NeuroPath: Neurobiology-Inspired Path Tracking and Reflection for Semantically Coherent Retrieval

Thank you for the detailed and insightful comments. We address the issues you raised below.

Issue 1: Motivation clarification: structural relevance vs. semantic coherence.

We have already provided a detailed query example and textual explanation in Figure 1 (Page 2), using comparisons across different methods to illustrate our motivation. We understand that further clarification of certain concepts can help better convey the motivation.

1. We define structural relevance as methods that retrieve graph nodes based on the graph topology (e.g., Personalized PageRank in HippoRAG, subgraph construction in LightRAG). However, such approaches risk retrieving nodes that are structurally close but semantically irrelevant.

2. In contrast, semantic coherence requires that the path formed by nodes and edges aligns with the logical intent of the query(e.g.,[Android]->(createdBy)->[Andy Rubin]->(founded)->[Essential Products]->(acquiredBy)->[Nothing] as illustrated in Figure 1 (c)).

Issue 2: Lack of analysis or evidence for semantic path learning

1. The case study (Section 4.3, Pages 7–8) shows example path sequences selected by Neuropath. Compared to HippoRAG and LightRAG, NeuroPath contain richer relational information and show clear logic connections to the query.

2. Figure 10 and Figure 11 (Appendix B, Page 20) provide concrete examples of Dynamic Path Tracking and Post-retrieval Completion

3. The ablation study (Section 4.4, Page 8) compares retrieval performance using only path tracking (i.e., without post-retrieval completion). The table below is based on Table 3 (Page 8) in the main paper and highlights the relative improvement (%) from path tracking alone under the recall@2 metric.

Table 1: Ablation study of the path tracking without post-retrieval completion. (Section 4.4, Page 8)

Issue 3-4：The methods and motivation background are unclear and inaccessible.

• Methods：

We describe our framework in detail in the following three aspects:

1. Algorithm 1 (Appendix B, Page 14) formalizes workflow with step-by-step pseudocode.
2. Figure 3 (Page 4) visually maps the interplay between static indexing, dynamic path tracking, and post-retrieval completion. Extended examples in Figure 10-11 (Pages 20-21) demonstrate concrete LLM outputs during path expansion. Appendix J (Page 19) provides detailed prompts.
3. Mathematical specifications (e.g., coreference resolution in Eq.1-2, path concatenation in Eq.4, pruning via Eq.5) anchor abstract concepts to computable operations.

We argue this multi-angle presentation meets reproducibility standards for complex systems, and we have also released our code and detailed reproduction steps on an anonymous GitHub repository (linked in the abstract).

• Motivation:

To aid readers unfamiliar with neuroscience, we summarize the relevant biological mechanisms and their parallels to retrieval, with further details in Appendix A (Page 13).

1. Limitation of current graph-based RAG: Methods like HippoRAG (PageRank-based) and LightRAG (subgraph expansion) construct static topologies, often introducing semantically incoherent or noisy results. In contrast, brain retrieval is dynamic and goal-directed.
2. In the brain, place cells support two key mechanisms: preplay for path planning and replay for memory consolidation. NeuroPath simulates this via LLM-based dynamic path tracking (navigation) and post-retrieval completion (consolidation).
3. Brain/Bio-inspired mechanisms have recently sparked new ideas in the LLM era. Our work aims to shift the RAG paradigm from “brain-like storage” (e.g., HippoRAG’s graph-based indexing) to “brain-like reasoning”, by enabling dynamic retrieval over structured knowledge graphs.

Issue 5-6: Lack of trade-off discussion for the method’s underperformance on HotpotQA.

Thank you for pointing out this limitation. We discuss it from three perspectives: 1. Shortcuts in HotpotQA, 2. Iter-based methods and tradeoffs, and 3. NeuroPath’s tradeoffs.

1. Shortcuts in HotpotQA: Prior work (e.g., HippoRAG [1], MuSiQue [2]) has shown that HotpotQA often allows shallow shortcut-based answering without full multi-hop reasoning. We have analyzed it in Appendix F.3 (Page 17). HippoRAG 2 achieves 9.3% more correct answers not backed by full evidence compared to NeuroPath (Table 10, Page 17). In fact, many HotpotQA questions can be effectively retrieved and answered using naive RAG methods [3], due to several inherent limitations:

   1. HotpotQA documents are short and dense.
   2. Most questions are synthetically constructed 2-hop queries with limited dependency complexity [2].
   3. Prior work [1] shows that distractor documents are often weak, making target passages easier to retrieve.

