Pith. sign in

REVIEW 3 major objections 5 minor 45 references

SABRE: A Multi-Agent Approach for Selecting Out-of-Distribution Detectors Under a Budget

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read An out-of-distribution detector that leads on natural images can invert on a specialized domain, and SABRE replaces that fixed choice with per-regime selection.

desk verdict The empirical finding that fixed OOD detectors invert across domains is solid and worth knowing; the SABRE framework is plausible but the paper overstates 'tracks the strongest detector' because it omits the obvious baseline of using the same calibration labels to simply pick the best single detector, which likely beats SABRE at lower cost. read the letter →

arxiv 2608.02959 v1 pith:2OMGVHZB submitted 2026-08-03 cs.MA

classification cs.MA
keywords out-of-distributiondetectiondomainshiftdetectorinversionreliabilitycalibrationvision-languagemodelsmulti-agentselectionbudgetedinferencepost-hocdetectors
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 claims that the standard practice of choosing one post-hoc out-of-distribution (OOD) detector on a benchmark and fixing it at deployment is unsafe: on a single frozen encoder, a detector that leads on natural images can invert on a specialized domain, scoring normal inputs as more anomalous than genuine outliers, and the best detector changes from one domain to the next. This matters because a deployed system cannot know in advance which domain it faces, so a fixed choice can be not merely suboptimal but systematically reversed. SABRE addresses this by using three language-model agents at inference: a Selector chooses which detectors to consult under a bounded budget, a Reporter summarizes evidence per image, and an Analyst weights detectors by reliability measured on a small labeled calibration set drawn from the deployment domain. The paper argues that reliability is therefore something to establish at deployment, not assume from a benchmark, and that SABRE does this automatically, recovering reliable detection where a fixed detector inverts and matching the best fixed detector where one already works.

What carries the argument

The load-bearing identity is the Analyst's reliability weight $w_m = \max(0, 2A_m(\mathcal{C})-1)$, computed from each detector's AUROC $A_m$ on a small labeled calibration set $\mathcal{C}$ drawn from the deployment domain and disjoint from the scored data. Detectors at or below chance ($A_m \le 0.5$) receive zero weight and are ruled out, preventing an inverted detector from cancelling sound ones in the aggregate. The final score is the weight-adjusted mean of calibrated positions, $s_{\text{SABRE}}(x) = \frac{\sum_{m \in R(x)} w_m \tilde{s}_m(x)}{\sum_{m \in R(x)} w_m}$; each raw score is mapped to the scale of positions among calibration in-distribution scores. This calibration loop, not the order in which the Selector consults detectors, carries the recovery, with the Selector and Reporter contributing budget efficiency and modest gains.

What would settle it

If a domain pair is found on a frozen encoder where no post-hoc detector inverts and the benchmark-preferred detector remains best, the central claim would collapse; concretely, compute AUROC of MSP, Energy, and MCM on natural images versus any specialized image domain, and if none of them falls below 50 while scoring above 50 on natural images, the inversion premise fails.

Watch

Extended reading notes

Core claim

On CLIP ViT-B/16, MCM scores 97.7 AUROC on iNaturalist and 31.2 on NCT-CRC; Energy drops from 98.5 to 43.6; MSP drops from 87.9 to 29.9, so the detector that ranks highest on natural-image benchmarks ranks below chance on histopathology. The paper establishes that no fixed post-hoc detector is reliable throughout, and the same pattern recurs on all five encoders, including a domain-specialized biomedical encoder that lifts every detector but leaves the per-domain ordering intact. SABRE's central claim is that reliability can be measured at deployment rather than assumed: three language-model agents select and weight detectors per regime, and the decisive component is the Analyst's reliability calibration, which assigns zero weight to detectors that are at or below chance on a small held-out labeled sample. The result is that SABRE reaches or approaches the best fixed detector in each domain without knowing in advance which detector that is, and a component analysis isolates the calibration step as the source of the recovery.

