acceptodds
Under review as a conference paper at ICLR 2027

Can MLLMs Decode the Creative Leap? Introducing C4 for Cross-Concept Understanding

Abstract

Creative capabilities of MLLMs matter in design, communication, education, and human-AI collaboration, yet remain difficult to evaluate because explicit targets and reward signals are scarce compared with accuracy-oriented tasks. Cross-concept understanding is a core cognitive capacity underlying receptive creativity. It enables a perceiver to recover intended meaning from non-obvious but meaningful conceptual relations. We operationalize item construction as cross-concept encoding and model inference as cross-concept decoding. We introduce C4, a cognition-inspired evaluation framework for Chengyu (Chinese idiom)-based Cross-Concept Creativity. Its encoding component maps target slots to imageable substitute concepts along bridge paths in a manually annotated and third-party-reviewed cross-concept network, enabling batch generation with explicit structure, difficulty indexed by bridge count and depth, and exact answers. Using this framework, we instantiate the C4 Evaluation Set (C4-Eval), comprising 184 synthetic items and 37 human-created cross-concept chengyu figures collected from online sources. We manually construct and review cross-concept relations, bridge paths, and reasoning processes for the collected figures. Each C4-Eval item is instantiated in five task settings, yielding 884 primary answer-recovery cases. Across ten evaluated MLLMs, the strongest closed models reach 50.7% and 48.0% primary accuracy, while open-source models remain substantially lower. Candidate constraints improve accuracy sharply, but bridge hints and explanation requests provide only modest gains. These results expose a substantial gap in how current MLLMs decode creatively encoded meaning through cross-concept relations. Release details are provided in the appendix and code and dataset can be found at https://anonymous.4open.science/r/C4-06BF.

Primary area:
datasets and benchmarks
Keywords:
Cross-Concept, Creativity, Multimodal Large Language Models, Evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_sLdzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces C4, a cognition-inspired framework for evaluating cross-concept understanding in MLLMs by treating creative interpretation as the reversal of a cross-concept encoding process. Under this framework, a target Chinese four-character idiom is mapped through manually annotated associative bridges into indirect visual clues, and the model is asked to recover the original idiom. The authors build C4-Eval, a benchmark of 184 synthetic items and 37 web-collected human-created figures covering 84 idioms, organized into four difficulty levels and evaluated under five task settings. Experiments show a top accuracy of 50.7%, with candidate constraints substantially improving performance and open-ended answer-space search emerging as a major bottleneck.

Critical strengths
  1. The benchmark construction pipeline is explicit and structured. The paper makes a substantial effort to turn an otherwise subjective creative task into a reproducible evaluation protocol. The authors also distinguish image recognition, bridge reconstruction, and answer-space search through different task settings.
  2. The empirical analysis goes beyond a single leaderboard number. The candidate-recognition comparison and the paired T4/T5 analysis help localize model failures rather than just report overall accuracy. The large gap between candidate recognition and unconstrained recovery suggests that answer-space search is a key difficulty in this benchmark. The analysis of slot count versus bridge depth is also informative, though the authors rightly note these comparisons are not causal, since the difficulty levels differ in targets and images.
Critical weaknesses
  1. The construct validity of C4 as a benchmark of “creativity” is not sufficiently established.

The construct validity of C4 as a benchmark of “creativity” or “creative understanding” is not sufficiently established. The central conceptual motivation relies heavily on the associative account of creativity, particularly Mednick's notion of remote conceptual connections. However, the paper moves rather quickly from remote association to receptive creative understanding and then to a benchmark consisting almost entirely of Chinese chengyu decoding. Remote association is certainly relevant to creativity, but it is not synonymous with creativity as a broader construct, which also involves dimensions such as novelty, appropriateness, generation, flexibility, and contextual usefulness. The current evaluation therefore seems to establish that models struggle with a particular form of culturally and linguistically mediated associative decoding, rather than establishing a general deficiency in multimodal creativity.

