{"id":"4d1b53f7-35b4-4bde-9633-16fa8aef9d49","arxiv_id":"2507.07415","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"EPIC uses similarity-gated cross-modal prompt interactions in frozen CLIP to improve text-image classification accuracy while training about 1% of parameters.","lead":"EPIC is a new way to fine-tune CLIP for image-text classification: small trainable prompt vectors are added to a few middle layers, and the text and image branches exchange information through a similarity-based Interaction Hub. The method reportedly trains about 1% of the model's parameters and beats prior prompt-based methods on two of three benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Efficiency claim in strongest_claim is contradicted by Table I: EPIC-base uses 17.63 GB training memory vs PMF's 15.29 GB, and 2.0M params vs P-CLIP's 0.1M; 'minimal' is unsupported.","rationale":"The reader's weakest assumption is test-set leakage from hyperparameter selection on UPMC-Food101, which is a valid methodological concern. However, the paper's own Table I provides an independent, internally checkable problem: the central claim explicitly asserts 'minimal GPU memory usage and trainable parameters,' and the table shows that EPIC-base is not minimal in either dimension among the prompt-based methods it compares with. This is not a question of generalizability or leakage; it is a factual mismatch between the claim and the reported evidence. The reader's conditional verdict already requires revisions, so my read does not move the verdict, but it identifies a more directly load-bearing reason for conditioning acceptance: the efficiency half of the headline claim must be weakened or precisely scoped. The concrete test settles whether the contradiction is an artifact of unstated measurement differences or a real overclaim; either way, the paper needs a clarifying sentence about what 'minimal' means.","tokens_in":8614,"tokens_out":5278,"duration_ms":59070,"concrete_test":"Recompute Table I's memory and parameter entries for EPIC-base, PMF, P-CLIP, and MaPLe* under one identical protocol (same batch size, same input resolution/text length, same GPU and memory-measurement tool), and report the exact measurement setup. If EPIC-base then has higher training memory than PMF and more trainable parameters than P-CLIP/MaPLe* (as Table I currently shows), the sentence 'minimal GPU memory usage and trainable parameters' must be replaced with a qualified efficiency claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing efficiency component of the central claim is contradicted by the paper's own Table I. EPIC-base reports 17.63 GB training memory, which is higher than PMF's 15.29 GB; its 2.0M trainable parameters exceed P-CLIP, PromptFuse, and BlindPrompt (0.1M each) and MaPLe* (0.8M). The abstract and conclusion state 'minimal GPU memory usage and trainable parameters,' but under the reported numbers EPIC is neither the lowest-memory nor the lowest-parameter prompt-based method. The claim could still be true if memory and parameter counts were measured under non-identical settings (e.g., different batch sizes or sequence lengths), but the paper provides no measurement protocol, so the 'minimal' claim is not supported as written. If corrected to 'competitive,' the performance claim is unaffected; if not, the central claim as stated fails.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EPIC, a prompt-based interaction method for text-image classification using a frozen CLIP ViT-L/14 backbone. Temporal prompts are inserted on selected intermediate layers of the vision and text branches and are exchanged through a lightweight 'Interaction Hub' that computes intra- and inter-modality similarity gates. The method reports 93.95% accuracy on UPMC-Food101, 73.45% on SNLI-VE, and 65.9/56.3 F1-micro/macro on MM-IMDB with 2.0M trainable parameters. The paper claims state-of-the-art results among prompt-based methods with the same foundation model at minimal GPU memory usage and trainable parameters, and includes an ablation study of the components and a sensitivity analysis of interaction layer choices and similarity functions.","tokens_in":8853,"tokens_out":5920,"duration_ms":60317,"significance":"The interaction mechanism is conceptually simple, and the ablation study in Table II shows monotonic gains from adding temporal prompts, prompt interaction, and the similarity-based gating, which gives internal support to the design. The reported accuracy wins over re-implemented prompt-based baselines on two of three datasets, and the parameter count of EPIC-base (2.0M) is far below full fine-tuning. However, the paper's central efficiency claim is contradicted by its own Table I, and the headline UPMC-Food101 accuracy is obtained after selecting interaction layers and similarity type on the same test set that is used for the main result. If these protocol issues are corrected, the method could be a useful parameter-efficient multimodal fusion baseline; as written, the strength of the claimed contribution is not fully established.","major_comments":[{"comment":"The central efficiency claim is not supported by the paper's own measurements. Table I reports EPIC-base with 17.63 GB training memory, which is higher than PMF's 15.29 GB, and 2.0M trainable parameters, which is higher than MaPLe* (0.8M), P-CLIP, PromptFuse, and BlindPrompt (all 0.1M). The abstract and conclusion state that EPIC achieves 'minimal GPU memory usage and trainable parameters', and Figure 3(c) repeats this. Since no measurement protocol (batch size, sequence length, hardware, gradient checkpointing) is provided, the claim as stated is directly contradicted by the reported numbers. The authors should either provide controlled comparisons and correct the claim to 'competitive', or supply additional evidence for a minimal-resource interpretation.","section":"Table I, Abstract, Conclusions"},{"comment":"The interaction layers (4,5,6) and cosine similarity are selected by inspecting results on UPMC-Food101, as shown in Figure 4, and the same dataset is then used as the headline benchmark in Table I. No held-out validation split is used. This means the reported 93.95% accuracy is the maximum over configurations evaluated on the test set, so the improvement over PMF* (93.95 vs 92.54) may be inflated by selection. The authors should either use a separate validation split for model selection or report results under a protocol that avoids test-set-based selection, for example by pre-registering the configuration or by reporting post-selection corrected estimates.","section":"Section IV-B and Table I"},{"comment":"The training loss for uni-label classification is written as binary cross-entropy with y_i in {0,1}, but the prediction is a softmax over K classes in Eq (7). For a K-class task such as UPMC-Food101, the correct loss is the negative log-likelihood of the true class, not the binary form in Eq (8). If a sigmoid activation and binary labels are used instead, Eq (7) should be changed accordingly. This inconsistency must be resolved for the training procedure to be reproducible.","section":"Section II-E, Eq (8)"},{"comment":"All results are reported as single numbers with no error bars, confidence intervals, or significance tests. The margins over PMF* are 1.01-1.41 accuracy points on two datasets; without repeated runs (e.g., 3-5 seeds) and variance reporting, it is not possible to assess whether these differences are reliable. The authors should report mean and standard deviation over multiple runs.","section":"Table I and Section III-B"}],"minor_comments":[{"comment":"The dimensions of the similarity matrices and the axis over which SoftMax is applied are not specified. Please clarify whether softmax is over prompt tokens, feature units, or both.","section":"Section II-D, Eqs (4)-(6)"},{"comment":"The row for EPIC-base has values run together without visible separators, making the table difficult to read. Please reformat the table so each metric is clearly aligned.","section":"Table I"},{"comment":"The introduction says the interaction happens 'without participation of modality features', but Eq (2) concatenates prompts with modality features before the encoder layer. Please rephrase to avoid this apparent contradiction.","section":"Section I and Section II-D"},{"comment":"The first temporal prompt for text is randomly initialized with a Gaussian and then used to produce the vision prompt via a linear layer. Please clarify whether the text prompt itself is also trained and how the initialization relates to the shared trainable parameters across layers.","section":"Section II-C"},{"comment":"The abstract says 'fewer trainable parameters (about 1% of the foundation model)' but gives no calculation; please specify the foundation model parameter count used for the percentage.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The test-set-based selection in Section IV-B is the main correctness risk: the reported UPMC-Food101 number is likely optimistically biased. The efficiency claim is also objectively wrong in the paper's own table. If the authors can rerun with a proper validation split, report multi-seed variance, correct the loss equation, and restate the efficiency claim as 'competitive', the paper could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about EPIC is that the core mechanism is genuinely new and the empirical story is mostly consistent, but the headline efficiency claim is not supported by the paper's own numbers. The Interaction Hub, with intra- and inter-modality softmax gates and shared parameters across layers, is a real departure from MaPLe and PMF, and the ablations show each added component moves accuracy. The gains on UPMC-Food101 and SNLI-VE over re-implemented baselines are consistent, and EPIC-small staying competitive at one interaction layer is a nice robustness check. That is honest, useful progress for prompt-based multimodal classification.\n\nThe soft spots are in proportion to how much they matter. First, the abstract and conclusion say 'minimal GPU memory usage and trainable parameters,' but Table I shows EPIC-base at 17.63 GB training memory, above PMF's 15.29 GB, and 2.0M parameters, above P-CLIP, PromptFuse, and BlindPrompt at 0.1M. That is not a minor wording issue; it is a load-bearing claim contradicted by their own results. 'Competitive' or 'lightweight' would be defensible; 'minimal' is not. Second, the interaction layers (4,5,6) and cosine similarity were selected by inspecting results on UPMC-Food101, the same dataset used as the headline benchmark. With no held-out validation split, the reported 93.95% is partly a selection artifact. Third, all results are single-run with no error bars or significance tests, and no code or training details are provided. These are all fixable, but they mean the central 'superior performance' claim is conditional.