Pith. sign in

REVIEW 3 major objections 1 cited by

Strong models must emit reasoning traces that weaker models can actually follow and check.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 21:26 UTC pith:OW6ZSSV4

load-bearing objection We do not have the legibility paper—only its abstract plus an unrelated camera-trap manuscript—so there is nothing solid to evaluate yet. the 3 major comments →

arxiv 2603.20508 v2 pith:OW6ZSSV4 submitted 2026-03-20 cs.MA cs.AIcs.CL

Measuring Weak-to-Strong Legibility of Reasoning Models

Dani Roytburg , Shreya Sridhar , Daphne Ippolito This is my paper
classification cs.MA cs.AIcs.CL
keywords weak-to-strong legibilityreasoning language modelschains of thoughtmulti-agent monitoringsafety oversightdistillationthoroughness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Reasoning language models now sit at the center of multi-agent systems: other models monitor them, distill from them, or cooperate with them. When the partners are weaker, the strong model’s chain-of-thought must be digestible by those weaker agents—what the authors name weak-to-strong legibility. Trustworthiness, and especially cheap safety oversight that relies on weak monitors, depends on this property. The paper’s central claim is that existing “efficiency” metrics only reward short, concise traces and therefore miss the thoroughness weaker monitors actually need.

Core claim

Existing efficiency-based metrics for the legibility of reasoning traces systematically fail to capture thoroughness; they focus on conciseness instead. As a result they are inadequate measures of the weak-to-strong legibility required when strong models must be monitored or distilled by weaker ones.

What carries the argument

Weak-to-strong legibility—the requirement that the shape of a strong model’s decision-making traces be accessible to weaker monitors or student models.

Load-bearing premise

That what weaker monitors mainly need is a particular thorough shape of the strong model’s trace, rather than smaller capability gaps, shared training data, or different monitor designs.

What would settle it

Show that weak models can successfully monitor or distill from strong-model traces even when those traces score poorly on any thoroughness-aware metric, or that ordinary efficiency metrics already predict weak-monitor success rates.

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

If this is right

  • Safety scaffolds that use cheap weak monitors become more reliable once strong models are optimized for thorough, digestible traces rather than short ones.
  • Distillation into smaller models will succeed more often when the teacher’s chain-of-thought is explicitly shaped for the student.
  • New evaluation metrics must score thoroughness of intermediate reasoning, not merely length or token efficiency.
  • Multi-agent systems with mixed capability tiers will need explicit legibility objectives during training of the stronger agents.

Where Pith is reading between the lines

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

  • A practical thoroughness test could check whether a weak model can reconstruct or verify the strong model’s final answer from the trace alone.
  • The size of the capability gap between strong and weak models almost certainly sets how much thoroughness is required; a fixed shape metric may not transfer across gaps.
  • Current training recipes that reward short chains of thought may be actively reducing weak-to-strong legibility.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. The manuscript (as provided in full) presents STREAMTRAP, a large-scale benchmark and unified study of camera-trap species recognition under temporal shift at fixed sites. It argues that the dominant cross-site / domain-generalization framing (e.g., iWildCam) mismatches practitioner needs: maintaining accuracy over chronological streams at a single deployment site as seasons, backgrounds, and species distributions change. The authors construct a streaming protocol over 546 camera traps derived from LILA BC, evaluate biological foundation models (e.g., BioCLIP 2), show that naive adaptation can fall below zero-shot, diagnose class imbalance and temporal shift as main drivers, and report that combining model updates with post-processing narrows but does not close the gap to an oracle upper bound. They also surface open questions about when zero-shot is sufficient and when updates are necessary.