The related-work discussion omits several directly relevant lines of research, weakening the paper's positioning of what C4 actually measures. First, it misses benchmarks of multimodal associative reasoning. [1]'s alternative operationalizations of this cognitive ability expose the key question of whether a benchmark captures general associative reasoning or merely a narrow task convention. [2] evaluates open-ended associative reasoning with human baselines, sensitivity analyses, and diversity across languages, cultures, and domains. Their inclusion of human comparisons and multi-lingual/cultural validation makes C4's lack of analogous validation especially consequential. Second, it overlooks work on associative reasoning in MLLMs such as [3] which explicitly studies this reasoning and its tension between creativity and faithfulness, and is therefore important for clarifying what capability C4 measures beyond existing evaluations of associative behavior. More broadly, the paper does not engage with recent work questioning the construct validity of applying human creativity tests to LLMs, such as [4] which shows that the link between a given test and a target creativity construct is highly test- and construct-dependent. This concern bears directly on C4: demonstrating that models struggle with one form of remote association does not by itself establish that C4 measures general multimodal creativity or "creative understanding". The authors should provide stronger construct validation and clarify the intended scope of their claims.

All targets in C4-Eval are Chinese four-character idioms, and the task design relies heavily on Chinese phonetic, lexical, cultural, and character-level associations. However, different languages have their own idioms, proverbs, fixed expressions, wordplay conventions, and culturally specific associative structures. It is unclear whether the reported failure reflects a general limitation of MLLMs in cross-concept understanding or difficulty with Chinese-specific linguistic and cultural encoding. The current evidence does not support the paper's broader framing of C4 as a general evaluation of creative understanding.

  1. The benchmark lacks a human baseline and does not sufficiently establish the naturalness or difficulty of its test items.

The paper evaluates ten MLLMs but does not establish how well humans can solve the same image-to-chengyu tasks under the same instructions. Consequently, a 50.7% score for the best model is difficult to interpret: it could indicate a substantial gap from human performance, near-human performance on a genuinely difficult puzzle, or simply that the benchmark contains highly artificial associations that are difficult even for humans.

Some of the intended associations appear highly non-obvious. Figure 5 relies on associations such as strawberry → grass and monk → wooden fish → wood. These are valid paths under the authors' annotation scheme, but it is not clear that they represent associations that humans would naturally infer from the image without substantial task-specific prompting or knowledge of the encoding convention. The benchmark may therefore measure the ability to reverse-engineer the authors' particular encoding system rather than a naturally occurring form of creative interpretation.

The exact evaluation prompts are not sufficiently exposed in the paper, and there is no systematic prompt-sensitivity study. The paper does report that T1-T4 use fixed prompts and that T1/T4 differences are used to examine prompt-form effects, but this does not establish robustness to alternative reasonable instructions. Given how unusual the task is, relatively small changes in instructions could plausibly have a large effect: for example, explicitly telling models to consider homophones, Chinese character decomposition, idiomatic expressions, or multi-hop associations could substantially change performance. The large improvement under candidate constraints already indicates that task formulation and answer-space restriction have a major impact. Without human baselines and prompt-sensitivity experiments, it is difficult to determine whether the reported low accuracy reflects a genuine model limitation or a benchmark-specific difficulty induced by the task interface.

[1] Li, Hong, et al. "The labyrinth of links: Navigating the associative maze of multi-modal llms." International Conference on Learning Representations. Vol. 2025. 2025.

[2] Huang, Zimeng, et al. "MM-OPERA: Benchmarking Open-ended Association Reasoning for Large Vision-Language Models." Advances in Neural Information Processing Systems 38 (2026).

[3] Lyu, Xinyu, et al. "Flexac: Towards flexible control of associative reasoning in multimodal large language models." Advances in Neural Information Processing Systems 38 (2026): 2473-25

[4] Schapiro, Samuel, et al. "Assessing the Creativity of Large Language Models: Testing, Limits, and New Frontiers." arXiv preprint arXiv:2605.13450 (2026).

Questions for authors
  1. How well do human participants perform on T1-T4 under exactly the same images and instructions? What is the human accuracy on image-only recovery, and how does it change with the bridge hint and candidate constraints?

