Textual Inversion for Efficient Adaptation of Open-Vocabulary Object Detectors Without Forgetting

Frank Ruis; Gertjan Burghouts; Hugo Kuijf

REVIEW 3 major objections 3 minor 41 references

Textual Inversion for Efficient Adaptation of Open-Vocabulary Object Detectors Without Forgetting

T0 review · 3 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper proposes that open-vocabulary object detectors can be adapted to novel or fine-grained classes by learning new text tokens while keeping the vision-language model's weights frozen, with as few as three examples, without losing ori

desk verdict Textual inversion is a sensible, cheap adaptation idea for open-vocabulary detectors, but the abstract alone doesn't back the 'matches or outperforms' claim; it deserves a referee and a careful look at the actual experiments. read the letter →

arxiv 2508.05323 v1 pith:FM2WVCIM submitted 2025-08-07 cs.CV

Frank Ruis , Gertjan Burghouts , Hugo Kuijf This is my paper

classification cs.CV

keywords textualinversionopen-vocabularyobjectdetectioncatastrophicforgettingfew-shotlearningtokenembeddingvision-languagemodelszero-shottransfer

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

The paper proposes adapting open-vocabulary object detectors by learning new word tokens—textual inversions—while keeping the vision-language model completely frozen. The authors claim that with as few as three example images, a learned token can make the detector recognize novel or fine-grained object classes, improve existing ones, and still retain the original model's benchmark performance and zero-shot transfer abilities, such as detecting a sketch after training only on photos. Because gradients and storage are limited to the token-embedding dimension, the adaptation is far cheaper than full fine-tuning. The central promise is a no-forgetting path to specialization.

What carries the argument

The central object is a learned text token—a pseudo-word embedding—inserted into the text-encoder vocabulary of the vision-language model. During adaptation, only this token embedding is optimized, driven by a detection loss over a few examples; the backbone and head weights stay frozen. The token acts as a compact, reusable pointer that links a novel visual concept to the model's existing text-conditioned detection pathway.

What would settle it

Train the method on a novel class that is visually far from the pretraining vocabulary (for example, a rare tool or unusual animal) using exactly three images, then measure detection AP on a held-out test set; if performance does not exceed a frozen random-token baseline, the claim that one token suffices is falsified.

Watch

Extended reading notes

Core claim

The central claim is that textual inversion, originally used to personalize text-to-image diffusion models, transfers to open-vocabulary object detection: learning a small number of token embeddings in the text encoder's vocabulary is sufficient to steer a frozen detector's outputs toward novel categories. The learned tokens are compatible with the original vision-language model weights, so the detector's base performance remains intact and its existing zero-shot and domain-transfer strengths are preserved. The paper reports quantitative and qualitative experiments comparing against baseline methods that do suffer from forgetting, aiming to show that token-only adaptation matches or outperfo

Load-bearing premise

The method depends on a single learned token embedding being able to shift a frozen detector's predictions enough to recognize a new class; if the frozen weights cannot put that token to use, the no-forgetting advantage disappears.

Editorial extensions

If this is right

New classes can be added to an existing detector from a handful of images without degrading its original benchmark performance.
Domain transfer like photo-to-sketch detection remains available after adaptation because the base weights are untouched.
Adaptation compute and memory scale with the embedding dimension, not the full model, making per-category specialization cheap.
Multiple learned tokens could be accumulated over time, offering a route to continual learning without catastrophic forgetting.

Reading between the lines

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

If a single token can carry a concept, then composing tokens may describe combined or fine-grained concepts, enabling compositional open-vocabulary detection without retraining.
The same token-inversion idea could plausibly transfer to other frozen VLM-based tasks such as open-vocabulary segmentation or visual grounding, since they share the text-conditioning mechanism.
A testable boundary: concepts that are hard to compress into one token, such as highly articulated or multi-component objects, may need more tokens or an additional learned projection, which would reveal the capacity limits of the approach.

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.

Desk Editor's Note

read the letter

Colleague,

My read of 2508.05323 is limited because the supplied full text is garbled (mojibake throughout), so I'm working from the abstract. That matters, and I'd rather say it plainly than pretend otherwise.