2. Iter-based methods and tradeoffs: These perform well on HotpotQA by matching dense chunks and refining queries via LLMs. However, they consume 22.8% more tokens (Appendix G) and are more sensitive to embedding models (Table 1, Page 7), resulting in inconsistent performance.

3. NeuroPath’s tradeoffs: Due to the limitations of HotpotQA, NeuroPath may underperform iter-based methods on this dataset. However, its path-based, fine-grained retrieval is better suited for complex multi-hop reasoning (e.g., MuSiQue and 2WikiMultiHopQA).

   It is also more robust to embedding model changes due to shorter and cleaner path token. It is worth noting that we evaluate the scalability of NeuroPath on other tasks:

   1. We demonstrate that our NeuroPath can retain the impressive performance of simple retrieval on the standard retrieval datasets PopQA and NaturalQuestions (Appendix D, Page 15).
   2. We also extended our evaluation to the more complex and uncertain sense-making task. (follow the settings and baselines in HippoRAG 2 paper) Our test results on the NarrativeQA dataset are as follows, and NeuroPath still maintains good performance.

Table 2: QA performance on sense-making task.

Given HotpotQA’s known limitations and NeuroPath’s broader scalability, we believe this trade-off is acceptable and will elaborate further in revised Appendix F.3.

Issue 7-8：Suggest using higher-quality datasets such as MultiHop-RAG.

Thank you for the suggestion. We conducted additional experiments on this dataset.

Setup: We selected 1,000 questions and split the documents into 512-token chunks. Contriever and BGE were used as embedding model, respectively.

Table 3: Retrieval results on MultiHop-RAG.

Results show that NeuroPath remains in the leading position. It is worth noting that other baselines show high sensitivity to the performance of the embedding model, whereas NeuroPath remains more robust in comparison.

Issue 9：The computational cost needs clearer quantification and comparison to existing methods.

We rigorously evaluated the cost and efficiency of NeuroPath against baselines (Appendix G, Page 18-19), and our results demonstrate a favorable tradeoff between overhead and performance gains.

Table 4: Comparison of graph indexing time with graph-based methods. (Page 19)

Table 5: Comparison of average retrieval time per question with iter-based methods. (Page 19)

1. NeuroPath’s indexing is 1.5× faster than graph/path-based methods due to single-pass KG extraction. It reduces token usage by 31.1% on average (Appendix G, Figure 9).

2. NeuroPath’s retrieval is 1~2s slower per query than iter-based methods, this is mainly due to the additional query processing. Path pruning allows embeddings to be computed in parallel on the GPU, with an average runtime of just 0.2 seconds per query. By calling LLM twice for each query, our token cost is reduced by 22.8% (Appendix G, Figure 9).

Limitation: Need more discussion of the proposed method's limitations and cost.

As we detailed in our response to Issues 5-9, we discussed the limitations of NeuroPath on HotpotQA, as well as the trade-offs between performance and scalability. We further validated the superiority of our method on a more challenging datasets MultiHop-RAG. Finally, we compared the computational cost of different methods and clarified the source of our method’s complexity.

[1] HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models

[2] MuSiQue: Multihop Questions via Single-hop Question Composition

[3] Compositional Questions Do Not Necessitate Multi-hop Reasoning

Thank you for appreciation of our motivation, method, and discussions. We address the issues you raised below.

Issue 1: Missing comparison to strong large-scale embedding models (e.g., baselines in HippoRAG 2).

Our NeuroPath aims to address the issues of semantic incoherence and noise in graph-based RAG by leveraging a goal-directed, dynamic semantic path tracking mechanism, which is independent of the embedding model size. Experimental results demonstrate that the LLM-driven path tracking mechanism itself brings significant performance gains when compared under the same embedding model conditions (e.g., Contriever and BGE-M3). Moreover, it exhibits scalability across LLMs of different sizes.

We appreciate the reviewer’s suggestion to conduct experiments with larger-scale embedding models. Following the baselines used in HippoRAG2, we selected three 7B-scale embedding models for evaluation, and set the embedding models for both HippoRAG2 and NeuroPath to NV-Embed-v2.

Table 1: Retrieval results using NV-Embed-v2 as the embedding model and GPT-4o-mini as the LLM for both HippoRAG 2 and NeuroPath.

The results show that NeuroPath remains at a leading level and achieves significant gains compared to NV-Embed-v2.

We also conducted experiments using open-source LLMs Qwen3-8B and Qwen3-4B.

Table 2: Retrieval results after replacing the LLM with Qwen3-8B and Qwen3-4B.

The results show that NeuroPath still outperform HippoRAG 2 on small-scale open-source models.