Load-bearing premise

The method assumes a small labeled calibration set from the deployment domain is available at deployment; without those labels the reliability weights cannot be computed and SABRE degrades to an unweighted selector that is near chance where detectors invert.

Editorial extensions

If this is right

  • A practitioner entering a new domain with a frozen encoder can measure which detectors to trust with roughly 50 to 150 labeled examples and can rule out detectors that invert there.
  • Benchmark comparison should not be used to lock in a single detector for deployment; the paper's evidence implies reliability must be re-established per domain.
  • Under a budget of three detector calls per image, SABRE reaches about 97% of full-pool quality while spending 2.6 calls per image on average.
  • SABRE tracks the best fixed detector rather than surpassing it, so gains concentrate where a fixed detector would fail; on the hardest tissue source, where no detector separates the classes, selection cannot create signal.
  • Using a domain-matched encoder and per-regime selection are complementary: a specialized encoder raises every detector's floor but does not remove the need to choose among them.

Reading between the lines

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

  • Editorial inference: the inversion phenomenon is likely broader than the detectors tested here; any fixed scoring rule reads a particular structure of a frozen representation, and a domain shift can reorganize which structure separates novelty, so the same failure could appear for other score families.
  • Editorial inference: the calibration weight $2A_m(\mathcal{C})-1$ is a monotone transform of per-domain AUROC, so the same logic could be used to choose among feature extractors or captioning models, not only among post-hoc detectors.
  • Editorial inference: the paper's random-selection control shows the Selector's ordering contributes cost savings rather than accuracy, suggesting the agent layer could be replaced by a cheaper deterministic scheduling rule whenever latency matters, leaving the Analyst's weights as the core mechanism.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper studies whether a post-hoc OOD detector selected on a benchmark remains reliable under domain shift for a frozen vision-language encoder. Using CLIP-family encoders and two domains (natural images and histopathology), the authors show that detectors such as MCM and Energy, which lead on natural images, invert on histopathology by scoring in-distribution tissue as more anomalous than outliers, and that the best detector changes by domain. They introduce SABRE, a multi-agent system with three LLM agents: a Selector chooses detectors under a bounded per-image budget, a Reporter consolidates evidence, and an Analyst computes reliability weights from a labeled calibration set disjoint from the test data. The paper proposes four multimodal density detectors and reports a component analysis showing that the Analyst's reliability weighting, rather than the selection order, drives recovery; random selection with the same weights matches SABRE, and SABRE reaches 76.3 AUROC on NCT-CRC with 2.6 mean detector calls.

Significance. The empirical finding that fixed detector reliability does not transfer across domains is clearly supported by the data and strengthened by careful controls: disjoint calibration and test splits, validation-split tracking that matches the test split, a random-selection control, and paired-bootstrap significance tests. However, the paper's central method claim, that SABRE tracks the strongest detector per domain, is not established because the most natural baseline using the same calibration labels, selecting the single detector with maximum calibration AUROC and deploying it, is never evaluated. Such a baseline costs one detector call per image and would likely approach the oracle 'Best single detector' entries that SABRE fails to reach on the main tissue datasets. The contribution of the LLM agents is therefore unclear beyond budget accounting, and the paper's method-level conclusions need to be re-examined.

