Thank you for your careful reading of our paper and for the insightful feedback. We clarify two points upfront:

1. Our work is proposed not as a competing platform to frameworks like AutoGen, Agentverse or Chatdev, but as a complementary set of 'plug-and-play' cognitive principles designed to enhance existing systems.
2. Our framework is designed to be model-agnostic: any LLM with certain social behaviors should benefit from it.

We have tried to address all of your concerns below and everything will be added into the final version of manuscript.

Response to Question: Cognitive Load Ceiling vs. Context Window

Just out of curiosity, how does the theoretical "cognitive load ceiling" you propose relate to or interact with the practical limitations imposed by context windows in LLMs?

It's an interesting question touching the heart of our theoretical framework. We'll answer it first to clarify foundational concepts.

1. The cognitive load ceiling is typically reached long before the context window is filled, especially for tasks with high element interactivity (e.g., a reasoning task requiring tracking logical steps throughout a long prompt). This overload isn't due to the "physical" size of the context but to the limitations of the attention mechanism. When a task requires the model to process and integrate many interdependent pieces of information, the required attention becomes too scattered across tokens. This can lead to "drifting attention" and an inability to complete the task successfully (We've added a section regarding Theoretical & Empirical Justification of LLM Working Memory at the end for your interest).
2. The context window is a "physical" limitation of the model as models are not trained beyond it. Thus the model's attention will naturally drift over the context window, even when the task is easy. Our theoretical framework well explains the issue of reaching context window limits, and we also correct the misconception that the context window as the LLM's working memory.

Response to Weakness 1: Lack of Architectural Comparison

...it lacks a detailed architectural comparison ... discern CoThinker's unique architectural contributions...

We provide a detailed comparison below. AutoGen is designed as an engineering platform for building multi-agent systems, aggregating functional components to meet a goal. Its design choices reflect this engineering purpose: agent roles are often functional (e.g., Coder, Planner), communication is frequently managed via a broadcast-style GroupChatManager, and memory is typically a coarse-grained conversation history.

In contrast, CoThinker provides a set of cognitive design principles derived from CLT, aimed at boosting the efficiency of agent collaboration itself. Our approach isn't a competing platform but rather an exploration of a new, complementary dimension of MAS design. These principles could, be integrated into frameworks like AutoGen to enhance their collaborative core.

In short, our work serves a different purpose. We are not competing with platforms like AutoGen but are instead proposing a cognitively-grounded approach to enhance agent cooperation, moving beyond purely functional roles.

Action: We'll add this detailed architectural comparison to Section 2 of the revised manuscript.

---

Clarification to Weakness 2: Exclusive Focus on Gemini Models

... focus on the Gemini series... generalizability...

Our core thesis is that CoThinker is a model-agnostic framework that coordinates a model's inherent abilities to reduce cognitive load. Our methods leverage emergent social behaviors possessed by the LLM rather than relying on its internal parameters. Therefore, any LLM exhibiting these general interactive capabilities should benefit.

For your information, we conducted new experiments across other model families. The results validate our thesis, showing consistent improvements. The enhancement is particularly significant for models with stronger reasoning capabilities, highlighting our work as a powerful amplifier for existing intelligence.

Response to Minor Suggestions

i. ...a more detailed breakdown for Table 1 is needed... A deeper, more granular discussion...

The aggregated performance in Table 1 can be misleading, as it includes tasks with low intrinsic cognitive load (e.g., Instruction Following). On these tasks, the overhead of a multi-agent framework can slightly lower the relative score. This is fully consistent with CLT. The true strength of CoThinker is revealed in high-load domains.

Action: To provide this granular analysis, we will revise Section 5.2 to discuss this nuance. Furthermore, we'll add qualitative case studies to the appendix, presenting a fine-grained comparison of agent outputs on a specific reasoning task.

---

Theoretical Justification of LLM working memory Here we provide details responding to the concern about the lack of formal proof or empirical evidence linking CLT to specific internal representations or mechanisms within LLMs.

1. We first identify the key characteristics of working memory. We then identify the dual characteristics in LLMs that corresponding to the working memory features.
2. Finally, we provide empirical evidence to support the existence of these dual characteristics in LLMs.

1. Key Characteristics of Working Memory By definition, working memory is a feature that temporarily holds and processes information simultaneously. By definition, cognitive load is the attention needed for interactive information within working memory (mental effort), which determines the easiness of task completion [1].

