{"id":"b67d8e89-2ad4-453e-bd1f-330af0a0e5da","arxiv_id":"2509.04480","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Recursive generation, evaluation, and refinement of discrete prompts tunes a black-box multimodal LLM to each user, raising personalized visual emotion recognition accuracy on Affection from 40.6% to 44.9%.","lead":"This paper tailors the text prompts given to a vision-language AI to match how each individual reacts emotionally to images, by repeatedly generating, testing, and refining prompts on that person's labeled photos. On a 15-user benchmark, the tuned prompts outscore several baselines, although the reported gains are within the noise band of the measurements.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No same-budget random-search control means the recursive-refinement gain may be selection noise; training ACC (Eq. 2) is the only signal, and reported gaps are within per-user SD.","rationale":"The reader's weakest_assumption is legitimate: prompt ranking by training ACC on ~260 images is noisy. I go a step further: even if ranking were reliable, the paper's causal attribution to recursive utilization requires a same-budget random-search baseline. The presented evidence (Table 3, Figure 2) cannot rule out that the same or better test accuracy would be obtained by simply drawing a larger pool of diverse prompts and selecting top-H by training ACC. The paper does provide useful structure, clear pseudocode, public dataset, and a limitations section, so the issue is not soundness but under-determination. The proposed test would settle it. Thus the appropriate disposition remains CONDITIONAL (the reader's verdict), with the condition being the random-search control and significance testing.","tokens_in":14523,"tokens_out":4234,"duration_ms":54150,"concrete_test":"Run a per-user random-search control on the same Affection split and budget: for each user, replace the refinement loop in Algorithm 1 with I3×I2 draws of N fresh prompts from GPT-4o (same t_init, no t_mod feedback), compute ACC on the same 30% training subset, keep the top H=5 by ACC, majority-vote on the same 70% test set, and report mean accuracy/ECC over the same 15 users. If mean test accuracy is within noise of or above 44.9% (or ECC 63.4%), the recursive-refinement claim is not supported. Also report a paired Wilcoxon signed-rank test (n=15) for Our method vs CM7; if p>0.05 the headline gain is not distinguishable from per-user noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that ACC(p) computed on 30% of a user's images (Eq. 2, Section III-B) is a reliable selector for the H prompts used at test time (Eq. 5), and that the iterative top-/worst-prompt feedback in Table 2 improves that signal. Both are untested. Figure 2 shows only training-set accuracy rising with iterations; since Algorithm 1 always keeps the argmax over a growing pool, training accuracy must rise under selection even when the selected prompts are no better than random candidates on the held-out 70%. The reported gains over the strongest ablation (CM7, single modified prompt) are 1.8 accuracy points and 1.6 ECC points, while per-user SDs are ~9.6 and ~6.6; no paired significance test is reported. A random-search control that spends the same number of MLLM evaluations without iterative refinement is absent, so the contribution of 'recursive utilization' cannot be separated from the trivial effect of evaluating many prompts and majority-voting.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a discrete prompt tuning method for personalized visual emotion recognition (VER) using a black-box multimodal large language model (MLLM). An LLM generates candidate natural-language prompts; each prompt is scored by its accuracy on 30% of a user's training images (Eq. 2); the top-k and worst-k prompts are fed back to the LLM to generate modified prompts iteratively (Algorithm 1); at test time, the H highest-scoring prompts are used for recognition and a majority vote yields the final label (Eqs. 5-6). Experiments on 15 Affection users report accuracy 44.9% ± 9.62%, ECC 63.4% ± 6.61%, outperforming CNN/Transformer baselines (CM1-CM2), a zero-shot MLLM (CM3), and ablation variants (CM4-CM7). The authors conclude that recursive discrete prompt tuning adapts the MLLM to individual users without access to gradients or model parameters.","tokens_in":14777,"tokens_out":4785,"duration_ms":62852,"significance":"If the empirical claims hold, the method offers a practical, interpretable, gradient-free way to personalize black-box MLLMs for VER, and the algorithmic template (LLM-guided prompt generation, evaluation, and