Pith. sign in

REVIEW 6 major objections 5 minor 24 references

ALAS: An Automatic Latent Alignment Score for Audio Language Models

T0 review · 6 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read ALAS proposes a training-free, layer-wise score for how tightly speech-LLM hidden states bind audio frames to text tokens, using Whisper's word timestamps as a reference path.

desk verdict Promising training-free alignment metric for speech LLMs, but the Whisper-shared encoder reference and thin evidence leave its validity unproven. read the letter →

arxiv 2505.19937 v3 pith:7GJHUTAR submitted 2025-05-26 cs.CL cs.SDeess.AS

Pooneh Mousavi , Yingzhi Wang , Mirco Ravanelli , Cem Subakan This is my paper · ORCID
classification cs.CLcs.SDeess.AS
keywords audio-textalignmentspeechlargelanguagemodelsspokenunderstandinglatentrepresentationsimilaritymonotonicsearchWhisperwordtimestampsquestionansweringemotionrecognition
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Speech-language models claim to understand spoken audio, but until now there was no standard way to check whether their internal representations actually tie audio frames to the words they transcribe. ALAS is a training-free metric that does this: for each transformer layer, it computes cosine similarity between audio and text hidden states, extracts the best monotonic audio-to-text path, and compares that path against word timestamps from Whisper. On two models and two tasks, ALAS shows alignment strengthening in later layers for open-ended spoken question answering, and weakening in the final layers for emotion recognition, matching the intuition that emotion is more acoustic than semantic. If ALAS is right, it gives the field a cheap, interpretable, model- and task-agnostic probe for one of the core properties multimodal speech LLMs are supposed to learn.

What carries the argument

The central machinery has three parts: layer-wise cosine similarity matrices $S^l_{i,j}$ between audio hidden states $a^l_i$ and text hidden states $t^l_j$; the monotonic alignment search algorithm applied to each matrix to produce a single audio-to-text path, the ALAS-path; and a reference path built from Whisper's word-level timestamps. ALAS is then the average absolute difference between these two paths, scaled by audio length. The comparison against Whisper turns a raw similarity pattern into a number with an interpretable scale, and the layer-by-layer computation reveals where alignment forms or dissolves inside the network.

What would settle it

A concrete test: apply ALAS to a speech LLM whose audio encoder is not Whisper, such as a wav2vec2- or HuBERT-based model, and compare ALAS scores against human-annotated word boundaries; if the scores stop tracking human alignment or become noisy, the reference is not independent. A sharper ablation would corrupt Whisper timestamps with increasing noise and check whether ALAS responds to that corruption more strongly than to semantic perturbation of the transcript.

Watch

Extended reading notes

Core claim

The paper's central claim is that ALAS is a valid automatic measure of semantic audio-text alignment in speech LLMs, defined as the mean absolute deviation between a monotonic alignment path found on the layer-wise cross-modal cosine similarity matrix and a Whisper-derived reference path. Applying ALAS to Qwen2-Audio and SALMONN across LibriSQA and IEMOCAP, the paper finds that alignment is not uniform across layers: it becomes crisper and closer to the Whisper path in later layers for semantic question answering, while the last few layers show decreasing alignment for emotion recognition. The paper also reports that ALAS tracks but does not duplicate task accuracy, which it interprets as evidence that the metric can expose models that perform well on benchmarks without genuinely grounding their answers in the audio.

Load-bearing premise

The metric's validity rests on Whisper's word timestamps being an independent and unbiased ground truth for audio-text alignment, even though both models being scored use Whisper as their audio encoder and the audio hidden states are built on Whisper features.

