{"id":"77571946-7ba4-44d6-9539-dc4caa805de1","arxiv_id":"2505.18053","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"FDBPL caches teacher soft labels offline and adds positive-negative region prompts, achieving 2.2x faster training and modest zero-shot gains over prior distillation-based prompt learning.","lead":"FDBPL speeds up distillation-based prompt learning for vision-language models by pre-computing and storing teacher supervision once, then reusing it during training. It also adds a negative-prompt branch that teaches the model to reject uninformative image regions, which may improve zero-shot recognition at lower training cost.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 2.2x training speedup may be an artifact of excluding offline RIL construction: 500 teacher-crops per image likely dominate total wall-clock vs KDPL's per-epoch inference.","rationale":"The reader's weakest_assumption targets stale cached soft labels as the student prompts drift. That is not the most load-bearing risk: offline teacher soft labels are generated independently of the student's prompt parameters, so they remain valid targets throughout training exactly as in standard knowledge distillation. The paper's actual vulnerability is the accounting of the speedup. FDBPL trades online teacher inference for a massive one-time offline computation (500 crops per image). If the offline phase is excluded from the reported training time, then the comparison to KDPL is apples-to-oranges, and the central contribution—efficiency—is unverified. This concern is concrete, externally checkable, and directly tied to the strongest claim. I therefore disagree with the reader's weakest_assumption, while agreeing that the result should remain conditional until the missing timing breakdown is provided. If the total-time check shows FDBPL is not faster end-to-end, the verdict should be revised downward.","tokens_in":20037,"tokens_out":5890,"duration_ms":58148,"concrete_test":"On ImageNet-1K base-to-new (16 shots/class), measure full wall-clock time from raw data to final prompt model for FDBPL, including offline teacher inference on all M=500 crops, RIL table serialization, and every epoch's I/O, and compare with KDPL's complete training time including its online teacher inference over 50 epochs. If FDBPL total time is not at least 2x lower than KDPL, the ``2.2x faster'' claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central speed claim, ``2.2x faster training,'' is based on per-epoch prompt-training time in Figure 8, but the method's RIL table is constructed offline by running the ViT-L/14 teacher on M=500 random crops per training image (Section 4.1.1). For ImageNet base-to-new with 16 shots/class, this is 16,000 images x 500 crops = 8,000,000 teacher forward passes, performed once. KDPL's online teacher inference over 50 epochs is 16,000 images x 50 = 800,000 teacher forward passes. FDBPL therefore requires 10x more teacher compute for a single student run. If offline RIL construction, storage writing, and subsequent I/O for reading millions of crop files are excluded from the reported ``training time,'' the headline speedup is not end-to-end and may even reverse. The paper also does not disclose how many of the 500 pre-stored crops are consumed per epoch, so the student-side training cost is ambiguous. Until total wall-clock time—offline generation plus student training plus I/O—is reported against KDPL's full training curve, the central efficiency claim is unsupported.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"FDBPL proposes an offline pre-computation and caching strategy for distillation-based prompt learning in CLIP adaptation. The method samples M random crops per training image, stores the teacher's soft outputs (with Top-K sparsification) in a Region Information Lookup (RIL) table, and then trains positive and negative learnable prompts on the student CLIP using entropy-weighted KL divergence, a uniform-distribution rejection loss, and first/second-order difference losses. The paper reports base-to-new, cross-dataset, and robustness evaluations over 11 datasets and claims an average 2.2x training speedup relative to KDPL.","tokens_in":20355,"tokens_out":8670,"duration_ms":74935,"significance":"If fully supported, the RIL idea would be useful: it separates expensive teacher inference from prompt optimization and provides a concrete space-for-time trade-off, while the dual positive-negative prompt design is an interesting way to exploit low-information crops instead of discarding them. The paper contains broad benchmark coverage and ablations of loss components, teacher capacity, and label sparsity. However, the significance is currently limited: the headline speedup does not account for the offline RIL construction, the accuracy advantage over existing strong baselines is small and in the aggregate negative relative to PromptSRC, and several central hyperparameters are missing. I do not see a circularity problem in the proposed losses. The concern that cached teacher targets become stale is also not, by itself, a correctness issue, because the teacher model is frozen and its output for a fixed crop is epoch-independent; the real issue is the cost and I/O of building and reading the