major comments (3)
  1. [Methodology, 'Analyst and the calibration mechanism' (Eq. 2); Table 3] The central claim that SABRE 'tracks the strongest detector in each domain' is not supported because the paper omits the calibration-argmax baseline. The labeled calibration set C used to compute w_m = max(0, 2A_m(C)-1) in Eq. (2) can equally be used to rank detectors by A_m(C), select m* = argmax_m A_m(C), and deploy that single detector. This baseline uses exactly the same supervision as SABRE, costs one detector call per image rather than SABRE's mean 2.6 calls (Table 3), and should approximately achieve the 'Best single detector' row: 79.9 on NCT-CRC where SABRE obtains 76.3, and 69.6 on CRC-VAL where SABRE obtains 64.4 (Table 2). The paper gives no evidence that SABRE reaches this baseline's level, so the method-level claim of tracking or converging to the strongest detector is unsubstantiated.
  2. [Experiments, 'What the Recovery Depends On' (Table 3)] Table 3 shows that random selection under the Analyst's weights (76.6) matches SABRE (76.3) on NCT-CRC, and that unweighted selection is near chance. Combined with the component analysis (Table 2), this means the entire accuracy recovery comes from the calibration weights, not from the agent-based selection order. The paper should either demonstrate a domain or dataset where the Selector and Reporter improve over random selection under the same budget and weights, or reframe SABRE as a calibration-weighted ensemble with a budget cap. As written, the multi-agent selection mechanism is not shown to contribute to accuracy.
  3. [Table 2 vs. Table 6; natural-image results] There is an inconsistency in the reported best single detector for NCT-CRC: Table 2 gives 79.6, while Table 3 and supplementary Table 6 give 79.9; this must be reconciled. In addition, the natural-image SABRE results are reported only in prose and Figure 4, with no table of AUROC values and no comparison to the calibration-argmax baseline; such a table is needed to support the claim that SABRE 'matches the strongest fixed detector' on that domain, especially because there the reliability weights are also derived from labeled calibration data.
minor comments (5)
  1. [Abstract and main text] There are formatting artifacts in the abstract and introduction, such as 'SABRE (Selective Agentic Budgeted Reliability Ensemble,)' and 'S elective Agentic BudgetedR eliabilityE nsemble', with stray spaces and punctuation; please correct these.
  2. [Supplementary D.6; main-text experiments] The selective gate with threshold theta = 0.15 is described in the supplementary material but never mentioned in the main-text experimental protocol; please clarify whether the reported mean calls and AUROC figures include this gate.
  3. [Table 13] Table 13 shows that MSP and MCM retain high selection shares even after their calibration weights become zero; since these calls do not affect the aggregate, please explain whether the Selector is aware of zero weights when spending budget, or report an effective weighted-call count.
  4. [Supplementary C.5] The four proposed detectors are strongly rank-correlated (Kendall tau between 0.72 and 0.99), which the main text should acknowledge when interpreting pool diversity and the Selector's choices.
  5. [GCHTID discussion] GCHTID is repeatedly called a 'difficulty ceiling', but the +Crit configuration reaches 58.0 against best single 60.7 on ViT-L/14 (Table 6); please define the term more carefully or weaken the ceiling language.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: SABRE's calibration weights are held-out and the QPM self-citation is not load-bearing.

full rationale

The derivation chain is not circular. The central inversion claim is an empirical benchmark result measured on external datasets (ImageNet-100, iNaturalist, SUN, Places, Textures, NCT-CRC, CRC-VAL, GCHTID), not derived from the method's own weights. The Analyst's reliability weights w_m = max(0, 2A_m(C)-1) are computed on a labeled calibration set C that is explicitly disjoint from the scored test data, so reporting test AUROC is a held-out evaluation rather than a fitted prediction. Eq. (3) combines detector scores with these held-out weights; no test label enters the computation. The only self-referential element is the QPM baseline, a static variant of the authors' own T-QPM (Naiknaware and Sekeh 2026), but it is one of nine pool members and the inversion result is independently established by MSP, Energy, and MCM on external benchmarks, so the self-citation is not load-bearing. The skeptical concern that a calibration-argmax single-detector baseline is missing is an evaluation or comparison gap rather than circularity: omitting a baseline does not make SABRE's reported AUROC equivalent to its inputs by construction. The paper also runs controls (random selection under the same weights, budget sweeps, disjoint validation splits) that separate calibration from selection order and confirm the configuration ordering was not selected on test data, further supporting the non-circular interpretation.