  2. What are the exact prompts used for T1-T5? Have the authors tested reasonable alternative instructions, such as explicitly encouraging phonetic, semantic, cultural, or multi-hop associations?

  3. Is there evidence that the proposed formulation transfers beyond Chinese chengyu? If not, how do the authors justify interpreting C4 as an evaluation of general MLLM creative understanding rather than Chinese-specific cross-concept decoding?

  4. How does C4 differ conceptually and empirically from recent multimodal association benchmarks such as [1] and [2]? In particular, what capability does C4 measure that cannot already be captured by those evaluations?

3 comments
AuthorsOpenReview ↗

Response to Reviewer sLdz (Part 1/2): Construct Validity & Related Work

We sincerely thank Reviewer sLdz for the deep, scholarly review and the highly relevant literature pointers, as well as for recognizing our structured pipeline and multi-view empirical analysis. Due to OpenReview's character limit, we organize our response in two parts: Part 1 addresses construct validity and related work, and Part 2 provides the exact T1T_1–T5T_5 prompts and our ongoing prompt-sensitivity and human-baseline studies.

1. Calibrating construct validity and scope of claims

We fully accept your critique regarding construct validity. Moving directly from Mednick's associative theory to general multimodal "creativity" overstates what a Chinese chengyu decoding benchmark alone can establish, as creativity encompasses broader generative, flexible, and context-sensitive dimensions (Runco & Jaeger, 2012; Schapiro et al., 2026), and chengyu items intertwine conceptual bridging with Chinese phonetic, lexical, and cultural conventions.

  • Ongoing work & revision plan: We will recalibrate the title, abstract, introduction, and conclusion to frame C4 precisely as a benchmark for linguistically and culturally grounded cross-concept associative decoding—a specific receptive sub-skill relevant to figurative/creative interpretation rather than a general measure of multimodal creativity. Empirically, we are also compiling a breakdown across the seven relation families in Table S2 to separate Chinese-specific phonetic/character operations from language-general visual, object, and role associations, and we will discuss how the slot-and-bridge encoding schema transfers to visual puns and rebus expressions in other languages.

2. Positioning with respect to recent multimodal association and creativity literature ([1]–[4])

Thank you for pointing out these directly relevant works, which we will integrate into a dedicated subsection in Related Work:

  • Compared with Labyrinth of Links (Li et al., 2025) and MM-OPERA (Huang et al., 2026), which evaluate associative reasoning across broader tasks, open-ended links, and multilingual/cultural settings with human baselines, C4 focuses on controlled forward-encoded slot substitutions where multi-hop latent bridge paths (L1L_1–L4L_4) are explicitly known by construction and paired across unconstrained recovery (T1,T2,T4T_1, T_2, T_4), candidate recognition (T3T_3), and known-answer path explanation (T5T_5).
  • We will engage FlexAC (Lyu et al., 2026) when analyzing the tension between faithful cue grounding and associative leaps in T4T_4/T5T_5 rationales, and cite Schapiro et al. (2026) to ground our explicit discussion of construct boundaries when adapting associative paradigms to MLLMs.

(Continued in Part 2/2 below.)

AuthorsOpenReview ↗

Response to Reviewer sLdz (Part 2/2): Exact Prompts & Human Evaluation

(Continued from Part 1/2.)

3. Exact prompts for T1T_1–T5T_5 and prompt-sensitivity study

In our submission, the exact Chinese prompts were included in the supplementary repository (eval.jsonl), and we provide all five verbatim below (with English translations):

  • T1T_1 (Image-only):

    你会看到一张图。请猜它对应的中文成语。只输出一个成语,不要解释。 (You will see an image. Guess the corresponding Chinese chengyu. Output only one chengyu without explanation.)

  • T2T_2 (Bridge hint):

    你会看到一张图。这张图可能通过谐音、拆字、替换、角色联想、物品联想等跨概念方式表达一个中文成语。请根据图像线索猜成语。只输出一个成语,不要解释。 (You will see an image. This image may express a Chinese chengyu via cross-concept relations such as homophony, character decomposition, substitution, role association, or object association. Guess the chengyu from the visual clues. Output only one chengyu without explanation.)