iterative refinement) could transfer to other affective computing tasks. The paper has notable strengths: a clear algorithm (Algorithm 1), a held-out test protocol, multiple ablation conditions (CM4-CM7) that isolate initial prompts, one-iteration tuning, cross-user transfer, and majority voting, plus an explicit limitations section. However, the central comparative claims are not yet compelling because the reported gains over the strongest ablations are small relative to per-user variability, no significance testing is reported, and there is no same-budget random-search control that would distinguish the benefit of recursive refinement from the effect of selecting among many random prompts. These are fixable with additional analysis and one baseline experiment.","major_comments":[{"comment":"The central claim that the proposed method achieves the highest accuracy and ECC rests on mean differences that are not accompanied by any significance testing. The proposed method beats CM7 by only 1.8 accuracy points (44.9 vs 43.1) and 1.6 ECC points (63.4 vs 61.8) while the per-user standard deviations are 9.62 and 6.61, respectively; the gain over CM5 is 3.3/2.5 points. With 15 users and seven comparisons, these gaps may be within noise. Please report per-user paired tests or confidence intervals for the proposed method against CM5 and CM7, and correct for multiple comparisons. Without this, the headline 'highest performance' is not established.","section":"§IV-B, Table 3"},{"comment":"The load-bearing premise is that ACC(p) computed on the 30% per-user training split is a reliable selector for the H prompts used on the held-out 70%. Because the pool grows and the algorithm always keeps the argmax over that pool, training accuracy necessarily rises with iterations even if the iterative feedback adds no signal; Fig. 2 shows only training accuracy. The claim in §IV-B that the improvement is 'not simply due to selection of high-accuracy prompts from a set of randomly generated candidates' needs a control that spends the same number of MLLM evaluations on randomly generated prompts (with the same majority-voting protocol) and compares test accuracy. Please add this random-search baseline, and ideally a validation-split-based selection, to separate the contribution of recursive utilization from selection noise.","section":"§III-B, §III-D, Algorithm 1 (Eqs. 2, 5)"},{"comment":"The method has seven free hyperparameters (N=6, T=5, k=3, I1=20, I2=2, I3=3, H=5), but no sensitivity analysis is reported. Since the final majority vote over H prompts is a major source of the reported gain over CM7, and k controls the LLM's feedback, the results may depend nontrivially on these choices. Please report ablations over at least H and k (and ideally I1), or provide a justification that the chosen values are not performance-critical. Currently the only iteration-related ablation is CM5, which is insufficient to assess robustness.","section":"§IV-A, hyperparameters N, T, k, I1, I2, I3, H"},{"comment":"The Affection user pool is restricted to 15 users who provided 'multiple images for each emotion label'; individuals with small or extremely large response counts were excluded. This filters the population and may affect the difficulty of personalization and the magnitude of the reported gains. The paper should state how many users were excluded, analyze the sensitivity of the conclusions to this filter, or at least discuss the potential bias in the limitations section. As written, the personalization claim is limited to a selected moderate-response subset rather than to Affection users generally.","section":"§IV-A, user selection"}],"minor_comments":[{"comment":"The definitions of CM5 and CM7 are ambiguous: CM5 is 'a prompt modified in one iteration' but §IV-B states CM5 'included a majority voting approach,' while CM7 is 'a single modified prompt without majority voting.' Please clarify exactly which prompts and voting rule each ablation uses, and make Table 3 annotations consistent.","section":"§IV-A, CM5 vs CM7"},{"comment":"Figure 2 and the accuracies in Table 4 are training-set accuracies. This should be stated explicitly in the captions and in the text, because the reader may otherwise mistake them for held-out performance.","section":"Fig. 2 / Table 4"},{"comment":"Please report the exact GPT-4o version, decoding parameters (e.g., temperature), and any random-seed handling for LLM and MLLM inference. Without these, the iterative prompt-generation process is not reproducible.","section":"§IV-A / reproducibility"},{"comment":"The phrase 'implicit gradients indicate the direction toward the optimal prompt' is metaphorical and could be misread as a formal gradient in prompt space. Consider rephrasing to describe the LLM's refinement feedback more literally.","section":"§III-C"},{"comment":"Minor formatting issues: 'Y ou' appears in several references (e.g., [33], [35]), 'Proceeding' is inconsistently used, and 'Mikel's wheel' should probably be 'Mikels' wheel' to match [46]. These should be corrected.