Pith. sign in

REVIEW 4 major objections 4 minor 35 references

Enhancing Interpretable Image Classification Through LLM Agents and Conditional Concept Bottleneck Models

T0 review · 4 major / 4 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 This paper claims that the optimal size of a concept bank for interpretable image classification can be discovered automatically, by an LLM agent that adds and removes concepts in response to feedback from a concept-bottleneck model…

desk verdict Novel combination of label-conditioned scoring and agent-based concept bank refinement, but the headline interpretability gain rests on an unvalidated, self-referential LLM metric plus two fixable formalism errors. read the letter →

arxiv 2506.01334 v1 pith:SPLEXO74 submitted 2025-06-02 cs.CL

Yiwen Jiang , Deval Mehta , Wei Feng , Zongyuan Ge This is my paper
classification cs.CL
keywords conceptbottleneckmodelsinterpretableimageclassificationLLMagentsdynamicbankconditionalscoringeditablematrixfew-shotfeedbackLLM-basedinterpretabilityevaluation
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

Concept Bottleneck Models (CBMs) make image classification interpretable by routing predictions through a bank of human-readable concepts, but previous methods fix the bank size in advance and rely on a single shared concept-scoring mechanism. The paper introduces Conditional Concept Bottleneck Models (CoCoBMs), which score each concept separately for each candidate label, and an editable matrix that zeroes out concept-label pairs the LLM deems factually impossible. It also introduces a Concept Agent that uses few-shot image feedback to prune redundant concepts and generate missing ones, converging on a compact bank of roughly one concept per label. On six datasets the method reports an average accuracy gain of 6.15% over the common LaBo baseline and roughly a 30% improvement on a newly proposed LLM-based interpretability metric. The broader claim is that concept banks should be treated as a live, feedback-tuned artifact rather than a static list generated once.

What carries the argument

The central mechanism is CoCoBMs, a Conditional Concept Bottleneck Model that computes a label-conditioned concept score matrix, $\mathbf{S} \in \mathbb{R}^{N \times M}$, where $s_j^k$ is the score of concept $c_k$ for label $y_j$, instead of a shared vector. It also includes an editable matrix $E$ that encodes factual compatibility of each concept-label pair, setting scores to zero when the LLM deems the concept impossible for that label. The second load-bearing piece is the Concept Agent, a memory-planning-action LLM agent that uses CoCoBMs as a perception tool: it reads concept activation patterns on K-means-selected few-shot instances, removes concepts with no contribution or with duplicate activation patterns, and generates new concepts for labels that remain indistinguishable.

What would settle it

Run a human-subject study in which people are shown the same ranked concept lists and asked whether the concepts match reality and whether they identify the correct class, then correlate those human ratings with the paper's LLM MCQ scores; if the correlation is weak or negative, the interpretability claim collapses. Alternatively, replace the evaluating LLM with a different model family (e.g., Claude or Llama) and check whether the scores remain similar; a large drop would indicate self-agreement rather than genuine interpretability.

Watch

Extended reading notes

Core claim

The paper's central claim is that the performance bottleneck of CLIP-based CBMs is not only the number of concepts but the way concepts are scored: a single shared score for each concept across all labels is too crude. CoCoBMs replace this with category-conditioned scoring, where each concept receives a score relative to each candidate label, producing a full label-by-concept matrix. An editable matrix then suppresses activations that contradict LLM factual knowledge, and a Concept Agent iteratively grows and prunes the concept bank based on activation patterns observed on a few selected training instances. The authors claim this dynamic grounding converges to roughly the number of labels while improving both accuracy and a quantitative interpretability score, which they measure by asking an LLM multiple-choice questions about the truthfulness and distinguishability of ranked concept lists.

Load-bearing premise

The headline 30% interpretability improvement rests on the assumption that an LLM judge's multiple-choice answers about 'truthfulness' and 'distinguishability' measure what humans find interpretable, even though the same LLM family generated and verified the concepts being judged.

Editorial extensions

If this is right

  • Concept-bank size can be automated: roughly one concept per label, chosen by feedback, can match or beat fixed banks several times larger.
  • LLM factual knowledge, encoded as an editable matrix, can correct false-positive concept activations during training without human annotation.
  • Category-conditioned scoring gives per-label concept attributions, making it possible to spot confusable label pairs (e.g., cat vs. dog, automobile vs. truck) and to debug them automatically.
  • The accuracy gap between concept-bottleneck models and black-box models shrinks to a few percent, making interpretable models a more practical default.
  • The proposed MCQ-based interpretability measure offers a reproducible, automatically generated evaluation, though it is not validated against human judgments.