Thus, the key characteristics to quantify working memory are the "attention on interactivity of information" and the "easiness of task completion". Interestingly, the attention mechanism in LLMs and its output perplexity naturally becomes the dual characteristics providing a way to quantify these two aspects.

2. Pre-experiment 1: Attention as a proxy for cognitive load

1. Justification:

$$\text{Attention Entropy} = -\sum_{i=1}^{N} a(s_i) \log a(s_i)$$

measures diversity of attention distribution across different tokens. A higher entropy indicates a more uniform distribution of attention, suggesting the model is considering multiple aspects of the input, which corresponds to a higher cognitive load in WM [2].

1. Experiment and results: We choose a set of Q&A pairs and construct a 4 level difficulty question set. We constrol the length of the Q&A for fair comparison. We then measure the average entropy of the attention distribution of each layer in the model. The results are shown in the table below:

1. Implication: The attention entropy increases with task complexity, indicating that the model have to use more pieces of information to process the task, which corresponds to higher cognitive load. The presence of reasoning steps consistently lowers attention entropy, indicating that reasoning helps the model to "offload" some of the cognitive load by focusing its attention on key pieces of information.

3. Pre-experiment 2: Perplexity as a proxy for working memory

1. Justification:

$$\text{Perplexity}=\exp\left(-\frac{1}{N}\sum_{i=1}^{N}\log p(s_i)\right)$$

measures the certainty of the model's generation, which means the "easiness" of task completion.

1. Experiment and results: Full details on its validity are in the Appendix. We choose a set of Q&A pairs (high quality answer) and construct 5 guiding instructions with increasing complexity. We then measure the perplexity of the on the answer for each instruction:

1. Implication: For hard tasks, perplexity first decreases then increases, indicating that the instructions help the model to focus on the task up to a certain point, and then become an additional burden on the model's cognitive load; For easy tasks, perplexity increases with instruction complexity showing no help. This measurement of "easiness" aligns with CLT.

Reference

[1] Sweller, John. "Cognitive load theory." Psychology of learning & motivation, 2011.

[2] Zhang, et al. Attention Entropy is a Key Factor: An Analysis of Parallel Context Encoding with Full-attention-based Pre-trained Language Models. ACL, 2025.

We appreciate your insightful feedback and are glad you found our application of Cognitive Load Theory (CLT) novel and our analysis inspiring. We address your questions and suggestions below.

Response to Questions

1. Is the working memory described in Section 3.1 also incorporates LLM’s own parametric knowledge?

To better answer your question, we first clarify two concepts. An LLM's parameters incorporate two aspects: 1) Parametric Knowledge, corresponding to human long-term memory, and 2) model's innate generation capabilities. In our framework, Working Memory (WM) is the mechanism of selective attention—a limited capacity to hold and manipulate information simultaneously—and isn't simply the in-context information itself.

Therefore, a model's parameters determine how effectively it selects information for its WM, while its parametric knowledge affects how well it can understand and integrate that information into its reasoning process (reflected in its attention patterns). WM is the resource used to process information, requiring the LLM to load both context and activated parametric knowledge for a task.

The reason that simple elicitation appears unrelated to WM is that it often involves sequentially giving facts, which consumes little WM as it doesn't require integrating many interdependent pieces of information. However, if asked to elicit a large volume of structured information, the LLM would need to use its WM to remember what it has outputted and what comes next, which can indeed lead to cognitive overload.

(Note: For mathematical justification of LLM WM and its link to attention, please see a note at the end.)

2. How is your theory and agent design related to the concept of “cognitive offloading”?

Our framework aligns with cognitive offloading, through our Transactive Memory System (TMS). The TMS maintains a consensus and "who knows what." We use agents' self-awareness so it can check the TMS to see what cognitive functions other agents have performed (e.g., "Agent B has done the fact-checking"), allowing it to offload that function and focus on other parts of the task.

3. For the network construction in method... how the information is generated and passed in the system?

The network updates dynamically each round. To illustrate the information flow, let's take Agent 1 as an example:

1. The Communication Moderator first selects current peers for Agent 1 to interact with, say Agent 2 & 3.
2. Using the TMS as a medium, Agent 1 accesses two key pieces of information: the group's current consensus and the "who knows what".
3. From the TMS, Agent 1 may learn that Agent 2 is performing fact-checking and Agent 3 is identifying relevant mathematical formulas. Realizing these parts are being handled, Agent 1 offloads these tasks. It can read Agent 3's output, take the proposed formulas to find alternative solutions.
4. The new outputs from all agents are sent back to the TMS and the Communication Moderator, informing the network for the subsequent round.