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable empirical contribution but the central claims are currently undersupported. The missing random-search control and significance testing are the key blockers; both are tractable with the existing experimental setup. The authors should also clarify the ablation definitions and hyperparameter sensitivity. If these are addressed, the paper could be suitable for publication; in its current form, the evidence does not justify the headline conclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a clean, honest application of black-box prompt optimization (Liu et al.'s OPRO loop, which the authors cite) to personalized visual emotion recognition. The new thing is the target task and the per-user evaluation setup, not the machinery. The gains look real but fragile: roughly +4 accuracy points over the best ablation on 15 Affection users, with per-user SDs of 8–10 points and no significance tests. I'd send it to a referee, but the referee should push for a random-search control and paired tests.\n\nWhat the paper does well: the method is fully specified (templates, Algorithm 1, hyperparameters), the dataset is public, and the authors include a limitations section that admits the anger class does not improve. The comparison is not circular — prompts are selected on a 30% training split and evaluated on the held-out 70%. The ablations (CM4–CM7) are sensible and isolate the contribution of iterative refinement, personalization, and majority voting. They also correctly cite [29] as the origin of the loop; this is not hidden.\n\nWhere it's soft: the central claim rests on small gaps. Our method at 44.9 ± 9.62 vs. CM7 at 43.1 ± 8.63 is 1.8 points; vs. CM4 it's 4.3 points. With 15 users, a paired test might be significant, but they don't report one. The load-bearing premise is that ACC(p) on roughly 260 training images reliably ranks prompts for the rest of that user's images. Figure 2 only shows training accuracy rising, which is guaranteed to happen under argmax selection over a growing pool, so it doesn't demonstrate that iterative refinement beats random search. Without a same-budget random-search baseline, the 'recursive utilization' contribution could be just selection plus majority voting. The user filtering (excluding users with small or very large response counts) also limits generality. These are addressable, not fatal.\n\nBottom line: a reasonable empirical contribution for the VER community, but the conclusion overclaims when it says 'significantly improve.' A serious referee could turn this into a solid paper by adding significance tests, a random-search control, and per-user analysis. Worth engaging, not worth taking on faith.","headline":"Honest OPRO-style application to personalized VER; small gains, no significance tests, but worth a referee.","tokens_in":15309,"tokens_out":2284,"would_cite":false,"duration_ms":27252,"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":"Per-user natural-language prompts, refined through a recursive generate-score-refine loop, shift a black-box multimodal model's emotion predictions toward an individual viewer's own labels — the paper's route to personalized visual emotion","keywords":["visual emotion recognition","personalized emotion recognition","multimodal large language models","discrete prompt tuning","black-box prompt optimization","prompt engineering","majority voting"],"falsifier":"Re-run the full tuning loop on the same 15 Affection users with a different random 30/70 percent split per user. If the tuned prompts' held-out accuracy advantage over the fixed initial prompt (CM4) and the single tuned prompt without voting (CM7) does not reproduce across resplits — or if per-user gains do not exceed the reported between-user standard deviation of about 9.6 percentage points — the gains are artifacts of the training-slice ranking rather than true personalization.","tokens_in":14397,"feed_emoji":"🎭","tokens_out":16059,"duration_ms":156657,"temperature":0.7,"pith_summary":"Visual emotion recognition is inherently personal: the same photo can excite one viewer and frighten another, yet multimodal large language models (MLLMs) are trained on general opinions