TellTale: Blending Multi-Instance LoRA Text Encoders and a Zero-Shot LLM Judge for Ambivalence/Hesitancy Recognition in Videos

Abdel-Karim Al-Tamimi; Ali Rodan

REVIEW 3 major objections 3 minor 15 references

TellTale: Blending Multi-Instance LoRA Text Encoders and a Zero-Shot LLM Judge for Ambivalence/Hesitancy Recognition in Videos

T0 review · 3 major / 3 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A text-only system can spot ambivalence and hesitancy in interview videos, beating the official vision-based baseline by a wide margin.

desk verdict Solid, honestly reported text-only result on BAH; the absolute Macro-F1 is credible, but the paper's central blending claim is undercut by its own noise estimate and should be softened or better supported. read the letter →

arxiv 2607.16635 v1 pith:BM5GCV4A submitted 2026-07-18 cs.CV cs.AI

Abdel-Karim Al-Tamimi , Ali Rodan This is my paper · ORCID

classification cs.CVcs.AI

keywords ambivalence/hesitancyrecognitionmultiple-instancelearningLoRAfine-tuningzero-shotLLMjudgetranscript-onlyvideoanalysissmooth-maxpoolingparticipant-groupedcross-validationaffectivecomputing

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

TellTale aims to show that ambivalence/hesitancy (A/H) in interview videos can be recognized from the transcript alone, without using video, audio, or facial cues. It combines three text-based probability streams: two fine-tuned language encoders trained with a multiple-instance objective, and one training-free large language model judge that scores hesitancy zero-shot. The key idea is that a video-level label should be driven by the single most A/H-like chunk of speech, not by averaging over all chunks. If this works, A/H detection becomes cheap, reproducible, and useful for real-time behavioral-change applications such as digital health coaching.

What carries the argument

The central mechanism is the smooth maximum (LogSumExp) pooling over transcript chunks: z_video = (1/r) log( (1/T) Σ exp(r z_t) ), with r=50, which behaves like max pooling while remaining differentiable, letting only the most A/H-like chunk drive the video-level score. This is combined with a weighted blend of three streams — two MIL-fine-tuned text encoders and a zero-shot LLM judge — where weights and threshold are chosen by grid search on participant-grouped out-of-fold predictions, using a flat-region midpoint rule for the threshold to ensure robustness.

What would settle it

Refit the blend weights and threshold on leave-one-participant-group-out folds while varying the positive-class prevalence of the held-out group, then measure Macro-F1; if the flat-region threshold shifts materially with prevalence or the blend stops beating its best single stream, the reported gain is an artifact of threshold selection on a particular distribution.

Watch

Extended reading notes

Core claim

TellTale claims that a weighted blend of three transcript-only streams achieves a Macro-F1 of 0.7364 and an average precision of 0.7940 on the private test set of 152 videos from unseen participants, compared with 0.2827 Macro-F1 for the official vision-based baseline. The paper attributes this success to two mechanisms: smooth-max multiple-instance pooling, which lets the most indicative transcript chunk determine the video-level prediction, and a diversity-aware blend of two LoRA-fine-tuned text encoders with a zero-shot instruction-tuned LLM judge. The blend weights (0.45, 0.30, 0.25) and the decision threshold (0.53) are selected on participant-grouped cross-validated predictions, and th

Load-bearing premise

The load-bearing premise is that the out-of-fold predictions used to select the blend weights and the 0.53 threshold have the same probability scale and positive-class prevalence as the private test set; if the test distribution shifts, the fixed combination may no longer be optimal.

Editorial extensions

If this is right

If the central claim holds, transcript-only models can outperform vision-based models for A/H recognition, suggesting that linguistic hedging and self-correction are the primary carriers of the signal in this interview setting.
The smooth-max multiple-instance objective alone provides a gain of about +0.066 OOF Macro-F1 over averaging chunk predictions with an identical encoder, implying that weakly-supervised localization of the crucial utterance is valuable.
A zero-shot LLM judge reaches an OOF Macro-F1 of 0.6970 without any training, close to the fully fine-tuned encoder stream, indicating that instruction-tuned LLMs can serve as effective, label-free raters for behavioral constructs.
The blend improves over its best single component even under a conservative leave-one-fold-out estimate (0.7213 vs 0.7147), supporting the idea that diverse error patterns across streams cancel out in a weighted average.
The evaluation protocol — participant-grouped cross-validation with early stopping on a separate grouped split — predicts private-test performance within ±0.015 Macro-F1, which may generalize to other video-level classification tasks with sparse labels.