Assumptions & free parameters 8 free parameters · 6 assumptions · 4 invented entities

The central claim rests on a small set of hand-chosen hyperparameters (gamma, delta, G, theta, B, baseline defaults) and on the availability of labeled calibration data from the deployment domain. The new entities are four OOD detector variants, all fully specified and benchmarked, so they carry independent evidence. No physics-style invented entities (new particles, forces, dimensions) appear.

free parameters (8)
  • gamma (caption weight in proposed detectors) = 2
    Fixed weight balancing the caption deficit against the density term in Eq. (6) and Eq. (10); set by hand, not fitted to data.
  • delta (MMCA cross-modal penalty weight) = 0.25
    Small weight on the MMCA cross-modal coupling penalty in Eq. (10); chosen by hand.
  • G (number of semantic groups) = 4
    Number of clusters used by SMAP, RCAP, and MMCA; fixed default in the agglomerative merging procedure.
  • theta (selective gate threshold) = 0.15
    Latency gate threshold in the deployment mode; a fixed option layered on the full system, not tuned on the test split.
  • B (query budget) = 3
    Maximum detector calls per image; a user-chosen cap, with a cost sweep showing quality plateaus at B=3.
  • MCM temperature T = 1
    Hyperparameter of the MCM baseline; the paper says it may be tuned on validation, but the default is used in the reported runs.
  • QPM beta weights = uniform 0.25
    Weights for the four prototype banks in the QPM baseline; defaults are uniform and may be tuned on validation.
  • calibration set fraction = 30% (150 images)
    Fraction of assembled samples held out for the Analyst's calibration; a sweep shows the method works down to 50 labeled examples.
assumptions (6)
  • domain assumption A labeled calibration set can be drawn from the deployment domain, disjoint from the scored set.
    The Analyst's weights w_m = max(0, 2A_m(C)-1) require labels on a sample from the target domain. The 'Analyst and the calibration mechanism' subsection states the calibration set is 'held out from the deployment domain and never scored at test time'.
  • domain assumption The captioner's output does not leak class identity.
    Supplement A.3 relies on a prompt forbidding class names and excludes biomedical captioners. If captions leaked labels, the caption term a(x) would carry supervised information and inflate the proposed detectors' apparent reliability.
  • domain assumption The calibration set is representative of the test distribution (same deployment domain).
    Weights are computed on the calibration set and applied to the test split; if the two diverge, the weights may not transfer. The paper draws both from the same deployment domain.
  • standard math AUROC is the correct threshold-free metric for comparing OOD detectors.
    All comparisons use AUROC, which is the standard protocol in the cited OOD literature; the paper does not discuss alternative operating points.
  • domain assumption LLM agents with greedy decoding and fixed seeds produce reproducible decisions.
    Supplement D.1 sets temperature 0, top-k=1, fixed seed, and token caps; hardware-level floating point nondeterminism remains, so identical text output is expected but not guaranteed across hardware.
  • ad hoc to paper The detector pool is static and finite; selection cannot create signal absent from all detectors.
    The Discussion states 'where no detector in the pool separates the classes... selection cannot create signal', bounding the method's scope to pools that contain at least one informative detector.
invented entities (4)
  • SMAP (Semantic Mahalanobis Anomaly Profile) independent evidence
    purpose: Proposed OOD detector: per-semantic-group covariance Mahalanobis distance plus a per-image caption agreement term.
    SMAP is fully specified in Supplement C.2 and benchmarked with AUROC values on two domains and five encoders (Tables 1, 7-10), so it is a testable method, not an unfalsifiable postulate.
  • RCAP (Regional Concentration Anomaly Profile) independent evidence
    purpose: Proposed OOD detector: pooled-covariance Mahalanobis distance across semantic groups plus the caption term.
    RCAP is specified in Supplement C.2 and benchmarked with AUROC values across domains and encoders, giving it an independent empirical handle.
  • MMCA (Multi-Modal Confidence Asymmetry) independent evidence
    purpose: Proposed OOD detector: per-group density plus an explicit cross-modal penalty when the caption names a group far from the visual embedding.
    MMCA is specified in Supplement C.2 and benchmarked across domains and encoders, so its behavior is externally checkable.
  • GMAP (Gated Mixture Anomaly Positioning) independent evidence
    purpose: Proposed OOD detector: per-class mixture density with a learned gate over a caption-space density.
    GMAP is described in Supplement C.3 and benchmarked across domains and encoders, giving it an independent empirical evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SABRE: A Multi-Agent Approach for Selecting Out-of-Distribution Detectors Under a Budget." pith.science (2026). https://pith.science/paper/2OMGVHZB