What's genuinely new: the paper takes textual inversion from text-to-image diffusion and applies it to open-vocabulary object detection. The token-only, frozen-weights formulation is a clean idea. If it works, it gives you a lightweight way to add novel or fine-grained classes without catastrophic forgetting and while preserving zero-shot transfer. That fills a real gap, and the connection to detection is not just a re-labeling: the region-text alignment head adds a real constraint that text-to-image TI doesn't have to respect.

The main soft spot is that the abstract gives no numbers. "We evaluated whether the method matches or outperforms the baseline methods" is a claim, not a result. No benchmarks, no baselines, no forgetting metrics, no error bars. I can't verify the central empirical assertion from what's here. That doesn't make the paper wrong - abstracts routinely omit numbers - but it does mean the abstract under-determines soundness. The stress-test worry about overfitting from three examples is a legitimate concern, not a demonstrated flaw. A single token vector is high-dimensional, and three examples give you little signal, especially for fine-grained classes. Whether the method gets past that depends on initialization, regularization, and how much the frozen text encoder constrains the learned token. Those are exactly the details I'd need from the full paper.

I can't audit the math or the citation pattern because the text didn't come through. From the abstract I see no obvious circularity: the evaluation is domain transfer and held-out detection, which is an external check if done honestly.

Who's this for? Anyone working on few-shot adaptation of vision-language models, particularly detection. I'd bring it to a reading group to discuss whether token-only adaptation is expressive enough. I would not cite it as established until I see the full results.

Recommendation: send it to peer review. The idea is worth referee time; either the experiments support the claim or they don't, and a referee should be the one to decide. I'd engage with the full paper.

Referee Report

3 major / 3 minor

Summary. The paper proposes applying Textual Inversion (TI) to open-vocabulary object detection: a single token embedding is learned while all VLM weights remain frozen, enabling detection of novel or fine-grained classes from as few as three examples and preserving the original model's benchmark performance and zero-shot capabilities. The abstract reports that the method 'matches or outperforms the baseline methods' but gives no numerical results, benchmark names, baselines, ablations, error bars, or experimental details. The body text in the provided manuscript is largely unreadable due to apparent character-encoding corruption, so no method description, equations, or tables can be audited.

Significance. If the central claims hold, the contribution is genuinely valuable: a compute-efficient adaptation technique for open-vocabulary detectors that avoids catastrophic forgetting and preserves zero-shot domain transfer would be of broad interest. However, the significance is currently only potential. The abstract's central assertion is empirical, and no supporting evidence is visible in the submitted material. The idea of learning a token embedding rather than updating detector weights is simple and falsifiable, which is a strength, but without a legible methods section and a quantitative evaluation the paper cannot be accepted or even fully assessed.

major comments (3)

[Abstract] The central quantitative claim—that the method 'matches or outperforms the baseline methods'—is asserted without reporting any metrics, dataset names, baseline methods, or error bars. This is the load-bearing claim of the paper: the whole contribution is an empirical adaptation method. As presented, the abstract contains no evidence that would allow a reader to verify even the direction of the effect, let alone its statistical reliability.
[Full text (passages after the abstract)] The body text of the provided manuscript is unreadable, consisting of replacement characters with no recoverable equations, tables, or experiment descriptions. Consequently I cannot audit the method formulation, the training objective, the token initialization strategy, or the evaluation protocol. In particular, the critical overfitting risk—learning a high-dimensional token embedding from three examples—is not addressed anywhere visible. Unless a readable version is provided, the technical content of the paper is effectively absent.
[Abstract, 'retaining the original model's benchmark performance'] The claim of 'without forgetting' requires a direct comparison of the frozen model's performance on its original vocabulary before and after token learning. No such forgetting metric, original-vocabulary benchmark, or quantitative result is reported. The abstract's phrasing 'We evaluated whether ...' is a promise of evaluation, not an evaluation; the paper must present the actual numbers, along with the number of examples, training iterations, and variance across seeds/classes.