Reading between the lines

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

Beyond the paper: the text-only result suggests that in settings where transcripts are unavailable or noisy, visual and acoustic cues might still carry some A/H signal, but exploiting them would likely require temporal models and larger datasets than those used here.
Beyond the paper: the flat-region threshold selection rule could be a reusable recipe for other low-prevalence classification tasks, where small threshold changes swing Macro-F1 drastically.
Beyond the paper: the same three-stream recipe — MIL-fine-tuned encoders plus a zero-shot LLM judge with a diversity-optimized blend — could transfer to other weakly-labeled video-language tasks where only video-level labels exist and the evidence is localized, such as deception detection or stance ambivalence.
Beyond the paper: a testable extension is to apply the same blend with the fixed 0.53 threshold to new participant groups with different A/H prevalence; if the threshold is not re-tuned and performance drops below the best single stream, the claimed advantage would depend on calibration match.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Worth reading mainly because the absolute result is a real measurement: text-only transcripts score 0.7364 Macro-F1 on the private test set versus 0.2827 for the official vision baseline. That is a useful benchmark for the BAH task and a practical demonstration that LoRA fine-tuning plus a zero-shot LLM judge can carry the signal. The authors also deserve credit for the evaluation discipline: participant-grouped OOF predictions, a conservative leave-one-fold-out estimate, a flat-region threshold rule, and an explicit statement of the noise level. Those choices make the headline number believable.

The soft spot is exactly where the stress-test note lands. The paper's advertising claim is that blending three streams is what makes TellTale work. On the private test set the full blend beats stream A alone by 0.0199 Macro-F1 and beats A+B by 0.0129. The paper itself says differences of a few thousandths are evaluation noise at n=152, but then treats these gaps as meaningful. Given their stated noise of ±0.015, a gap of 0.0129 is not clearly outside noise, and there are no confidence intervals or significance tests. The conservative OOF comparison is even thinner: full blend vs stream A alone is 0.7213 vs 0.7147, and A+B is actually slightly below stream A alone. So the absolute number is solid, but the causal claim that weighted averaging of diverse errors is what produces the win is not established by the reported data. That needs to be either backed with error bars or reframed as a plausible but unconfirmed effect.

The secondary weaknesses are real but smaller. The claim that visual and acoustic features add no signal is asserted without showing any modality ablation. Since the text-only nature is part of the pitch, that table should be included. The OOF-fitted weights and threshold are acknowledged as mildly optimistic, and the leave-one-fold-out estimate is a reasonable mitigation; I do not think that is a fatal flaw.

This paper deserves a serious referee. It reports a new result on a public challenge, with reproducible components and unusually transparent selection details. I would send it to review and ask for a significance statement on the blend margin and for the missing multimodal ablation. After that, it would be a useful addition to the affective computing literature.

Referee Report

3 major / 3 minor

Summary. The paper presents TellTale, a text-only system for ambivalence/hesitancy (A/H) recognition in interview videos. It combines three transcript-based probability streams: two LoRA-fine-tuned text encoders (multilingual-e5-large and mDeBERTa-v3-base) trained with a smooth-max multiple-instance objective, and a zero-shot quantized Qwen3:14B LLM judge. The three probabilities are combined by a weighted average with weights (0.45, 0.30, 0.25) and a decision threshold of 0.53, both selected on participant-grouped out-of-fold predictions. On the organizer-scored private test set of 152 videos, TellTale reports Macro-F1 0.7364 and AP 0.7940, versus 0.2827 Macro-F1 for the official vision-based baseline. The paper includes component ablations and a leave-one-fold-out conservative estimate to address the optimism of selecting weights on the pooled OOF data.

