Weakness 1. Data quality: The construction of the steering vector depends on hallucinated input-response pairs, which are derived via blurred inputs and specific prompting. This may introduce distributional biases, e.g., capturing only specific hallucination types. Besides that, top-k selection can be sensitive to outlier data. Some robustness or sensitivity analysis of the vector construction would strengthen the approach.

Responses:

• Hallucination Distribution: We analyzed our data and confirmed it contains a rich mixture of both object-level (86.35%) and relation-level (94.8%) hallucinations, ensuring it is not biased toward a specific type.
• Sensitivity Analysis of Top-K: Our method demonstrates robustness in two key aspects: the top-k selection is not sensitive to outliers, and the optimal k value is generalizable across datasets.
  • Generalizability of k=50: As shown in the table I and our original appendix (Figure. 17), k=50 is a robust choice that delivers near-peak performance across all tested benchmarks (CHAIR, VDAT, MME, HallusionBench), making it a reliable and practical default setting.
  • Robustness to Outliers: The method is not sensitive to including a large number of instances (and potential outliers). For instance, on MME (Perception), performance degrades gracefully from the peak at k=50 (1427.8) to k=2000 (1390.0), rather than collapsing. This confirms the robustness of our vector construction.

Table I: Sensitivity analysis of Top-K selection.

---

Weakness 2. Connection between hallucination vector and associative reasoning: While the extracted steering vector clearly controls the factual grounding, it's less certain whether it directly reflects a model’s associative reasoning capability. The paper seems to assume that hallucinated features correspond to associative reasoning, but this isn't explicitly validated. For instance, in Figure 9 (b)–(c), FlexAC-P (presumably more powerful on associative reasoning) performs similarly or worse than FlexAC-C on reasoning-heavy tasks like MMMU and Math. Therefore, it seems to me this method is for hallucination control instead of associative reasoning control.

Responses: We believe there may be a critical misunderstanding of our notation which, once clarified, resolves the core concern.

• A Crucial Clarification: The Reviewer's Premise on P vs. C is Inverted. The reviewer’s analysis mistakenly assumes that FlexAC-P enhances associative reasoning, which contradicts our definition. In fact, the correct premise is as follows: FlexAC-C (Creativity, α = 1) is designed to ENHANCE associative reasoning; FlexAC-P (Precision, α=-1) is designed to SUPPRESS it.

• Re-interpreting Figure 9 with the Correct Premise： With the correct premise, the data in Figure 9 now strongly supports our claim. It shows that FlexAC-C (the creativity-enhanced version) maintains stable performance on logical reasoning tasks (MMMU/Math). This is a key feature: our method precisely targets associative reasoning without disrupting the model's separate, core logical reasoning capabilities.

• Direct Validation on Creativity Tasks. The fact that our vector controls more than just hallucination is validated by FlexAC-C's strong performance on creativity-focused benchmarks: It generates abstract, metaphorical content on Creation-MMBench. It improves divergent thinking scores on VDAT, a result confirmed by our human study (Appendix, Sec. 4.1).

---

Weakness 3. Model Architecture: The empirical analysis on layer-wise impact on associative reasoning are demonstrated only on LLaVA model. It remains unclear whether the same behavior holds for other architectures, particularly mixture-of-experts (MoE) models.

Responses: Our submitted appendix (App. 6.4, Fig. 30) already provides feature analysis for Qwen-VL and the MoE-based DeepSeek-VL2. To provide even more direct evidence, the following intervention analysis for DeepSeek-VL2 shows each layer's impact; a larger '$\Delta$' value signifies a greater influence on the final result:

The table's "$\Delta$" column peaks at Layer 6 (-1.4), confirming our paper's finding that the middle layers (4-7) are the critical control region. This result from the MoE model is consistent with our main analysis. The full plots will be added to the camera-ready version.

---

Question 1. Since the non-middle layers show minimal change on hallucination, have the authors explored applying steering vectors to all layers as an ablation?

Responses: Per your advice, we conducted an ablation study applying the steering vector to all layers.

Table II: Comparison of Middle-Layer vs. All-Layer Intervention in LLaVA.

While the 'all-layers' intervention drastically reduces CHAIR scores, it comes at the cost of catastrophic damage to the model's generative ability. This is evidenced by the collapse in performance (e.g., LLaVA's Recall dropped from 75.0 to 7.1) and the frequent generation of meaningless outputs like empty strings or gibberish. This result confirms that all-layer intervention is harmful, thus validating our design choice to precisely target only the middle layers. The possible reason is that all-layer intervention disrupts both low-level representations and high-level semantics, leading to unstable decoding and degraded output.