cache.","major_comments":[{"comment":"The claimed 2.2x faster training is only for the student prompt-training phase, not for the full procedure. Section 4.1.1 states that M=500 random crops are generated per training image for the RIL table; under the 16-shot ImageNet base-to-new setup this is about 16,000 x 500 = 8,000,000 teacher forward passes on crops, while KDPL's online teacher inference over 50 epochs is about 16,000 x 50 = 800,000 teacher forward passes on images. If Figure 8 excludes RIL construction, storage writing, and crop-reading I/O, the reported speedup is not end-to-end and may even reverse on total wall-clock time. Please report end-to-end timing for both methods, state how many of the 500 stored crops are actually consumed per epoch, and tabulate the per-dataset training times that currently appear only in Figure 8.","section":"Sec. 4.1.1 / Fig. 8"},{"comment":"The abstract's 'superior performance' claim is not supported by the aggregate result: the average harmonic mean in Table 2 is 74.57 for FDBPL but 74.97 for PromptSRC, and the average advantage over KDPL is only 0.58 HM. The paper reports three seeds but no standard deviations or significance tests, so differences below one point cannot be distinguished from noise. Please report mean plus/minus standard deviation over the three seeds for all main tables, and clarify whether the comparison target is KDPL specifically or all listed methods.","section":"Table 2 / Sec. 4.1.2"},{"comment":"Equation (16) as written is degenerate for pairs assigned to the same pseudo-label: in Eq. (14), D^(2)_{c,c'} is the zero vector when c=c', so the normalized direction Dhat^(2) is undefined and the hinge term contributes a constant delta. If the indicator 1[y_i,y_j] is intended to select only pairs of different classes, this must be stated explicitly; otherwise the second-order PCD loss is not well-defined. This issue bears directly on the claimed benefit of the PCD module in Table 6 and Section 5.5.","section":"Eqs. (14)-(16)"},{"comment":"The final objective in Eq. (17) depends on five hyperparameters: alpha in Eq. (13), delta in Eq. (16), and lambda_neg, lambda_diff1, and lambda_diff2. None of these values is given in the experimental setup or in the ablations. Without them the method is not reproducible and the ablation results in Section 5.5 cannot be interpreted quantitatively. Please report the chosen values and, if they vary across the 11 datasets, list them per dataset.","section":"Secs. 3.4-3.5 / Sec. 4.1.2"}],"minor_comments":[{"comment":"The text refers to 'Regional Image Logits (RIL)' while Section 3.2 defines 'Region Information Lookup (RIL)'; please unify the terminology.","section":"Sec. 4.2.1"},{"comment":"The Figure 8 caption says results are averaged over seeds 1, 2, and 3, while Section 4.1.2 states that three runs use seeds 0, 1, and 2; please correct the inconsistency.","section":"Fig. 8"},{"comment":"The second contribution bullet contains a typo: 'Prompot' should be 'Prompt'.","section":"Contributions list"},{"comment":"The text cites 'CoOpOp [60]' but the reference is CoCoOp; please correct the name.","section":"Sec. 5.2"},{"comment":"The caption says 'using CoCo and PromptSRC base models'; this should presumably read 'CoOp and PromptSRC'.","section":"Fig. 9"},{"comment":"The negative loss mixes notation for 'uniform predictions' and the expression 1/|Y| log((1/|Y|)/pbar_i(c)); please clarify that this is the KL divergence from the uniform distribution to the negative-prompt distribution.","section":"Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The core idea is coherent and worth pursuing, but the efficiency claim is the main risk: the reported speedup needs to be re-measured end-to-end, including offline RIL construction and I/O. The accuracy comparison should also be framed against PromptSRC, since Table 2's average HM is lower for FDBPL. I recommend major revision with a clear timing protocol, per-dataset hyperparameters, and error bars before any acceptance decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read of arXiv:2505.18053. The core idea is worth a look: lift the space-for-time trick from FKD/FerKD into distillation-based prompt learning by precomputing teacher soft labels for random crops and storing them in a lookup table, then training student prompts from the cache. That removes the per-epoch teacher forward pass, and the paper honestly credits FKD and FerKD. Adding a 'negative prompt' for low-information regions is a reasonable twist, and the numbers do show a small but consistent gain over KDPL (average HM 74.57 vs 73.99; ImageNet new-class +1.13). So the work is not a rehash.\n\nBut the headline '2.2x faster training' does not survive scrutiny. The RIL table is built by running ViT-L/14 on M=500 random crops per training image. For ImageNet 16-shot (16k images) that is 8M teacher forward passes once; KDPL's online inference over 50 epochs is 800K. FDBPL uses 10x more teacher compute for a single student run. If the offline table construction and its I/O are excluded from the measured training time (and they appear to be, since Figure 8 shows only per-epoch prompt-training time), the 2.2x is not end-to-end and could easily reverse. The paper never reports total wall-clock including offline generation, and never says how many of the 500 crops are consumed per epoch. That is a load-bearing omission.