Significance. If the empirical findings hold under the stated protocol, the work is a useful reorientation of camera-trap CV toward site-level temporal reliability rather than static cross-domain leaderboards. Strengths include the scale of the processed benchmark (546 traps, multi-continent sources), an explicit streaming evaluation that mirrors deployment lifecycles, a FAIR-oriented data pipeline, and concrete practitioner-facing guidance plus open questions for algorithm developers. The joint-training-on-history baseline correctly isolates temporal shift from artificial storage constraints common in continual learning. These contributions are of clear interest to ecological ML and continual learning communities, provided the manuscript under review is the STREAMTRAP paper rather than the mismatched abstract title.

major comments (3)
  1. Title/abstract vs. full text mismatch: the submission header and abstract describe arXiv:2603.20508 on 'weak-to-strong legibility of reasoning models' (cs.MA), while the entire full manuscript is the unrelated STREAMTRAP camera-trap paper (arXiv:2603.20509, cs.CV). No definitions, metrics, experiments, or results for weak-to-strong legibility appear. This is a load-bearing integrity issue: the central claim of the stated paper cannot be evaluated from the supplied text.
  2. Assuming the intended manuscript is STREAMTRAP: Sec. 3 and Fig. 1b define streaming evaluation with joint training on all past data as the update recipe, yet the abstract and findings claim 'naive adaptation can even degrade below zero-shot.' The manuscript needs a clearer separation of which update methods (fine-tuning recipe, PEFT, loss, post-processing) produce degradation versus improvement, with per-interval tables so the claim is falsifiable rather than aggregated.
  3. Finding (3) attributes difficulty to severe class imbalance and temporal shift, but the main text (as provided) does not report quantitative shift measures (e.g., species-distribution TV distance or background feature drift between consecutive intervals) correlated with accuracy drops. Without those measurements or ablations that isolate imbalance vs. shift, the causal diagnosis remains under-supported relative to the claim.

Circularity Check

0 steps flagged

No significant circularity: the supplied full manuscript is an empirical camera-trap benchmark study with no derivation that reduces a claimed prediction to its inputs by construction.

full rationale

The cacheable full text is STREAMTRAP (camera-trap species recognition over time; arXiv:2603.20509), not a first-principles or fitted-parameter derivation paper. Its load-bearing content is (1) construction of a chronological streaming benchmark from LILA BC, (2) zero-shot and adaptation experiments under that protocol, and (3) empirical findings on imbalance and temporal shift. There are no equations in which a fitted scale, uniqueness theorem, or ansatz is renamed as an independent prediction. Self-citations (e.g., authors’ prior PEFT / fine-tuning / continual-learning work) appear as background method choices, not as load-bearing uniqueness results that force the central claims. The abstract/title metadata about weak-to-strong legibility (2603.20508) does not match the manuscript body, so no legibility metric or thoroughness claim can be checked for definitional circularity either. Honest non-finding: score 0; steps empty.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 1 invented entities

Abstract-only review of 2603.20508. No free parameters, formal axioms, or invented physical entities appear. The main load-bearing conceptual moves are definitional: introducing “weak-to-strong legibility” and asserting that efficiency metrics omit thoroughness. These are domain assumptions / paper-local definitions, not fitted constants. The mismatched full manuscript (camera-trap STREAMTRAP) is not used as evidence for this paper’s claims.

axioms (3)
  • domain assumption Strong models’ intermediate chains of thought can be made (or fail to be) digestible by weaker models in multi-agent monitoring/distillation settings.
    Stated as the premise for defining weak-to-strong legibility in the abstract; no formal proof or measurement given in available text.
  • ad hoc to paper Existing efficiency-based legibility metrics focus on conciseness and fail to capture thoroughness needed by weaker monitors.
    Central critique in the abstract; treated as given without cited metric definitions or empirical comparison in the provided material.
  • domain assumption Adoption of weak monitors is a realistic and desirable reliability scaffold for safety oversight under budget constraints.
    Abstract motivates the work with this safety-practice claim; it is a field assumption, not derived here.
invented entities (1)
  • weak-to-strong legibility no independent evidence
    purpose: Name the property that strong models’ decision traces remain accessible/usable to weaker monitors or student models.
    Paper-local term introduced in the abstract; independent evidence would require a falsifiable metric and experiments not present in the available text.

pith-pipeline@v1.1.0-grok45 · 22024 in / 2580 out tokens · 34757 ms · 2026-07-13T21:26:02.606832+00:00 · methodology