---

Question 2. Does FlexAC operate solely on the language-side layers, or is the vision encoder also considered? A brief discussion would clarify the scope of parameter mixing

Responses: FlexAC operates solely on the language-side layers, a deliberate choice based on the standard MLLM architecture.

In MLLMs, the vision encoder functions as the perceptual "eyes," while the LLM is the cognitive "brain" responsible for reasoning[1]. Since hallucination and creativity are cognitive phenomena that occur in the "brain," our intervention is precisely targeted at the LLM's layers where this reasoning takes place

[1] A survey on multimodal large language models (National Science Review 2024)

---

Question 3. Equations (7) and (9) suggest both unnormalized and normalized vector mixing. Given the presence of LayerNorm in the transformer, can the authors comment on whether there are redundant computations involved?

Responses: No redundant computation is involved. Our normalization (Eq. 9) and the transformer's LayerNorm serve distinct, complementary purposes:

• Our Eq. 9 (Norm Preservation):
  • Target: A single hidden state vector after we have steered it.
  • Purpose: To reset the steered vector's magnitude to match the original vector's magnitude. This is a stability measure ensuring our intervention only changes the representation's direction, not its strength. This is a common practice in activation steering[1,2,3].
• Transformer's LayerNorm:
  • Target: The entire set of all token vectors in the sequence.
  • Purpose: To rescale the distribution of this entire group of vectors (e.g., to zero mean, unit variance) to ensure a stable input for the next network layer.

[1] Reducing hallucinations in large vision-language models via latent space steering (ICLR 2025)

[2] Steering away from harm: An adaptive approach to defending vision language model against jailbreaks(CVPR 2025)

[3] Safety arithmetic: A framework for test-time safety alignment of language models by steering parameters and activations(EMNLP 2024)

---

Question 4. While most prior efforts on steerable LLMs focus on trainable approaches, they are still relevant and should be discussed in the related work section. Including these works would help position FlexAC more clearly within the broader landscape of controllable and steerable LLM/MLLMs.

Responses: We agree and will update the "Related Work" section to better position FlexAC relative to trainable steering methods. The discussion will include:

• Preference-Based Alignment: Methods like Reinforcement Learning from Human Feedback (RLHF)[1] and Direct Preference Optimization (DPO)[2], which fine-tune the model on preference data.
• Targeted Fine-tuning: Approaches like CogSteer[3], which fine-tunes lightweight adapter modules in specific middle layers to steer behavior.
• Auxiliary Module Training: Frameworks that train a separate module to guide the LLM, such as SCAV[4] (trains a linear classifier to find a steering vector) or TSV[5] (trains a lightweight vector to be added during inference).

[1] Training language models to follow instructions with human feedback (NeurIPS 2022)

[2] Direct preference optimization: Your language model is secretly a reward model (NeurIPS 2023)

[3] Cogsteer: Cognition-inspired selective layer intervention for efficiently steering large language models. (ACL Finding 2024)

[4] Uncovering Safety Risks of Large Language Models through Concept Activation Vector (NeurIPS 2024)

[5] Steer LLM Latents for Hallucination Detection (ICML 2025)

W1 (Data quality): Thank you for the clarification; this has resolved my initial concern. The empirical analysis on the sensitivity of top-k is convincing and demonstrates the robustness of your results for k>1. I suggest you incorporate the specific breakdown of hallucinations ('object-level: 86.35%' and 'relation-level: 94.8%') into the final paper, as it adds valuable detail.

W2 (FlexAC-C vs. FlexAC-P Performance): I appreciate the clarification on the distinction between FlexAC-C and FlexAC-P. However, the question remains regarding the performance divergence. Given that FlexAC-C maintains stable performance on logical reasoning tasks like MMMU and Math, why does it underperform on the MMStar logical reasoning benchmark? Furthermore, the performance comparison between P v.s. C in the "Science & Technology" and "Instance Reasoning" categories are also notable. Could you provide a systematic discussion on these discrepancies between the 'C' and 'P' variants as shown in Figure 9?

Furthermore, my concern about the conceptual link between the 'hallucination vector' and 'associative reasoning' persists. 'Associative reasoning' is a broad cognitive concept that is difficult to formalize, and I am not convinced that it can be directly equated with the identified hallucination vector. The paper would be stronger if this claim were either better substantiated or framed more cautiously.