\n\nThe paper deserves a serious referee, because the method is coherent, the mechanism is novel enough, and the empirical gains survive re-implemented baselines under a reasonable reading. But the review must push for corrected efficiency claims, a proper validation split for hyperparameters, error bars, and code release. As written, I would not cite the efficiency numbers, and I would be cautious about the accuracy numbers until the protocol is tightened.","headline":"Solid incremental prompt-interaction method with real gains, but the 'minimal memory/parameters' claim doesn't survive its own Table I and test-set tuning.","tokens_in":9354,"tokens_out":1347,"would_cite":false,"duration_ms":15476,"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":"EPIC shows that gated prompt exchange between frozen vision and text encoders beats prior prompt-based methods on food and visual-entailment classification while training only 2.0M parameters.","keywords":["efficient prompt interaction","text-image classification","prompt tuning","cross-modal interaction","Interaction Hub","CLIP","multimodal classification","parameter-efficient fine-tuning"],"falsifier":"Split UPMC-Food101 (or SNLI-VE) into training and validation, tune interaction-layer indices, similarity type, and prompt length only on the validation split, then evaluate the chosen configuration once on the untouched test set; if EPIC's margin over PMF shrinks to noise or reverses, the reported SOTA is an artifact of test-set-based selection.","tokens_in":8444,"feed_emoji":"⚡","tokens_out":8233,"duration_ms":82233,"temperature":0.7,"pith_summary":"This paper tries to show that cross-modal information flow in a frozen vision-language model can be carried entirely by trainable prompt vectors, leaving the backbone untouched. Its method, EPIC, inserts temporal prompts into intermediate layers of both a vision and a text encoder and exchanges information between them through a lightweight similarity-based Interaction Hub. On standard benchmarks the reported accuracy is the best among prompt-tuning baselines on UPMC-Food101 (93.95%) and SNLI-VE (73.45%), and comparable on MM-IMDB, while training only about 2.0M parameters — roughly 1% of the frozen CLIP ViT-L/14 model. The significance, if the numbers hold, is that a cheap prompt-only mechanism can replace heavy fine-tuning and explicit fusion modules for multimodal classification.","feed_headline":"Prompt gating tops prior prompt methods with 1% of the parameters","feed_subtitle":"Similarity-gated prompt exchange gives 93.95% on UPMC-Food101 and 73.45% on SNLI-VE with minimal training cost.","key_machinery":"The Interaction Hub is the module that carries the argument. Given the temporal prompts $\\hat{p}^l_m$ and $\\hat{p}^l_{m'}$ of the two modalities on an interaction layer, it projects the other modality's prompt into the current one's space, computes intra-modality and inter-modality similarity maps with cosine similarity under ReLU, softmaxes each map into a gate, and outputs $p^{l+1}_m = z^l_m \\cdot \\hat{p}^l_m + (1-z^l_m)\\cdot r^l_m \\cdot \\tilde{p}^l_{m'}$. The gates decide how much context to keep and how much cross-modal content to accept token by token. Because the prompt and projection parameters are shared across all interaction layers, adding more interaction layers does not increase the trainable parameter count.","core_discovery":"EPIC's central claim is that prompt-level interaction alone is sufficient for effective vision-language fusion: temporal prompts that take part in each encoder layer's self-attention can simultaneously act as a communication channel between the modalities. The Interaction Hub measures cosine similarity within and across the two prompt streams, turns these maps into softmax gates, and forms the next layer's prompt as a gated blend that keeps the modality's own context and admits the other modality's information only where correspondences are strong. With this mechanism the paper reports the best accuracy among prompt-based methods that share the same CLIP ViT-L/14 foundation model: 93.95% on UPMC-Food101, 73.45% on SNLI-VE, and 65.9/56.3 F1-micro/macro on MM-IMDB, using a 2.0M-parameter trainable set and minimal GPU memory during training.","pith_inferences":["A testable extension the authors do not run is to replace the cosine-similarity gates with a scalar learned mixing weight; if accuracy drops, token-level correspondence gating is doing the work, and if it does not, a simpler mechanism suffices.","Because the interaction-layer choice (4,5,6) came from the UPMC-Food101 sweep in Figure 4, re-tuning per dataset on a true validation split could shift the ranking on SNLI-VE and MM-IMDB.","The Interaction Hub is effectively a content-addressable exchange between prompts, so a natural probe is whether high gate values coincide with human-judged alignment between image and text.","Since only prompts and the projection MLP are trained, EPIC could be a substrate for continual learning, where new tasks reuse the frozen backbone and exchange only the small prompt set."],"forward_implications":["If the reported results hold, frozen-backbone prompt interaction can replace heavy fine-tuning and explicit fusion modules