CAM: A Constructivist View of Agentic Memory for LLM-Based Reading Comprehension

Thanks for your careful and insightful comments. Below, we provide point-by-point responses to each concern raised.

Q1: Potential LLM summarization drift.

Thank you for raising this point. Using LLM to generate summaries has become a common practice in current agentic memory systems (e.g., ReadAgent, RAPTOR, GraphRAG, and MemTree). As we discussed in the Limitation section (Appendix I), we acknowledge that LLM-generated summaries may carry risks of hallucination propagation, especially in deeper memory layers. This limitation could be mitigated by introducing more agentic behaviors (such as self-questioning and reflection) to reduce the hallucination issue and improve the summary fidelity, which we believe is a promising direction for future memory research.

Q2: Inference-time latency from "Prune‑and‑Grow" retrieval strategy.

To investigate this point, we compare CAM's Prune-and-Grow retrieval with two common retrieval strategies (used in the baseline methods): Hierarchical Traversal (HT) and Global Retrieval (GR). To fully address your concern, we also isolate the impact of LLM selection by evaluating CAM with two non-LLM graph traversal strategies: Breadth-First Search (BFS) and Personalized PageRank (PPR). We report the inference performance (F1 for MH-RAG, R-L for others) / average time (in seconds) below.

We observe that while the LLM-based retrieval incurs moderate overhead in response time, it brings consistently favorable performance gains beyond the baselines and non-LLM variants. The Prune-and-Grow process can effectively filter irrelevant memory nodes, resulting in more focused activation sets and improved downstream generation. Once the activation set stabilizes, the memory exploration then stop early to avoid unnecessary computation.

Moreover, we would like to clarify that CAM is not tied to a specific retrieval strategy; rather, it is inherently compatible with various retrieval methods, allowing users to choose the most suitable one based on their latency and performance requirements in real-world applications.

Q3: Trade-off between positional bias and semantic relevance.

This balance is explicitly controlled by the hyperparameter $\alpha$ in the similarity scoring function (Equation 2). When $\alpha = 1$, the influence of positional bias is fully removed, and chunk relevance is computed purely based on semantic similarity, which would facilitate the retrieval of spatially distant yet topically related segments.

Incorporating positional bias helps maintain discourse coherence and improves retrieval of local context, aligning with human reading behavior where adjacent sentences are semantically cohesive. To quantify this effect, we conduct a sensitivity analysis across different $\alpha$ values, summarized below:

To more thoroughly address your concern about hyperparameter robustness, we also provide a broader sensitivity analysis in our response to your Q7.

Q4: Handling of contradictory information.

We agree that the ability to detect and reconcile conflicting or contradictory information is a critical aspect of human cognition. Consistent with prior LLM-based memory systems such as GraphRAG, RAPTOR, and MemTree, our current framework assumes that source information is generally consistent and focuses on the memory development process. We consider memory reconciliation as a highly promising direction for future memory research, and promise to include a clear discussion of this point in our final version.

Q5: Discussion on the effects of "Ego-Centric Disentanglement".

Effects. Ego-Centric Disentanglement is central to CAM’s ability to form coherent and nuanced memory abstractions. Its core motivation lies in capturing the semantic multifacetedness of complex text chunks---each chunk often relates to multiple themes, topics, or concepts. By replicating nodes based on their local semantic neighborhoods, CAM allows a single chunk to contribute to multiple higher-level abstractions, enabling more comprehensive and context-sensitive summarization.

Does it lead to redundancy in the final retrieved context? No. At inference time, chunk-level duplicates are automatically de-duplicated (based on unique chunk IDs). While multiple higher-level nodes may reference the same original chunk, these abstract nodes originate from distinct semantic clusters, making them meaningfully differentiated. To quantify this, we compute pairwise semantic similarity among all node embeddings in the memory structure, and the average results demonstrate low semantic redundancy:

Does managing a replica graph impact retrieval efficiency or memory footprint? No. The replica graph is a logical concept used to disentangle semantic roles during memory summarization. It does not result in physically duplicating memory content or embedding storage.

A Qualitative Example. To fully address your concern, we further present a qualitative example to demonstrate the flexibility enabled by Ego-Centric Disentanglement. The example is drawn from the long-form novel Jane Eyre and involves a comprehensive question that requires integrating multiple semantically distant chunks of information, as detailed below.