Significance. If the result holds, TellTale demonstrates that a text-only, parameter-efficient pipeline can substantially outperform the official vision-based baseline on this A/H recognition task. The evaluation protocol is a strength: participant-grouped cross-validation, out-of-fold predictions, an external private test set, detailed hyperparameters, and explicit acknowledgment of selection optimism. The proposed combination of MIL-fine-tuned encoders and a zero-shot LLM judge is interesting and reproducible. However, the core methodological claim that the three-stream blend is what drives the improvement is only weakly supported: the private-test gains over components are within the paper's own stated noise level, the conservative OOF gaps are much smaller, and the secondary AP metric actually favors a component. This weakens the central narrative and requires additional statistical evidence or a more cautious interpretation.

major comments (3)

[Section 3, Table 2] The full blend's private-test Macro-F1 advantage over Stream A alone (0.7364 vs 0.7165, Δ=0.0199) and over Streams A+B (Δ=0.0129) is comparable to the paper's own stated ±0.015 noise level quoted in Section 2.2 and Section 3. No confidence intervals, bootstrap estimates, or paired significance tests are provided. The conservative OOF estimates in Table 2 show a much smaller gap (0.7213 vs 0.7147, Δ=0.0066) and place Streams A+B at 0.7141, below Stream A alone (0.7147). Thus the ordering is not robust, and the claims in Sections 2.5 and 4 that the blend improves over its best component are not yet supported. Please add uncertainty quantification (e.g., bootstrap CIs for Macro-F1 and AP, or McNemar-style tests on the decisions) or explicitly soften the blending claim.
[Section 2.5] The description of the 'conservative' leave-one-fold-out estimate is underspecified. It is not clear whether, for each held-out fold, the entire selection procedure (the weight grid search and the flat-region threshold rule) is refit using only the other folds' OOF predictions, or whether only the weights are refit with the threshold fixed at 0.53. This distinction matters: the final system uses a fixed configuration (0.45, 0.30, 0.25, 0.53), whereas the 0.7213 figure appears to estimate the expected performance of the selection procedure. Please specify the exact nested protocol and report the per-fold performance and variance.
[Section 3, Table 2] On the secondary AP metric, the full blend scores 0.7940, which is lower than Stream A alone (0.8119) and Streams A+B (0.8137). The paper does not discuss this discrepancy. Since AP is reported as a secondary metric and the abstract highlights it, the claim that blending makes errors cancel is not uniformly supported. Please address this, or explicitly justify why Macro-F1 is the primary lens for the blending conclusion.

minor comments (3)

[Equation (2)] The text says 'at r=1 it is close to an average,' but LogSumExp at r=1 is log(1/T Σ exp(z_t)), which equals the arithmetic mean only when all z_t are equal. The limiting behavior as r→0 is what approaches the mean; please clarify.
[Table 2 footnote] The table footnote is ambiguous about which rows use the pooled OOF estimate and which use the leave-one-fold-out estimate. Stream A alone appears to be pooled, while blends are labeled conservative. Please state this explicitly for each row so the comparison is not misleading.
[References] Reference [4] has 'Video-LLaV A' which should be 'Video-LLaVA'. Also, the official baseline description would benefit from stating that it is zero-shot vision-only, since that context is important for interpreting the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central result is an organizer-scored external private-test measurement; blend weights and threshold are model-selection choices on development OOF predictions, not renamed predictions.