W3 (MoE architecture analysis): The additional analysis on MoE is clear and addresses my questions. Thank you.

Q1 (Generative Ability): The observation that intervention across all layers leads to catastrophic damage to the model's generative ability is an important finding. I recommend including this analysis in the final version of the paper

Q2 & Q3 & Q4: Thank you for the clarifications; my questions on these points are resolved.

• 1.2 For tasks that benefit from flexible matching and creative problem-solving, FlexAC-C performs better than FlexAC-P.:The solution paths for these tasks are often not fixed and may require connecting complex visual patterns to abstract concepts or discovering non-obvious "shortcuts."

  • Tasks: 'Coarse/Fine-grained Perception', 'Math'.

  • Why does FlexAC-C perform better?: FlexAC-C enhances associative abilities to promote "divergent thinking." This allows the model to think outside the box, connecting seemingly unrelated visual elements to form a meaningful whole or devising novel approaches to solve a problem.

  • Why does FlexAC-P perform worse?: FlexAC-P's high focus becomes a form of "cognitive rigidity" in this context. It may over-focus on the literal, local details of an image while failing to grasp its holistic, abstract meaning. It can also get stuck on standard solutions, unable to adapt flexibly.

  • Specific Task Examples:

    • 'Coarse/Fine-grained Perception':
      • Task Requirement: Identify an object's abstract category (e.g., car brand, painting style) based on visual cues.
      • Scenario: Judge a car's brand from a blurry close-up photo of only its front grille.
      • How FlexAC-C (Creativity Mode) succeeds: It can creatively match multiple weak cues—the blurry shape of the grille, the faint outline of its pattern, the reflection of light—with a wide net of associations in its knowledge base, like "BMW's kidney grille" or "Audi's single-frame grille." It synthesizes these weak signals into a correct identification even with incomplete information.
      • Why FlexAC-P (PrecisionMode) fails: It over-focuses on the image's physical details, asking, "How many pixels are black?" or "What is the exact curvature of this line?" Because the image is blurry, these precise details are useless. By suppressing association, it cannot make the conceptual leap from these imperfect local features to the abstract concept of a brand, likely concluding "unknown" due to the lack of a precise match.
    • 'Math':
      • Task Requirement: Solve visual math problems, especially geometry tasks that require a "flash of insight."
      • Scenario: Calculate the area of an irregular shape in a geometry problem.
      • How FlexAC-C (Creativity Mode) succeeds: Through association, it might recall strategies of "cutting" and "rearranging." It could creatively try adding an "auxiliary line" to the diagram, decomposing the irregular shape into a simple rectangle and a triangle. This act of adding a line is a non-linear, creative problem-solving step.
      • Why FlexAC-P (Precision Mode) fails: It would see the irregular shape and try to directly apply the standard area formulas it knows (for rectangles, circles, etc.). Finding that none apply, it gets stuck. Its focused nature makes it difficult to perform "destructive" modifications on the original image (like adding a line) because it prefers to analyze the given visual information as-is, thereby missing the key to the solution.

In summary, the performance divergence on MMStar is not a defect but a demonstration of the fundamental trade-off between creative, divergent thinking and precise, convergent reasoning. We will add this detailed analysis to the final manuscript.

Thank you for your thoughtful review and for acknowledging that our previous rebuttal has resolved your concerns regarding data quality (W1), MoE architecture analysis (W3), and your other questions. We are greatly encouraged by your positive feedback.

We would like to offer further clarification on the two insightful final points you raised.

1. Systematic Discussion on the Performance of 'C' and 'P' Variants on MMStar

We appreciate this detailed question, as it allows us to present a more systematic discussion. As shown in Figure 9, the performance divergence between FlexAC-C and FlexAC-P is an insightful and, indeed, an expected outcome of our work. The performance of each variant is directly tied to how well its reasoning style aligns with the demands of a specific task.

• 1.1 For tasks requiring highly focused, convergent thinking, FlexAC-P performs better than FlexAC-C: What these tasks have in common is that they typically have a single, objective correct answer that must be reached through rigorous, step-by-step reasoning. The model must suppress irrelevant associations and focus strictly on the given visual evidence.

  • Tasks: 'Logical reasoning', 'Instance reasoning', 'Science & technology'.

  • Why does FlexAC-P perform better?: FlexAC-P suppresses association to enhance focus. This allows the model to act like a meticulous analyst, strictly following instructions and logical chains. It concentrates on the most direct evidence at hand, excelling in tasks that demand precision and objectivity.