4. Is it possible to mix agents of different possibilities in the same network?

Yes, our framework implements a form of this. The "Agent Parallel Thinking" module assigns each agent a specific cognitive focus (a "thinking style"), creating a network of agents with different specialized abilities. To demonstrate the effectiveness of this component, we conducted an ablation study (w/ vs. w/o the thinking style module). Below, assigning thinking styles consistently improves performance.

(Note: In w/o Styles condition, identical agents were used, which required a non-zero temperature to encourage diversity. This setting gives the baseline an advantage over a purely greedy approach, yet the module still demonstrates clear benefits.)

Action: We'll add this ablation study to appendix, together with case studies showing how different agents collaborate.

---

Response to Weaknesses

1. The paper’s writing on the network construction part could be further improved...

We'll incorporate detailed information flow provided in our answer to Question 3 into the main manuscript.

2. The model used for testing only involves gemini model which is limited. Also, it’s possible to further conduct ablation study on different LLM configuration and each component’s role.

To clarify, our framework doesn't compete with a model's native ability but to coordinate it effectively. As long as model possesses certain emergent social behaviors, our framework can enhance its performance by reducing cognitive load generally. Our approach is designed to be model-agnostic, coordinating a model's capabilities. For your information, we have also added results from other model families.

Regarding component ablations, the role of the Communication Moderator is analyzed in the paper's main ablation studies (Fig. 4), which investigate its key parameters (N and β). The ablation on different agents is provided above rebutal and will be added to appendix.

---

Response to Limitations

Limitations could be discussed from the role of each component...what are some misaligned aspect between the original CL theory and the LLM side, and what are some error cases...

These are excellent suggestions for a more robust discussion.

Action: We'll revise our Limitations and Future Work section to specifically include:

1. An analysis of assumptions related to each component's role.
2. A discussion of misalignments in the CLT-LLM analogy.
3. An examination of common error cases to better define boundaries.

---

Theoretical Justification of LLM WM Here we provide details responding to your second concern on the lack of formal proof or empirical evidence linking CLT to specific internal representations or mechanisms within LLMs.

1. We'll first identify the key characteristics of WM. We'll then identify the dual characteristics in LLMs that corresponding to the WM features.
2. Finally, we'll provide empirical evidence to support the existence of these dual characteristics in LLMs.

1. Key Characteristics of WM By definition, WM is a feature that temporarily holds and processes information simultaneously. By definition, cognitive load is the attention required to handle interactive information within WM (mental effort), which determines the easiness of task completion [1].

Thus, the key characteristics to quantify WM are the "attention on interactivity of information" and the "easiness of task completion". Interestingly, the attention mechanism in LLMs and its output perplexity naturally becomes the dual characteristics providing a way to quantify these two aspects.

2. Pre-experiment 1: Attention as a proxy for cognitive load

1. Justification:

$$\text{Attention Entropy} = -\sum_{i=1}^{N} a(s_i) \log a(s_i)$$

represents the diversity of attention distribution across different tokens. A higher entropy indicates a more uniform distribution of attention, suggesting model is considering multiple aspects of the input, corresponding to a higher cognitive load in WM [2].

1. Experiment and results: We choose a set of Q&A pairs and construct a 4 level difficulty question set. We constrol the length of the Q&A for fair comparison. We measure average entropy of the attention distribution of each layer in model.

1. Implication: The attention entropy increases with task complexity, indicating that model have to use more pieces of information to process the task, corresponding to higher cognitive load. Adding reasoning steps lowers entropy, indicating that reasoning helps "offload" cognitive load by focusing attention.

3. Pre-experiment 2: Perplexity as a proxy for WM

1. Justification:

$$\text{Perplexity}=\exp(-\frac{1}{N}\sum_{i=1}^{N}\log p(s_i))$$

represents model's certainty, measuring the "easiness" of task completion.

1. Experiment and results: We'll briefly describe the experiment here and details of validity of experiment to be found in Appendix. We choose a set of Q&A pairs and construct 5 guiding instructions with increasing complexity. We then measure the perplexity of the on the answer for each instruction:

1. Implication: For hard tasks, perplexity initially decreases and then increases, showing that instructions help up to a point before becoming a cognitive burden; While for easy tasks, the perplexity increases with complexity showing no help. This measurement of "easiness" perfectly aligns with the CLT.