Editorial extensions

If this is right

  • Any practitioner with a frozen speech LLM and an off-the-shelf ASR can compute ALAS per layer, making audio-text alignment auditable without fine-tuning or manual forced alignment.
  • ALAS can flag ungrounded fluency: models that answer benchmarks accurately while their internal alignment path diverges from the reference, separating benchmark competence from genuine audio grounding.
  • If alignment strength tracks task semantics, model developers can use ALAS to decide which layers to trust or adapt for a given task, and to test whether instruction tuning actually improves grounding.
  • The same pipeline extends to new tasks and datasets without manual transcription, provided paired audio and text are available.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because both evaluated LLMs use Whisper-v3 as their audio encoder, an alternative reading is that ALAS partly measures how faithfully each LLM preserves Whisper's temporal structure; testing on a model with a non-Whisper audio encoder would tell whether the score isolates semantic grounding.
  • Whisper timestamps are themselves estimates, so ALAS inherits any ASR boundary errors; comparing against human-annotated word boundaries on a small subset would quantify how much metric noise comes from the reference.
  • The sentence-embedding filter that discards audio/transcript pairs with dissimilar responses likely selects for easy, well-aligned examples; reporting unfiltered ALAS values would reveal how strongly this selection shapes the score.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 5 minor

Summary. The paper proposes ALAS (Automatic Latent Alignment Score), a metric intended to measure how well a speech-LLM internally aligns audio representations with text representations. The pipeline extracts layer-wise hidden states for both a speech input and its transcription, computes cross-modal cosine similarity matrices, and applies a monotonic alignment search to obtain a latent alignment path. This path is compared against a reference alignment derived from Whisper word timestamps, yielding a numeric score per layer. The authors evaluate ALAS on two models, SALMONN-7B and Qwen2-Audio-Instruct-7B, and on two datasets, LibriSQA (spoken question answering) and IEMOCAP (emotion recognition), and report qualitative similarity maps and quantitative layer-wise score curves.

Significance. If rigorously validated, ALAS would provide a training-free, task-agnostic probing tool for speech-LLMs, requiring only a forward pass and an off-the-shelf ASR reference. The open-source release of the library (mentioned in the abstract) and the automatic nature of the pipeline are commendable. However, the current evidence does not yet establish that ALAS measures genuine semantic alignment rather than an artifact of the shared audio encoder between the reference and the probed features; the experimental documentation also lacks the statistical detail needed to support the paper's central claims.