  • Why does FlexAC-C perform worse?: FlexAC-C's strength lies in activating a broad network of associations, which becomes a form of "cognitive noise" here. It can introduce irrelevant, distracting ideas that cause the model to deviate from the correct logical path and make flawed judgments based on these unrelated associations.

  • Specific Task Examples:

    • Instance reasoning:

      • Task Requirement: Precisely compare specific attributes of two or more objects (e.g., quantity, color, size).
      • Scenario: The question is, "Is the number of yellow squares in Image A greater than in Image B?"
      • How FlexAC-P (Precision Mode) succeeds: It strictly executes the instructions: 1) Locate all squares in Image A; 2) Filter for the yellow ones; 3) Count them. It repeats this process for Image B and then compares the numbers. Its attention is locked onto the core attributes of "yellow" and "square."
      • Why FlexAC-C (Creativity Mode) fails: When seeing the yellow squares in Image A, it might associate them with "cheese" or "sunshine." When viewing the squares in Image B, their arrangement might make it think of "buildings" or "Tetris." These irrelevant associations interfere with the core task of counting, leading to miscounts or a flawed comparison.

    • Logical Reasoning:

      • Task Requirement: Follow step-by-step reasoning based on a chart, code, or a set of visual rules.
      • Scenario: A flowchart instructs: "If the shape is a circle, follow the red arrow; if it's a square, follow the blue arrow."
      • How FlexAC-P (Precision Mode) succeeds: It meticulously executes the logical judgment: 1) Identify the current shape as a "circle." 2) Look up the rule corresponding to "circle." 3) Strictly follow the instruction to "follow the red arrow."
      • Why FlexAC-C (Creativity Mode) fails: It is prone to "over-interpretation." Seeing a red-colored circle, it might incorrectly force an association between the object's color and the "red arrow," even though the rule only specifies shape. Alternatively, it might observe the overall layout of the flowchart and guess the path based on visual "harmony" or "symmetry" rather than the explicit rule, causing the reasoning to break down.

    • Special Case: (Science & technology) Performance is on par here because the task requires both precise visual observation (benefiting from P-mode's focus) and the flexible connection of phenomena to abstract scientific principles (a divergent skill). Therefore, the 'P' and 'C' variants have similar performance.

Weakness 1. Associative Strength Spectrum Is Oversimplified.

Responses: We agree this is a simplification, but argue it's an effective operational model rather than an oversimplification.

• Theoretical Grounding: Inspired by neuroscience research linking creativity and hallucination[1], we propose an 'associative strength' spectrum to model this connection. Drawing on theories of convergent and divergent thinking[2], we frame faithfulness as low associative strength (convergent) and creativity as high (divergent), with hallucination seen as an extreme form of uncontrolled divergence
• Empirical Validation: Critically, Fig. 3 validates this model. By adjusting a single coefficient, α, we make it success to control both hallucination (CHAIR: 38.8→53.6) and associative strength (VDAT: 83→87.9). This provides strong empirical evidence that the Associative Strength Spectrum is an effective and practical abstraction for steering between creativity and hallucination
• Beyond a Single Spectrum: Our framework goes beyond a unidimensional two-end simplification. We acknowledge the existence of diverse creative trajectories between creativity and hallucination. To capture this complexity, we introduce task-specific associative vectors that guide the model along multiple creative pathways, thereby addressing the inherently multidimensional nature of creativity

[1] To create or to recall? Neural mechanisms underlying the generation of creative new ideas(NeuroImage 2014)

[2] The standard definition of creativity (Creativity research journal 2012)

---

Weakness 2. This work is essentially a direct application of existing representation steering concepts to MLLMs.

Responses: While our work utilizes established steering principles, it introduces two novel contributions specifically for controlling associative reasoning in MLLMs:

• A Novel Control Signal: We are the first to discover that the direction of hallucinated representations can be used to steer associative reasoning. This offers a new perspective on the relationship between hallucination and creativity.
• Multi-Directional Steering for Creativity: Recognizing the multi-dimensional nature of creativity, we introduce Directional Integration. This technique combines both general $v_{l}^{gen}$ and task-specific vectors $v_{l}^{task}$ in Eq.7 to adapt to varied creative tasks, a significant advance over conventional single-vector steering

---

Weakness 3. VDAT is a narrow and potentially misleading proxy for divergent thinking. A model could achieve a high score by simply ignoring the visual input, which undermines the goal of image-grounded creativity.