and default to majority reactions. The paper's claim is that this individual bias can be corrected without fine-tuning or seeing the model's internals, by tuning the natural-language prompt itself. An LLM proposes candidate prompts; the black-box MLLM scores each prompt by how often its predicted emotion matches the user's own labels on a training slice of that user's photos; and the best- and worst-scoring prompts are handed back to the LLM to write better ones. After several rounds, the highest-scoring prompts are each run on the target image, majority-voted into a final label, and this recursive discrete prompt tuning reaches 44.9 percent accuracy on 15 Affection users — against 39.1 percent for a zero-shot MLLM prompt and 40.6 percent for a fixed one — while keeping personalization human-readable and requiring no gradient access. The paper itself notes the gains are uneven across emotions: anger barely improves, which it attributes to class imbalance.","feed_headline":"Prompt rewrite loop personalizes a black-box emotion AI","feed_subtitle":"A generate-score-refine loop lifts per-user emotion recognition from 39.1 to 44.9 percent accuracy, with no gradients needed.","key_machinery":"The load-bearing object is the recursive discrete prompt-tuning loop of Algorithm 1, built from two parts. First, an accuracy scorer ACC(p) (Eq. 2): each candidate prompt runs through the MLLM on the user's training images, and the fraction of the user's labels it matches is its score. Second, the LLM modifier t_mod(Ppos, Pneg): the LLM receives the top-k and worst-k prompts with their scores and returns T new prompts, drawn closer to the high-scoring examples and away from the low-scoring ones — an implicit gradient in language space. Iteration builds a prompt pool; the H highest-scoring prompts run on the target image and are majority-voted. The scorer keeps the method black-box compatible","core_discovery":"The central claim: a black-box multimodal LLM can be personalized per viewer purely by discrete prompt tuning — rewriting the instruction text — in a recursive loop powered by the LLM's self-correction. Each candidate prompt is scored by how often the model's predictions match the user's own labels on a slice of that user's photos. The top-k and worst-k prompts, with their scores, return to the LLM, which writes new prompts more like the good ones and less like the bad ones — an 'implicit gradient' toward the user's optimal prompt. On 15 Affection users this beat every baseline and ablation, indicating genuinely personal prompts rather than random winners.","pith_inferences":["The decisive check the paper does not run is split-robustness: prompts are chosen by accuracy on one 30% slice of each user's photos, so rerunning the loop on different random splits would show whether the ranking signal is stable or partly selection noise; this is directly testable with the same dataset and models.","The same loop should transfer to other subjective visual judgments where a viewer's own response is ground truth — aesthetic preference, humor, trustworthiness, or 'would this appeal to me' — since the mechanism only needs per-user labels and an MLLM that can score prompts.","The method's ceiling is set by the two models' abilities rather than by the procedure: a stronger refining LLM or a stronger recognition MLLM should improve the resulting prompts without any change to the algorithm, a scaling prediction that could be tested by swapping either model.","Because winning prompts drift toward role-playing instructions (choreographer, emotional interpreter), the loop appears to discover personas as a way around the MLLM's majority-opinion default; making persona induction an explicit objective is a natural, testable extension."],"forward_implications":["Personalized emotion recognition becomes available for closed, API-only multimodal models: no parameter access, no loss gradients, and no soft-prompt embeddings are required.","The tuned prompts beat a zero-shot MLLM prompt (44.9% vs 39.1% accuracy) and a fixed initial prompt (40.6%), so the gain comes from the iterative refinement itself rather than from switching to an MLLM.","Prompts tuned to another user underperform user-specific prompts (CM6, 40.4%), so the tuned prompts carry genuinely individual information, not just a generically good instruction.","Running five optimized prompts and majority-voting outperforms a single tuned prompt by about 1.8 accuracy points (43.1% → 44.9%), a stability gain the authors attribute to averaging away per-prompt misclassifications.","The final prompts remain readable natural language with role