major comments (6)
  1. [Abstract vs. body] The abstract claims evaluation on four open-source Speech-LLMs (AF3, Qwen2-Audio, Qwen-Omni, SALMONN) across three datasets (IEMOCAP, LibriSQA, MMAU-speech), but the body reports experiments on only two models, SALMONN and QWEN, and two datasets, LibriSQA and IEMOCAP. The abstract also claims calibration to an interpretable uniform baseline, but no baseline is defined in the body. Please align the abstract with the reported experiments or provide the missing results.
  2. [Sec. 2.1, Step 3; Sec. 3.2] The reference alignment is obtained from Whisper-v3 (Step 3), and Section 3.2 states that both evaluated models, Qwen2-Audio and SALMONN, use Whisper-v3 as their audio encoder. The audio-side hidden states used in Eq. (1) therefore originate from the same system that produced the ground-truth alignment. This creates a confound: ALAS may largely reflect how faithfully the LLM preserves the temporal structure of Whisper's representations rather than measuring genuine cross-modal semantic binding. The authors' observation that layer 0 already scores well because Whisper outputs 'semantically informed representations' supports this worry. To validate the metric, please test with models that use non-Whisper encoders (e.g., HuBERT-based) or use a reference alignment from an independent forced aligner such as MFA.
  3. [Fig. 3] The quantitative results in Figure 3 show average ALAS curves without any error bars, sample counts, or statistical tests. The paper does not state how many utterances are evaluated, how the average is computed, or whether differences between layers/tasks are significant. As a proposed metric, ALAS needs at least confidence intervals or significance testing across a subsample to support the claim that it captures meaningful layer-wise patterns.
  4. [Eq. (2)] Equation (2) defines ALAS as the mean absolute difference between two alignment paths measured in token-index units. Since QWEN encodes audio with a 20 ms frame rate while SALMONN uses 330 ms frames (Figure 2 caption), one token-index step corresponds to very different time spans in the two models. Consequently, the absolute ALAS values are not comparable across models, and the claim that the metric is 'comparable across tasks' is unsupported without a normalized or calibrated score.
  5. [Sec. 2.1, Step 2] Step 2 filters out input pairs based on a Sentence-BERT response-similarity threshold of 0.7, but the paper reports neither the number of discarded pairs for each model/task nor the sensitivity of the results to this threshold. This filtering is a free parameter that could substantially bias which examples are scored, so a sensitivity analysis is needed.
  6. [Abstract and Sec. 3] The abstract states that ALAS 'tracks but does not duplicate task accuracy, exposing models that score well without genuinely grounding in the audio,' but the body does not present any such comparison. The IEMOCAP accuracy numbers cited in Section 3 are taken from Wang et al. (2024), not from the authors' own runs, and no joint analysis of ALAS and task accuracy is provided. This claim should be backed by direct evidence or removed.
minor comments (5)
  1. [Figure 3 caption] The figure caption refers to 'QWENN' instead of 'QWEN'; please correct the typo.
  2. [Sec. 2.3] Section 2.3 states that the similarity matrix S_l is in [0,1]^{T×A}, while Eq. (1) allows cosine similarities in [-1,1]. Please clarify whether negative similarities are clamped or otherwise handled.
  3. [References] The reference list entry for 'Taori et al.' is incomplete (missing year); also, the citation in the introduction lacks a year.
  4. [Appendix A, Figure 4] The appendix prompts in Figure 4 contain placeholder typos (e.g., the QWEN text-only audio prompt shows '<|audio_bos|>audio_eos|>' without the speech placeholder). Please correct the prompt diagrams.
  5. [Abstract] The abstract mentions an open-source library release, but no URL or repository name is given in the body; please add the link.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ALAS is a transparent distance-to-reference metric with no fitted parameters; the shared-Whisper-encoder issue is a validity confound, not a circular reduction.

full rationale

ALAS is defined as an L1 distance (Eq. 2) between a monotonic-alignment-search path derived from cosine similarities (Eq. 1) and a Whisper word-timestamp reference. No parameter is fitted to the evaluated models, and no load-bearing claim rests on a self-citation. The only substantive concern is that both evaluated models use Whisper-v3 as their audio encoder while Whisper-v3 also supplies the reference timestamps, so the score partly measures encoder self-consistency. This is an independence and construct-validity limitation of the two experiments, not an equation-level reduction: the MAS output is not forced to equal the Whisper path by construction, and the paper's layer-wise observations are empirical rather than definitional. The ALAS formula would be applied identically to models with different encoders. Hence no circular step can be quoted and no specific reduction exhibited.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on one hand-picked threshold and several domain assumptions. No new physical or conceptual entities are introduced.

free parameters (1)
  • Response similarity filtering threshold = 0.7
    Samples where the Sentence-BERT similarity between audio and transcript responses falls below 0.7 are discarded (Section 2.1, Step 2). This threshold is chosen by hand and affects which samples enter the ALAS computation, potentially biasing results.
assumptions (4)
  • domain assumption Cosine similarity between audio and text hidden states reflects semantic alignment.
    Used in Equation 1 to build the similarity matrix; the paper interprets the resulting alignment path as semantic alignment, but cosine similarity measures direction, not necessarily linguistic grounding.
  • domain assumption Whisper-v3 word timestamps are the ground-truth alignment path.
    The ALAS score is defined as deviation from Whisper's path (Equation 2). Whisper is an ASR system optimized for transcription, not for semantic grounding, and its timestamps may not match human perception or LLM semantics.
  • domain assumption The text-only branch, after filtering, produces representations comparable to the audio branch.
    Step 2 passes only the transcription and then filters by response similarity, assuming the remaining pairs have equivalent task-relevant information in both modalities.
  • standard math Monotonic alignment search finds the correct alignment path on the similarity matrix.
    Adopted from Glow-TTS (Kim et al., 2020); assumes audio and text proceed monotonically in time, which is standard for speech but may fail in tasks with non-linear semantic correspondences.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ALAS: An Automatic Latent Alignment Score for Audio Language Models." pith.science (2026). https://pith.science/paper/7GJHUTAR