Responses:

• VDAT and Divergent Thinking: VDAT is a principled, multimodal adaptation of the validated Divergent Association Task (DAT), a paradigm from cognitive science[1]. Divergent thinking, a core component of creativity, is defined as the ability to generate diverse ideas from a single stimulus[2]. Research has established that the semantic distance between these ideas serves as a robust, objective measure of this ability[1,3]. VDAT operationalizes this principle by using an image as the stimulus.
• High VDAT ≠ Ignoring the Image: Outputs with high VDAT often include objects that are relevant to the image's scene but are not visually present. As shown in Fig. 23 in Appendix, an indoor cooking scene includes "chair" and "table"—typical kitchen items not pictured.
• Consistency Performance with Creation-MMBench: For a holistic assessment, we rely on Creation-MMBench[4], which evaluates overall image-grounded creative generation in terms of VFS scores. Higher associative strength of our FlexAC on VDAT (Table 2) directly translates to higher-quality, image-grounded creative generation on Creation-MMBench[4] (Table 3) with high VFS scores. This consistency validates that VDAT effectively measures a fundamental mechanism contributing to broader creative intelligence.

[1] Naming unrelated words predicts creativity (National Academy of Sciences 2021)

[2] Creativity: Yesterday, today and tomorrow (Journal of Creative Behavior 1967)

[3] Probing the “creativity” of large language models: Can models produce divergent semantic association (EMNLP 2023)

[4] Creation-MMBench: Assessing Context-Aware Creative Intelligence in MLLM (ICCV 2025)

---

Question 1. Does the Vector Reflect Creativity or Context-Ignorance? Why Not Use Human Data?

Responses:

1. On Why Hallucination is a Valid Signal for Creativity, Not "Context-Ignorance"
   • Our use of hallucinations as a signal source is grounded in cognitive science: studies show that the brain regions involved in creative thinking significantly overlap with those that produce hallucinations[1]. This suggests hallucinations may reflect an intensified form of divergent thinking—semantically rich, contextually relevant, yet unconstrained by reality
   • The ultimate proof is empirical. If the vector merely captured "context-ignorance," it would degrade performance on functional creative tasks. Instead, FlexAC achieves a +10.92 Reward on Creation-MMBench, while competing methods are negative (Table 3). This provides strong evidence that our vector is steering towards genuine, functional creativity
2. On Using Human-Generated Content: We did experiment with using human-written captions to extract creativity vectors. Annotators were instructed to “be creative,” resulting in diverse styles—narrative, humorous, metaphorical, and anthropomorphic. While imaginative, this diversity introduced high variance, making the averaged vectors noisy and inconsistent.

As shown in Table I, models guided by these human-derived vectors performed worse (Reward: -22.25), indicating their instability as guidance signals. In contrast, hallucination-guided vectors, though model-generated, provided more consistent and controllable cues, leading to improved image-grounded creativity. This may be because human creativity often relies on loosely grounded rhetorical styles that are less suitable for generating coherent guidance.

Table I: Performance of human-generated vs. hallucination-derived steering vectors on Qwen-VL

[1] To create or to recall? Neural mechanisms underlying the generation of creative new ideas (NeuroImage 2014)

---

Question 2. How does VDAT performance correlate with grounded creative tasks? Is there evidence that pushing VDAT scores higher doesn't simply train the model to ignore visual input, undermining its utility as a multimodal model?

Responses: As stated in our paper (line 189), VDAT is a diagnostic benchmark designed to measure a model's associative reasoning strength, which serves to verify our control mechanism's effectiveness. The evaluation of functional creativity quality is primarily conducted using Creation-MMBench. Our results show:

• Positive Correlation with Functional Creative Tasks: Higher VDAT scores of our FlexAC (Table 2) consistently lead to better creative performance on the grounded tasks in Creation-MMBench (Table 3). This consistent trend validates that the capability measured by VDAT translates to better performance on practical creative tasks.
• Evidence Against Ignoring Visual Input: We explicitly test for "ignoring visual input" using the Visual Fidelity Score (VFS) in Creation-MMBench, which measures the visual-text alignment degree. Table 3 shows our FlexAC improves creative Reward while VFS remains high and comparable to the baseline (6.25 vs 6.1). This provides strong, quantitative evidence that our method enhances creativity without sacrificing visual grounding

---

Question 3. The method requires identifying optimal control layers for each model. How sensitive is performance to layer selection? Is there a principled way to select layers for new MLLMs without exhaustive testing?