Question: Why does Jane leave Thornfield after her wedding is interrupted?
Evidence Pieces: 
- Jane discovers that Rochester is already married (factual revelation).
- She experiences intense internal moral conflict about staying (psychological theme).
- She reflects on religious and spiritual values (spiritual/ethical dimension).
- She asserts her personal independence and dignity (character development arc).

[Without Flexibility]
Response: Jane leaves Thornfield because she finds out Rochester is already married. ❌
(This answer is semantically shallow, missing Jane’s inner struggle, moral reasoning, and identity preservation---these themes are essential to the novel’s integrity.)

[With Flexibility]
Response: Jane leaves Thornfield not only because of Rochester’s hidden marriage, but also because staying would betray her moral and spiritual convictions. Despite her love, she chooses self-respect and integrity, fearing the loss of her identity. ✔️

Q6: Dynamic robustness of CAM to batch order.

Following your advice, we conduct a new experimental analysis to evaluate CAM's dynamic robustness to random batch order. We report the memory statistics (branching factor) / inference performance (F1 for MH-RAG, R-L for others) under each configuration, which are averaged over five runs.

We observe that:
(1) Even with batch size = 1, CAM yields more stable memory structures and inference performance than MemTree, confirming the effectiveness of flexible assimilation and dynamic accommodation in our framework.
(2) As the batch size increases, CAM inherently benefits from a broader contextual view, thereby further reducing the memory variations across different batch orders.

Q7: Hyperparameter sensitivity analysis.

To assess the robustness of CAM with respect to the key hyperparameters (i.e., $\alpha$, $\sigma$, and $\theta$), we conduct a sensitivity analysis using 100 randomly sampled validation instances from each dataset. The performance results (averaged over three runs) across different hyperparameter combinations are summarized below:

We observe that:
(1) For narrative datasets (e.g., NovelQA and FABLES), incorporating positional proximity (lower $\alpha$) is beneficial, as the adjacent sentences are semantically cohesive.
(2) The performance is relatively robust to $\sigma$ across datasets, and the sharper decay setting ($\sigma$ = 1.0) generally outperforms the flatter alternative ($\sigma$ = 2.0).
(3) $\theta$ controls the density of the foundational semantic network and shows the dataset-specific sensitivity.

We sincerely promise to incorporate the above experiments and discussions into our final version. We hope that these responses have addressed your concerns, and would greatly appreciate it if you could kindly raise the scores accordingly. Your support really means a great deal to us.

Dear Reviewer vP7d,

Thank you for taking the time to acknowledge our rebuttal.

During the rebuttal phase, we have made great efforts to address each of your concerns. In particular, we would like to kindly ask whether our responses have addressed your concerns regarding the additional analysis on inference efficiency and dynamic robustness.

If you have any further questions or suggestions, we would be happy to continue the discussion. Your support really means a great deal to us, and we sincerely look forward to your feedback.

Best regards,
Authors

Thanks for your careful and insightful comments. Below, we provide point-by-point responses to each concern raised.

Q1: Clarification on the ablations of the proposed principles.

We would like to respectfully clarify that we have conducted comprehensive ablation studies to evaluate the contribution of the three proposed principles (i.e., Hierarchy, Flexibility, and Dynamicity), as detailed in Sections 5.3 and 5.4:

• For Hierarchy and Flexibility, we present two ablation variants (w/o Hierarchy and w/o Flexibility) in Table 3, evaluated on four long-context reading comprehension benchmarks.
• The w/o Hierarchy variant omits the hierarchical clustering stage entirely and relies solely on the foundational semantic network.
• The w/o Flexibility variant bypasses ego-centric disentanglement and directly applies non-overlapping label propagation for clustering.
• For your convenience, we re-include their comparison with the full CAM below. We can see that both variants result in clear performance degradation, thereby confirming the importance of hierarchy and flexibility in our design principles.

• Moreover, the benefits of Dynamicity are extensively investigated in Section 5.3.
• Specifically, Figure 3 illustrates how the dynamicity of CAM contributes to both processing efficiency and inference stability.
• In terms of processing efficiency (Figure 3a), CAM can dynamically integrate newly arrived chunks in batches, offering over 4× speedup compared to methods constrained to offline (RAPTOR, GraphRAG) or entry-wise online processing (MemTree).
• In terms of inference stability (Figure 3b), CAM maintains stable and competitive performance across varying batch sizes. These results collectively validate the benefits and necessity of our dynamicity principle.