@misc{pith2026260802959,
  author       = {Pith},
  title        = {Pith review of: SABRE: A Multi-Agent Approach for Selecting Out-of-Distribution Detectors Under a Budget},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2OMGVHZB}},
  note         = {Machine review of arXiv:2608.02959}
}
read the original abstract

Post-hoc out-of-distribution (OOD) detection for vision-language models assumes that a detector chosen on a benchmark stays reliable once deployed. We show this fails across domains: on a single frozen encoder, a detector that leads in one domain can invert in another, scoring in-distribution inputs as more anomalous than genuine outliers, and the best detector changes from domain to domain, so no fixed choice is reliable throughout. We introduce SABRE (Selective Agentic Budgeted Reliability Ensemble,) which replaces this fixed choice with per-regime selection at inference. Three language-model agents reason over a library of post-hoc detectors under a bounded query budget: a Selector chooses which detector to consult next, a Reporter consolidates the evidence for each input, and an Analyst calibrates detector reliability on a small labeled sample held out from the deployment domain and disjoint from the test data, weighting selection and aggregation without ever observing a scored input's label. The library includes four multimodal density detectors we propose. Inferring the operating regime from data, SABRE tracks the strongest detector in each domain without prior knowledge of it, recovering reliable detection where a conventional detector inverts and converging to that detector where it is sound. A component analysis shows the agents are complementary: the Reporter's feedback yields consistent gains, and the Analyst's calibration is decisive against inversion, ruling out unreliable detectors so that aggregation no longer cancels the sound ones. Since no fixed rule can be trusted across domains, reliability must be established at deployment rather than assumed from a benchmark, and SABRE shows this can be done automatically.

Figures

Figures reproduced from arXiv: 2608.02959 by the authors.

Figure 1
Figure 1. SABRE. The Selector consults detectors from the pool one at a time under a budget [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Detector AUROC on CLIP ViT-B/16, from natural [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 5
Figure 5. Recovery across the general-purpose encoders [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The Analyst’s reliability weights wm for the same system in the two domains (CLIP ViT-B/16). The confidence and concept detectors carry the weight on natural images and are switched off on tissue (marked off), and the density detectors reverse. The reweighting is infer…
Figure 6
Figure 6. Figure 6: Domain-specialized encoder ablation (NCT-CRC, [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Per-detector AUROC (%) for all five encoders (mean of five seeds), natural and tissue blocks separated by the rule. On [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 40 canonical work pages

  1. [1]

    Hendrycks and K

    D. Hendrycks and K. Gimpel. A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks. ICLR. 2017

  2. [2]

    Liang and Y

    S. Liang and Y. Li and R. Srikant. Enhancing the Reliability of Out-of-Distribution Image Detection in Neural Networks. ICLR. 2018

  3. [3]

    Lee and K

    K. Lee and K. Lee and H. Lee and J. Shin. A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks. NeurIPS. 2018

  4. [4]

    Liu and X

    W. Liu and X. Wang and J. D. Owens and Y. Li. Energy-based Out-of-Distribution Detection. NeurIPS. 2020

  5. [5]

    Sun and C

    Y. Sun and C. Guo and Y. Li. ReAct : Out-of-Distribution Detection with Rectified Activations. NeurIPS. 2021

  6. [6]

    Sun and Y

    Y. Sun and Y. Ming and X. Zhu and Y. Li. Out-of-Distribution Detection with Deep Nearest Neighbors. ICML. 2022

  7. [7]

    Sun and Y

    Y. Sun and Y. Li. DICE : Leveraging Sparsification for Out-of-Distribution Detection. ECCV. 2022

  8. [8]