Reference

[1] Sweller, John. "Cognitive load theory." Psychology of learning & motivation, 2011.

[2] Zhang, et al. Attention Entropy is a Key Factor: An Analysis of Parallel Context Encoding with Full-attention-based Pre-trained Language Models. ACL, 2025.

Thank you for your detailed review. After carefully considering your comments, we clarify some facts on Sec. 3: Our analogy isn't only based on observations but also grounded in theoretical and empirical evidence. It begins with a foundational analogy: human brain can only attend to a limited amount of information at once, a feature known as working memory (Sec. 3.1). The architecture of LLM, due to its inherent attention mechanism, similarly restricts the amount of information it can focus on at once, where we refer this LLM feature as LLM's working memory (Response 1 for theoretical justification).

We then show how Cognitive Load Theory (CLT) effectively manages WM feature (Limited focused information capacity), thereby guiding human interaction to enhance efficient group collaboration in solving complex problems involving diverse knowledge integration. CLT provides a design principles to manage effective information within limited WM (section 3.2).

Since MAS simulates human interactions to solve complex problems involving diverse knowledge integration, the design principles extracted from human CLT can be reused to inform the design of MAS. These design principles guide our framework's setting of Communication Moderator, Transactive Memory System, and Parallel Thinking Modules, which differentiates us from existing MAS (Response 2 for details).

Response to W1, W2 & Q1, Q2: The theoretical analogy...is overgeneralized...No formal proof or empirical evidence...key assumption...not verified...subdividing tasks reduces cognitive load? What metrics...

We address these concerns through a unified logic chain that establishes WM in LLM, validates the CLT analogy, and proves our core assumption with direct empirical evidence.

Step 1: Establish WM in LLM & Measure Cognitive Load We identify the key characteristics of WM, then dual characteristics in LLM, finally empirical evidence.

1. Key Characteristics of WM: By definition, WM is a feature that temporarily holds and processes information simultaneously and cognitive load is the attention required to handle information within WM, which determines the easiness of task completion [1]. Thus, the key characteristics to quantify WM are "attention on interactivity of information" and "easiness of task completion". LLM attention mechanism & perplexity naturally become dual characteristics quantifying these aspects.

2. Pre-experiment 1: Attention as a proxy for WM

$$\text{Attention Entropy} = -\sum_{i=1}^N a_i \log a_i$$

measures attention distribution diversity. Higher entropy indicates uniform distribution, suggesting the model considers multiple aspects of the input, corresponding to higher cognitive load [7].

We use Q&A pairs (AMPS-Hard) w/ 4-level difficulty, then control length of input for fair comparison and measure attention entropy on the answers:

Attention entropy increases with complexity, indicating the model have to use more information pieces to process the task, indicating higher cognitive load. Presence of reasoning steps lowers attention entropy, indicating that reasoning helps the model to "offload" partial cognitive load by focusing on key informations.

1. Pre-experiment 2: Perplexity as a proxy for WM

$$\text{Perplexity}=\exp\left(-\frac{1}{N}\sum_{i=1}^{N}\log p(s_i)\right)$$

measures certainty of the model's generation, which is the "easiness" of task completion [6].

We choose Q&A pairs (FLASK) w/ 5-level guiding instructions complexity, then measure the perplexity on answers:

For hard tasks, the perplexity decreases then increases, showing that instructions help LLM to focus on the task up to a certain point, and then become additional burdens on the model's cognitive load; While for easy tasks, the perplexity increases, the instructions showing no help. This measurement of "easiness" aligns with CLT (Sec. 3.2).

Step 2: Validate CLT Analogy Similar to CLT initially grounded in behavioral observations & experiments [1], our analogy uses observed LLM limitations [2,3] & previous "WM capacity" experiments [4,5], which consists with cognitive science research. Furthermore, with established metrics above, we show LLM exhibit the same fundamental WM limitations.

Step 3: Subtasking Reduces Cognitive Load & Show CLT Principles Work Using established metrics, CoThinker's "subtasking" means each agent attends to smaller, focused info sets, reflected by lower Attention Entropy and Perplexity, providing quantitative evidence of reduced CL. This causal mechanism directly accounts for performance gains in ablation studies (Fig 4): System performs better only on high CL tasks, where subtasking is helpful.

Clarification on Module-Level Measurement: Our modules are functional tools/algorithms, not cognitive agents. CL measurement only makes sense for individual agents, not the tools themselves.