0 comments
read the original abstract

Reasoning language models (RLMs) and the intermediate chains of thought they emit play an increasingly central role in multi-agent setups such as inter-model monitoring or distillation into smaller models. When agents at different capability tiers must cooperate, strong models need to produce traces digestible by weaker ones. We refer to this goal as "weak-to-strong legibility". Trustworthiness of large models depends in part on this legibility property. For safety oversight in particular, adoption of weak monitors may become a standard for reliability scaffolds on a healthy budget. Legibility requires that the shape of these decision-making traces takes some form accessible to weaker monitors. Existing efficiency-based metrics for legibility fail to capture "thoroughness", instead focusing on conciseness.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. CLORE: Content-Level Optimization for Reasoning Efficiency

    cs.AI 2026-05 unverdicted novelty 6.0

    CLORE augments correct on-policy rollouts by deleting repetitive and irrelevant segments then optimizes with auxiliary DPO to improve accuracy-efficiency trade-off on math benchmarks.

Reference graph

Works this paper leans on

62 extracted references · 7 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Lila bc: Labeled information library of alexandria: Biology and conservation.https://lila.science/. 2, 4

  2. [2]

    Gradient based sample selection for online continual learning.Advances in neural information processing sys- tems, 32, 2019

    Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Ben- gio. Gradient based sample selection for online continual learning.Advances in neural information processing sys- tems, 32, 2019. 1

    2019

  3. [3]

    Monitoring the mammalian fauna of ur- ban areas using remote cameras and citizen science.Journal of Urban Ecology, 4(1):juy002, 2018

    Victor Anton, Stephen Hartley, Andre Geldenhuis, and Heiko U Wittmer. Monitoring the mammalian fauna of ur- ban areas using remote cameras and citizen science.Journal of Urban Ecology, 4(1):juy002, 2018. 8

    2018

  4. [4]

    The MegaDetector: Large-scale deployment of computer vision for conservation and biodiversity monitor- ing

    Sara Beery. The MegaDetector: Large-scale deployment of computer vision for conservation and biodiversity monitor- ing. InAI for Social Impact. Cambridge University Press,

  5. [5]

    Recognition in terra incognita

    Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. InProceedings of the European confer- ence on computer vision (ECCV), pages 456–473, 2018. 6, 8

    2018

  6. [6]

    Recognition in terra incognita

    Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. InProceedings of the European confer- ence on computer vision (ECCV), pages 456–473, 2018. 1

    2018

  7. [7]

    The iwildcam 2018 challenge dataset.arXiv preprint arXiv:1904.05986, 2019

    Sara Beery, Grant Van Horn, Oisin Mac Aodha, and Pietro Perona. The iwildcam 2018 challenge dataset.arXiv preprint arXiv:1904.05986, 2019. 3, 1

    Pith/arXiv arXiv 2018

  8. [8]

    The iwildcam 2021 competition dataset.arXiv preprint arXiv:2105.03494, 2021

    Sara Beery, Arushi Agarwal, Elijah Cole, and Vighnesh Birodkar. The iwildcam 2021 competition dataset.arXiv preprint arXiv:2105.03494, 2021. 2, 3, 1

    Pith/arXiv arXiv 2021

  9. [9]

    Deep learning-based ecological analysis of camera trap images is impacted by training data quality and quantity

    Peggy A Bevan, Omiros Pantazis, Holly Pringle, Guil- herme Braga Ferreira, Daniel J Ingram, Emily Madsen, Liam Thomas, Dol Raj Thanet, Thakur Silwal, Santosh Rayama- jhi, et al. Deep learning-based ecological analysis of camera trap images is impacted by training data quality and quantity. arXiv preprint arXiv:2408.14348, 2024. 3, 2

    arXiv 2024

  10. [10]

    Pelagic Publishing Ltd, 2016

    Luigi Boitani.Camera trapping for wildlife research. Pelagic Publishing Ltd, 2016. 3, 1

    2016

  11. [11]

    Automated wildlife image classification: An active learning tool for ecological applica- tions.Ecological Informatics, 77:102231, 2023

    Ludwig Bothmann, Lisa Wimmer, Omid Charrakh, Tobias Weber, Hendrik Edelhoff, Wibke Peters, Hien Nguyen, Caryl Benjamin, and Annette Menzel. Automated wildlife image classification: An active learning tool for ecological applica- tions.Ecological Informatics, 77:102231, 2023. 1

    2023

  12. [12]