    Wang and Z

    H. Wang and Z. Li and L. Feng and W. Zhang. ViM : Out-of-Distribution with Virtual-Logit Matching. CVPR. 2022

Show all 45 references
  1. [9]

    Wei and R

    H. Wei and R. Xie and H. Cheng and L. Feng and B. An and Y. Li. Mitigating Neural Network Overconfidence with Logit Normalization. ICML. 2022

  2. [10]

    Ren and S

    J. Ren and S. Fort and J. Liu and A. G. Roy and S. Padhy and B. Lakshminarayanan. A Simple Fix to Mahalanobis Distance for Improving Near- OOD Detection. arXiv preprint arXiv:2106.09022. 2021

  3. [11]

    Lakshminarayanan and A

    B. Lakshminarayanan and A. Pritzel and C. Blundell. Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles. NeurIPS. 2017

  4. [12]

    Yang and P

    J. Yang and P. Wang and D. Zou and others. OpenOOD : Benchmarking Generalized Out-of-Distribution Detection. NeurIPS. 2022

  5. [13]

    Yang and K

    J. Yang and K. Zhou and Y. Li and Z. Liu. Generalized Out-of-Distribution Detection: A Survey. International Journal of Computer Vision. 2024

  6. [14]

    Radford and J

    A. Radford and J. W. Kim and C. Hallacy and others. Learning Transferable Visual Models from Natural Language Supervision. ICML. 2021

  7. [15]

    Zhai and B

    X. Zhai and B. Mustafa and A. Kolesnikov and L. Beyer. Sigmoid Loss for Language Image Pre-Training. ICCV. 2023

  8. [16]

    Zhang and Y

    S. Zhang and Y. Xu and N. Usuyama and others. Large-Scale Domain-Specific Pretraining for Biomedical Vision-Language Processing. arXiv preprint arXiv:2303.00915. 2023

  9. [17]

    Ming and Z

    Y. Ming and Z. Cai and J. Gu and Y. Sun and W. Li and Y. Li. Delving into Out-of-Distribution Detection with Vision-Language Representations. NeurIPS. 2022

  10. [18]

    Esmaeilpour and B

    S. Esmaeilpour and B. Liu and E. Robertson and L. Shu. Zero-Shot Out-of-Distribution Detection Based on the Pre-trained Model CLIP. AAAI. 2022

  11. [19]

    Wang and Y

    H. Wang and Y. Li and H. Yao and X. Li. CLIPN for Zero-Shot OOD Detection: Teaching CLIP to Say No. ICCV. 2023

  12. [20]

    Miyai and Q

    A. Miyai and Q. Yu and G. Irie and K. Aizawa. LoCoOp : Few-Shot Out-of-Distribution Detection via Prompt Learning. NeurIPS. 2023

  13. [21]

    Jiang and F

    X. Jiang and F. Liu and Z. Fang and H. Chen and T. Liu and F. Zheng and B. Han. Negative Label Guided OOD Detection with Pretrained Vision-Language Models. ICLR. 2024

  14. [22]

    Miyai and Q

    A. Miyai and Q. Yu and G. Irie and K. Aizawa. GL-MCM : Global and Local Maximum Concept Matching for Zero-Shot Out-of-Distribution Detection. International Journal of Computer Vision. 2025

  15. [23]

    Cao and Z

    C. Cao and Z. Zhong and Z. Zhou and Y. Liu and T. Liu and B. Han. Envisioning Outlier Exposure by Large Language Models for Out-of-Distribution Detection. ICML. 2024

  16. [24]

    Ming and Y

    Y. Ming and Y. Li. How Does Fine-Tuning Impact Out-of-Distribution Detection for Vision-Language Models?. International Journal of Computer Vision. 2023

  17. [25]

    Xu and K

    R. Xu and K. Ding. Large Language Models for Anomaly and Out-of-Distribution Detection: A Survey. arXiv preprint arXiv:2409.01980. 2024

  18. [26]