@misc{pith2026250519937,
  author       = {Pith},
  title        = {Pith review of: ALAS: An Automatic Latent Alignment Score for Audio Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7GJHUTAR}},
  note         = {Machine review of arXiv:2505.19937}
}
read the original abstract

Large Language Models (LLMs) are extended into Speech-LLMs, and the quality of the audio--text alignment they learn affects most downstream Spoken Language Understanding (SLU) behavior. Yet despite a growth of fusion strategies, there is no standard way to measure how well a Speech-LLM internally binds audio frames to text tokens. We introduce ALAS (Automatic Latent Alignment Score), a model and task-agnostic metric that probes the LLM's per-layer hidden states, scoring the cross-modal cosine similarity between audio and text representations against a Whisper-derived reference. ALAS needs only a frozen forward pass and an off-the-shelf ASR reference, with no training or fitted classifier, and is calibrated to an interpretable uniform baseline comparable across tasks. Applying ALAS to four open-source Speech-LLMs (AF3, Qwen2-Audio, Qwen-Omni, SALMONN) across emotion recognition (IEMOCAP), open-ended SQA (LibriSQA), and multi-choice audio understanding (MMAU-speech), we find that the depth and strength of alignment reflect each model's audio-encoder design and the acoustic-versus-semantic demands of the task, and that ALAS tracks but does not duplicate task accuracy, exposing models that score well without genuinely grounding in the audio. We release ALAS as an open-source library so that practitioners can probe their own Speech-LLMs or try it on new tasks.

Figures

Figures reproduced from arXiv: 2505.19937 by the authors.

Figure 1
Figure 1. The Automatic Latent Alignment Score (ALAS) pipeline: (Step 1) Obtaining latent representations of the audio input,(Step 2) Obtaining latent representations of the transcription input, (Step 3) Extracting reference alignment from Whisper, and (Step 4) Calculating the ALAS score using cross-modal similarities and the alignment path between the two modalities. have the corresponding transcription. We begin by selectin… view at source ↗
Figure 2
Figure 2. Qualitative comparison of the alignments obtained for Qwen2-Audio-Instruct-7B (left) and Salmonn-7B (right). In each sub-figure, the horizontal axis represents the audio time series, with each frame corresponding to 20 ms for Qwen2-Audio-Instruct-7B and 330 ms for Salmonn-7B. The vertical axis represents the sequence of text tokens, where Qwen2-Audio-Instruct-7B utilizes word-level tokens while Salmonn-7B employs BP… view at source ↗
Figure 3
Figure 3. Quantitative comparison of the alignments through Automatic Latent Alignment Score (ALAS) for QWENN and SALMONN for LibriSQA and IEMOCAP datasets. The horizontal axes represent the layer-id, and the vertical axes represent the corresponding ALAS score. For the Emotion Recognition task, the models respectively obtain scores of 21.56%, and 49.30% (Wang et al., 2024) on the IEMOCAP dataset. 3.1. Qualitative Analysis In… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The prompt structure different models (SALMONN and QWEN) for the Spoken-Question-Answering (SQA), and Emotion Recognition tasks. The tags <Question>, <Transcription>, <Task>, <SpeechHere> are replaced with the associated entries for each prompt instance. 7 [PITH_FULL_…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 5 canonical work pages

  1. [1]

    L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774,

    arXiv

  2. [3]

    D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language models are few-shot learners. Advances in neural information processing systems , 33: 1877–1901,

    work page 1901

  3. [5]