  • T3T_3 (Four-way candidates):

    你会看到一张图。它通过跨概念方式表达下面候选成语之一。 A. [Cand 1] B. [Cand 2] C. [Cand 3] D. [Cand 4] 请只输出最可能的成语,不要输出选项字母,不要解释。 (You will see an image expressing one of the following candidate chengyu via cross-concept relations. [A–D] Output only the most likely chengyu, without option letters or explanation.)

  • T4T_4 (Free answer + JSON rationale):

    你会看到一张图。这张图是某个中文成语的跨概念创意表达。请先给出你认为的成语,再解释你如何从图像线索推到这个成语。必须返回 JSON,不要输出 JSON 以外的文字,格式如下: (You will see an image. This image is a cross-concept creative expression of a Chinese chengyu. First give the chengyu you believe it is, and then explain how you infer this chengyu from the visual clues. You must return valid JSON and output no text outside the JSON, in the following format:)

    {
      "answer": "成语 (chengyu)",
      "perceptual_cues": ["你看见的关键图像线索 (key visual clues you observe)"],
      "bridge_analysis": [
        {
          "visual_or_textual_cue": "线索 (cue)",
          "bridged_concept": "被联想到的概念 (associated concept)",
          "idiom_part": "对应的成语片段 (aligned chengyu fragment)"
        }
      ],
      "final_rationale": "一句话说明整体推理 (one-sentence summary of the overall reasoning)"
    }
    

    (Note: parenthetical English glosses above are added here for readability; the model-facing prompt contains only the Chinese placeholder strings.)

  • T5T_5 (Gold answer + JSON rationale):

    你会看到一张图。已知它表达的成语是“{gold_chengyu}”。请解释图像中的关键线索如何通过跨概念桥接到这个成语。必须返回 JSON,不要输出 JSON 以外的文字,格式如下: (You will see an image. Given that the chengyu it expresses is "{gold_chengyu}", explain how the key clues in the image bridge across concepts to this chengyu. You must return valid JSON and output no text outside the JSON, in the following format: [followed by the exact same JSON schema as T4T_4, with "answer" pre-filled with "{gold_chengyu}"].)

Note that T2T_2 already explicitly instructs models to consider homophones, character decomposition, substitution, and role/object associations, yielding modest differences (−2.2-2.2 to +6.8+6.8 percentage points over T1T_1). We are running broader prompt-sensitivity experiments (step-by-step multi-hop CoT instructions, English vs. Chinese prompts, and text-only few-shot demonstrations) and will move all exact prompts and sensitivity results into the revised manuscript.

4. Independent human baseline and item naturalness

We completely agree that examples such as strawberry -> grass or monk -> wooden fish -> wood require empirical verification of whether independent humans naturally infer them without prior exposure to our annotation convention. We are actively conducting an independent human baseline across T1T_1–T4T_4 under identical instructions alongside human naturalness/plausibility ratings, and we will update the revised manuscript with results both on the full cleaned set and stratified by human recoverability.

Thank you for your detailed response. I appreciate your acknowledgment of the construct-validity concerns. However, the main concerns remain unresolved. The human baseline and item-naturalness evaluations are still ongoing. Thus, I maintain my score.

Reviewer_bQkUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores a very interesting Chinese creativity situation and proposes a benchmark to evaluate existing model creativity.