---

Clarification on W2: ...implementation...is distinct from actual human teams...differs from existing MAS...

We clarify that Transactive Memory System (TMS) and managing group cognitive load (Communication Moderator (CM) ) are established theoretical concept from human teams, describes as human emergent behavior in social and cognitive psychology [8,9]. These aren't engineering choices but rather design principles derived from CLT. Our framework explicitly grounds in CLT, actively using design principles to operationalize these emergent behaviors into modules to guide the MAS effectively. Each design choice follows CLT principles, distinguishing us from existing MAS. Please refer to Clarification 4 & 5 for details on TMS & CM.

---

Clarification to W6 & Q5: Benchmark baseline missing...

Our framework doesn't compete with a model's native reasoning but to coordinate it effectively. As long as the model possesses certain emergent social behaviors, our framework can enhance its performance by reducing cognitive load generally. Our approach is model-agnostic, coordinating a model's capabilities rather than changing them. FYI, results from other model families:

Results on these SOTAs show greater improvement. Our framework enhances reasoning efficiency. DeepSeek-R1 model struggled w/ mid-level math problems within 4k output context window; CoThinker enabled it to find solutions.

---

Clarification on W4 & Q3: The retrieval process... not explained...

Agent receives input from both TMS & CM. Two-channel design is designed to manage cognitive load:

1. TMS provides high-level, synthesized, long-term context (group progress, consensus), giving agent strategic awareness over rounds w/ low cognitive burden.
2. CM provides specific, low-volume, current peer outputs for immediate reaction & refinement.

"Who knows what" can be seen as long-term info for peers to quickly build consensus, guiding agent's current understanding of its peers. (e.g. How much should I rely on this peer? What detail this peer ignores so I should explore?)

---

Clarification on W5 & Q4: Communication topology & algorithm determined by pre-defined rules...

Rules for communication algorithm are direct operationalization of CLT principles:

1. Rule 1: Fixed In-Degree (N): Directly enforces CLT principle of limited WM. By capping number of incoming peer messages, we prevent any single agent from high extraneous cognitive load from info overload.
2. Rule 2: Adaptive Rewiring ($\beta$): This rule is designed to create a adaptive network by managing the trade-off between exploiting similar ideas (lower $\beta$ reduces extraneous load) and exploring diverse ones (high $\beta$ distributes intrinsic load). This is principled approach supported by network science [10,11].

---

Response to W7 & Q6: CoThinker is sensitive to hyperparameters...

We view sensitivity not as weakness, but as empirical confirmation of our thesis: parameters directly correspond to cognitive trade-offs predicted by CLT. Existence of optimal, task-dependent range is what our theory expects. Future work involves estimating task's cognitive load at runtime & dynamically adjusting hyperparameters, creating self-regulating system.

References

[1] Sweller, J. "Cognitive load theory." Psychology of Learning and Motivation, 2011.

[2] Liang et al. "Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate." EMNLP, 2024.

[3] Huang et al. "Large language models cannot self-correct reasoning yet." ICLR, 2024.

[4] Gong et al. "Working memory capacity of ChatGPT: An empirical study." AAAI, 2024.

[5] Zhang et al. "Working memory identifies reasoning limits in language models." EMNLP, 2024.

[6] Xiong et al. "Can LLMs Express Their Uncertainty? " ICLR, 2024.

[7] Zhang et al. "Attention Entropy is a Key Factor." ACL, 2025.

[8] Zhang et al. "Exploring Collaboration Mechanisms for LLM Agents: A Social Psychology View." ACL, 2024.

[9] Kirschner et al. "A cognitive load approach to collaborative learning." Educational Psychology Review, 2009.

[10] Almaatouq et al. "Adaptive social networks promote the wisdom of crowds." PNAS, 2020.

[11] Rand et al. "Dynamic social networks promote cooperation in experiments with humans." PNAS, 2011

Thank you for your thoughtful and constructive review. We are delighted that you found our work's grounding in Cognitive Load Theory (CLT) to be "solid, highly interesting and has potential" and that you found our explanations and empirical evaluations to be clear and correct. Your suggestions are invaluable for improving the final version of our paper.

Below, we address each of your points and outline the specific changes we will make in the final manuscript.

Response to Weakness 1 & Question 1: On the Motivation for Benchmark Selection

What were the motivations or selection criteria for the evaluation benchmarks?