minor comments (3)

[Abstract] 'From as little as three examples' is an impressive but underspecified claim. The number of classes, per-class example count, and performance variability across random choices of the three examples should be stated.
[Abstract] The phrase 'We evaluated whether the method matches or outperforms the baseline methods' should be replaced with a concrete summary of the evaluation—e.g., 'We report mAP on X and Y, showing a gain of Z over baselines'—or with a pointer to a specific table in the paper.
[Manuscript readability] The submitted file appears to have a character-encoding problem. If this is a submission artifact, the authors should resubmit a readable PDF; as it stands, the manuscript is not reviewable beyond the abstract.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; frozen-weight performance retention is a logical consequence, and the supplied full text is unreadable, so no specific reduction can be exhibited.

full rationale

The paper's proposed method is: keep the VLM weights frozen and optimize only token embeddings for novel concepts. The claim that benchmark performance is retained follows directly from the decision to freeze weights, so it is a logical consequence of the method rather than a fitted-parameter prediction or a self-referential derivation. The central empirical claims—detecting novel objects from three examples and zero-shot domain transfer—are presented as experimental evaluations against external benchmarks, which would provide independent checks. The abstract does not report numerical results, and the supplied full text is corrupted (mojibake), so no equations or specific evaluation protocols can be inspected. Under the hard rule that circularity must be demonstrated with a specific quoted reduction, no circular step can be identified from the readable material. The absence of quantitative results is a completeness/evidence gap, not a circularity. No load-bearing self-citations or imported uniqueness theorems are visible. Therefore the score is 0.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

The review is abstract-only. The visible axioms are the domain assumptions that textual inversion works for detection and that frozen weights remain compatible. No free parameters or invented entities are identifiable from the abstract.

assumptions (2)

domain assumption Textual inversion formulations transfer from text-to-image diffusion to open-vocabulary object detection.
The entire method rests on the premise that optimizing token embeddings can effectively adapt a detector, not just a generative model. This is asserted in the abstract.
domain assumption Frozen VLM weights can be extended with new tokens without degrading original capabilities.
The abstract claims retention of benchmark performance and zero-shot transfer while keeping the VLM frozen; the method assumes the frozen weights remain compatible with added tokens.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Textual Inversion for Efficient Adaptation of Open-Vocabulary Object Detectors Without Forgetting." pith.science (2026). https://pith.science/paper/FM2WVCIM

@misc{pith2026250805323,
  author       = {Pith},
  title        = {Pith review of: Textual Inversion for Efficient Adaptation of Open-Vocabulary Object Detectors Without Forgetting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FM2WVCIM}},
  note         = {Machine review of arXiv:2508.05323}
}

read the original abstract

Recent progress in large pre-trained vision language models (VLMs) has reached state-of-the-art performance on several object detection benchmarks and boasts strong zero-shot capabilities, but for optimal performance on specific targets some form of finetuning is still necessary. While the initial VLM weights allow for great few-shot transfer learning, this usually involves the loss of the original natural language querying and zero-shot capabilities. Inspired by the success of Textual Inversion (TI) in personalizing text-to-image diffusion models, we propose a similar formulation for open-vocabulary object detection. TI allows extending the VLM vocabulary by learning new or improving existing tokens to accurately detect novel or fine-grained objects from as little as three examples. The learned tokens are completely compatible with the original VLM weights while keeping them frozen, retaining the original model's benchmark performance, and leveraging its existing capabilities such as zero-shot domain transfer (e.g., detecting a sketch of an object after training only on real photos). The storage and gradient calculations are limited to the token embedding dimension, requiring significantly less compute than full-model fine-tuning. We evaluated whether the method matches or outperforms the baseline methods that suffer from forgetting in a wide variety of quantitative and qualitative experiments.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 30 canonical work pages

[1]

Kamath, A., Singh, M., LeCun, Y., Synnaeve, G., Misra, I., and Carion, N., ``Mdetr-modulated detection for end-to-end multi-modal understanding,'' in [ Proceedings of the IEEE/CVF International Conference on Computer Vision 0.1em ] , 1780--1790 (2021)

work page 2021
[2]

Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M., ``Hierarchical text-conditional image generation with clip latents,'' arXiv preprint arXiv:2204.06125 (2022)

arXiv 2022
[3]

A., and Nematzadeh, A., ``Measuring progress in fine-grained vision-and-language understanding,'' (2023)

Bugliarello, E., Sartran, L., Agrawal, A., Hendricks, L. A., and Nematzadeh, A., ``Measuring progress in fine-grained vision-and-language understanding,'' (2023)

work page 2023
[4]

Wei, J., Wang, X., Schuurmans, D., Bosma, M., Chi, E., Le, Q., and Zhou, D., ``Chain of thought prompting elicits reasoning in large language models,'' arXiv preprint arXiv:2201.11903 (2022)

arXiv 2022
[5]

Chan, S. C. Y., Santoro, A., Lampinen, A. K., Wang, J. X., Singh, A., Richemond, P. H., McClelland, J., and Hill, F., ``Data distributional properties drive emergent in-context learning in transformers,'' (2022)

work page 2022
[6]

Mialon, G., Dessì, R., Lomeli, M., Nalmpantis, C., Pasunuru, R., Raileanu, R., Rozière, B., Schick, T., Dwivedi-Yu, J., Celikyilmaz, A., Grave, E., LeCun, Y., and Scialom, T., ``Augmented language models: a survey,'' (2023)

work page 2023
[7]

Li*, L. H., Zhang*, P., Zhang*, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.-N., Chang, K.-W., and Gao, J., ``Grounded language-image pre-training,'' in [ CVPR 0.1em ] , (2022)

work page 2022
[8]

and Cohen, N

McCloskey, M. and Cohen, N. J., ``Catastrophic interference in connectionist networks: The sequential learning problem,'' in [ Psychology of learning and motivation 0.1em ] , 24 , 109--165, Elsevier (1989)

work page 1989

Show all 41 references

[9]

this is my unicorn, fluffy

Cohen, N., Gal, R., Meirom, E. A., Chechik, G., and Atzmon, Y., ``“this is my unicorn, fluffy”: Personalizing frozen vision-language representations,'' in [ Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part XX 0.1em...

2022
[10]

H., Chechik, G., and Cohen-Or, D., ``An image is worth one word: Personalizing text-to-image generation using textual inversion,'' (2022)

Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A. H., Chechik, G., and Cohen-Or, D., ``An image is worth one word: Personalizing text-to-image generation using textual inversion,'' (2022)

2022
[11]

C., and Liu, Z., ``Conditional prompt learning for vision-language models,'' in [ IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 0.1em ] , (2022)

Zhou, K., Yang, J., Loy, C. C., and Liu, Z., ``Conditional prompt learning for vision-language models,'' in [ IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 0.1em ] , (2022)

2022
[12]

Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I., ``Learning transferable visual models from natural language supervision,'' in [ ICML 0.1em ] , (2021)

2021
[13]

Gu, X., Lin, T.-Y., Kuo, W., and Cui, Y., ``Open-vocabulary detection via vision and language knowledge distillation,'' arXiv preprint arXiv:2104.13921 (2021)

2021 arXiv
[14]

Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., Wang, X., Zhai, X., Kipf, T., and Houlsby, N., ``Simple open-vocabulary object detection with vision transformers,'' arXiv preprint arXiv:22...

2022 arXiv
[15]

H., Dai, X., Wang, L., Yuan, L., Hwang, J.-N., and Gao, J., ``Glipv2: Unifying localization and vision-language understanding,'' arXiv preprint arXiv:2206.05836 (2022)

Zhang, H., Zhang, P., Hu, X., Chen, Y.-C., Li, L. H., Dai, X., Wang, L., Yuan, L., Hwang, J.-N., and Gao, J., ``Glipv2: Unifying localization and vision-language understanding,'' arXiv preprint arXiv:2206.05836 (2022)

2022 arXiv
[16]

Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll \'a r, P., and Zitnick, C. L., ``Microsoft coco: Common objects in context,'' in [ Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V ...