Critical strengths
  • The proposed task is valuable.
  • The performance gap between multiple-choice and open-ended provides meaningful insight.
Critical weaknesses
  • Since the main contribution of this paper is to propose a benchmark, the authors should pay more attention to the benchmark validity. The authors have illustrated that every conception path is reasonable for humans; however, it does not show that humans can answer correctly. The authors should provide human evaluations about it. For example, the authors could recruit some human evaluators to do this benchmark and evaluate the performance. This can distinguish the model's ability and question itself.
  • The authors claim that 'Open answer-space search is the dominant bottleneck' because the performance of T3 is much better than T1. However, I think the results could not fully support it. I think T3 not only reduces the answer space, but also provides external semantic information from the candidate choices.
  • The results of Qwen3.5 show that the format of the answer, the final answer extraction, would influence the performance. (The accuracy of Qwen is 22.2%; however, 152 out of 221 instances already give the answer segment.)
Questions for authors

please see the weakness above

1 comment
AuthorsOpenReview ↗

Response to Reviewer bQkU

We thank Reviewer bQkU for the insightful comments and for recognizing the value of our task and the open-vs.-candidate diagnostic comparison. Below we clarify the Qwen3.5-9B result with existing diagnostic data and describe the additional experiments we are actively working on for the revised manuscript.

1. Clarification and diagnostic analysis on Qwen3.5-9B (22.2%22.2\% vs. 152/221152/221)

First, we would like to clarify a table-column detail in Appendix G (Table S5): 22.2%22.2\% (49/22149/221) is Qwen3.5-9B's accuracy on candidate-constrained T3T_3, whereas 152/221152/221 (68.8%68.8\%) is the count for T5T_5, where the gold chengyu is already supplied in the prompt and the model only needs to output a JSON explanation (in the remaining 69 T5T_5 cases, Qwen3.5-9B exhausted its native 262k context window during reasoning, finish_reason=length, before emitting the final JSON).

Second, we completely agree with your broader point that answer format and reasoning-trace termination affect thinking models. In fact, we conducted a post-hoc diagnostic extraction on Qwen3.5-9B that recovers the last explicit answer candidate from the reasoning trace whenever no final output was emitted:

Metric Policy (Qwen3.5-9B)T1T_1T2T_2T3T_3T4T_4Primary (T1T_1–T4T_4)
Strict Final Answer (Main Table)6.3% (14/221)4.1% (9/221)22.2% (49/221)4.1% (9/221)9.2% (81/884)
Post-hoc Reasoning Extraction (Diagnostic)12.7% (28/221)11.3% (25/221)34.8% (77/221)10.0% (22/221)17.2% (152/884)

Extracting answers from unfinished reasoning traces raises Qwen3.5-9B's primary score from 9.2%9.2\% to 17.2%17.2\%, confirming your observation that strict output parsing penalizes over-reasoning. Crucially, however, even at 17.2%17.2\%, Qwen3.5-9B remains within the open-model tier (≤18.1%\le 18.1\%) and far below top closed models (50.7%50.7\%), and the large gap between T3T_3 (34.8%34.8\%) and open recovery (10.0%10.0\%–12.7%12.7\%) still holds. We will include this side-by-side diagnostic comparison in the revised paper.

2. Disentangling search-space reduction vs. semantic cues in T3T_3

This is a very sharp observation. Although we selected T3T_3 distractors that share characters with the gold chengyu (Appendix E) to reduce trivial elimination, we agree that the four candidate strings themselves inject external semantic anchors that help models align visual cues backward, rather than solely shrinking the search space.

  • Ongoing work & revision plan: We will refine our wording in RQ2 so as not to attribute the T3T_3–T1T_1 gap purely to search-space size, and we are conducting controlled ablations varying candidate pool size (e.g., K∈{4,16,84}K \in \{4, 16, 84\}) alongside structural/slot-masked hints without full candidate idioms to update in the revised manuscript.

3. Independent human evaluation

We fully agree that human solvability cannot be assumed solely from forward path plausibility. We are actively conducting an independent human study on T1T_1–T4T_4 under identical instructions and will update the revised manuscript with these human reference results if given the opportunity.

Reviewer_jLPnOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces C4, a framework for evaluating cross-concept understanding in multimodal language models through Chinese idiom image puzzles. Its evaluation set contains 184 synthetic images constructed using annotated association paths and 37 human-created images collected online. Ten models are evaluated across five task settings, covering open-ended answering, candidate-based recognition, and explanation. The results show low open-ended accuracy and substantial gains when candidate answers are provided, which the authors interpret as evidence of difficulties in cross-concept decoding and open answer-space search.