We believe smaller embedding models better reflect the performance gap between different RAG methods, particularly graph-based methods, and their sensitivity to embedding quality. In our experiments, other methods showed up to 20% gap on 2WikiMultiHopQA when changing embeddings. In contrast, our method remains stable across all three datasets (Section 4.2, Page 6).

Additionally, Contriever and BGE-M3 require only 1–5 GB of GPU memory to support most tasks, while larger models like NV-Embed-v2 (7B) need up to 26 GB in full precision. In our experiments, even with float16, it still required at least 20 GB, making it impractical for real-world use. We agree that using diverse embedding models helps evaluate robustness, but in future agentic applications, allocating resources to LLMs is likely more beneficial. Our method already achieves substantial gains with lightweight embedding models.

Issue 2: Does the method generalize beyond multi-hop retrieval?

We have verified that NeuroPath exhibits strong scalability on both standard retrieval and sense-making tasks:

(i) We have already evaluated the scalability of NeuroPath on standard retrieval tasks (PopQA and NaturalQuestions) in Appendix D (Page 15). The results demonstrate that our NeuroPath still outperforms graph-based methods (achieving 15% and 7% improvements over HippoRAG2 in recall@2 and recall@5, respectively) and remains highly competitive even when compared to iter-based methods like Iter-RetGen (with 14% and 6% gains on the PopQA dataset, respectively).

Table 3: Retrieval performance on standard retrieval tasks. (Page 15)

(ii) In addition to standard and multi-hop tasks, we also validated the scalability of NeuroPath on the sense-making task (the ability to interpret larger, more complex, or uncertain context). Our experiments followed the setup and baselines from the HippoRAG 2, using the NarrativeQA dataset. This dataset requires understanding and reasoning over large-scale narratives, such as long novels.

Table 4: QA performance on sense-making task.

Although not originally designed for sense-making, NeuroPath performs best on NarrativeQA. However, our focus was on factual multi-hop retrieval tasks. Experiments show that NeuroPath can be extended to more complex or uncertain tasks. We attribute this capability to the dynamic tracking process of the LLM, where the model can autonomously choose the most plausible path under uncertain context.

Thank you for appreciating the clear motivation, detailed description, and substantial improvements of NeuroPath. Below, we will address the weaknesses you point out.

Weakness 1: Should explore the framework's scalability and generalization

1. We have already evaluated the scalability of NeuroPath on standard retrieval tasks (PopQA and NaturalQuestions) in Appendix D (Page 15). The results demonstrate that our NeuroPath still outperforms graph-based methods and remains highly competitive compared to iter-based methods.

   Table 1: Retrieval performance on standard retrieval tasks. (Page 15)

   	PopQA		NQ		Avg.
   Method	R@2	R@5	R@2	R@5	R@2	R@5
   Contriever	27.0	43.2	29.1	54.6	28.1	48.9
   Iter-RetGen	35.4	46.6	40.9	68.6	38.1	57.6
   IRCoT	30.2	38.2	35.1	55.8	32.6	47.0
   HippoRAG	36.5*	52.7	22.5	45.7	29.5	49.2
   HippoRAG 2	34.2	47.4	32.1	59.5	33.2	53.4
   NeuroPath	40.5	49.3*	35.8*	65.2*	38.2	57.2*

2. We conducted additional experiments on the MultiHop-RAG benchmark, a challenging dataset specifically designed for evaluating multi-hop retrieval and reasoning. MultiHop-RAG contains 2,556 questions requiring 2 to 4 hops of reasoning. It features longer candidate documents and noisier supporting evidence.

   Table 2: Comparison of the average number of tokens per document.

   	MuSiQue	2Wiki	HotpotQA	MultiHop-RAG
   Average tokens per document	110	105	128	2,289 (chunk to 512)

   We selected 1,000 questions for retrieval, and the results are as follows:

   Table 3: Retrieval results on MultiHop-RAG.

   	MultiHop-RAG (Contriever)	MultiHop-RAG (BGE)
   Method	R@2 / R@5	R@2 / R@5
   Contriever	9.5 / 20.4	-
   BGE-M3	-	25.4 / 44.7
   Iter-RetGen	17.8 / 30.9	28.8 / 45.8
   HippoRAG 2	15.7 / 29.5	21.4 / 40.8
   NeuroPath	23.7 / 39.0	26.1 / 46.8

   Our method outperforms typical baselines, especially when using the smaller Contriever. Notably, other methods still exhibit sensitivity to the choice of embedding model, whereas NeuroPath shows relatively stable performance differences between Contriever and BGE-M3 (consistent with the conclusions in Section 4.2, Page 6).

