM token-reduction findings; ELF's capability claim is conditional on collapsing the entire 12-lead 5-second signal into one embedding.
  • Training example cap = 400,000
    Applied to all LLM/dataset combinations; this budget affects how fairly encoders and ELF are compared.
  • ECG instance normalization range = [0,1]
    Diverges from ECG-Byte's global percentile normalization; changes the ECG-Symbol baseline and may affect fairness.
  • Patch size Lp for Patch ELF = unspecified
    Defined symbolically but never given numerically; Patch ELF results are not fully specified.
axioms (4)
  • domain assumption Autoregressive maximum-likelihood training with exact-match accuracy is a valid measure of ECG-language task performance.
    Section 3.3-3.4; the paper later questions this for single-verify/choose items, so the assumption is load-bearing and partly self-undermined.
  • domain assumption Flattening a 12-lead 5-second ECG into one d-dimensional vector preserves task-relevant information.
    Figure 1 and Section 3.3 design of ELF; if false, parity claims would be artifacts.
  • ad hoc to paper Excluding comparison-style questions from ECG-QA does not change relative conclusions.
    Section 3.1 says comparison-style questions are not considered; no count or sensitivity analysis is provided.
  • domain assumption The 70:30 data split is leakage-free without patient-level partitioning.
    Section 3.1 states only 70:30 training/testing splits; no patient-level split is mentioned, so performance may be optimistic.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of ELF: A Family of Encoder-Free ECG-Language Models." pith.science (2026). https://pith.science/paper/HQXHW5HP

@misc{pith2026260118798,
  author       = {Pith},
  title        = {Pith review of: ELF: A Family of Encoder-Free ECG-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HQXHW5HP}},
  note         = {Machine review of arXiv:2601.18798}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

ECG-Language Models (ELMs) extend recent advances in Multimodal Large Language Models (MLLMs) to automated ECG interpretation. However, most existing ELMs inherit Vision-Language Model (VLM) design choices and rely on pretrained ECG encoders, introducing substantial architectural and training complexity. Inspired by encoder-free VLMs, we introduce ELF, a family of three encoder-free ELMs that remain competitive with, and often outperform, prior state-of-the-art ELMs across two datasets despite substantially simpler architectures and training pipelines. All code and data are available at github.com/ELM-Research/ECG-Language-Models.

Figures

Figures reproduced from arXiv: 2601.18798 by Chaojing Duan, Ding Zhao, Emerson Liu, Michael A. Rosenberg, Tony Chen, William Han, Xiaoyu Song, Yihang Yao, Yuzhe Yang.

Figure 1
Figure 1. Figure 1: The Architecture of ELF. Given an ECG signal X ∈ R C×L, where C is the number of leads and L the signal length, we flatten it to x ∈ R CL and project it into a d-dimensional latent vector z = W x + b, with W ∈ R d×(CL) and b ∈ R d . The projected ECG embedding is inserted into the textual query embeddings E ∈ R T ×d and processed autoregressively by an LLM to generate outputs y1:T . LLM embedding space wit… view at source ↗
Figure 2
Figure 2. Figure 2: Averaged accuracy across three random seeds [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: We compare ELF’s performance against strong ELMs with dedicated encoders when evaluating across increasing steps trained. All results are averaged across three random seeds and trained/evaluated on the PTB-XL ECG-QA dataset. visual tokens, even down to a single token, can still preserve strong performance (Li et al., 2025b; Lau￾rençon et al., 2024). We provide the conversation template and system prompt us… view at source ↗
Figure 4
Figure 4. Figure 4: We train and inference ELF under three train [PITH_FULL_IMAGE:figures/full_fig_p007_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

4 extracted references · 3 linked inside Pith

  1. [2]

    Preprint, arXiv:2306.06340

    Ecgbert: Understanding hidden language of ecgs with self-supervised representation learning. Preprint, arXiv:2306.06340. Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. 2024. Unveil- ing encoder-free vision-language models.Preprint, arXiv:2406.11832. Haiwen Diao, Xiaotong Li, Yufeng Cui, Yueze Wang, Haoge Deng, Ting Pan, We...

    Pith/arXiv arXiv 2024

  2. [2023]

    Seokmin Choi, Sajad Mousavi, Phillip Si, Haben G

    Fuyu-8b: A multimodal architecture for ai agents. Seokmin Choi, Sajad Mousavi, Phillip Si, Haben G. Yhdego, Fatemeh Khadem, and Fatemeh Afghah

  3. [2024]

    An- swer:

    Ecg-chat: A large ecg-language model for car- diac disease diagnosis.Preprint, arXiv:2408.08849. Jianwei Zheng, Jianming Zhang, Sidy Danioko, Hai Yao, Hangyuan Guo, and Cyril Rakovski. 2020. A 12-lead electrocardiogram database for arrhythmia re- search covering more than 10,000 patients.Scientific Data, 7:48. Hongling Zhu, Cheng Cheng, Hang Yin, Xingyi L...

    Pith/arXiv arXiv 2020

  4. [2025]

    William Han, Chaojing Duan, Michael A

    Signal, image, or symbolic: Exploring the best input representation for electrocardiogram-language models through a unified framework.Preprint, arXiv:2505.18847. William Han, Chaojing Duan, Michael A. Rosenberg, Emerson Liu, and Ding Zhao. 2024. Ecg-byte: A to- kenizer for end-to-end generative electrocardiogram language modeling.Preprint, arXiv:2412.1437...

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.