Critical strengths

1. Explicit paths provide a useful basis for analysis. Beyond image–answer pairs, the dataset records target fragments, substitute concepts, and the paths used to construct the images. These annotations make the construction process inspectable and could support more detailed analysis of relation-specific difficulties and errors in path reconstruction.

2. The task enables useful comparisons across different levels of support. The five settings cover open-ended answering, general hints, candidate-based recognition, and explanation with or without the answer provided. This design offers a broader view of model behavior than a single accuracy score.

Critical weaknesses

1. I am concerned about how much its scores reflect general cross-concept understanding rather than specific skills such as Chinese idiom knowledge, homophone substitution, and character or word manipulation. Appendix C, Table S2 distinguishes seven relation types, but the paper does not report performance by type. It is therefore unclear whether the low overall scores are driven primarily by difficulties with particular relations. For example, models might struggle with homophone substitutions while performing better on other relations. The current experiments do not distinguish these possibilities and thus do not sufficiently support attributing low overall scores to a general deficit in cross-concept understanding.

2. The paper does not report performance by independent human participants. Incorrect model answers do not necessarily indicate a capability deficit. They may instead reflect insufficient visual clues or ambiguous associations. An explicit forward construction path does not guarantee that a solver can recover it from the final image. Independent human evaluation would provide an important reference, including chengyu recovery without the answer, path recovery given the answer, and judgments of the intended paths’ plausibility. If participants generally fail to recover the answers or paths and substantially disagree on the plausibility of the intended associations, the items’ recoverability requires further validation. Given the limited dataset size, I consider such an evaluation necessary to establish the benchmark’s validity.

3. I am concerned about the quality control of the final images. My limited inspection of the supplementary materials revealed two clear cases of answer leakage: syn_l1_a512e3d3e7bd and syn_l4_c627d90a0972. Both images explicitly contain their answers in the image text, allowing models to answer through OCR without cross-concept decoding. These cases show that the review process did not fully eliminate answer leakage. The authors should systematically audit all final images, report the affected items and how they are handled, and recompute the evaluation results after cleaning the dataset.

Questions for authors
  1. Please address the main concerns raised in the Critical Weaknesses section.

  2. What were the sources and selection criteria for the 37 web-collected figures? Given the abundance of image-based chengyu puzzles online, why does this subset constitute such a small portion of the evaluation set?

2 comments
AuthorsOpenReview ↗

Response to Reviewer jLPn

We thank Reviewer jLPn for the careful inspection of our dataset and constructive feedback, as well as for recognizing the value of our explicit association-path annotations and five-setting diagnostic design. We are actively conducting the full image audit and additional evaluations outlined below, and we will incorporate all updated results and analyses into the revised manuscript if given the opportunity to revise.

1. Answer leakage in rendered images (syn_l1_a512e3d3e7bd and syn_l4_c627d90a0972)

Thank you for catching these two cases. While we manually reviewed every text scene description prior to rendering to eliminate direct mentions of the gold chengyu (Section 3.2), the image generator (GPT Image 2) occasionally hallucinated unprompted Chinese text—such as speech bubbles or decorative inscriptions containing target characters—into the rendered image, and these two instances slipped past our post-rendering check.

  • Ongoing work & revision plan: We have initiated a systematic two-stage audit (OCR text extraction + dual-annotator visual inspection) across all 221 final images to identify any image containing target characters or unintended textual hints. In the revised manuscript and dataset release, we will report the full list of affected items, replace leaked synthetic images with clean text-free regenerations, and update all evaluation tables (alongside a sensitivity comparison excluding those items).

2. Performance breakdown by the seven relation families (Table S2)

We agree that aggregate scores alone do not separate difficulties in homophone/character manipulation from conceptual, visual, or cultural bridging.