Q2: Hyperparameter sensitivity analysis.

Thank you for raising this important point. To assess the robustness of CAM with respect to the key hyperparameters (i.e., $\alpha$, $\sigma$, and $\theta$), we conduct a sensitivity analysis using 100 randomly sampled validation instances from each dataset. We perform grid search over the following values: $\alpha\in$ {$0.5,0.7,0.9$}, $\sigma\in$ {$1.0,2.0$}, and $\theta\in$ {$0.5,0.7$}. The performance results (averaged over three runs) across different hyperparameter combinations are summarized below:

We observe that:
(1) For narrative datasets (e.g., NovelQA and FABLES), incorporating positional proximity (lower $\alpha$) is beneficial, as the adjacent sentences are semantically cohesive.
(2) The performance is relatively robust to $\sigma$ across datasets, and the sharper decay setting ($\sigma$ = 1.0) generally outperforms the flatter alternative ($\sigma$ = 2.0).
(3) $\theta$ controls the density of the foundational semantic network and shows the dataset-specific sensitivity.

Q3: The relationship between the ablations in Table 3 and the proposed principles.

As you noticed, Table 3 comprehensively presents the performance of CAM with different configurations and two ablation variants.

As described in Section 5.4, we group the rows into four categories for clarity:

• LLM Backbones: We evaluate CAM using different language models, including the default commercial model GPT-4o-mini and two popular open-source alternatives: Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct. The results suggest that CAM's core strengths primarily stem from the effective memory design, rather than relying on expensive LLM backbones.
• Embedding Models: We assess the effect of different embedding encoders, including text-embedding-3-small (default), text-embedding-3-large, and E5-Mistral-7B-Instruct. The results show that the use of more advanced embedding models can improve the performance of CAM, since they deliver more accurate similarity measurements for the construction of memory foundational networks.
• Retrieval Strategy: This section compares our Prune-and-Grow strategy with two alternatives: Hierarchical Traversal and Global Retrieval. We observe that both alternatives lead to lower performance, demonstrating the effectiveness of our Prune-and-Grow strategy.
• Ablation Variants: As elaborated in our response to your Q1, these two variants are designed to demonstrate the importance of the Hierarchy and Flexibility principles in CAM's design. The w/o Hierarchy variant removes the hierarchical abstraction, and the w/o Flexibility variant disables the ego-centric disentanglement. Both variants result in consistent performance degradation, thereby supporting the benefits and rationality of these two principle designs.

Q4: Plan to release the source code of the prototype.

Sincerely thanks for your interest. We have provided the implementation code in an anonymous repository (linked in the footnote on the first page) to support reproducibility during the review process. Upon publication, we will open-source the full codebase, and further develop it into a general-purpose constructivist memory framework to benefit the broader research community.

We sincerely promise to incorporate the above experiments and discussions into our final version. We hope that these responses have addressed your concerns, and would greatly appreciate it if you could kindly raise the scores accordingly. Your support really means a great deal to us.

Dear Reviewer 4RE6,

Thank you very much for taking the time to acknowledge our rebuttal.

During the rebuttal phase, we have made great efforts to address each of your concerns. In particular, we would like to kindly ask whether our responses have clarified your misunderstandings regarding our ablation studies.

If you have any further questions or suggestions, we would be happy to continue the discussion. Your support really means a great deal to us, and we sincerely look forward to your feedback.

Best regards,
Authors

Dear Reviewer 4RE6,

Thank you very much for your feedback and support on our paper. We sincerely promise to include the above description and analysis in our final version.

Best regards,
Authors

Thanks for your careful and insightful comments. Below, we provide point-by-point responses to each concern raised.

Q1: Hyperparameter sensitivity analysis.