2014
[17]

Lester, B., Al-Rfou, R., and Constant, N., ``The power of scale for parameter-efficient prompt tuning,'' in [ Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing 0.1em ] , 3045--3059 (2021)

2021
[18]

Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., and Yu, Q., ``Llama-adapter: Efficient fine-tuning of language models with zero-init attention,'' arXiv preprint arXiv:2303.16199 (2023)

2023 arXiv
[19]

Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., Li, H., and Qiao, Y., ``Llama-adapter v2: Parameter-efficient visual instruction model,'' arXiv preprint arXiv:2304.15010 (2023)

2023 arXiv
[20]

C., and Liu, Z., ``Learning to prompt for vision-language models,'' International Journal of Computer Vision 130 (9), 2337--2348 (2022)

Zhou, K., Yang, J., Loy, C. C., and Liu, Z., ``Learning to prompt for vision-language models,'' International Journal of Computer Vision 130 (9), 2337--2348 (2022)

2022
[21]

Jia, M., Tang, L., Chen, B.-C., Cardie, C., Belongie, S., Hariharan, B., and Lim, S.-N., ``Visual prompt tuning,'' in [ Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part XXXIII 0.1em ] , 709--727, Springer (2022)

2022
[22]

and Liu, B., ``Lifelong machine learning,'' Synthesis Lectures on Artificial Intelligence and Machine Learning 12 (3), 1--207 (2018)

Chen, Z. and Liu, B., ``Lifelong machine learning,'' Synthesis Lectures on Artificial Intelligence and Machine Learning 12 (3), 1--207 (2018)

2018
[23]

Ding, Y., Liu, L., Tian, C., Yang, J., and Ding, H., ``Don't stop learning: Towards continual learning for the clip model,'' arXiv preprint arXiv:2207.09248 (2022)

2022 arXiv
[24]

Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al., ``Overcoming catastrophic forgetting in neural networks,'' Proceedings of the national academy of sciences 114 (13), 3521--35...

2017
[25]

Meng, K., Bau, D., Andonian, A., and Belinkov, Y., ``Locating and editing factual associations in GPT ,'' Advances in Neural Information Processing Systems 36 (2022)

2022
[26]

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y., ``Generative adversarial networks,'' Communications of the ACM 63 (11), 139--144 (2020)

2020
[27]

a henb \

Zhu, J.-Y., Kr \"a henb \"u hl, P., Shechtman, E., and Efros, A. A., ``Generative visual manipulation on the natural image manifold,'' in [ Computer Vision--ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14 0.1em ] , 5...

2016
[28]

Zhu, J., Shen, Y., Zhao, D., and Zhou, B., ``In-domain gan inversion for real image editing,'' in [ Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XVII 16 0.1em ] , 592--608, Springer (2020)

2020
[29]

Isola, P., Zhu, J.-Y., Zhou, T., and Efros, A. A., ``Image-to-image translation with conditional adversarial networks,'' in [ Proceedings of the IEEE conference on computer vision and pattern recognition 0.1em ] , 1125--1134 (2017)

2017
[30]

Li, D., Li, J., Le, H., Wang, G., Savarese, S., and Hoi, S. C. H., ``Lavis: A library for language-vision intelligence,'' (2022)

2022
[31]

Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., and Zhang, L., ``Grounding dino: Marrying dino with grounded pre-training for open-set object detection,'' (2023)

2023
[32]

Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., and De Mello, S., `` ODISE: Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models ,'' arXiv preprint arXiv: 2303.04803 (2022)

2022 arXiv
[33]

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K., ``Bert: Pre-training of deep bidirectional transformers for language understanding,'' arXiv preprint arXiv:1810.04805 (2018)

2018 arXiv
[34]

Dai, X., Chen, Y., Xiao, B., Chen, D., Liu, M., Yuan, L., and Zhang, L., ``Dynamic head: Unifying object detection heads with attentions,'' in [ Proceedings of the IEEE/CVF conference on computer vision and pattern recognition 0.1em ] , 7373--7382 (2021)