3. In addition to multi-hop and standard tasks, we also evaluated the scalability for the sense-making task (the ability to interpret larger, more complex, or uncertain context). We followed the setup and baselines from the HippoRAG 2, using the NarrativeQA dataset. This dataset requires understanding and reasoning over large-scale narratives, such as long novels.

   Table 4: QA performance on NarrativeQA.

   Method	NarrativeQA (F1 Score)
   NV-Embed-v2 (7B)	24.2
   GraphRAG	20.9
   LightRAG	9.0
   HippoRAG	16.1
   HippoRAG 2	25.2
   NeuroPath	32.7

The results show that NeuroPath can be extended to more complex or uncertain tasks. We attribute this capability to the dynamic path tracking process of the LLM, where the model can autonomously choose the most likely path under uncertain context.

Weakness 2: Reliance on coreference resolution may cause errors on less structured or unclean data.

While fundamentally improving graph construction is not the goal of our work, we mitigate coreference-related errors at runtime.

The quality of generated KG data depends on the LLM's performance. While existing methods depend on multi-step LLM extraction, we simplify this with one-pass KG extraction and efficient embedding-based coreference resolution.

Although we simplified the process to achieve faster indexing, we still adopt a two-stage solution to address errors in coreference linkage:

1. Path tracking: For nodes with the same surface form but different referents, each path expansion step collects a series of triple fragments that carry relational context. The LLM uses these fragments to disambiguate and filter irrelevant coreferences.
2. Post-retrieval completion: If seed nodes or their coreference sets fail to capture key nodes during expansion, the final path may lack information. We address this with an additional embedding-based retrieval step using the output of the last expansion, which includes requests for missing information. (Ablation results: Section 4.4, Page 8.)

Despite the simplified KG construction, our NeuroPath still outperforms other graph-based baselines, demonstrating the strength of our retrieval framework.

We also verify whether LLMs can benefit from identifying phrases in the coreference set by changing the number k of expanded coreferent phrases (0, 1, 5).

Table 5: Impact of the number k of expanded coreferent phrases on retrieval.

In our entity extraction prompts (Figure 12, Page 22), we explicitly instruct the LLM to use noun phrases with explicitly referents. As a result, it tends to extract consistent phrases for the same entity, which explains why the performance gains from k = 0, 1, and 5 are relatively small. Nonetheless, the results show that using coreference sets still brings improvements, indicating that the LLM can recognize and leverage them.

Weakness 3: Method may still be resource-intensive due to the dynamic nature of the path-tracking and post-retrieval steps.

This insightful comment aligns with our goal of balancing performance and cost. We respectfully clarify that the main source of resource consumption (i.e., LLM calls) lies in the path tracking stage. The post-retrieval completion stage simply extracts the output from the final step of path tracking and performs fast vector retrieval. Its purpose is to compensate for any missing information that may have been overlooked during the path tracking process. In fact, it was our intentional resource control that resulted in a 22.8% reduction in token consumption over the iter-based methods.

In Dynamic Path Tracking stage, we use 2 strategies to control the token cost:

1. Candidate paths are pruned before being fed into the LLM. We use an embedding model to compute the similarity between each candidate path and the query, retaining only the top 30 paths. (default setting)
2. The LLM is used to further select valid paths, in order to prevent the expansion of irrelevant paths in the next step (if any).

Table 6: Ablation study of the pruning strategy. (Section 4.4, Page 8)

Our ablation studies (Section 4.4, Page 8) systematically evaluate the impact of path pruning on computational efficiency. Specifically, we compare three configurations: 1) no pruning, 2) retaining top-20 paths, and 3) retaining top-30 paths (default). The results demonstrate that:

1. Pruning to 20 paths reduces token consumption by 17.3%-52.6%
2. Our default setting (30 paths) achieves a 8.4%-45.7% reduction in token usage Notably, both pruning configurations maintain comparable recall performance, with the 30-path setting even showing slight improvements in some cases.

In real-world applications, the maximum number of expansion hops and pruning thresholds can be flexibly adjusted.

As noted in Appendix H (Page 19), LLM calls are the runtime bottleneck. Nevertheless, our method demonstrates the feasibility of leveraging LLMs for path tracking to address semantic incoherence and noise issues, leading to substantial improvements in retrieval quality for multi-hop queries. It also shows scalability on different tasks, such as standard factual retrieval task and sense-making task. Based on the above, we consider the tradeoff between performance and cost to be reasonable.

Weakness 4: The main concern is its reliance on high-quality KGs, which may limit real-world applicability.

