Mindscape-Aware RAG

• The paper introduces MiA-RAG, embedding a global 'mindscape' summary within retrieval and generation processes to enhance evidence-based reasoning.
• MiA-RAG constructs its mindscape via hierarchical summarization, converting overlapping document chunks into a unified global semantic representation.
• Empirical evaluations show significant performance gains, with improved recall and accuracy in long-context and multilingual benchmarks.

Mindscape-Aware Retrieval Augmented Generation (MiA-RAG) is a Retrieval-Augmented Generation (RAG) framework that incorporates explicit global semantic context—termed the mindscape—into both retrieval and generation modules. Motivated by psychological studies of human reading comprehension, MiA-RAG models the process by which humans activate a global semantic schema to organize, retrieve, and synthesize dispersed evidence within long and complex texts. The method utilizes hierarchical summarization to construct a document-level semantic representation and conditions both the retriever and generator on this mindscape, yielding significant gains in long-context and evidence-based reasoning tasks (Li et al., 19 Dec 2025).

1. Psychological Motivation and Theoretical Framework

MiA-RAG is grounded in psychological and neuroscientific findings (e.g., Bartlett 1932; Tulving 1973; Ralph et al. 2017) that human readers form an abstract "gist" (schema) or global semantic representation when engaging with text. This "mindscape" mediates selective retrieval of relevant details and supports the coherent integration of information as new evidence is encountered. In contrast, conventional RAG systems process each query against a corpus in isolation, relying on local signals and potentially neglecting cross-document dependencies or global disambiguation.

The central hypothesis is that explicit representation of a mindscape—via a global summary $S$—enables enriched contextual understanding, more targeted evidence retrieval, and integrative reasoning in tasks requiring long-context semantic alignment.

2. Hierarchical Summarization: Construction of the Mindscape

The MiA-RAG pipeline builds the mindscape through hierarchical summarization, leveraging a pre-trained LLM summarizer $\mathcal{M}_s$ (specifically, GPT-4o):

1. Chunk-Level Summarization: The document $D$ is segmented into overlapping chunks $\{c_1, \dots, c_n\}$ (e.g., 1,200 tokens each). Each chunk is summarized:

$$s_i = \mathcal{M}_s\bigl([\texttt{INST}]_{sum_c},\,c_i\bigr)$$

The summarizer can be viewed as minimizing the loss

$$\mathcal{L}_{sum\_c} = -\sum_i \log P\bigl(s_i \mid c_i; \mathcal{M}_s\bigr)$$

2. Global Summarization: All chunk summaries $\{s_1,\ldots,s_n\}$ are concatenated in document order and summarized again to yield $S$:

$$S = \mathcal{M}_s\bigl([\texttt{INST}]_{sum_g},\,[s_1,\ldots,s_n]\bigr)$$

With corresponding loss:

$$\mathcal{L}_{sum\_g} = -\log P\bigl(S \mid s_1,\ldots,s_n; \mathcal{M}_s\bigr)$$

The overall summarization objective is $$\mathcal{L}_{summ} = \mathcal{L}_{sum\_c} + \mathcal{L}_{sum\_g}$$. In MiA-RAG, GPT-4o is deployed as a black-box prompting model for both steps.

3. Mindscape-Aware Retriever (MiA-Emb)

The retriever employs a bi-encoder architecture ($\mathcal{E}$, instantiated as Qwen3-Emb-8B with LoRA) that encodes the query $q$ in conjunction with the mindscape $S$: $\mathbf{q}' = f_{ret}(q, S) \in \mathbb{R}^d$

Retrieval Workflow

• Input Formatting: The input sequence is

$$Q = [[\texttt{INST}]_{emb};\,q;\,d_q;\,S;\,d_n;\,d_c]$$

where $d_q$ marks the query endpoint and $(d_n, d_c)$ specify retrieval modes.

• Encoding and Integration: The bi-encoder outputs hidden states $\mathbf{H}$, from which token representations $\mathbf{h}_q$ (at $d_q$) and $\mathbf{h}_t$ (at the retrieval mode token) are extracted. They are integrated via a residual:

$$\tilde{\mathbf{q}} = \delta\;\mathbf{h}_q + (1-\delta)\;\mathbf{h}_t\,,\quad \delta\in[0,1]$$

• Scoring: Retrieval is conducted by measuring cosine similarity with static chunk (or node) embeddings $\mathbf{d}$ in the corpus:

$$\mathrm{score}(q, S; d) = \mathrm{sim}(\tilde{\mathbf{q}}, \mathbf{d})$$

This approach enables the retriever to generate contextually enriched query embeddings, guided by the global semantic frame defined by $S$.