for text-image classification, cutting training memory to a small fraction of full fine-tuning.","The same gated prompt-exchange pattern transfers to other multimodal classification tasks built on frozen encoders, including visual entailment and multi-label sentiment analysis.","Because prompt and projection parameters are shared across interaction layers, increasing the number of interaction layers can improve accuracy without increasing the trainable parameter count.","The single-interaction-layer variant, EPIC-small, already beats several prompt baselines on all three datasets, indicating the mechanism does not depend on stacking many interaction layers."],"supporting_citations":[{"why":"MaPLe, the multimodal prompt-learning baseline that EPIC re-implements under its own settings and claims to beat on two benchmarks.","marker":"[14]"},{"why":"PMF, the interactive-prompting method that is the closest prompt-based competitor on all three datasets.","marker":"[9]"},{"why":"PromptFuse and BlindPrompt, modular prompt-fusion baselines whose limited information exchange motivates the Interaction Hub design.","marker":"[8]"},{"why":"CLIP, the frozen ViT-L/14 foundation model that supplies both vision and text encoders for every compared method.","marker":"[23]"},{"why":"CMA-CLIP, a cross-modal attention CLIP baseline whose published results on these datasets set the target for prompt-only methods.","marker":"[1]"},{"why":"Visual Prompt Tuning, the source of the visual prompt idea that EPIC extends to temporal prompts on both modality branches.","marker":"[11]"},{"why":"UPMC-Food101, the primary uni-label benchmark where EPIC reports its headline 93.95% accuracy.","marker":"[20]"},{"why":"MM-IMDB, the multi-label benchmark where EPIC is compared using F1-micro/F1-macro.","marker":"[21]"},{"why":"SNLI-VE, the visual-entailment benchmark where EPIC claims the best prompt-based accuracy of 73.45%.","marker":"[22]"}],"fun_headline_variants":["Prompt-only fusion beats prior prompt methods using just 1% of weights","Similarity-gated prompts achieve top results with 1% trainable params","EPIC: 1% of model parameters, top accuracy on image-text tasks","Prompt gating cuts training cost, boosts multimodal accuracy with 1% params"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The performance claim rests on hyperparameters being chosen honestly: the interaction layers (4,5,6), the cosine similarity metric, and the prompt length were selected after inspecting results on the UPMC-Food101 test set, and the same test set is the headline benchmark, with no held-out validation split reported.","fun_headline_variants_meta":{"raw":{"variants":["Prompt-only fusion beats prior prompt methods using just 1% of weights","Similarity-gated prompts achieve top results with 1% trainable params","EPIC: 1% of model parameters, top accuracy on image-text tasks","Prompt gating cuts training cost, boosts multimodal accuracy with 1% params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001051,"raw_usage":{"total_tokens":4394,"prompt_tokens":907,"completion_tokens":3487,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":3404}},"tokens_in":523,"tokens_out":3487,"duration_ms":27358,"temperature":1.0,"reasoning_tokens":3404,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:41:08.390878+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Split UPMC-Food101 (or SNLI-VE) into training and validation, tune interaction-layer indices, similarity type, and prompt length only on the validation split, then evaluate the chosen configuration once on the untouched test set; if EPIC's margin over PMF shrinks to noise or reverses, the reported SOTA is an artifact of test-set-based selection.","supporting_citations":[{"cited_title":"Maple: Multi-modal prompt learning,","cited_arxiv_id":null,"evidence_quote":"MaPLe, the multimodal prompt-learning baseline that EPIC re-implements under its own settings and claims to beat on two benchmarks."},{"cited_title":"Efficient multimodal fusion via interactive prompting,","cited_arxiv_id":null,"evidence_quote":"PMF, the interactive-prompting method that is the closest prompt-based competitor on all three datasets."},{"cited_title":"Modular and Parameter-Efficient Multimodal Fusion with Prompting","cited_arxiv_id":"2203.08055","evidence_quote":"PromptFuse and BlindPrompt, modular prompt-fusion baselines whose limited information exchange motivates the Interaction Hub design."},{"cited_title":"Learning transferable visual models from natural language supervision,","cited_arxiv_id":null,"evidence_quote":"CLIP, the frozen ViT-L/14 foundation model that supplies both vision and text encoders for every compared method."},{"cited_title":"Cma-clip: Cross- modality attention clip for text-image classification,","cited_arxiv_id":null,"evidence_quote":"CMA-CLIP, a cross-modal attention CLIP baseline whose published results on these datasets set the target for prompt-only methods."},{"cited_title":"Visual prompt tuning,","cited_arxiv_id":null,"evidence_quote":"Visual Prompt Tuning, the source of the visual prompt idea that EPIC extends to temporal prompts on both modality branches."},{"cited_title":"Recipe recognition with large multimodal food dataset,","cited_arxiv_id":null,"evidence_quote":"UPMC-Food101, the primary uni-label benchmark where EPIC reports its headline 93.95% accuracy."}],"review_version":1}