Responses: Layer selection is a principled and simple procedure, not an exhaustive search.

• On Sensitivity to Layer Selection: The sensitivity analysis for Qwen-VL, LLaVA, and DeepSeek-VL is shown in Fig. 10 (main paper) and expanded in Fig. 18–19 (appendix).
• On the Method for Identifying Optimal Layers: Layer selection is not an exhaustive search, but a principled, simple procedure. For any new model, we simply perform a single forward pass with associative/non-associative pairs ($f^{(a)}$, $f^{(n)}$) and identify the three layers where the cosine distance $D_{\cos}$ between hidden states peaks $L^∗ = \{l^∗,l^∗+1,l^∗+2\}, \quad \text{where} \quad l^∗ = \arg\max_{l} \sum_{i=0}^{2} D_{\cos}(f_{l+i}^{(a)}, f_{l+i}^{(n)})$

---

Question 4. How does VDAT performance correlate with functional, grounded creative tasks? What evidence shows that improving VDAT scores doesn't cause the model to ignore visual input?

Responses:

• Transferring Across Creative Tasks: Transferability is directly validated by its strong performance on Creation-MMBench, which comprises 51 diverse creative tasks across varied domains (e.g., Art, Animation, UI).
• Genuine Creativity vs. Reduced Factual Grounding:
  • Quantitative Evidence: Concerns about reduced factual grounding are refuted by the Visual Fidelity Score (VFS), which evaluates image-text alignment in Creation-MMBench (Table 3): despite a +10.92 gain in creativity reward, our method maintains a high VFS (6.25), slightly above the baseline (6.10), showing creativity improves without compromising factual grounding.
  • Qualitative Evidence: As shown in Fig.26 (dinosaur story), FlexAC-C creates a novel, coherent scene by adding plausible elements (e.g., "butterfly", "worm") and using them in specific actions and dialogue to develop character—demonstrating genuine narrative creation.

Weakness 1. Similar findings in prior work: Although well-executed empirical analysis, the observation that middle layers control hallucination has been reported in previous studies[1]. This paper builds upon these existing findings by adding steering vector techniques.

Response: We would like to clarify that FlexAC is an independent study conducted without prior knowledge of [1]. We will add a discussion to the Related Work section to elaborate on the key differences between the two approaches.

1. Different Research Goals and Findings with [1].
   • Different Research Objectives: The goal of [1] is diagnostic—to investigate "where and how do hallucinations arise?". In contrast, FlexAC is the first work to explore hallucination's intrinsic link with creativity, and provide a new method for bidirectional control. Our focus is not on controlling hallucination, but rather on understanding and leveraging its generative potential.
   • Different Core Findings: The finding in [1] is that "middle layers control hallucination." Our core finding is different: we found that the state of hallucination can be harnessed to guide creative behavior.
2. Methodological Novelty Beyond "Adding Steering Vectors": Our method is more than adding steering vectors. We introduce three key innovations:
   • Hallucination-Guided Steering: We introduce a novel method to derive association steering vectors by leveraging hallucinated states to enhance the model's associative capabilities.
   • Steering Intensity Calibration (SIC): To prevent over-steering, we designed the SIC module, which adaptively scales steering vectors.
   • Directional Integration: To further support tasks requiring multi-dimensional associations, we augment general associative vector $v_{l}^{\mathrm{gen}}$ in Eq.7 with task-specific associative vectors $v_{l}^{\mathrm{task}}$ in Eq.7 derived from GPT-4o-generated, high-association samples.

[1] Wang, Sudong, et al. "Towards understanding how knowledge evolves in large vision-language models." CVPR. 2025.

---

Weakness 2. Model-specific tuning: Optimal layers vary dramatically (DeepSeek: 4-6, Qwen: 15-17), requiring manual search for each architecture.

Response: We clarify that our method does not require manual search. For any MLLM, the optimal layers can be automatically identified via a single forward pass by maximizing the cosine distance between features from associative and non-associative inputs (Section 2.1 and Appendix 6.4). $L^∗ = \{l^∗, l^∗ + 1, l^∗ + 2\}, \quad \text{where} \quad l^∗ = \arg\max_{l} \sum_{i=0}^{2} D_{\cos}(f_{l+i}^{(a)}, f_{l+i}^{(n)})$ Here, $f^{(a)}$ and $f^{(n)}$ are features from associative and non-associative inputs, and $D_{\cos}$ denotes cosine distance.