Reading between the lines

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

  • Since the same LLM family generates, verifies, and evaluates the concepts, the reported 30% interpretability gain may partly reflect self-consistency; a human study correlating the MCQ scores with people's judgments would clarify the real gain.
  • The dynamic grounding loop could transfer to other discrete-vocabulary decision pipelines beyond CBMs, such as compositional classifiers or retrieval-augmented models that must decide how many human-readable units to keep.
  • Seeding the Concept Agent with concepts from a different LLM family or a knowledge graph would test whether the converged bank reflects dataset structure or the generator's priors.
  • The few-shot feedback depends on K-means centroids, so the method's sensitivity to instance selection could be probed by perturbing the number of clusters or the chosen samples.
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

4 major / 4 minor

Summary. The paper proposes Conditional Concept Bottleneck Models (CoCoBMs) with label-conditioned concept scoring via learnable prompts, an LLM-populated editable matrix for suppressing factually incompatible concept activations, and a Concept Agent that dynamically expands/prunes the concept bank from environment feedback. It also introduces an LLM-based quantitative interpretability metric in which GPT-4-turbo answers truthfulness and distinguishability MCQs over ranked concept lists. The empirical claims are an average accuracy gain of 6.15% over LaBo-n, 5.97% over LM4CV-n, and 3.21% over LM4CV-2n, together with an approximately 30% improvement in interpretability over LM4CV-2n, evaluated on six image classification datasets.

Significance. If the claims are sustained, the CoCoBM architecture and the dynamic grounding loop would be a useful step toward concise yet sufficient concept banks and category-specific concept scoring, both of which are relevant to the interpretable-vision community. The paper has concrete strengths: it compares against public baselines and their released concept banks, it includes ablations for dynamic grounding and the editable matrix, and it provides a detailed CIFAR-10 case study of the agent's iterative concept-bank evolution. However, the headline interpretability gain rests on an unvalidated, self-referential LLM-judge metric, and several parts of the formal derivation are internally inconsistent. The accuracy comparisons also lack uncertainty quantification. Because these issues affect the two central contributions, the manuscript needs substantial revision before the claims can be accepted.

