TecoPrompt: Temporal-Conservative Prompt Learning for Vision-Language Models
Zeyi Shao, Haowen Hua, Jiaxin Zhang, John See, Zeyd Boukhers, Cong Yang
cs.CV
Quick verdict
Temporal-gated OT label rewriting yields strong high-noise gains for CLIP prompt learning, though the theory is assumption-heavy and key thresholds unreported.
Summary
TecoPrompt is a method paper on robust prompt learning for CLIP-style vision-language models under noisy few-shot labels. It couples entropic optimal transport (Sinkhorn) pseudo-labeling in the CLIP embedding space with a temporal verification gate: a noisy label is hard-rewritten only if the OT pseudo-label stays unchanged for K consecutive epochs (K=8 default) and an EMA-smoothed confidence remains above a threshold, after a warm-up period. Rewritten labels feed back into training via a tri-group objective — cross-entropy, generalized cross-entropy, and MAE applied to clean, mid, and noisy subsets partitioned by confidence thresholds. Appendix theorems give assumption-based lower bounds showing rewrite precision increases with K. Experiments on seven 16-shot benchmarks with 12.5–75% symmetric/asymmetric noise, Food101N, and SUN397 show consistent gains over CoOp, GCE, JoAPR, and NLPrompt, largest at high noise (Flowers102 75% asymmetric: 66.6% vs 55.3%; OxfordPets 50% asymmetric: 84.3% vs 77.5%), and the framework transfers to MaPLe and PromptSRC. Limitations: gains shrink or invert at low noise (trails GCE/NLPrompt on DTD and Caltech101 at 12.5% symmetric), the theory rests on stipulated stability-separation assumptions, several key thresholds are never reported numerically, no code release is mentioned, and only 3-seed means without variance are given.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 4
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Temporal stability of OT pseudo-labels over a K-epoch window, gated by EMA confidence, is a high-precision criterion for conservative hard rewriting of noisy labels in few-shot CLIP prompt learning (94.87% rewrite precision on OxfordPets at 50% noise).
- The tri-group objective (CE/GCE/MAE on clean/mid/noisy subsets) is necessary for robustness; an improper loss assignment collapses at 70% noise (Tab. 4 row e: 58.62%).
- TecoPrompt outperforms CoOp, GCE, JoAPR, and NLPrompt on seven 16-shot benchmarks under 12.5-75% symmetric/asymmetric noise and on Food101N (78.67% vs 76.46%), with the largest margins at high noise.
- Under the paper's assumptions, rewrite precision increases monotonically with window size K and converges to 1-eps* (Appendix B, Theorems 1-3).
- The framework is plugin-compatible with other prompt learners, improving MaPLe and PromptSRC on UCF101 at every noise rate (Tab. 5).
Strengths
Falsifiable empirical motivation for the core gate
Fig. 1 is a controlled 50%-noise experiment on four datasets showing noisy samples sustain long stable-correct OT pseudo-label runs (including the 20+ epoch bin) while clean samples rarely sustain stable-wrong runs — a direct, checkable premise for the K-epoch rewriting criterion.
Large gains exactly in the high-noise regime that motivates the work
Tab. 1: OxfordPets 75% sym 84.6 vs NLPrompt 70.8 (+13.8pp); Flowers102 75% asym 66.6 vs 55.3 (+11.3pp). Sec. 5.8 quantifies the mechanism: 234 rewrites at 94.87% precision covering 75% of the 296 noisy labels on OxfordPets at 50% noise.
Generality beyond the base prompt learner
Candid failure analysis and cost accounting
Ablations cover the main design axes, including a failure mode
Weaknesses
Theory is assumption-laden and near-circular
No dispersion or significance reporting; low-noise losses glossed over
Key hyperparameters unreported; no code release
Baseline provenance and protocol inconsistencies
Internal inconsistency in the main ablation table
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is TecoPrompt and what problem does it solve?
TecoPrompt is a 2026 arXiv method for adapting CLIP-style vision-language models via prompt learning when some of the few-shot training labels are wrong. It uses optimal transport to propose corrected label candidates, rewrites a noisy label only if the candidate has stayed identical for K=8 consecutive epochs with high smoothed confidence, and then trains with different losses (cross-entropy, generalized cross-entropy, MAE) on clean, mid-confidence, and noisy sample groups.
How much does TecoPrompt improve accuracy under label noise?
On OxfordPets with 50% asymmetric noise it reaches 84.3% accuracy versus 77.5% for the strongest prior method (NLPrompt); on Flowers102 with 75% asymmetric noise it reaches 66.6% versus 55.3%. Gains are largest at high noise rates and shrink at low noise (12.5%), where the method can be slightly below simpler baselines such as GCE.
Is TecoPrompt code publicly available?
The arXiv paper does not mention a code release, and several hyperparameters (confidence gate theta, grouping thresholds tau_clean/tau_mid, entropic regularization epsilon, and loss weights) are not reported numerically, so exact reproduction currently requires contacting the authors.
What does the K-epoch stability window do in TecoPrompt?
It is a conservative gate for label correction: a sample's optimal-transport pseudo-label must remain unchanged for K consecutive epochs (default K=8, after a warm-up equal to K epochs), and its EMA-smoothed confidence must stay above a threshold, before the observed noisy label is overwritten. The authors report 94.87% rewrite precision on OxfordPets at 50% noise with K=8.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1, 4, 5, and 6 are available only through captions and surrounding prose; quantitative trends (e.g., few-shot curves in Fig. 4, rewrite-count trajectories in Fig. 5) could not be independently verified from the extracted text.
- Table 4's rows (b) and (c) appear identically configured in the extraction (both EMA checked, Rewrite dash); this may be a PDF-extraction artifact or a genuine table error — flagged rather than adjudicated.
- Hyperparameters beta and K are inferable from the ablation tables, but theta, tau_clean/tau_mid, eps, phi_0, and loss weights are not stated anywhere in the extracted text; 'not reported' judgments rest on the full embedded text including the appendix.
- Minor PDF-extraction garbling of subscripts and math (e.g., Eq. 8 constraint vectors) was reconstructed from context; no substantive reading appears affected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.