  • Ongoing work & revision plan: Because every synthetic item records its exact slot-to-substitute bridge_chain, we are aligning all item paths with the seven relation families in Appendix C (Table S2) to add a fine-grained breakdown of T1T_1–T5T_5 accuracy by relation family (specifically contrasting phonetic/lexical operations against semantic, object, role, part–whole, and cultural relations), and we will refine our claims accordingly so we do not overgeneralize language-specific errors.

3. Independent human evaluation

We fully agree that reviewed forward construction paths do not automatically guarantee reverse recoverability by independent solvers.

  • Ongoing work & revision plan: We are organizing an independent human evaluation with native speakers who had no involvement in dataset construction, using the exact same images and T1T_1–T4T_4 instructions given to models, supplemented by post-answer path-recovery and plausibility ratings to establish item-level recoverability in the revised version.

4. Sources and selection criteria for the 37 web-collected figures

Although image-based chengyu puzzles are abundant online, the vast majority violate our cross-concept formulation (Eq. 4): most are either (i) literal depictions of an idiom's surface story rather than cross-concept substitutions, (ii) multi-panel collages or pure homophonic text memes, (iii) low-resolution images with embedded answer text or watermarks, or (iv) widely circulated puzzles prone to pretraining memorization. We therefore applied strict filtering to retain 37 single-panel, genuinely cross-concept human illustrations as an ecological anchor, while relying on the 184 bridge-controlled synthetic items (L1L_1–L4L_4) for systematic stratification across slot count and bridge depth. We will document the exact sources, filtering criteria, and exclusion statistics in Section 3.3 and Appendix D of the revised manuscript.

Thank you for your response and for acknowledging my concerns. I agree with your proposed plan, but correcting the dataset, updating many results reported in the paper, and completing key experiments, particularly the human evaluation that several reviewers called for, would require more work than can reasonably be completed in the current review cycle. Therefore, my recommendation remains unchanged.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Area Chair and All Reviewers

Dear Area Chair and Reviewers,

We sincerely thank Reviewers jLPn, bQkU, and sLdz for their constructive, detailed, and insightful reviews. We are encouraged that all reviewers recognized the value of C4's explicit bridge-path annotations and five-setting (T1T_1–T5T_5) diagnostic design for localizing model failures beyond a single accuracy score.

In our individual responses, we have:

  1. Clarified specific questions using existing artifacts and diagnostic results, including:

    • The exact verbatim T1T_1–T5T_5 evaluation prompts (Reviewer sLdz);
    • The strict filtering criteria behind the 37 web-collected figures (Reviewer jLPn);
    • The Table S5 distinction between T3T_3 (22.2%22.2\%) and T5T_5 (152/221152/221), along with our post-hoc reasoning-extraction diagnostic showing Qwen3.5-9B moves from 9.2%9.2\% to 17.2%17.2\% primary accuracy while leaving all cross-model and cross-task conclusions intact (Reviewer bQkU).
  2. Initiated additional audits, human evaluations, and analyses, which we are actively carrying out and will incorporate into the revised manuscript whenever the revision opportunity permits:

    • Full image audit & dataset cleaning (Reviewer jLPn): Auditing all 221 rendered images via OCR + dual human review to eliminate unprompted generator text hallucinations (such as the two leaked items identified by jLPn), regenerating clean images, and recomputing all results.
    • Independent human evaluation (Reviewers jLPn, bQkU, sLdz): Measuring independent human accuracy on T1T_1–T4T_4 under identical instructions, plus human plausibility and path-recoverability ratings of the intended bridge paths.
    • Construct scope, fine-grained relation breakdown, and related work (Reviewers sLdz, jLPn, bQkU): Recalibrating our claims from general "creativity" to linguistically and culturally grounded cross-concept associative decoding, reporting accuracy across the seven relation families in Table S2, adding ablations on candidate-pool size vs. semantic priming and prompt sensitivity, and integrating recent literature on multimodal association and creativity construct validity (Li et al., 2025; Huang et al., 2026; Lyu et al., 2026; Schapiro et al., 2026).

We deeply appreciate the reviewers' time and guidance, and we will gladly share any interim experimental updates should the discussion period allow.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.