\n\nOther issues, in decreasing severity: the hyperparameters alpha (Eq. 13), delta (Eq. 16), and all three lambda weights in Eq. 17 are never specified; the loss in Eq. 16 is ambiguous for the c=c' case since D^(2)_c,c is zero; no error bars or standard deviations are reported despite three seeds; no code is provided. The abstract's 'superior performance' overclaims given that average HM is below PromptSRC (74.97), though the fair comparison is to distillation-based KDPL, where the improvement is real.\n\nOn balance the central efficiency claim is unsupported as stated, but the underlying approach is sound enough to check. If the authors report end-to-end time (offline + training + I/O), fix the hyperparameters, and clarify the losses, this becomes a useful incremental contribution to the prompt-learning subfield. I'd send it to peer review with a request for major revision, not desk-reject it. The idea is worth a serious referee's time; the numbers just need to be presented honestly.","headline":"A plausible caching idea for distillation-based prompt learning, but the 2.2x speedup likely excludes the dominant offline teacher cost; the accuracy gains are real but modest.","tokens_in":20805,"tokens_out":3747,"would_cite":true,"duration_ms":32013,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Distillation-based prompt learning can run 2.2x faster without losing accuracy.","keywords":["vision-language models","prompt learning","knowledge distillation","zero-shot generalization","CLIP adaptation","training efficiency","region-aware prompts","soft label caching"],"falsifier":"Train FDBPL on ImageNet with the Region Information Lookup table refreshed every few epochs by re-running the teacher on the same crops, and compare new-class accuracy to the one-time-cache version. If fresh supervision improves accuracy by more than the run-to-run variance reported, the cached labels are stale and the 2.2x speed-up trades away some of the claimed generalization gain.","tokens_in":19899,"feed_emoji":"⚡","tokens_out":5966,"duration_ms":57968,"temperature":0.7,"pith_summary":"FDBPL is a prompt-learning method that adapts a frozen CLIP model to downstream tasks by distilling knowledge from a larger CLIP teacher. The paper's central claim is that the teacher's soft supervision can be computed once, stored in a Region Information Lookup table, and reused across every training epoch, removing the repeated online teacher inference that makes earlier distillation-based prompt learning slow. On this basis FDBPL adds a region-aware dual-prompt scheme: positive prompts learn real class semantics from high-information cropped regions, while negative prompts learn to reject low-information regions. The result, the paper reports, is a 2.2x average training speed-up over KDPL with comparable or better zero-shot accuracy, including +1.13 new-class accuracy on ImageNet.","feed_headline":"Cached teacher labels make prompt learning 2.2x faster","feed_subtitle":"FDBPL stores teacher supervision once, then trains a CLIP student to say yes to real classes and no to low-information regions.","key_machinery":"The central mechanism is the Region Information Lookup (RIL) table: a precomputed store of crop coordinates, augmentation types, and teacher soft labels that lets the student fetch supervision instead of recomputing it. Around it sit two components: Region-Aware Dual-Prompt (RADP) learning, which uses an entropy-derived information weight $w$ to route high-information regions to a positive prompt and low-information regions to a negative prompt; and Prompt-Cascaded Difference (PCD) learning, which forms first-order differences between positive and negative text features and second-order differences between classes, aligning intra-class diversity and inter-class structure. The Top-K label sparsification strategies (MS and MR) keep the table compact enough for fast I/O.","core_discovery":"The discovery is that the generalization benefit of distillation-based prompt learning does not require the teacher to be invoked during student training. By pre-computing teacher soft labels, spatial coordinates, and augmentation metadata for randomly cropped regions and storing them in the Region Information Lookup table, FDBPL turns distillation into a fast table lookup. Two label-sparsification strategies, Marginal Smoothing with Top-K and Marginal Re-Norm with Top-K, compress stored soft labels so I/O does not become the new bottleneck. The paper further claims that low-information crop regions, normally discarded, can be exploited through a learnable negative prompt that says 'no,' and that a two-level difference space between positive and negative prompts captures intra-class and inter-class structure, improving zero-shot recognition on novel classes.","pith_inferences":["The caching design suggests an asynchronous teacher-update protocol: refreshing the table every few epochs with a stronger or ensemble teacher might recover any staleness without giving back the full online-inference cost.","The negative-prompt 'say no' behavior is effectively a lightweight out-of-distribution detector; it could be evaluated on open-set