major comments (4)
  1. [Section 4.2] The interpretability metric is not validated as a measure of human interpretability, yet it supports the headline claim of 'around a 30% improvement in interpretability.' The concepts are generated and fact-verified by GPT-4o and then judged by GPT-4-turbo MCQs that see only ranked concept names; the judge receives neither the input images nor the model's concept scores or logits. High truthfulness may therefore reflect agreement between two models from the same LLM family rather than agreement with human judgment or with visual evidence. No human study or correlation with an established interpretability benchmark is reported, and the Limitations section itself states that evaluating LLM internal knowledge 'remains open challenges.' I recommend either adding a human-subject evaluation (e.g., concept ranking or factual-relevance judgments) or explicitly reframing the 30% figure as 'LLM-judged concept quality' rather than 'interpretability.'
  2. [Section 3.1 (Eq. 4, Eq. 5) and Section 3.4] The editable matrix is defined inconsistently. Equation (4) sets Ejk = 1 for concept–label pairs that are factually incompatible (ck ∉ yj), whereas Section 3.4 states that Ejk is set to 1 when the concept is judged 'critical' or 'occasionally present,' i.e., for compatible pairs. These are opposite conventions. Moreover, Equation (5), s_j_ck = min(s_j_ck, 0), clamps positive scores to zero but leaves negative scores unchanged; it does not 'set the concept score to zero' as the surrounding prose claims. Since Table 2's ablation of the editable matrix depends on this matrix, the reported interpretability drop without E is implementation-dependent until the convention and the masking operation are clarified.
  3. [Section 3.1, Eq. (6)] The objective function as printed cannot train the model as described. The sum over j=1..N is taken over terms Wp·y_i·log(ŷ_i) + (1−y_i)·log(1−ŷ_i), where neither y_i nor ŷ_i carries a j index, so every term in the sum is identical and the loss reduces to a constant multiple of one term. This does not provide label-specific training signal for the conditional scores in Eq. (2). Please rewrite the loss with per-label targets y_ij and per-label predictions ŷ_ij, and clarify how the positive-class weight Wp is applied.
  4. [Section 4.5 and Figure 3] The accuracy comparisons are reported as single values with no error bars, standard deviations, or significance tests, even though the training process, LLM concept generation, and agent iterations are stochastic. The claimed margins, e.g., 6.15% over LaBo-n, 5.97% over LM4CV-n, and 3.21% over LM4CV-2n, may be within run-to-run variance. Please report multiple seeds (at least for the main comparison) and specify whether the reported numbers are means. Also, Section 4.5 says 'across five datasets' while the abstract and introduction say 'across 6 datasets'; please reconcile this discrepancy.
minor comments (4)
  1. [Section 4.3] The implementation details say 'The number of selected concepts equals the number of prompt labels,' but Section 3.5 describes an insufficiency criterion that expands concepts for unidentifiable labels and a redundancy criterion that deletes concepts. Please explain how the final concept count is determined and how this is consistent with a fixed initial size equal to the number of labels.
  2. [Section 4.2 / Section 4.3] The evaluation repeats each MCQ three times and takes a majority vote, but the sampling temperature and exact GPT-4-turbo version are not reported; the paper says only that MCQs 'remain fixed for each dataset.' Please report the API model identifier and decoding parameters for reproducibility.
  3. [Table 1 and Table 2] The arrow notation 'Sta→Dyn' and 'w/ E→w/o E' in the table captions is ambiguous; please state explicitly which direction corresponds to the reported accuracy/interpretability values, e.g., 'static → dynamic' means the first column is the static result and the second is the dynamic result.
  4. [Figure 4] The x-axis label 'number of samples' should specify whether these are the K-means cluster centroids (βN instances) described in Section 3.4, and the value of β used in the few-shot feedback phase should be stated.

Circularity Check

1 steps flagged · score 6.0 of 10

Interpretability improvement is a closed-loop LLM self-assessment: GPT-4-turbo judges concepts generated and fact-checked by GPT-4o, so the 30% gain measures LLM self-consistency, not human interpretability.

  1. self definitional [Section 4.2 (Interpretability; Truthfulness) and Section 4.3 (Implementation Details; also Section 3.4 Fact Verification)]
    "Truthfulness. This metric evaluates whether the concepts that support the predicted labels are consistent with objective real-world facts. ... The Concept Agent prompts the GPT-4o API (Hurst et al., 2024) for concept generation and verification. ... For evaluation, GPT-4-turbo is prompted to answer MCQs, each repeated 3 times, with the majority vote taken as the final result."

    The 'objective facts' are the parametric knowledge of the same LLM family (GPT-4o) that generated and fact-checked the concept bank; the judge (GPT-4-turbo) shares that prior. Truthfulness therefore measures GPT-4o-to-GPT-4-turbo self-consistency, not an independent human-interpretability standard. The claimed approximately 30% interpretability improvement over LM4CV-2n is thus a closed-loop validation: the evaluator's knowledge is the generator's knowledge. No human study or correlation with an established interpretability metric is reported, and the paper's Limitations section concedes that 'evaluating the internal knowledge of LLMs ... remains open challenges.'

full rationale

The accuracy comparisons (e.g., 6.15% average gain over LaBo-n) are benchmarked against external baselines and are not circular: the dynamic concept bank is refined on validation feedback, while final reported accuracy is an independent test-set result. The central circularity is confined to the interpretability claim. The truthfulness and distinguishability metrics are answered by GPT-4-turbo on concept lists produced and fact-verified by GPT-4o; the 30% interpretability gain is therefore not evidence about human interpretability but about within-family LLM agreement. The paper's own Limitations section flags that evaluating LLM internal knowledge remains an open challenge, which weakens the metric's validity. A separate non-circularity issue: Section 3.4 states Ejk=1 for critical/occasional concepts while Eq. (4) defines Ejk=1 for incompatible concepts, making the editable-matrix ablation implementation-dependent; this is a correctness/reproducibility risk, not a circularity. Overall, the interpretability result reduces, in effect, to an LLM self-consistency score, while the accuracy result retains independent content—hence partial circularity, not full collapse.