---

Weakness 3. Heavy reliance on task-specific non-associative features: The method depends on non-associative features extracted from specific hallucination scenarios, but these features vary significantly across tasks, limiting generalization to different applications.

Response: Our method does not rely on features from "specific" scenarios; its core is a general steering vector designed for broad applicability.：

• General Hallucination Scenarios: Our primary steering vector is computed from a large, diverse set of images (randomly sampled from COCO) to capture a fundamental and universal associative pattern of the model. This ensures it has strong generalization for new tasks.
• Broad Experiments Directly Prove Generalization: Its generalization is directly proven by our results on Creation-MMBench[1] (Table 3). This benchmark includes 51 distinct creative tasks, where our method achieves a significant +10.92 Reward, while competing methods show performance degradation (-3.86 and -1.63).

[1] Fang, Xinyu, et al. "Creation-MMBench: Assessing Context-Aware Creative Intelligence in MLLM." ICCV. 2025.

---

Weakness 4. Limited evaluation on general benchmarks: The authors only evaluate on 3 general benchmarks (MME, MMMU, MMStar), which is insufficient to fully assess the impact on model's general capabilities.

Response: In direct response, we conducted a comprehensive new evaluation on 11 diverse benchmarks, including those you recommended, to assess the impact on the model's general capabilities.

Table I: Performance of FlexAC on 11 general-purpose benchmarks with the Qwen-VL model.

Across all 11 benchmarks, FlexAC maintains performance closely comparable to the baseline. This provides strong evidence that our targeted control mechanism does not harm the model's general capabilities.

---

Question 1. Model generalization and comprehensive evaluation: My biggest concern is the generalization ability of FlexAC. As mentioned in W3 and W4, could you provide evaluation results on more diverse benchmarks to better assess the impact on general capabilities? (You don't need to evaluate on all of these, but results from a representative subset would help demonstrate that the non-associative steering doesn't harm performance on diverse tasks requiring different types of reasoning and association.) - General multimodal benchmarks: MMBench, SEED-Image, MMVet. - Vision-centric benchmarks: CVBench, RealWorldQA. - OCR and document understanding: TextVQA, ChartQA, DocVQA.

Response: Thank you. We have addressed this directly in our response to Weakness 4 and present new results on 11 benchmarks in Table I. These comprehensive results confirm our method generalizes well without harming core capabilities.

---

Question 2. Architectural dependence: The optimal layers vary dramatically across models. How would you determine optimal control layers for a new model architecture? Is manual search always required, or can this be automated?

Response: As detailed in our response to Weakness 2, this process is straightforward. For any new model, we use a lightweight analytical method (a single forward pass and cosine distance check) to identify optimal layers, requiring no manual search.

---

Question 3. Steering vector stability: How stable are the derived steering vectors across different prompts and domains? Do they need to be recomputed frequently, or can a fixed set be reused effectively?

Response:

• The vector is highly stable. We proved this by successfully applying a single, fixed vector, i.e., $v_{l}^{\mathrm{gen}}$ in Eq.7 of manuscript—without modification—across our entire range of diverse benchmarks (CHAIR, MME, MMMU, MMStar).
• Our vector was computed only once from diverse data (COCO) and then used for all subsequent experiments, demonstrating that frequent re-computation is not necessary. A fixed set can be reused effectively.

---

Question 4. VDAT validity: Can you provide more rigorous validation of VDAT beyond CLIP similarity? How does it correlate with human judgments of creativity quality? Automatic metrics may not fully capture quality differences in creative tasks.

Response: To answer your question, we would like to clarify your misconception regarding the distinction between VDAT evaluation and creativity assessment.

As stated in the paper (line 189), VDAT is designed as a diagnostic benchmark to assess associative reasoning ability, rather than as a direct measure of functional creativity. Creativity is a multifaceted construct; therefore, we supplement VDAT with Creation-MMBench[1], which evaluates open-ended, task-level creative performance (Table 3).

• Empirical Validation via User Study: To assess the validity of VDAT as a measure of associative creativity, we conducted a human evaluation comparing FlexAC with several baselines (Appendix 4.1).
• Correlation with Human Judgments: A user study involving 15 raters demonstrated a strong correlation between VDAT scores and human assessments of associative ability (Appendix 4.1).
• Foundation of VDAT's Validity: Our VDAT builds upon the DAT benchmark [3], which employs automatic metrics to evaluate associative reasoning in LLMs. DAT has been both theoretically [2] and empirically [3] validated as effective. Moreover, models (e.g., FlexAC) with higher VDAT scores (Table 2) consistently outperformed others on Creation-MMBench (Table 3), indicating that VDAT reliably captures key aspects of creative quality in open-ended tasks.