    Class-balanced loss based on effective number of samples, 2019

    Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples, 2019. 7, 2

    2019

  13. [13]

    Class-balanced loss based on effective number of samples

    Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268–9277,

  14. [14]

    A continual learning survey: Defying for- getting in classification tasks.IEEE transactions on pattern analysis and machine intelligence, 44(7):3366–3385, 2021

    Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ale ˇs Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying for- getting in classification tasks.IEEE transactions on pattern analysis and machine intelligence, 44(7):3366–3385, 2021. 1

    2021

  15. [15]

    Multimodal foundation models for zero-shot animal species recognition in camera trap images.arXiv preprint arXiv:2311.01064,

    Zalan Fabian, Zhongqi Miao, Chunyuan Li, Yuanhan Zhang, Ziwei Liu, Andr´es Hern´andez, Andr´es Montes-Rojas, Rafael Escucha, Laura Siabatto, Andr ´es Link, et al. Multimodal foundation models for zero-shot animal species recognition in camera trap images.arXiv preprint arXiv:2311.01064,

    Pith/arXiv arXiv

  16. [16]

    A brief review of domain adaptation

    Abolfazl Farahani, Sahar V oghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: pro- ceedings from ICDATA 2020 and IKE 2020, pages 877–894,

    2020

  17. [17]

    Wildclip: Scene and animal attribute retrieval from camera trap data with domain-adapted vision-language models.International Journal of Computer Vision, 132(9): 3770–3786, 2024

    Valentin Gabeff, Marc Rußwurm, Devis Tuia, and Alexan- der Mathis. Wildclip: Scene and animal attribute retrieval from camera trap data with domain-adapted vision-language models.International Journal of Computer Vision, 132(9): 3770–3786, 2024. 3, 1

    2024

  18. [18]

    Geodesic flow kernel for unsupervised domain adaptation

    Boqing Gong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain adaptation. In2012 IEEE conference on computer vision and pattern recognition, pages 2066–2073. IEEE, 2012. 4

    2066

  19. [19]

    Bioclip 2: Emergent properties from scaling hierarchi- cal contrastive learning.arXiv preprint arXiv:2505.23883,

    Jianyang Gu, Samuel Stevens, Elizabeth G Campolongo, Matthew J Thompson, Net Zhang, Jiaman Wu, Andrei Kopanev, Zheda Mai, Alexander E White, James Balhoff, et al. Bioclip 2: Emergent properties from scaling hierarchi- cal contrastive learning.arXiv preprint arXiv:2505.23883,

    arXiv

  20. [20]

    A baseline for detect- ing misclassified and out-of-distribution examples in neural networks.arXiv preprint arXiv:1610.02136, 2016

    Dan Hendrycks and Kevin Gimpel. A baseline for detect- ing misclassified and out-of-distribution examples in neural networks.arXiv preprint arXiv:1610.02136, 2016. 9

    Pith/arXiv arXiv 2016

  21. [21]

    Parameter-efficient transfer learning for nlp, 2019

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp, 2019. 7, 2

    2019

  22. [22]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022. 7, 2

    2022

  23. [23]

    Idaho camera traps.https://lila.science/datasets/idaho- camera-traps/

    Idaho Department of Fish and Game. Idaho camera traps.https://lila.science/datasets/idaho- camera-traps/. 8

  24. [24]

    Northern and central annamites camera traps 2.0

    IUCN SSC Asian Wild Cattle Specialist Group’s Saola Working Group. Northern and central annamites camera traps 2.0. Dataset, 2021. SWG (2021). 6

    2021

  25. [25]

    Vi- sual prompt tuning, 2022

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi- sual prompt tuning, 2022. 7, 2

    2022

  26. [26]

    Wilds: A benchmark of in-the- wild distribution shifts

    Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubra- mani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the- wild distribution shifts. InInternational conference on machine learning, pages 5637–5664. PMLR, 2021. 2, 3, 1 10

    2021

  27. [27]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6

    2014

  28. [28]

    Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning

    Zheda Mai, Ruiwen Li, Hyunwoo Kim, and Scott San- ner. Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3589–3599,

  29. [29]

    Online continual learning in image classification: An empirical survey.Neurocomputing, 469:28–51, 2022