recognition benchmarks beyond the robustness tests reported here.","Because the table stores teacher outputs rather than model parameters, the method may transfer to other vision-language model families or to tasks such as detection and segmentation where region-level soft supervision is natural.","The paper's efficiency claim is wall-clock measured on one GPU setting; the practical speed-up will depend on disk and CPU I/O speed and dataset size, so the 2.2x figure is likely workload-dependent."],"forward_implications":["Distillation-based prompt learning can be decoupled from a specific teacher deployment: the table is built offline once, so different or stronger teachers can be swapped in without retraining the student.","Training time on large-scale datasets such as ImageNet-1K drops from a KDPL-level slowdown to near-native prompt learning, because per-epoch teacher inference is eliminated.","Low-information regions, previously discarded in FKD-style pipelines, become training signal that teaches the student to reject uncertain input, improving novel-class and cross-dataset accuracy.","The RIL table's storage and I/O costs scale with the number of classes; Top-K sparsification keeps this manageable and yields an optimal operating point near K=20.","FDBPL is compatible with existing prompt-learning bases such as CoOp and PromptSRC, so the acceleration can be layered onto current adaptation pipelines."],"supporting_citations":[{"why":"Supplies the frozen CLIP dual-encoder backbone used as both student and teacher.","marker":"[1]"},{"why":"Defines the soft-prompt learning paradigm that FDBPL adapts and uses as a primary baseline.","marker":"[41]"},{"why":"Provides the base-to-new evaluation protocol used across the main experiments.","marker":"[42]"},{"why":"KDPL is the distillation-based prompt-learning baseline that FDBPL accelerates and compares against for accuracy.","marker":"[25]"},{"why":"Introduces random cropping, offline alignment objective storage, and label compression that FDBPL builds on.","marker":"[29]"},{"why":"Extends offline distillation efficiency with multi-level management, the contrast for FDBPL's use of low-information crops.","marker":"[30]"},{"why":"PromptSRC serves as the base prompt-learning model in cross-dataset and PCD ablation experiments.","marker":"[44]"}],"fun_headline_variants":["Distill teacher once, train faster: FDBPL","FDBPL: cached teacher labels, 2.2x faster training","Skip repeat teacher inference for 2.2x faster prompt learning","Region-aware prompt learning with teacher caching speeds up 2.2x","Reuse teacher supervision, cut prompt learning time 2.2x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The frozen teacher's soft labels, computed once before training, remain valid supervision for every student prompt update across all 50 training epochs, so student prompt drift does not make the cached targets stale.","fun_headline_variants_meta":{"raw":{"variants":["Distill teacher once, train faster: FDBPL","FDBPL: cached teacher labels, 2.2x faster training","Skip repeat teacher inference for 2.2x faster prompt learning","Region-aware prompt learning with teacher caching speeds up 2.2x","Reuse teacher supervision, cut prompt learning time 2.2x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000279,"raw_usage":{"total_tokens":1673,"prompt_tokens":977,"completion_tokens":696,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":603}},"tokens_in":593,"tokens_out":696,"duration_ms":6143,"temperature":1.0,"reasoning_tokens":603,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:35:55.404744+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train FDBPL on ImageNet with the Region Information Lookup table refreshed every few epochs by re-running the teacher on the same crops, and compare new-class accuracy to the one-time-cache version. If fresh supervision improves accuracy by more than the run-to-run variance reported, the cached labels are stale and the 2.2x speed-up trades away some of the claimed generalization gain.","supporting_citations":[{"cited_title":"Radford, J","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen CLIP dual-encoder backbone used as both student and teacher."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the soft-prompt learning paradigm that FDBPL adapts and uses as a primary baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the base-to-new evaluation protocol used across the main experiments."},{"cited_title":"Mistretta, A","cited_arxiv_id":null,"evidence_quote":"KDPL is the distillation-based prompt-learning baseline that FDBPL accelerates and compares against for accuracy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces random cropping, offline alignment objective storage, and label compression that FDBPL builds on."},{"cited_title":"Shen, Ferkd: Surgical label adaptation for efficient dis- tillation, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp","cited_arxiv_id":null,"evidence_quote":"Extends offline distillation efficiency with multi-level management, the contrast for FDBPL's use of low-information crops."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PromptSRC serves as the base prompt-learning model in cross-dataset and PCD ablation experiments."}],"review_version":1}