To assess the robustness of CAM w.r.t. the key hyperparameters (i.e., $\alpha$, $\sigma$, and $\theta$), we conduct a sensitivity analysis using 100 randomly sampled validation instances from each dataset. Since top-k serves as a soft cap to avoid memory over-density and only takes effect under extreme redundancy, we fix it to 10 throughout. For the other three hyperparameters, the results of different combinations (averaged over three runs) are reported below:

We observe that:
(1) For narrative datasets (NovelQA and FABLES), incorporating positional proximity (lower $\alpha$) is beneficial, as the adjacent sentences are semantically cohesive.
(2) The performance is relatively robust to $\sigma$ across datasets, and the sharper decay setting ($\sigma$ = 1.0) generally outperforms the flatter one ($\sigma$ = 2.0).
(3) $\theta$ controls the density of the foundational semantic network and shows the dataset-specific sensitivity.

Q2: Global coherence checking for inconsistent information.

We agree that the ability to detect and reconcile inconsistent or contradictory information (e.g., when a document transitions to a new theme) is a critical aspect of human cognition. Consistent with prior LLM-based memory systems such as GraphRAG, RAPTOR, and MemTree, our work assumes that source information is generally consistent and focuses on the memory development process. We believe that memory reconciliation is a highly promising direction for future memory research, and will include a clear discussion of this point in our final version.

Q3: Discussion on LLM-based text summarization.

Experimental Setup. We respectfully clarify that the detailed settings are provided in Section 5.1 and Appendix F: we use GPT-4o-mini with temperature = 0, and text-embedding-3-small as the embedding model for CAM and all baseline methods. Additionally, all LLM prompt templates have been included in the anonymous repository (linked in the footnote on the first page) since the initial submission. For the convenience of readers, we will follow your suggestion to add these prompts to the final version.

LLM-Based Summarization. Using LLM to generate summaries has become a common practice in current agentic memory systems (ReadAgent, RAPTOR, GraphRAG, and MemTree). As we discussed in the Limitation section (Appendix I), we acknowledge that LLM-generated summaries may carry risks of hallucination propagation. This limitation could be mitigated by introducing more agentic behaviors (such as self-questioning and reflection) to reduce the hallucination issue and improve the summary fidelity.

Comparing Summarization Strategies Across Different LLMs. We evaluate CAM and RAPTOR under abstractive and extractive summarization strategies. The results (averaged over three runs) are reported below:

We observe that extractive memory systems generally outperform abstractive ones, as extractive summarization offer greater factual fidelity and lower risk of hallucination, which is important for ensuring accurate retrieval. Moreover, CAM consistently outperforms RAPTOR under both strategies, confirming the effectiveness of our design principles.

Q4: Concern about LLM robustness and efficiency in memory retrieval.

LLM Retrieval Robustness. While LLMs carries the potential risk of hallucination, they inherently offer strong advantages in semantic association and reasoning, which are valuable in complex long-context scenarios. Many recent IR works [1] also show that LLM-based re-ranking or selection can enhance retrieval accuracy beyond what simple embedding similarity or rule-based methods can offer.

LLM Retrieval Efficiency. To further address this concern, we compare Prune-and-Grow retrieval with two graph traversal strategies: Breadth-First Search (BFS) and Personalized PageRank (PPR). We report the inference performance / average time (seconds) as follows.

We observe that while the LLM pruning incurs moderate overhead in response time, it brings consistently favorable performance gains beyond the non-LLM variants. Moreover, we respectfully clarify that CAM is inherently compatible with various retrieval methods, allowing users to choose the most suitable one based on their application needs.

Q5: Qualitative analysis.

A Qualitative Example. We follow your suggestion to analyze a representative question involving the long-form novel Jane Eyre (from the NovelQA dataset), which requires integrating two evidence chunks located over 11K tokens apart.

Question: It's someone's part to receive the company in Chapter XVII, then what did he/she did after being summoned?

Evidence Chunks:
- [211] Mrs. Fairfax assumed ... it was her part to receive the company ...
- [233] Mrs. Fairfax was summoned to give information respecting the resources of the house in shawls, dresses, draperies of any kind ...