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

The central accuracy result relies on the standard CLIP scoring assumption and on the agent's thresholds. The interpretability headline relies on the unvalidated assumption that LLM MCQs capture human interpretability, which is the main circular element.

free parameters (4)
  • t_a (activation threshold) = 0.1
    Threshold in Eq. 9 that decides whether a concept is activated for a label; empirically chosen in Section 4.3. No sensitivity analysis is reported.
  • t_m (redundancy distance threshold) = 0.3
    Manhattan distance below which two concepts with identical activation patterns are considered redundant; set empirically in Section 4.3.
  • q (learnable prompt tokens) = 8
    Number of learnable condition tokens in Eq. 3, following Zhou et al. (2022); fixed across datasets.
  • beta (K-means clusters per label) = implicit, 16 total instances
    Section 3.4 selects beta cluster centroids per label as feedback instances; Section 4.3 says 16 samples are used, so beta varies with dataset. This affects which images the agent sees.
assumptions (4)
  • domain assumption CLIP text-image dot products are a valid scoring mechanism for visual concepts.
    Section 3.1 assumes s_ck = I(x_i) dot T(c_k) measures concept presence; all scores and the feedback loop rely on this.
  • domain assumption LLM multiple-choice fact verification correctly determines whether a concept can apply to a label.
    The editable matrix E in Eq. 4 is filled by GPT-4o judgments treated as ground truth; no human verification of these facts is provided.
  • domain assumption LLM-based MCQs in Section 4.2 measure interpretability.
    Truthfulness and distinguishability are scored by GPT-4-turbo; this is the only quantitative interpretability metric, and no correlation with human ratings is shown.
  • ad hoc to paper The original CBM is a special case of CoCoBM by collapsing the label-conditioned matrix.
    Section 3.1 claims collapsing along the label dimension recovers the original CBM, but the label name is always part of the prompt, so it is not shown that the shared-score CBM is actually a limiting case of Eq. 2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Interpretable Image Classification Through LLM Agents and Conditional Concept Bottleneck Models." pith.science (2026). https://pith.science/paper/SPLEXO74

@misc{pith2026250601334,
  author       = {Pith},
  title        = {Pith review of: Enhancing Interpretable Image Classification Through LLM Agents and Conditional Concept Bottleneck Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SPLEXO74}},
  note         = {Machine review of arXiv:2506.01334}
}
read the original abstract

Concept Bottleneck Models (CBMs) decompose image classification into a process governed by interpretable, human-readable concepts. Recent advances in CBMs have used Large Language Models (LLMs) to generate candidate concepts. However, a critical question remains: What is the optimal number of concepts to use? Current concept banks suffer from redundancy or insufficient coverage. To address this issue, we introduce a dynamic, agent-based approach that adjusts the concept bank in response to environmental feedback, optimizing the number of concepts for sufficiency yet concise coverage. Moreover, we propose Conditional Concept Bottleneck Models (CoCoBMs) to overcome the limitations in traditional CBMs' concept scoring mechanisms. It enhances the accuracy of assessing each concept's contribution to classification tasks and feature an editable matrix that allows LLMs to correct concept scores that conflict with their internal knowledge. Our evaluations across 6 datasets show that our method not only improves classification accuracy by 6% but also enhances interpretability assessments by 30%.

Figures

Figures reproduced from arXiv: 2506.01334 by the authors.

Figure 1
Figure 1. Architectural comparison of CBMs and Co [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Left: Modular components with intra-module and inter-module workflows in the Concept Agent. Right: The planning module informs the action module to iteratively generate and refine concepts based on feedback. specific ones. The computation of SY is redefined as: ⃗ s j c = [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison with state-of-the-art CBMs and black-box models. The legend follows the format: method [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Effect of sample size on the classification accu [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Case Study. Left: Concept bank evolution on CIFAR-10 (3rd iteration omitted), concluding after 4 rounds with 9 final concepts. The x-axis of the Action module denotes the number of labels used to prompt the LLM, total generated concepts, and selected concepts. Right: L…
Figure 6
Figure 6. Figure 6: Prompt templates used in the Concept Agent’s action module, and interpretability evaluation templates. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 9 canonical work pages

  1. [1]

    David Arthur and Sergei Vassilvitskii. 2007. http://dl.acm.org/citation.cfm?id=1283383.1283494 k-means++: the advantages of careful seeding . In Proceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2007, New Orleans, Louisiana, USA, January 7-9, 2007 , pages 1027--1035. SIAM

    arXiv 2007

  2. [2]

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. 2014. https://doi.org/10.1007/978-3-319-10599-4\_29 Food-101 - mining discriminative components with random forests . In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI , volume 8694 of Lecture Notes in Computer Science, pages 446-...

    doi:10.1007/978-3-319-10599-4 2014

  3. [3]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gr...