[1] Fang, Xinyu, et al. "Creation-MMBench: Assessing Context-Aware Creative Intelligence in MLLM." ICCV. 2025.

[2] Jay A. Olson, Johnny Nahas, Denis Chmoulevitch, Simon J. Cropper, and Margaret E. Webb. Naming unrelated words predicts creativity. Proceedings of the National Academy of Sciences, 2021.

[3] Honghua Chen and Nai Ding. Probing the “creativity” of large language models: Can models produce divergent semantic association? In EMNLP, 2023.

Weakness 1. While the paper identifies middle layers as pivotal for associative behavior, it does not explain why these layers exhibit this property.

Response: We thank the reviewer for this insightful question. We can conceptualize the MLLM's information flow as a process from Perception, to Concept, and finally to Expression:

1. Shallow Layers (Perception): The initial layers are focused on processing raw inputs. In MLLMs, they primarily inject general visual features into the text representations to form a basic cross-modal understanding [1,2]. At this stage, high-level associations have not yet been formed.
2. Middle Layers (Concept): This is where abstract concepts are formed and factual knowledge is localized[3,4]. Research shows that these layers integrate specific, relevant visual details with the linguistic context [2] and are where the probabilities of factually correct tokens sharply increase[4]. This suggests the middle layers function as the primary locus for knowledge association, where the model transitions from low-level feature processing to high-level semantic reasoning.
3. Deeper Layers (Expression): The final layers are less about forming new associations and more about refining the already-integrated information for coherent linguistic output[1,2]. For instance, they handle tasks like syntactic formatting and ensuring the final answer is grammatically correct[2].

---

Question 1. Why was no comparison made with closed-source models like GPT-4V?

Response: Our method requires 'white-box' access to intervene on internal hidden states, which is impossible with closed-source models like GPT-4V. We explicitly acknowledge this limitation in our conclusion.

Moreover, a direct performance comparison is not meaningful due to the significant disparity in model scale between GPT-4V and our 7B-scale models (Qwen-VL, LLaVA-1.5, and DeepSeek-VL2). And model scale is a key factor influencing the associative reasoning capabilities of MLLMs. This effect is confirmed in Table I, within the same model family (Qwen-VL), the performance gap between the 7B and 72B variants is substantial. Table I. Compare with GPT-4V on Creation-MMBench.

---

Limitations 1. The experiments primarily relied on the COCO dataset (object-centric static images) and did not validate performance in video temporal reasoning, abstract art comprehension, or cross-cultural scenarios. If the model fails in dynamic or multimodal associative tasks , would FlexAC's modulation mechanism remain effective?

Response: We would like to clarify that our evaluation on Creation-MMBench[5]—a diverse benchmark comprising 51 creative tasks and no any COCO data—already includes subtasks covering Video Temporal Reasoning, Abstract Art Comprehension, and Cross-Cultural Scenarios:

• Video Temporal Reasoning: This is directly addressed by tasks like "story continue" of Creation-MMBench (Reward Score: 20.0). In this task, the model is presented with multiple sequential frames from an animation and must continue the narrative. This directly tests the model's ability to understand the temporal progression of events in the visual input.
• Abstract Art Comprehension: The benchmark includes the "Art inspired prose" task (Reward Score: 30.0) and a dedicated "Art" image category (5% of the dataset), which involves interpreting various paintings.
• Cross-Cultural Scenarios: Creation-MMBench is rich in cross-cultural contexts, featuring a "History & Culture" image category and tasks like "social media travel content". (Reward Score: 16.7)

[1] Yin, Hao, Guangzong Si, and Zilei Wang. "Lifting the Veil on Visual Information Flow in MLLMs: Unlocking Pathways to Faster Inference." CVPR. 2025.

[2] Zhang, Zhi, et al. "Cross-modal information flow in multimodal large language models." CVPR. 2025.

[3] Meng, Kevin, et al. "Locating and editing factual associations in gpt." Advances in neural information processing systems 35 (2022): 17359-17372.

[4] Chuang, Yung-Sung, et al. "DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models." The Twelfth International Conference on Learning Representations. 2023.

[5] Fang, Xinyu, et al. "Creation-MMBench: Assessing Context-Aware Creative Intelligence in MLLM." ICCV. 2025.