Global Retrieval:
- [159] I should have liked something clearer; but Mrs. Fairfax either could not ... ❌
- [209] The party were expected to arrive on ... ❌
- [211] Mrs. Fairfax assumed ... it was her part to receive the company ... ✔️
- [653] Jane observes Mr. Rochester's strange behavior and Mrs. Fairfax ... (Summary node covering [159], [211]) ❌
- [701] Leah and Mrs. Fairfax are preparing for company at Thornfield. There is a sense of ... (Summary node covering [209], [211]) ✔️

Direct LLM Reasoning (without Prune-and-Grow): 
She took the company on a tour of the gardens and the surrounding estate. ❌
(This hallucinated answer is plausible but incorrect, as Chunk [233], which contains the actual answer, is missing)

CAM’s Prune-and-Grow Trace:
[Pruning Step 1] Chunks [211] and [701] are retained based on LLM judgment.
[Growing Step 1] Neighbors are explored to form the candidate set that includes [233].
[Pruning Step 2] The LLM selects [233] and retains [211] as contextually important.
[Growing Step 2] Neighbors are explored to form the candidate set.
[Pruning Step 3] The activation set stabilizes with [211] and [233].

Final LLM Reasoning (with Prune-and-Grow): 
She gave information respecting the resources of the house in shawls, dresses, draperies of any kind. ✔️
(This answer is fully grounded and faithful to the original text)

Error Analysis. To fully address your concern，we further provide a detailed error analysis of 100 errors made by CAM on the NovelQA dataset. These errors can be categorized into three main types:

• LLM Reasoning Errors: All relevant nodes were retrieved, but the LLM failed to compose or interpret them correctly, typically due to distraction from less relevant context or hallucinated inference.
• LLM Selection Errors: Relevant nodes are present in the candidate pool but not selected into the final activation set.
• Retrieval Coverage Failures: Relevant nodes are not explored at all. This often occurs when the input question is ambiguous or underspecified, making initial localization difficult despite the presence of relevant evidence.

We will include these in our final version to provide deeper insight into CAM’s failure modes.

Q6: Positioning constructivist theory within the broader landscape?

We appreciate the insightful comment. While CAM centers around a hierarchical structure, the constructivist perspective is not inherently limited to this form. At its core, constructivist theory highlights two crucial memory operations (assimilation and accommodation) over (hybrid-)structured schemata, where the structure may vary in type. This idea can naturally inform the design of episodic or hybrid memory systems. For example, episodic memory systems may benefit from flexible assimilation by associating a single experience with multiple semantic cues, while hybrid systems may use accommodation mechanisms to reconcile conflicting memory streams. We promise to include more discussions in our final version to better position the constructivist design principles within the broader landscape of memory types.

We sincerely hope that our responses have addressed your concerns. Your support really means a great deal to us.

[1] Zhu et al. Large language models for information retrieval: A survey. 2024.

Dear Reviewer mkPe,

Thank you very much for your feedback and acknowledgement.

During the rebuttal phase, we have made great efforts to address each of your concerns. We sincerely promise to incorporate the above discussions into our final version, including analysis on hyperparameter sensitivity, retrieval efficiency, and qualitative results. In addition, we will also expand the Limitation section to discuss the potential drawbacks of using LLMs for memory summarization, which is the de facto standard in current memory research, and outline directions for future works.

If you have any further questions or suggestions, we would be happy to continue the discussion. We sincerely thank you for your support.

Best regards,
Authors

Thanks for your careful and insightful comments. Below, we provide point-by-point responses to each concern raised.

Q1: Evaluation on more reading comprehension datasets.

Dataset Selection Rationale. Our selection of evaluation datasets is guided by three criteria:

• Recency of Release: The selected datasets (e.g., NovelQA from ICLR 2025 [1], FABLES from COLM 2024 [2]) are newly released and absent from LLM pretraining corpora, thereby minimizing the risk of test-time data leakage and ensuring a clean evaluation environment. In contrast, many conventional benchmarks (e.g., NarrativeQA, QASPER, QuALITY) have been released over three years, and thus pose a higher risk of data contamination [1-2].
• Long-Context Property: These datasets feature extremely long input texts, often exceeding 150K tokens, which are ideal for evaluating the scalability of agentic memory systems. By comparison, conventional benchmarks such as QuALITY and NarrativeQA feature relatively short contexts, with average lengths around 5K and 50K tokens respectively [1-2].
• Task Diversity: We deliberately include a broad range of task types—question answering, query-based summarization, and claim verification—to comprehensively test the generalizability of CAM.

Evaluation on MHQA Datasets (Appendix G). To align with standard practices, we have already included evaluations on three widely-used multi-hop QA benchmarks---HotpotQA, 2WikiMultiHopQA, and MuSiQue---as presented in Appendix G. CAM achieves superior performance across the three datasets, confirming that our design not only excels on long narrative inputs but also generalizes well to classical MHQA settings.