    2020

  4. [4]

    Khyathi Raghavi Chandu, Yonatan Bisk, and Alan W. Black. 2021. https://doi.org/10.18653/V1/2021.FINDINGS-ACL.375 Grounding 'grounding' in NLP . In Findings of the Association for Computational Linguistics: ACL/IJCNLP 2021, Online Event, August 1-6, 2021 , volume ACL/IJCNLP 2021 of Findings of ACL , pages 4283--4305. Association for Computational Linguistics

    doi:10.18653/v1/2021.findings-acl.375 2021

  5. [5]

    Ruth Fong and Andrea Vedaldi. 2018. https://doi.org/10.1109/CVPR.2018.00910 Net2vec: Quantifying and explaining how concepts are encoded by filters in deep neural networks . In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018 , pages 8730--8738. Computer Vision Foundation / IEEE Computer Society

    work page arXiv 2018

  6. [6]

    Lijie Hu, Chenyang Ren, Zhengyu Hu, Hongbin Lin, Cheng-Long Wang, Hui Xiong, Jingfeng Zhang, and Di Wang. 2025. https://arxiv.org/abs/2405.15476 Editable concept bottleneck models . Preprint, arXiv:2405.15476

    arXiv 2025

  7. [7]

    Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Noah Brown, Tomas Jackson, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter. 2022. https://arxiv.org/abs/2207.05608 Inner monologue: Embodied reasoning through planning with language models . Pre...

    arXiv 2022

  8. [8]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. https://arxiv.org/abs/2410.21276 Gpt-4o system card . Preprint, arXiv:2410.21276

    arXiv 2024

Show all 35 references
  1. [9]

    Cai, James Wexler, Fernanda B

    Been Kim, Martin Wattenberg, Justin Gilmer, Carrie J. Cai, James Wexler, Fernanda B. Vi \' e gas, and Rory Sayres. 2018. http://proceedings.mlr.press/v80/kim18d.html Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCAV) . In P...

    2018

  2. [10]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. http://arxiv.org/abs/1412.6980 Adam: A method for stochastic optimization . In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings

    2015 arXiv

  3. [11]

    Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. 2020. https://proceedings.mlr.press/v119/koh20a.html Concept bottleneck models . In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceed...

    2020

  4. [12]

    Alex Krizhevsky. 2009. https://www.cs.toronto.edu/ kriz/learning-features-2009-TR.pdf Learning multiple layers of features from tiny images . Technical report, University of Toronto, Toronto, Ontario

    2009

  5. [13]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://arxiv.org/abs/1907.11692 Roberta: A robustly optimized bert pretraining approach . Preprint, arXiv:1907.11692

    2019 arXiv

  6. [14]

    Max Losch, Mario Fritz, and Bernt Schiele. 2019. https://arxiv.org/abs/1907.10882 Interpretability beyond classification output: Semantic bottleneck networks . Preprint, arXiv:1907.10882

    2019 arXiv

  7. [15]

    Deval Mehta, Yiwen Jiang, Catherine L Jan, Mingguang He, Kshitij Jadhav, and Zongyuan Ge. 2025. https://arxiv.org/abs/2503.02917 Interpretable few-shot retinal disease diagnosis with concept-guided prompting of vision-language models . Preprint, arXiv:2503.02917

    2025 arXiv

  8. [16]

    Maria - Elena Nilsback and Andrew Zisserman. 2008. https://doi.org/10.1109/ICVGIP.2008.47 Automated flower classification over a large number of classes . In Sixth Indian Conference on Computer Vision, Graphics & Image Processing, ICVGIP 2008, Bhubaneswar, India, 16-19 Decembe...