instructions (e.g., 'as a choreographer expressing feelings through movement'), so what changed for each user can be inspected and edited by a human."],"supporting_citations":[{"why":"Supplies the core mechanism the paper adapts: using a language model as a black-box optimizer that iteratively rewrites prompts for vision-language models.","marker":"[29]"},{"why":"The zero-shot MLLM-based VER baseline (CM3) the method must beat, and the previous MLLM-emotion prompting approach it extends.","marker":"[18]"},{"why":"The Affection dataset, which provides per-viewer emotion labels and defines the personalized VER task and the 15-user evaluation.","marker":"[45]"},{"why":"Supplies the ECC and EMC evaluation metrics and the emotional-distance weighting used to measure recognition quality.","marker":"[51]"},{"why":"Supplies LLaVA-v1.6-Mistral-7B, the black-box MLLM used for emotion recognition throughout the experiments.","marker":"[49]"},{"why":"Provides the gradient-free prompt-refinement precedent (feedback-guided iterative prompt editing) motivating the evaluation-and-modification loop.","marker":"[31]"},{"why":"WSCNet, the CNN-based VER baseline (CM1) the proposed method must outperform in the comparison.","marker":"[36]"},{"why":"Swin Transformer, the Transformer-based baseline (CM2) used in the comparison.","marker":"[50]"}],"fun_headline_variants":["No-gradient prompt loop personalizes emotion AI","Recursive prompt rewrite tailors emotion AI per user","Discrete prompt rewriting personalizes black-box emotion AI","Prompt self-correction loop tailors emotion AI"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Each user's final prompts are chosen by ranking candidate prompts on recognition accuracy over a 30% training slice of that user's photos (about 260 images), and the method's effect rests on that ranking transferring to the user's remaining 70% of images — if high training accuracy reflects overfitting to the slice rather than stable user-specific signal, the reported gains over the ablations could be selection noise.","fun_headline_variants_meta":{"raw":{"variants":["No-gradient prompt loop personalizes emotion AI","Recursive prompt rewrite tailors emotion AI per user","Discrete prompt rewriting personalizes black-box emotion AI","Prompt self-correction loop tailors emotion AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001806,"raw_usage":{"total_tokens":6921,"prompt_tokens":692,"completion_tokens":6229,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":436,"completion_tokens_details":{"reasoning_tokens":6169}},"tokens_in":436,"tokens_out":6229,"duration_ms":47137,"temperature":1.0,"reasoning_tokens":6169,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:22:22.055293+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the full tuning loop on the same 15 Affection users with a different random 30/70 percent split per user. If the tuned prompts' held-out accuracy advantage over the fixed initial prompt (CM4) and the single tuned prompt without voting (CM7) does not reproduce across resplits — or if per-user gains do not exceed the reported between-user standard deviation of about 9.6 percentage points — the gains are artifacts of the training-slice ranking rather than true personalization.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the core mechanism the paper adapts: using a language model as a black-box optimizer that iteratively rewrites prompts for vision-language models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The zero-shot MLLM-based VER baseline (CM3) the method must beat, and the previous MLLM-emotion prompting approach it extends."},{"cited_title":"Achlioptas, M","cited_arxiv_id":null,"evidence_quote":"The Affection dataset, which provides per-viewer emotion labels and defines the personalized VER task and the 15-user evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ECC and EMC evaluation metrics and the emotional-distance weighting used to measure recognition quality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies LLaVA-v1.6-Mistral-7B, the black-box MLLM used for emotion recognition throughout the experiments."},{"cited_title":"Pryzant, D","cited_arxiv_id":null,"evidence_quote":"Provides the gradient-free prompt-refinement precedent (feedback-guided iterative prompt editing) motivating the evaluation-and-modification loop."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"WSCNet, the CNN-based VER baseline (CM1) the proposed method must outperform in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Swin Transformer, the Transformer-based baseline (CM2) used in the comparison."}],"review_version":1}