Evaluation on More Datasets. To further strengthen the scope of our work, we follow your suggestion to conduct more experiments on six conventional benchmarks. We compare CAM against four representative baselines using consistent experimental setups (same LLMs, embedding models, etc.) to ensure fairness. For your convenience, we also re-include results on the MHQA datasets below. These comprehensive results confirm CAM's robust performance across tasks and domains, significantly broadening the scope of our work.

Q2: In-depth fine-grained performance analysis.

We appreciate your suggestion to provide a more nuanced comparison of CAM against prior methods. To this end, we perform fine-grained performance analysis on the NovelQA dataset by leveraging its available question-type annotations. Specifically, the queries are categorized along two orthogonal axes:

(1) Complexity-Based Categories, including Single-hop (42.8%), Multi-hop (35.0%), and Detail (22.2%) questions, which reflect the reasoning depth required to arrive at an answer;

(2) Aspect-Based Categories, covering Time, Meaning, Span, Setting, Relation, Character, and Plot, which reflect the semantic focus of each query.

We compare CAM against four representative baselines (ReadAgent, RAPTOR, GraphRAG, MemTree) under identical settings.

[Performance on Complexity Categories]

[Performance on Aspect Categories]

The results demonstrate that CAM exhibits particularly strong gains on complex questions (e.g., Multi-hop, Times, and Span), which require integrating semantically distant evidence and maintaining narrative coherence.

Q3: A qualitative example on the advantages of Ego-Centric Disentanglement.

Ego-Centric Disentanglement is central to CAM’s ability to form coherent and nuanced memory abstractions. Its core motivation lies in capturing the semantic multifacetedness of complex text chunks---each chunk often relates to multiple themes, topics, or concepts. By replicating nodes based on their local semantic neighborhoods, CAM allows a single chunk to contribute to multiple higher-level abstractions, enabling more comprehensive and context-sensitive summarization.

To fully address your concern, we follow your suggestion and present a qualitative example to demonstrate the flexibility enabled by Ego-Centric Disentanglement. The example is drawn from the long-form novel Jane Eyre and involves a comprehensive question that requires integrating multiple semantically distant chunks of information, as detailed below.

Question: Why does Jane leave Thornfield after her wedding is interrupted?
Evidence Pieces: 
- Jane discovers that Rochester is already married (factual revelation).
- She experiences intense internal moral conflict about staying (psychological theme).
- She reflects on religious and spiritual values (spiritual/ethical dimension).
- She asserts her personal independence and dignity (character development arc).

[Without Flexibility]
Response: Jane leaves Thornfield because she finds out Rochester is already married. ❌
(This answer is semantically shallow, missing Jane’s inner struggle, moral reasoning, and identity preservation---these themes are essential to the novel’s integrity.)

[With Flexibility]
Response: Jane leaves Thornfield not only because of Rochester’s hidden marriage, but also because staying would betray her moral and spiritual convictions. Despite her love, she chooses self-respect and integrity, fearing the loss of her identity. ✔️

Q4: Improve the clarity of Sections 4.1.2 and 4.1.3 with a simple example.

Thank you for the suggestion. We will include a clear, step-by-step toy example in the revision to illustrate the processes in Sections 4.1.2 and 4.1.3. This example will be aligned with Figure 2 to better explain how memory nodes are disentangled and updated during the construction process.

We sincerely promise to incorporate the above experiments and discussions into our final version. We hope that these responses have addressed your concerns, and would greatly appreciate it if you could kindly raise the scores accordingly. Your support really means a great deal to us.

[1] Novelqa: Benchmarking question answering on documents exceeding 200k tokens. ICLR 2025.
[2] Fables: Evaluating faithfulness and content selection in book-length summarization. COLM 2024.

Dear Reviewer FDCi,

Thank you very much for taking the time to acknowledge our rebuttal.

During the rebuttal phase, we have made great efforts to address each of your concerns. In particular, we would like to kindly ask whether our responses have addressed your concerns regarding the evaluation on more conventional datasets and the fine-grained performance analysis.

If you have any further questions or suggestions, we would be happy to continue the discussion. Your support really means a great deal to us, and we sincerely look forward to your feedback.

Best regards,
Authors