4. Mindscape-Aware Generator (MiA-Gen)

MiA-Gen (Qwen2.5-14B Instruct, fully fine-tuned) is conditioned on both the global summary $S$ and retrieved evidence $\widehat{C}_{ret}$:

$$p(\mathbf{y}\mid\widehat{C}_{ret}, S, q) = \prod_{t=1}^{|\mathbf{y}|} p\bigl(y_t \mid y_{<t},\,\widehat{C}_{ret},\,S,\,q\bigr)$$

Training employs standard autoregressive cross-entropy loss: $$\mathcal{L}_{MiA\text{-}Gen} = -\sum_{i}\sum_{t=1}^{|y_i|}\log\,p_{\theta}(y_{i,t} \mid y_{i,<t},\,\widehat{C}_{ret,i},\,S, q_i)$$

This dual conditioning ensures that generation processes the retrieved content within the coherent thematic context established by the mindscape.

5. Multi-Task Training and Optimization

MiA-RAG utilizes joint training objectives for the retriever and generator.

Retriever

• Positive (Silver-Standard) and Negative Evidence: Silver-standard units $U^+_i$ are generated via query augmentation, ensemble retrieval, and LLM filtering; negatives $U^-_i$ are sampled both randomly and adversarially.
• Contrastive InfoNCE Losses: For task $t \in \{c, n\}$:

$$\mathcal{L}_t = -\frac{1}{|B|}\sum_{j=1}^{|B|} \log\frac{\exp(\mathrm{sim}(\tilde{\mathbf{q}}_j, \mathbf{d}^+_j)/\tau)}{\sum_{d \in \{d^+_j\} \cup U^-_j} \exp(\mathrm{sim}(\tilde{\mathbf{q}}_j, \mathbf{d})/\tau)}$$

• Joint Loss: $$\mathcal{L}_{MiA\text{-}Emb} = \beta \mathcal{L}_c + (1-\beta) \mathcal{L}_n$$

Generator

• Autoregressive Cross-Entropy Loss: As described above.

Training both modules under the shared mindscape aligns retrieval and generative modules within a unified semantic frame.

6. Empirical Evaluation and Performance

MiA-RAG was evaluated across long-context and multilingual benchmarks:

Benchmark	Type	Avg. Length	Key MiA-RAG Gain
NarrativeQA	Free-form QA (EN)	~83K tokens	Recall@5: 54.5% → 75.9%; F1=53.5 (vs. 45.5 for vanilla 72B)
∞Bench-EN.MC	MCQA (EN)	~184K	Accuracy: ~80.8% → ~86.5%
DetectiveQA-ZH/EN	Bilingual QA	~118K	Consistent bilingual gains
NoCha	Claim verification	~139K	Multi-format improvements

MiA-RAG-14B outperformed a vanilla Qwen2.5-72B model on long-context tasks in average rank. On NarrativeQA, an absolute gain of +16.2 F1 points over a vanilla 14B model was observed. These results indicate substantial improvements in both retrieval localization and end-to-end reasoning.

7. End-to-End Workflow and Advantages in Long-Context Understanding

The MiA-RAG pipeline operates as follows:

Input: Long document D, query q // 1) Hierarchical summarization Partition D → {c_1…c_n} for each c_i: s_i ← SummarizeChunk(c_i) # GPT-4o prompt [INST]_sum_c S ← SummarizeGlobal([s_1…s_n]) # GPT-4o prompt [INST]_sum_g // 2) Mindscape-aware retrieval Q_repr = MiA_Emb.encode([INST_emb; q; d_q; S; d_n; d_c]) q* = δ·h_q + (1−δ)·h_t Retrieve top-k chunks: C_ret = arg top-k_d score(q*, d) // 3) Mindscape-aware generation x_gen = [INST_gen; S; C_ret; q] y* = MiA_Gen.generate(x_gen) Output: y*

The method excels in long-context scenarios due to:

• Global Anchor: The global summary $S$ narrows the retrieval space, supporting efficient localization in corpora exceeding 100K tokens.
• Symmetric Conditioning: Both retriever and generator leverage $S$, eliminating mismatches where only one module possesses the global summary.
• Multi-Task Supervision: Combined training over chunk- and node-level evidence plus varied end-tasks (QA, claim verification) yields robust, generalizable embeddings.

8. Contributions and Analytical Insights

MiA-RAG presents the first explicit computational model of human mindscape awareness for RAG. Its hierarchical summarization method is scalable and simple, facilitating the efficient creation of global semantic scaffolds that improve localization and reasoning across long-document inputs. Dual conditioning on $S$ addresses persistent challenges in aligning retrieval and generation. Both embedding space analyses and novel diagnostic metrics (e.g., MCEA, residual/attention pattern analyses) demonstrate that MiA-RAG internalizes authentic global semantics rather than relying on spurious dataset-specific shortcuts (Li et al., 19 Dec 2025).