    Yao and J

    S. Yao and J. Zhao and D. Yu and N. Du and I. Shafran and K. Narasimhan and Y. Cao. ReAct : Synergizing Reasoning and Acting in Language Models. ICLR. 2023

  19. [27]

    Shinn and F

    N. Shinn and F. Cassano and E. Berman and A. Gopinath and K. Narasimhan and S. Yao. Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS. 2023

  20. [28]

    Yao and D

    S. Yao and D. Yu and J. Zhao and I. Shafran and T. L. Griffiths and Y. Cao and K. Narasimhan. Tree of Thoughts: Deliberate Problem Solving with Large Language Models. NeurIPS. 2023

  21. [29]

    Wu and G

    Q. Wu and G. Bansal and J. Zhang and others. AutoGen : Enabling Next-Gen LLM Applications via Multi-Agent Conversation. COLM. 2024

  22. [30]

    Hong and M

    S. Hong and M. Zhuge and J. Chen and others. MetaGPT : Meta Programming for a Multi-Agent Collaborative Framework. ICLR. 2024

  23. [31]

    J. N. Kather and N. Halama and A. Marx. 100,000 Histological Images of Human Colorectal Cancer and Healthy Tissue. 2018. doi:10.5281/zenodo.1214456

  24. [32]

    J. N. Kather and J. Krisam and P. Charoentong and others. Predicting Survival from Colorectal Cancer Histology Slides Using Deep Learning: A Retrospective Multicenter Study. PLoS Medicine. 2019

  25. [33]

    Van Horn and O

    G. Van Horn and O. Mac Aodha and Y. Song and Y. Cui and C. Sun and A. Shepard and H. Adam and P. Perona and S. Belongie. The iNaturalist Species Classification and Detection Dataset. CVPR. 2018

  26. [34]

    Xiao and J

    J. Xiao and J. Hays and K. A. Ehinger and A. Oliva and A. Torralba. SUN Database: Large-Scale Scene Recognition from Abbey to Zoo. CVPR. 2010

  27. [35]

    Zhou and A

    B. Zhou and A. Lapedriza and A. Khosla and A. Oliva and A. Torralba. Places: A 10 Million Image Database for Scene Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2018

  28. [36]

    Cimpoi and S

    M. Cimpoi and S. Maji and I. Kokkinos and S. Mohamed and A. Vedaldi. Describing Textures in the Wild. CVPR. 2014

  29. [37]

    Huang and Y

    R. Huang and Y. Li. MOS : Towards Scaling Out-of-Distribution Detection for Large Semantic Space. CVPR. 2021

  30. [38]

    Naiknaware and S

    A. Naiknaware and S. Sekeh. T-QPM: Enabling Temporal Out-Of-Distribution Detection and Domain Generalization for Vision-Language Models in Open-World. arXiv preprint arXiv:2603.18481. 2026

  31. [39]

    Deng and W

    J. Deng and W. Dong and R. Socher and L.-J. Li and K. Li and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. CVPR. 2009

  32. [40]

    Lou and J

    S. Lou and J. Ji and H. Li and X. Zhang and Y. Jiang and M. Hua and K. Chen and K. Ge and Q. Zhang and L. Wang and P. Han and L. Cao. A Large Histological Images Dataset of Gastric Cancer with Tumour Microenvironment Annotation for AI. Scientific Data. 2025

  33. [41]

    J. N. Kather and J. Krisam and P. Charoentong and others. Predicting Survival from Colorectal Cancer Histology Slides Using Deep Learning: A Retrospective Multicenter Study. PLOS Medicine. 2019

  34. [42]

    ImageNet100

    ambityga. ImageNet100. 2026

  35. [43]

    NCT-CRC-HE-100K

    imrankhan77. NCT-CRC-HE-100K. 2026

  36. [44]

    CRC-VAL-HE-7K

    merveatasoy1. CRC-VAL-HE-7K. 2026

  37. [45]

    Gastric Cancer Histopathology Tissue Image Dataset

    orvile. Gastric Cancer Histopathology Tissue Image Dataset. 2026

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.