    Zheda Mai, Ruiwen Li, Jihwan Jeong, David Quispe, Hyun- woo Kim, and Scott Sanner. Online continual learning in image classification: An empirical survey.Neurocomputing, 469:28–51, 2022. 3, 4, 1

    2022

  30. [30]

    Fine-tuning is fine, if cali- brated.Advances in Neural Information Processing Systems, 37:136084–136119, 2024

    Zheda Mai, Arpita Chowdhury, Ping Zhang, Cheng-Hao Tu, Hong-You Chen, Vardaan Pahuja, Tanya Berger-Wolf, Song Gao, Charles Stewart, Yu Su, et al. Fine-tuning is fine, if cali- brated.Advances in Neural Information Processing Systems, 37:136084–136119, 2024. 2, 3, 7, 1

    2024

  31. [31]

    Lessons and insights from a unifying study of parameter-efficient fine-tuning (peft) in visual recognition

    Zheda Mai, Ping Zhang, Cheng-Hao Tu, Hong-You Chen, Quang-Huy Nguyen, Li Zhang, and Wei-Lun Chao. Lessons and insights from a unifying study of parameter-efficient fine-tuning (peft) in visual recognition. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 14845–14857, 2025. 2, 7

    2025

  32. [32]

    Two-phase training mitigates class imbalance for camera trap image classifica- tion with cnns.arXiv preprint arXiv:2112.14491, 2021

    Farjad Malik, Simon Wouters, Ruben Cartuyvels, Erfan Ghadery, and Marie-Francine Moens. Two-phase training mitigates class imbalance for camera trap image classifica- tion with cnns.arXiv preprint arXiv:2112.14491, 2021. 3, 2

    Pith/arXiv arXiv 2021

  33. [33]

    Trail camera images of new zealand animals.https://lila.science/datasets/nz- trailcams

    New Zealand Trailcams. Trail camera images of new zealand animals.https://lila.science/datasets/nz- trailcams. 8

  34. [34]

    Mohammad Sadegh Norouzzadeh, Anh Nguyen, Margaret Kosmala, Alexandra Swanson, Meredith S Palmer, Craig Packer, and Jeff Clune. Automatically identifying, count- ing, and describing wild animals in camera-trap images with deep learning.Proceedings of the National Academy of Sci- ences, 115(25):E5716–E5725, 2018. 3, 1

    2018

  35. [35]

    Snap- shot safari: A large-scale collaborative to monitor africa’s remarkable biodiversity.South African Journal of Science, 117(1-2):1–4, 2021

    Lain E Pardo, Sara Bombaci, Sarah E Huebner, Michael J Somers, Herve Fritz, Colleen Downs, Abby Guthmann, Robyn S Hetem, Mark Keith, Aliza le Roux, et al. Snap- shot safari: A large-scale collaborative to monitor africa’s remarkable biodiversity.South African Journal of Science, 117(1-2):1–4, 2021. 8

    2021

  36. [36]

    Har- nessing artificial intelligence to fill global shortfalls in biodi- versity knowledge.Nature Reviews Biodiversity, pages 1–17,

    Laura J Pollock, Justin Kitzes, Sara Beery, Kaitlyn M Gaynor, Marta A Jarzyna, Oisin Mac Aodha, Bernd Meyer, David Rolnick, Graham W Taylor, Devis Tuia, et al. Har- nessing artificial intelligence to fill global shortfalls in biodi- versity knowledge.Nature Reviews Biodiversity, pages 1–17,

  37. [37]

    Learning transferable visual models from natural language supervision, 2021

    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. Learning transferable visual models from natural language supervision, 2021. 2

    2021

  38. [38]

    Balanced meta-softmax for long-tailed visual recog- nition.Advances in neural information processing systems, 33:4175–4186, 2020

    Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recog- nition.Advances in neural information processing systems, 33:4175–4186, 2020. 2, 3

    2020

  39. [39]

    Balanced meta-softmax for long-tailed visual recognition, 2020

    Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition, 2020. 7, 2

    2020

  40. [40]

    A broad review on class imbalance learning techniques.Applied Soft Computing, 143:110415, 2023

    Salim Rezvani and Xizhao Wang. A broad review on class imbalance learning techniques.Applied Soft Computing, 143:110415, 2023. 3, 2