2021
[35]

H., Zhang, P., Aneja, J., Yang, J., Jin, P., Hu, H., Liu, Z., Lee, Y

Li, C., Liu, H., Li, L. H., Zhang, P., Aneja, J., Yang, J., Jin, P., Hu, H., Liu, Z., Lee, Y. J., and Gao, J., ``Elevater: A benchmark and toolkit for evaluating language-augmented visual models,'' Neural Information Processing Systems (2022)

2022
[36]

M., Vedaldi, A., Zisserman, A., and Jawahar, C

Parkhi, O. M., Vedaldi, A., Zisserman, A., and Jawahar, C. V., ``Cats and dogs,'' in [ IEEE Conference on Computer Vision and Pattern Recognition 0.1em ] , (2012)

2012
[37]

Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., and Aberman, K., ``Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation,'' (2022)

2022
[38]

Zhu, P., Wen, L., Du, D., Bian, X., Fan, H., Hu, Q., and Ling, H., ``Detection and tracking meet drones challenge,'' IEEE Transactions on Pattern Analysis and Machine Intelligence , 1--1 (2021)

2021
[39]

Steiner, A., Kolesnikov, A., Zhai, X., Wightman, R., Uszkoreit, J., and Beyer, L., ``How to train your vit? data, augmentation, and regularization in vision transformers,'' (2022)

2022
[40]

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al., ``Pytorch: An imperative style, high-performance deep learning library,'' Advances in neural information processing systems 32 (2019)

2019
[41]

write newline

" write newline "" before.all 'output.state := FUNCTION blank.sep after.quote 'output.state := FUNCTION fin.entry output.state after.quoted.block = 'skip 'add.period if write newline FUNCTION new.block output.state before.all = 'skip output.state after.quote = after.quoted.blo...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Kamath, A., Singh, M., LeCun, Y., Synnaeve, G., Misra, I., and Carion, N., ``Mdetr-modulated detection for end-to-end multi-modal understanding,'' in [ Proceedings of the IEEE/CVF International Conference on Computer Vision 0.1em ] , 1780--1790 (2021)

work page 2021

[2] [2]

Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M., ``Hierarchical text-conditional image generation with clip latents,'' arXiv preprint arXiv:2204.06125 (2022)

arXiv 2022

[3] [3]

A., and Nematzadeh, A., ``Measuring progress in fine-grained vision-and-language understanding,'' (2023)

Bugliarello, E., Sartran, L., Agrawal, A., Hendricks, L. A., and Nematzadeh, A., ``Measuring progress in fine-grained vision-and-language understanding,'' (2023)

work page 2023

[4] [4]

Wei, J., Wang, X., Schuurmans, D., Bosma, M., Chi, E., Le, Q., and Zhou, D., ``Chain of thought prompting elicits reasoning in large language models,'' arXiv preprint arXiv:2201.11903 (2022)

arXiv 2022

[5] [5]

Chan, S. C. Y., Santoro, A., Lampinen, A. K., Wang, J. X., Singh, A., Richemond, P. H., McClelland, J., and Hill, F., ``Data distributional properties drive emergent in-context learning in transformers,'' (2022)

work page 2022

[6] [6]

Mialon, G., Dessì, R., Lomeli, M., Nalmpantis, C., Pasunuru, R., Raileanu, R., Rozière, B., Schick, T., Dwivedi-Yu, J., Celikyilmaz, A., Grave, E., LeCun, Y., and Scialom, T., ``Augmented language models: a survey,'' (2023)

work page 2023

[7] [7]

Li*, L. H., Zhang*, P., Zhang*, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.-N., Chang, K.-W., and Gao, J., ``Grounded language-image pre-training,'' in [ CVPR 0.1em ] , (2022)

work page 2022

[8] [8]

and Cohen, N

McCloskey, M. and Cohen, N. J., ``Catastrophic interference in connectionist networks: The sequential learning problem,'' in [ Psychology of learning and motivation 0.1em ] , 24 , 109--165, Elsevier (1989)

work page 1989