full rationale

The paper's central claim is the private-test Macro-F1 of 0.7364 and AP 0.7940, evaluated by challenge organizers on 152 videos from 30 unseen participants. This is an external, held-out measurement: the models and the LLM judge produce probabilities on those videos, and the blend weights (0.45, 0.30, 0.25) and threshold (0.53) were selected on participant-grouped out-of-fold predictions from the 1,427 labeled development videos. That is standard model selection, not a renamed prediction. The paper explicitly acknowledges the mild optimism of fitting on the full OOF pool and provides a conservative leave-one-fold-out estimate (0.7213 vs 0.7147 for stream A alone), and even that estimate is a development-side diagnostic rather than the reported private-test result. The LLM judge stream is zero-shot and training-free, so it cannot leak training labels. The cited references include the BAH dataset, ABAW workshop, LoRA, MIL, and pretrained encoders; none is a load-bearing self-citation by the present authors. The closest thing to a circularity concern is that the threshold is selected to maximize OOF Macro-F1 and then used on the private test, which can be mildly optimistic under distribution shift, but this is a known selection-bias/calibration issue and not a definitional reduction of the reported result to the fitted inputs. No equation in the paper reduces the private-test score to the OOF fit, and no fitted parameter is presented as a prediction of held-out labels. Accordingly, no circular step is identifiable under the stated standards.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a few free parameters (blend weights, threshold, pooling temperature) that are fitted to validation predictions, and on domain assumptions about transcript quality, the MIL pooling hypothesis, and the trustworthiness of the challenge evaluation. No new theoretical entities are introduced.

free parameters (5)

Blend weight w_A (stream A) = 0.45
Selected by exhaustive grid search over OOF predictions (Section 2.5); directly determines the final blend.
Blend weight w_B (stream B) = 0.30
Selected by grid search on OOF predictions; the paper notes stream B is weaker but contributes diversity.
Blend weight w_J (LLM judge) = 0.25
Selected by grid search on OOF predictions; values the judge's independent errors.
Decision threshold = 0.53
Chosen by a flat-region midpoint rule on the OOF Macro-F1 curve (Section 2.5). Affects the reported private-test result.
LogSumExp temperature r = 50
Chosen after an OOF sweep of pooling functions (Section 2.3); controls how close pooling is to a true maximum.

assumptions (4)

domain assumption The provided BAH transcript chunks accurately reflect the spoken content and preserve A/H cues.
The entire system is text-only; if ASR errors delete hedging words or self-corrections, the method would be handicapped. The paper does not report transcript quality metrics.
domain assumption The most A/H-like chunk determines the video-level label (smooth-max pooling assumption).
Equation (2) uses LogSumExp with a high temperature to approximate the maximum. The paper gives empirical support (MIL improves OOF F1 over averaging) but this remains a modeling choice about the nature of A/H in a video.
domain assumption The official challenge evaluation is correct and the private test set is properly held out.
The reported Macro-F1 and AP depend on trustworthy organizer scoring; the paper cannot independently verify this.
domain assumption The zero-shot LLM judge's raw 0-100 score divided by 100 is usable as a probability in a weighted average.
No calibration step is applied to the LLM output; the blend and threshold are fit afterward, but the functional form assumes a linear combination of stream probabilities.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TellTale: Blending Multi-Instance LoRA Text Encoders and a Zero-Shot LLM Judge for Ambivalence/Hesitancy Recognition in Videos." pith.science (2026). https://pith.science/paper/BM5GCV4A

@misc{pith2026260716635,
  author       = {Pith},
  title        = {Pith review of: TellTale: Blending Multi-Instance LoRA Text Encoders and a Zero-Shot LLM Judge for Ambivalence/Hesitancy Recognition in Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BM5GCV4A}},
  note         = {Machine review of arXiv:2607.16635}
}

read the original abstract

We present TellTale, a text-only approach to ambivalence/hesitancy (A/H) recognition in interview videos, evaluated on the BAH dataset as part of the 3rd A/H Video Recognition Challenge (11th ABAW Workshop, ECCV 2026). Although the dataset provides video, audio, facial crops, and transcripts, TellTale relies on the transcript alone and combines three probability streams. Two text encoders, multilingual-e5-large and mDeBERTa-v3-base, are fine-tuned with parameter-efficient LoRA adapters under a multiple-instance learning (MIL) objective, in which transcript chunks are scored individually and pooled with a smooth maximum so that only the video-level label is needed for supervision. The third stream requires no training: a quantized 14B instruction LLM is prompted, zero-shot, to rate each transcript for A/H. The three probabilities are combined by a weighted average and a single decision threshold, both selected on participant-grouped cross-validated predictions. On the organizer-scored private test set of 152 videos from unseen participants, TellTale achieves a Macro-F1 of 0.7364 and an average precision of 0.7940, compared with 0.2827 Macro-F1 for the official vision-based baseline.