    2023

  41. [41]

    Extend- ing the wilds benchmark for unsupervised adaptation.arXiv preprint arXiv:2112.05090, 2021

    Shiori Sagawa, Pang Wei Koh, Tony Lee, Irena Gao, Sang Michael Xie, Kendrick Shen, Ananya Kumar, Weihua Hu, Michihiro Yasunaga, Henrik Marklund, et al. Extend- ing the wilds benchmark for unsupervised adaptation.arXiv preprint arXiv:2112.05090, 2021. 2, 3

    Pith/arXiv arXiv 2021

  42. [42]

    Catalog: A camera trap language-guided contrastive learning model

    Julian D Santamaria, Claudia Isaza, and Jhony H Giraldo. Catalog: A camera trap language-guided contrastive learning model. In2025 IEEE/CVF Winter Conference on Applica- tions of Computer Vision (WACV), pages 1197–1206. IEEE,

  43. [43]

    Online class- incremental continual learning with adversarial shapley value

    Dongsub Shim, Zheda Mai, Jihwan Jeong, Scott San- ner, Hyunwoo Kim, and Jongseong Jang. Online class- incremental continual learning with adversarial shapley value. InProceedings of the AAAI Conference on Artificial Intelligence, pages 9630–9638, 2021. 1

    2021

  44. [44]

    Domain adaptation: challenges, methods, datasets, and applications.IEEE access, 11:6973–7020,

    Peeyush Singhal, Rahee Walambe, Sheela Ramanna, and Ketan Kotecha. Domain adaptation: challenges, methods, datasets, and applications.IEEE access, 11:6973–7020,

  45. [45]

    Bioclip: A vision foundation model for the tree of life

    Samuel Stevens, Jiaman Wu, Matthew J Thompson, Eliza- beth G Campolongo, Chan Hee Song, David Edward Carlyn, Li Dong, Wasila M Dahdul, Charles Stewart, Tanya Berger- Wolf, et al. Bioclip: A vision foundation model for the tree of life. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 19412–19424,

  46. [46]

    Snapshot serengeti, high-frequency annotated camera trap images of 40 mammalian species in an african savanna.Scientific data, 2(1):1–14, 2015

    Alexandra Swanson, Margaret Kosmala, Chris Lintott, Robert Simpson, Arfon Smith, and Craig Packer. Snapshot serengeti, high-frequency annotated camera trap images of 40 mammalian species in an african savanna.Scientific data, 2(1):1–14, 2015. 1, 6, 8

    2015

  47. [47]

    Machine learning to classify ani- mal species in camera trap images: Applications in ecology

    Michael A Tabak, Mohammad S Norouzzadeh, David W Wolfson, Steven J Sweeney, Kurt C VerCauteren, Nathan P Snow, Joseph M Halseth, Paul A Di Salvo, Jesse S Lewis, Michael D White, et al. Machine learning to classify ani- mal species in camera trap images: Applications in ecology. Methods in Ecology and Evolution, 10(4):585–590, 2019. 6, 8

    2019

  48. [48]

    Use of camera traps for wildlife studies: a review.Biotechnologie, Agronomie, Soci ´et´e et En- vironnement, 18(3), 2014

    Franck Trolliet, C ´edric Vermeulen, Marie-Claude Huynen, and Alain Hambuckers. Use of camera traps for wildlife studies: a review.Biotechnologie, Agronomie, Soci ´et´e et En- vironnement, 18(3), 2014. 3, 1

    2014

  49. [49]

    Holistic trans- fer: towards non-disruptive fine-tuning with partial target data.Advances in Neural Information Processing Systems, 36:29149–29173, 2023

    Cheng-Hao Tu, Hong-You Chen, Zheda Mai, Jike Zhong, Vardaan Pahuja, Tanya Berger-Wolf, Song Gao, Charles Stewart, Yu Su, and Wei-Lun Harry Chao. Holistic trans- fer: towards non-disruptive fine-tuning with partial target data.Advances in Neural Information Processing Systems, 36:29149–29173, 2023. 2, 3, 7 11

    2023

  50. [50]