    Qwen2-audio technical report

    Chu, Y ., Xu, J., Yang, Q., Wei, H., Wei, X., Guo, Z., Leng, Y ., Lv, Y ., He, J., Lin, J., et al. Qwen2-audio technical report. arXiv preprint arXiv:2407.10759,

    arXiv

  4. [7]

    Glm: General language model pretrain- ing with autoregressive blank infilling

    Du, Z., Qian, Y ., Liu, X., Ding, M., Qiu, J., Yang, Z., and Tang, J. Glm: General language model pretrain- ing with autoregressive blank infilling. arXiv preprint arXiv:2103.10360,

    arXiv

  5. [8]

    H., Karlinsky, L., and Glass, J

    Gong, Y ., Luo, H., Liu, A. H., Karlinsky, L., and Glass, J. Listen, think, and understand. arXiv preprint arXiv:2305.10790,

    arXiv

  6. [10]

    Less peaky and more accurate ctc forced alignment by label priors

    Huang, R., Zhang, X., Ni, Z., Sun, L., Hira, M., Hwang, J., Manohar, V ., Pratap, V ., Wiesner, M., Watanabe, S., et al. Less peaky and more accurate ctc forced alignment by label priors. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 11831–11835. IEEE,

    work page 2024

  7. [12]

    Li, J., Li, D., Savarese, S., and Hoi, S

    doi: 10.48550/arXiv.2306.05425. Li, J., Li, D., Savarese, S., and Hoi, S. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pp. 19730–19742. PMLR,

    arXiv doi:10.48550/arxiv.2306.05425

  8. [13]

    Montreal forced aligner: Trainable text- speech alignment using kaldi

    McAuliffe, M., Socolof, M., Mihuc, S., Wagner, M., and Sonderegger, M. Montreal forced aligner: Trainable text- speech alignment using kaldi. In Interspeech, volume 2017, pp. 498–502,

    work page 2017

Show all 24 references
  1. [15]

    The 5 ALAS: Measuring Latent Speech-Text Alignment For Spoken Language Understanding In Multimodal LLMs kaldi speech recognition toolkit

    Povey, D., Ghoshal, A., Boulianne, G., Burget, L., Glembek, O., Goel, N., Hannemann, M., Motlicek, P., Qian, Y ., Schwarz, P., Silovsky, J., Stemmer, G., and Vesely, K. The 5 ALAS: Measuring Latent Speech-Text Alignment For Spoken Language Understanding In Multimodal LLMs kald...

    2011

  2. [18]

    Ssr: Alignment-aware modality connector for speech lan- guage models

    Tan, W., Inaguma, H., Dong, N., Tomasello, P., and Ma, X. Ssr: Alignment-aware modality connector for speech lan- guage models. arXiv preprint arXiv:2410.00168,

    arXiv

  3. [19]

    Salmonn: Towards generic hearing abilities for large language models.arXiv preprint arXiv:2310.13289,

    Tang, C., Yu, W., Sun, G., Chen, X., Tan, T., Li, W., Lu, L., Ma, Z., and Zhang, C. Salmonn: Towards generic hearing abilities for large language models.arXiv preprint arXiv:2310.13289,

    arXiv

  4. [20]

    stanford

    URL https://crfm. stanford. edu/2023/03/13/alpaca. html . Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozi`ere, B., Goyal, N., Hambro, E., Azhar, F., et al. Llama: Open and efficient foundation lan- guage models. arXiv preprint arXiv:2302.13971,

    2023 arXiv

  5. [21]

    Wu, J., Gaur, Y ., Chen, Z., Zhou, L., Zhu, Y ., Wang, T., Li, J., Liu, S., Ren, B., Liu, L., et al

    URL https://arxiv.org/abs/2406.16020. Wu, J., Gaur, Y ., Chen, Z., Zhou, L., Zhu, Y ., Wang, T., Li, J., Liu, S., Ren, B., Liu, L., et al. On decoder-only architecture for speech-to-text and large language model integration. In 2023 IEEE Automatic Speech Recognition and Unders...