Figures

Figures reproduced from arXiv: 2607.16635 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 2 canonical work pages

[1]

BAH dataset for ambiva- lence/hesitancy recognition in videos for behavioural change,

S. Belharbiet al., “BAH dataset for ambiva- lence/hesitancy recognition in videos for behavioural change,”arXiv preprint arXiv:2505.19328, 2025. doi: 10.48550/arXiv.2505.19328. 4

arXiv doi:10.48550/arxiv.2505.19328 2025
[2]

ABAW: Valence-Arousal Estimation, Expression Recognition, Action Unit Detection & Emotional Reaction Intensity Estimation Challenges

D. Kolliaset al., “ABAW: Affective behavior analysis in-the-wild,” inIEEE/CVF Conference on Computer Vi- sion and Pattern Recognition Workshops, 2023. doi: 10.48550/arXiv.2303.01498

work page Pith review arXiv doi:10.48550/arxiv.2303.01498 2023
[3]

From Affect to Complex Behavior: Advancing Multimodal Human-Centered AI at the 10th ABAW Workshop & Competition

D. Kollias, P. Tzirakis, A. Cowen, S. Zafeiriou, I. Kotsia, E. Granger, M. Pedersoli, S. Bacon, D. Madsen, S. Bel- harbi, M. H. Aslam, J. Shao, and G. Hu, “From af- fect to complex behavior: Advancing multimodal human- centered ai at the 10th ABAW workshop & compe- tition,”arXiv preprint arXiv:2605.27451, 2026. doi: 10.48550/arXiv.2605.27451

work page Pith review arXiv doi:10.48550/arxiv.2605.27451 2026
[4]

Video- LLaV A: Learning united visual representation by align- ment before projection,

B. Lin, B. Zhu, Y . Ye, M. Ning, P. Jin, and L. Yuan, “Video- LLaV A: Learning united visual representation by align- ment before projection,”arXiv preprint arXiv:2311.10122,

arXiv
[5]

LoRA: Low-rank adaptation of large language models,

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inInternational Conference on Learning Representations, 2022. doi: 10.48550/arXiv.2106.09685

arXiv doi:10.48550/arxiv.2106.09685 2022
[6]

Solving the multiple instance problem with axis-parallel rectangles,

T. G. Dietterich, R. H. Lathrop, and T. Lozano-P ´erez, “Solving the multiple instance problem with axis-parallel rectangles,”Artificial Intelligence, vol. 89, no. 1-2, pp. 31– 71, 1997. doi: 10.1016/S0004-3702(96)00034-3

doi:10.1016/s0004-3702(96)00034-3 1997
[7]

Attention- based deep multiple instance learning,

M. Ilse, J. Tomczak, and M. Welling, “Attention- based deep multiple instance learning,” inInterna- tional Conference on Machine Learning, 2018. doi: 10.48550/arXiv.1802.04712

arXiv doi:10.48550/arxiv.1802.04712 2018
[8]

Qwen3 technical report,

Qwen Team, “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025. doi: 10.48550/arXiv.2505.09388

arXiv doi:10.48550/arxiv.2505.09388 2025

Show all 15 references

[9]

Robust speech recog- nition via large-scale weak supervision,

A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recog- nition via large-scale weak supervision,” inInterna- tional Conference on Machine Learning, 2023. doi: 10.48550/arXiv.2212.04356

2023 doi arXiv
[10]

Cross-validatory choice and assessment of sta- tistical predictions,

M. Stone, “Cross-validatory choice and assessment of sta- tistical predictions,”Journal of the Royal Statistical So- ciety: Series B, vol. 36, no. 2, pp. 111–133, 1974. doi: 10.1111/j.2517-6161.1974.tb00994.x

1974
[11]

Scikit-learn: Machine learn- ing in Python,

F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V . Dubourg,et al., “Scikit-learn: Machine learn- ing in Python,”Journal of Machine Learning Research, vol. 12, pp. 2825–2830, 2011

2011
[12]