    Perspectives in machine learning for wildlife conservation.Nature communications, 13(1):792,

    Devis Tuia, Benjamin Kellenberger, Sara Beery, Blair R Costelloe, Silvia Zuffi, Benjamin Risse, Alexander Mathis, Mackenzie W Mathis, Frank Van Langevelde, Tilo Burghardt, et al. Perspectives in machine learning for wildlife conservation.Nature communications, 13(1):792,

  51. [51]

    The inaturalist species classification and de- tection dataset

    Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and de- tection dataset. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 8769–8778,

  52. [52]

    Reliable and efficient integration of ai into camera traps for smart wildlife monitoring based on continual learning.Eco- logical Informatics, 83:102815, 2024

    Delia Velasco-Montero, Jorge Fern ´andez-Berni, Ricardo Carmona-Gal´an, Ariadna Sanglas, and Francisco Palomares. Reliable and efficient integration of ai into camera traps for smart wildlife monitoring based on continual learning.Eco- logical Informatics, 83:102815, 2024. 1

    2024

  53. [53]

    An evaluation of platforms for processing camera-trap data using artificial intelligence

    Juliana V ´elez, William McShea, Hila Shamon, Paula J Castiblanco-Camacho, Michael A Tabak, Carl Chalmers, Paul Fergus, and John Fieberg. An evaluation of platforms for processing camera-trap data using artificial intelligence. Methods in Ecology and Evolution, 14(2):459–477, 2023. 8

    2023

  54. [54]

    Robust fine-tuning of zero-shot models

    Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gon- tijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 7959–7971, 2022. 2, 8

    2022

  55. [55]

    Generalized out-of-distribution detection: A survey.Inter- national Journal of Computer Vision, 132(12):5635–5662,

    Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey.Inter- national Journal of Computer Vision, 132(12):5635–5662,

  56. [56]

    Identifying and compensating for feature deviation in imbalanced deep learning.arXiv preprint arXiv:2001.01385,

    Han-Jia Ye, Hong-You Chen, De-Chuan Zhan, and Wei-Lun Chao. Identifying and compensating for feature deviation in imbalanced deep learning.arXiv preprint arXiv:2001.01385,

    Pith/arXiv arXiv 2001

  57. [57]

    Pro- crustean training for imbalanced deep learning

    Han-Jia Ye, De-Chuan Zhan, and Wei-Lun Chao. Pro- crustean training for imbalanced deep learning. InProceed- ings of the IEEE/CVF international conference on computer vision, pages 92–102, 2021. 7

    2021

  58. [58]

    Identifying and compensating for feature deviation in imbalanced deep learning, 2022

    Han-Jia Ye, Hong-You Chen, De-Chuan Zhan, and Wei-Lun Chao. Identifying and compensating for feature deviation in imbalanced deep learning, 2022. 7, 2

    2022

  59. [59]

    Automated identification of animal species in camera trap images

    Xiaoyuan Yu, Jiangping Wang, Roland Kays, Patrick A Jansen, Tianjiang Wang, and Thomas Huang. Automated identification of animal species in camera trap images. EURASIP Journal on Image and Video Processing, 2013(1): 52, 2013. 3, 1

    2013

  60. [60]

    Deep long-tailed learning: A survey.IEEE transactions on pattern analysis and machine intelligence, 45(9):10795–10816, 2023

    Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey.IEEE transactions on pattern analysis and machine intelligence, 45(9):10795–10816, 2023. 3, 7, 2

    2023

  61. [61]

    Domain generalization: A survey.IEEE transactions on pattern analysis and machine intelligence, 45(4):4396–4415, 2022

    Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey.IEEE transactions on pattern analysis and machine intelligence, 45(4):4396–4415, 2022. 2, 3

    2022

  62. [62]

    Class incremental learning for wildlife biodiversity monitoring in camera trap images.Ecological Informatics, 71:101760, 2022

    Haowei Zhu, Ye Tian, and Junguo Zhang. Class incremental learning for wildlife biodiversity monitoring in camera trap images.Ecological Informatics, 71:101760, 2022. 1 12 Lessons and Open Questions from a Unified Study of Camera-Trap Species Recognition Over Time Supplementary Material In this supplementary material, we provide more details and experiment...

    2022