    2008 doi

  9. [17]

    Oikarinen, Subhro Das, Lam M

    Tuomas P. Oikarinen, Subhro Das, Lam M. Nguyen, and Tsui - Wei Weng. 2023. https://openreview.net/forum?id=FlCg47MNvBA Label-free concept bottleneck models . In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net

    2023

  10. [18]

    A \" a ron Van Den Oord, Oriol Vinyals, and Koray Kavukcuoglu. 2017. https://proceedings.neurips.cc/paper/2017/hash/7a98af17e63a0ac09ce2e96d03992fbc-Abstract.html Neural discrete representation learning . In Advances in Neural Information Processing Systems 30: Annual Conferen...

    2017

  11. [19]

    Berkay Celik, and Ananthram Swami

    Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z. Berkay Celik, and Ananthram Swami. 2017. https://doi.org/10.1145/3052973.3053009 Practical black-box attacks against machine learning . In Proceedings of the 2017 ACM on Asia Conference on Computer and Communic...

    2017

  12. [20]

    Parkhi, Andrea Vedaldi, Andrew Zisserman, and C

    Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. 2012. https://doi.org/10.1109/CVPR.2012.6248092 Cats and dogs . In 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI, USA, June 16-21, 2012 , pages 3498--3505. IEEE Computer Society

    2012

  13. [21]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. http://proceedings.mlr.press/v139/radford21a.html Learning transferable visual models...

    2021

  14. [22]

    Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. 2024. https://doi.org/10.1109/CVPR52733.2024.01049 Incremental residual concept bottleneck models . In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, US...

    2024

  15. [23]

    Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. https://doi.org/10.1609/AAAI.V31I1.11164 Conceptnet 5.5: An open multilingual graph of general knowledge . In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, February 4-9, 2017, San Francisco, Ca...

    2017 doi

  16. [24]

    Belongie

    Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge J. Belongie. 2011. https://api.semanticscholar.org/CorpusID:16119123 The caltech-ucsd birds-200-2011 dataset . California Institute of Technology

    2011

  17. [25]

    Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, Wayne Xin Zhao, Zhewei Wei, and Jirong Wen. 2024. https://doi.org/10.1007/S11704-024-40231-1 A survey on large language model based autonomous agents . Frontie...

    2024 doi

  18. [26]

    Chi, Quoc V

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. http://papers.nips.cc/paper\_files/paper/2022/hash/9d5609613524ecf4f15af0f7b31abca4-Abstract-Conference.html Chain-of-thought prompting elicits reasoning...

    2022

  19. [27]

    An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Yang Wang, Jingbo Shang, and Julian J. McAuley. 2023 a . https://doi.org/10.1109/ICCV51070.2023.00287 Learning concise and descriptive attributes for visual recognition . In IEEE/CVF International Conferenc...

    2023

  20. [29]

    Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison - Burch, and Mark Yatskar. 2023. https://doi.org/10.1109/CVPR52729.2023.01839 Language in a bottle: Language model guided concept bottlenecks for interpretable image classification . In IEEE/CVF Conferen...

    2023

  21. [30]

    Narasimhan, and Yuan Cao

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2023. https://openreview.net/forum?id=WE\_vluYUL-X React: Synergizing reasoning and acting in language models . In The Eleventh International Conference on Learning Representations, ...

    2023

  22. [31]

    u ksekg \

    Mert Y \" u ksekg \" o n \" u l, Maggie Wang, and James Zou. 2023. https://openreview.net/forum?id=nA5AZ8CEyow Post-hoc concept bottleneck models . In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net

    2023

  23. [32]

    Bolei Zhou, Yiyou Sun, David Bau, and Antonio Torralba. 2018. https://doi.org/10.1007/978-3-030-01237-3\_8 Interpretable basis decomposition for visual explanation . In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, ...

    2018 doi

  24. [33]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. https://doi.org/10.1007/S11263-022-01653-1 Learning to prompt for vision-language models . International Journal of Computer Vision, 130(9):2337--2348

    2022 doi

  25. [34]

    Xizhou Zhu, Yuntao Chen, Hao Tian, Chenxin Tao, Weijie Su, Chenyu Yang, Gao Huang, Bin Li, Lewei Lu, Xiaogang Wang, Yu Qiao, Zhaoxiang Zhang, and Jifeng Dai. 2023. https://arxiv.org/abs/2305.17144 Ghost in the minecraft: Generally capable agents for open-world environments via...

    2023 arXiv

  26. [35]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  27. [36]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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