    2023 arXiv

  6. [22]

    and Wu, C

    Xie, Z. and Wu, C. Mini-omni2: Towards open-source gpt-4o with vision, speech and duplex capabilities. arXiv preprint arXiv:2410.11190,

    arXiv

  7. [23]

    Anygpt: Unified multimodal llm with discrete sequence modeling

    Zhan, J., Dai, J., Ye, J., Zhou, Y ., Zhang, D., Liu, Z., Zhang, X., Yuan, R., Zhang, G., Li, L., et al. Anygpt: Unified multimodal llm with discrete sequence modeling. arXiv preprint arXiv:2402.12226,

    arXiv

  8. [24]

    Video-LLaMA: An instruction-tuned audio-visual language model for video understanding

    Zhang, H., Li, X., and Bing, L. Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In Feng, Y . and Lefever, E. (eds.), Pro- ceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , pp. 543...

    2023

  9. [2006]

    Wavllm: Towards robust and adaptive speech large language model

    Hu, S., Zhou, L., Liu, S., Chen, S., Meng, L., Hao, H., Pan, J., Liu, X., Li, J., Sivasankaran, S., et al. Wavllm: Towards robust and adaptive speech large language model. arXiv preprint arXiv:2404.00656,

    arXiv

  10. [2008]

    X-llm: Bootstrapping advanced large language models by treating multi-modalities as foreign languages

    Chen, F., Han, M., Zhao, H., Zhang, Q., Shi, J., Xu, S., and Xu, B. X-llm: Bootstrapping advanced large language models by treating multi-modalities as foreign languages. arXiv preprint arXiv:2305.04160,

    arXiv

  11. [2017]

    Listen: Learning soft token embeddings for neural audio llms

    Mousavi, P., Gupta, S., Subakan, C., and Ravanelli, M. Listen: Learning soft token embeddings for neural audio llms. In Interspeech, arXiv:2505.18517, 2025a. Mousavi, P., Maimon, G., Moumen, A., Petermann, D., Shi, J., Wu, H., Yang, H., Kuznetsova, A., Ploujnikov, A., Marxer, ...

    arXiv

  12. [2019]

    Rubenstein, P

    URL https://arxiv.org/ abs/1908.10084. Rubenstein, P. K. et al. Audiopalm: A large lan- guage model that can speak and listen. arXiv preprint arXiv:2306.12925,

    1908 arXiv

  13. [2020]

    Li, B., Zhang, Y ., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., and Liu, Z

    URL https://arxiv.org/ abs/2005.11129. Li, B., Zhang, Y ., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., and Liu, Z. MIMIC-IT: Multi-Modal In-Context Instruction Tuning. arXiv e-prints, art. arXiv:2306.05425, June

    2005 arXiv

  14. [2022]

    org/abs/2212.04356

    URL https://arxiv. org/abs/2212.04356. Reimers, N. and Gurevych, I. Sentence-bert: Sentence em- beddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natu- ral Language Processing. Association for Computational Linguistics, 11

    2019 arXiv

  15. [2023]

    C., Dale, D., Dong, N., Duppenthaler, M., Duquenne, P.-A., Ellis, B., Elsahar, H., Haaheim, J., et al

    Barrault, L., Chung, Y .-A., Meglioli, M. C., Dale, D., Dong, N., Duppenthaler, M., Duquenne, P.-A., Ellis, B., Elsahar, H., Haaheim, J., et al. Seamless: Multilingual expres- sive and streaming speech translation. arXiv preprint arXiv:2312.05187,

    arXiv

  16. [2024]

    Dai, W., Li, J., Li, D., Tiong, A. M. H., Zhao, J., Wang, W., Li, B., Fung, P., and Hoi, S. Instructblip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500,

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.