We appreciate the reviewer’s concern about how KG quality may affect NeuroPath in real-world scenarios. To enhance real-world applicability, we will extend NeuroPath in two directions:

(a) Task-adaptive structured information processing. While the current KG directly reflects raw document content, higher-level tasks (e.g., abstraction or summarization) may require restructuring knowledge granularity. This flexibility will broaden task scalability.

(b) Efficient high-quality path expansion. Building upon optimized KGs, we will explore accelerated path-tracking mechanisms to alleviate LLM bottlenecks while maintaining retrieval precision.

Question: Explore the framework's scalability and generalization for more diverse and large-scale datasets.

As we detailed in our response to Weakness 1, our framework not only extends to noisier and more challenging multi-hop retrieval tasks, but also generalizes to standard retrieval tasks and more complex, uncertain sense-making tasks. Experimental results demonstrate its strong scalability and performance. We attribute this to the autonomy of the LLM in dynamic tracking.

Thank you for recognizing the novelty and experimental rigor of NeuroPath. We provide the following responses and perspectives regarding the weakness you pointed out.

Weakness: Graph construction offline is costly.

We agree that offline KG construction incurs upfront cost. However, NeuroPath and other RAG systems are designed for scenarios requiring repeated queries over static corpora. Here, one-time indexing amortizes cost, while online retrieval achieves 22.8% lower token consumption vs. iter-based methods (Appendix G, Page18).

Table 1: Comparison of graph indexing time with graph-based methods. (Page 19)

Our KG construction is 1.5× faster than graph/path-based RAG baselines (39 min vs. 52–72 min for HippoRAG/LightRAG/PathRAG on 2WikiMultihopQA, Table 11, Page 19) through:

1. Single-pass KG extraction: LLM generates entities and relations per document in one call (vs. multi-stage pipelines in baselines).
2. Embedding-based coreference resolution and similarity computation: Avoids LLM calls.

Although the graph construction stage inevitably incurs some cost, experiments show that NeuroPath brings significant gains over graph-based and iter-based methods under different LLMs and embedding models, demonstrating the effectiveness of our framework.

It is worth noting that finer-grained path-based retrieval is less sensitive to the performance of the embedding model (Section 4.2). Even small embedding models (e.g., Contriever) can effectively capture path semantics. In contrast, document-level retrieval in iter-based methods is highly sensitive to embedding model performance, with up to a 20% performance gap observed across different embedding models on the 2WikiMultiHopQA dataset (Table 1, Page 7).

Additionally, we further validated that NeuroPath can be extended beyond multi-hop retrieval to other tasks:

1. Standard retrieval tasks (PopQA and NaturalQuestions)

   We validated the generalizability of NeuroPath on standard retrieval tasks in Appendix D (Page 15). The results demonstrate that our NeuroPath still outperforms graph-based methods (achieving 15% and 7% improvements over HippoRAG2 in recall@2 and recall@5, respectively) and remains highly competitive even when compared to iter-based methods like Iter-RetGen (with 14% and 6% gains on the PopQA dataset, respectively).

   Table 2: Retrieval performance on standard retrieval tasks. (Page 15)

   	PopQA		NQ		Avg.
   Method	R@2	R@5	R@2	R@5	R@2	R@5
   Contriever	27.0	43.2	29.1	54.6	28.1	48.9
   Iter-RetGen	35.4	46.6	40.9	68.6	38.1*	57.6
   IRCoT	30.2	38.2	35.1	55.8	32.6	47.0
   HippoRAG	36.5*	52.7	22.5	45.7	29.5	49.2
   HippoRAG 2	34.2	47.4	32.1	59.5	33.2	53.4
   NeuroPath	40.5	49.3*	35.8*	65.2*	38.2	57.2*

2. More complex, uncertain retrieval scenarios such as sense-making tasks (NarrativeQA).

   We follow the experimental setup and baselines from the HippoRAG 2 paper, using NV-Embed-v2 as the embedding model. The results show that our method surprisingly outperforms the other baselines. Although our method was not originally designed for sense-making tasks, we seem to be able to demonstrate scalability. We attribute this to LLM actively choosing possible paths during path tracking when facing ambiguous or uncertain context.

   Table 3: QA performance on NarrativeQA (sense-making tasks)

   Method	NarrativeQA (F1 Score)
   NV-Embed-v2 (7B)	24.2
   GraphRAG	20.9
   LightRAG	9.0
   HippoRAG	16.1
   HippoRAG 2	25.2
   NeuroPath	32.7

We appreciate your recognition of these strengths and hope this clarifies the balance between efficiency, scalability, and performance.