Our selection of LiveBench and CommonGen-Hard was a deliberate strategy to test our hypothesis under two distinct, complementary conditions of high cognitive load.

1. LiveBench as a SOTA, comprehensive "Real-World" Challenge: We chose LiveBench because it is a well-known and SOTA benchmark for 2025 that contains comprehensive, real-world tasks, for example, high school math competition, zebra puzzles, and word puzzles. These tasks require a high cognitive load for human to solve. Its key advantages for our study are:

   • Broad-Spectrum Evaluation: It is built upon a diverse set of established and difficult benchmarks (e.g., Big-Bench Hard, AMPS [Math], IFEval [Instruction Following]) and real-world problems (e.g. International Mathematical Olympiad, Kaggle Competition), ensuring our framework is tested across a wide range of domains (e.g., math, reasoning, instruction following, data analysis and more).
   • Fair and Rigorous Assessment: Its frequent updates are specifically designed to minimize data contamination, ensuring that results reflect true reasoning capabilities rather than memorization, suitable to test SOTA models.
   • High Cognitive Load Scenario: The tasks are intentionally "hard"—even the most capable models struggle. This makes LiveBench an ideal proxy for the high cognitive load scenarios we aim to address, allowing us to probe the upper limits of a model's inherent problem-solving capabilities.

2. CommonGen-Hard as a Controlled, "Experimental" Challenge: We selected CommonGen-Hard because it functions as a more systematic, experimental test of a core CLT principle: information interactivity. As a classic language task, its "Hard" version is naturally constructed to impose high cognitive load by forcing the model to integrate a few target concepts from a large pool of distractors. This provides a focused, controlled environment to validate that our architecture successfully manages precisely the kind of information interactivity predicted by CLT to cause performance degradation.

Action: We will add a dedicated paragraph in Section 5.1 (Evaluation Benchmarks) to explicitly state this dual rationale.

Response to Weakness 2 & Question 2: On the Scope of Model Evaluation

You only evaluate gemini models. Could you elaborate on that decision and how these results might or might not transfer to other model families?

Our core thesis is that CoThinker is a model-agnostic framework that coordinates a model's inherent reasoning abilities to reduce cognitive load. The framework's effectiveness does not depend on a model's internal parameters but rather on its ability to exhibit emergent social behaviors, such as following directed prompts, reacting to peer outputs, and synthesizing information. Therefore, any model that possesses these fundamental interaction capabilities should benefit from our method.

To empirically validate this, we have conducted new experiments with models from other leading families. The results confirm our thesis, showing consistent performance gains. Notably, the improvement is more significant for models with stronger native reasoning, demonstrating that our framework enhances, rather than replaces, their capabilities.

Action: We will incorporate these new results and the expanded analysis into Section 5.2 to robustly demonstrate the framework's generalizability.

---

Response to Weakness 3: On the Resolution of Ablation Studies

Ablation studies have very low resolution (only 3 values per hyperparameter).

This is an accurate observation of our experimental design. The primary objective of the ablation studies was not to perform an exhaustive hyperparameter sweep, but to empirically validate our CLT-based theory. By showing that task-dependent optimal ranges exist—for instance, that Reasoning tasks benefit from a higher exploration rate (β) while Instruction Following does not (Figure 4)—we provide direct evidence for the cognitive trade-offs our framework is designed to manage. The resolution was sufficient to establish this principle.

Action: We will refine the language in Section 5.4 (Analysis of Ablation Studies) to clarify that the study's scope was intentionally focused on validating these theoretical principles.

---

Response to Weakness 4 & Question 3: On the Placement of Limitations

The limitations addressed in the appendix should be in the main paper.

We agree that this discussion is integral to the paper. To better contextualize our contributions and transparently outline the path for future research, the discussion of limitations and future directions belongs in the main text.

Action: In the final version, we will integrate the Limitations and Future Work discussion from Appendix E into the main body of the paper as a new Section 7.

---

Response to the optional suggestion

What were the motivations or selection criteria for the evaluation benchmarks? I can guess roughly from the rest of the paper, but it would be nice to have it explicitly estated, maybe also referencing one or two benchmarks that might seem suitable to inexpert eyes (this last bit is just a suggestion, so highly optional).

We appreciate your suggestion to clarify the motivations behind our benchmark selection and we add some examples of the actual datasets used in the benchmarks (See Response to Weakness 1 & Question 1 above). Please let us know if this addresses your suggestion.