Decoupled weight decay reg- ularization,

I. Loshchilov and F. Hutter, “Decoupled weight decay reg- ularization,” inInternational Conference on Learning Rep- resentations, 2019. doi: 10.48550/arXiv.1711.05101

2019 doi arXiv
[13]

Multilingual e5 text embeddings: A techni- cal report,

L. Wang, N. Yang, X. Huang, L. Yang, R. Majumder, and F. Wei, “Multilingual e5 text embeddings: A techni- cal report,”arXiv preprint arXiv:2402.05672, 2024. doi: 10.48550/arXiv.2402.05672

2024 doi arXiv
[14]

DeBERTaV3: Improv- ing DeBERTa using ELECTRA-style pre-training with gradient-disentangled embedding sharing,

P. He, J. Gao, and W. Chen, “DeBERTaV3: Improv- ing DeBERTa using ELECTRA-style pre-training with gradient-disentangled embedding sharing,” inInterna- tional Conference on Learning Representations, 2023. doi: 10.48550/arXiv.2111.09543. 5

2023 doi arXiv
[2023]

doi: 10.48550/arXiv.2311.10122

doi arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

BAH dataset for ambiva- lence/hesitancy recognition in videos for behavioural change,

S. Belharbiet al., “BAH dataset for ambiva- lence/hesitancy recognition in videos for behavioural change,”arXiv preprint arXiv:2505.19328, 2025. doi: 10.48550/arXiv.2505.19328. 4

arXiv doi:10.48550/arxiv.2505.19328 2025

[2] [2]

ABAW: Valence-Arousal Estimation, Expression Recognition, Action Unit Detection & Emotional Reaction Intensity Estimation Challenges

D. Kolliaset al., “ABAW: Affective behavior analysis in-the-wild,” inIEEE/CVF Conference on Computer Vi- sion and Pattern Recognition Workshops, 2023. doi: 10.48550/arXiv.2303.01498

work page Pith review arXiv doi:10.48550/arxiv.2303.01498 2023

[3] [3]

From Affect to Complex Behavior: Advancing Multimodal Human-Centered AI at the 10th ABAW Workshop & Competition

D. Kollias, P. Tzirakis, A. Cowen, S. Zafeiriou, I. Kotsia, E. Granger, M. Pedersoli, S. Bacon, D. Madsen, S. Bel- harbi, M. H. Aslam, J. Shao, and G. Hu, “From af- fect to complex behavior: Advancing multimodal human- centered ai at the 10th ABAW workshop & compe- tition,”arXiv preprint arXiv:2605.27451, 2026. doi: 10.48550/arXiv.2605.27451

work page Pith review arXiv doi:10.48550/arxiv.2605.27451 2026

[4] [4]

Video- LLaV A: Learning united visual representation by align- ment before projection,

B. Lin, B. Zhu, Y . Ye, M. Ning, P. Jin, and L. Yuan, “Video- LLaV A: Learning united visual representation by align- ment before projection,”arXiv preprint arXiv:2311.10122,

arXiv

[5] [5]

LoRA: Low-rank adaptation of large language models,

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inInternational Conference on Learning Representations, 2022. doi: 10.48550/arXiv.2106.09685

arXiv doi:10.48550/arxiv.2106.09685 2022

[6] [6]

Solving the multiple instance problem with axis-parallel rectangles,

T. G. Dietterich, R. H. Lathrop, and T. Lozano-P ´erez, “Solving the multiple instance problem with axis-parallel rectangles,”Artificial Intelligence, vol. 89, no. 1-2, pp. 31– 71, 1997. doi: 10.1016/S0004-3702(96)00034-3

doi:10.1016/s0004-3702(96)00034-3 1997

[7] [7]

Attention- based deep multiple instance learning,

M. Ilse, J. Tomczak, and M. Welling, “Attention- based deep multiple instance learning,” inInterna- tional Conference on Machine Learning, 2018. doi: 10.48550/arXiv.1802.04712

arXiv doi:10.48550/arxiv.1802.04712 2018

[8] [8]

Qwen3 technical report,

Qwen Team, “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025. doi: 10.48550/arXiv.2505.09388

arXiv doi:10.48550/arxiv.2505.09388 2025