MiA-RAG: Mindscape-Aware Retrieval Augmented Generation

• MiA-RAG is a retrieval-augmented generation framework that builds a global 'mindscape' through hierarchical summarization to capture a document’s overall gist.
• It integrates global context into both query and chunk embeddings, overcoming fragmentation and local bias common in standard RAG systems.
• Empirical evaluations show that MiA-RAG delivers significant improvements in retrieval recall and coherent, evidence-based generation across diverse benchmarks.

Mindscape-Aware Retrieval Augmented Generation (MiA-RAG) is a framework for long-context understanding that equips retrieval-augmented generation (RAG) systems with explicit global context awareness. By constructing a global semantic summary, or "mindscape," through hierarchical summarization, MiA-RAG conditions both retrieval and generation on this representation. This architecture systematically addresses the fragmentation and local bias limitations characteristic of standard RAG when applied to long and complex documents, enabling more coherent and human-like retrieval and reasoning (Li et al., 19 Dec 2025).

1. Theoretical Foundations and Motivation

MiA-RAG draws on established psychological and neuroscientific models of human cognition that highlight the use of a holistic semantic "mindscape" in understanding, retrieving, and reasoning over complex information. Psychological theories—Bartlett’s schema theory, Fuzzy-Trace Theory, and the Encoding Specificity Principle—demonstrate that humans form gist-level summaries (schemas) that are reactivated in familiar contexts. Neuroscientific evidence from Controlled Semantic Cognition and neuroimaging studies further supports the role of global semantic representations in guiding attention, interpretation, and retrieval. The cognitive functions attributed to the mindscape include:

• Enriched understanding: Filling in missing details and resolving ambiguities in new information.
• Selective retrieval: Biasing search towards knowledge globally relevant to the full context, not just local matches.
• Integrative reasoning: Enabling coherence across dispersed evidence by aligning local details with the global semantic structure.

Conventional RAG systems encode queries and document chunks independently, disregarding global semantics, and leading to context fragmentation and non-coherent or hallucinatory generation when operating over long documents.

2. System Architecture and Data Flow

MiA-RAG operates through three principal components: a hierarchical summarizer, a mindscape-guided retriever (MiA-Emb), and a mindscape-conditioned generator (MiA-Gen). The procedural flow is outlined in the following pseudocode:

Input: Long document D, query q 1. Split D into chunks c₁,…,cₙ 2. S ← HierarchicalSummarizer({cᵢ}) # Global mindscape 3. For each chunk cᵢ: compute embedding e_dᵢ = f_d(cᵢ, S) 4. Compute enriched query embedding q̃ = f_q(q, S) 5. Score each chunk: scoreᵢ = q̃ᵀ e_dᵢ; choose top-K → R = {r₁,…,r_K} 6. Form generator input X = [INST; S; R; q] 7. y ← MindscapeConditionedGenerator(X) Output: y

• Hierarchical Summarizer: Builds global summary $S$ from all chunks, capturing document gist.
• MiA-Emb: Computes context-aware query and chunk embeddings using $S$ and retrieves top-K relevant chunks.
• MiA-Gen: Generates the final output, attending jointly to the mindscape and retrieved evidence.

This unified data flow ensures both retrieval and generation are globally informed, overcoming query-chunk independence and generator asymmetry.

3. Hierarchical Summarization Mechanism

The hierarchical summarization module recursively summarizes input text chunks to generate a compact, multi-level semantic representation:

• Let $s^{(0)}_i = c_i$ for $i=1,\ldots,n$.
• For each level $l=1,\ldots,L$, recursively generate:

$$s^{(l)}_i = \mathrm{Summarize}\bigl(s^{(l-1)}_{2i-1},\,s^{(l-1)}_{2i}\bigr)$$

• The final mindscape is $S = s^{(L)}_1$, containing a global context vector for the document.

This bottom-up process ensures that $S$ encodes the document’s gist and is used as auxiliary input for both retrieval and generation stages.

4. Mindscape-Guided Retrieval and Generation

Retrieval

MiA-Emb implements a dual encoder that integrates $S$ into both query and chunk representations via a residual-fusion mechanism. Specifically:

• Query embedding: $\tilde{q} = f_q(q, S)$
• Chunk embedding: $e_d = f_d(d, S)$ for all $d$ in the set of chunks
• The retrieval score is: $\mathrm{score}(q, d) = \tilde{q}^\top e_d$

Residual integration combines the original and task-induced hidden states:

$$\tilde{q} = \delta\,h_q + (1-\delta)\,h_t,\quad \delta\in[0,1]$$

Top-K chunks are selected by sorting these scores.

Generation

Given retrieved passages $R = (r_1,\ldots,r_K)$ and the mindscape $S$, a transformer-based generator maximizes:

$$P(y\mid R, S) = \prod_{t=1}^T P(y_t\mid y_{<t}, R, S)$$

The input prompt to the LLM is $[INST]_{gen}; S; r_1;\ldots;r_K; q$, allowing the generation process to attend to both summary and evidence jointly, supporting integrative reasoning.

5. Training Objectives

Training is conducted through modular and joint objectives:

• Retriever (MiA-Emb):
  • Silver-standard supervision: Aligns queries with chunks/nodes automatically.
  • Contrastive InfoNCE losses: For chunks ($\mathcal{L}_c$) and nodes ($\mathcal{L}_n$), defined by:

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

  with $t\in\{c, n\}$. - Multi-task retrieval loss: $$\mathcal{L}_{\mathrm{MiA-Emb}} = \beta\,\mathcal{L}_c + (1-\beta)\,\mathcal{L}_n$$.

• Generator (MiA-Gen):
  • Supervised fine-tuning: Cross-entropy loss on datasets such as NarrativeQA, CLIPPER:

  $$\mathcal{L}_{\mathrm{MiA-Gen}} = -\sum_{i}\sum_{t=1}^{|y_i|}\log P(y_{i,t}\mid y_{i,<t}, x^{\mathrm{gen}}_i)$$

• Unified joint training (optional):

$$\mathcal{L}_{\mathrm{MiA-RAG}} = \lambda_r\,\mathcal{L}_{\mathrm{MiA-Emb}} + \lambda_g\,\mathcal{L}_{\mathrm{MiA-Gen}}$$

6. Empirical Evaluation

MiA-RAG is evaluated on long-context, evidence-based, and bilingual benchmarks:

Dataset	Task/Metric	Avg. Context (tokens)
NarrativeQA	English long-form QA (F1, EM)	≈83,000
∞Bench-EN.MC	Multiple-choice (Accuracy)	≈184,000
DetectiveQA	Chinese & English (Accuracy)	≈118,000
NoCha	Claim verification (Pairwise Acc.)	≈139,000

Key findings:

• Retrieval Recall@K (K=3/5/10) on NarrativeQA: MiA-Emb-8B achieves 62.7/75.9/88.1%, substantially outperforming vanilla RAG (41.8/54.5/71.1%).
• End-to-end performance: MiA-RAG-14B achieves +16.2% average score improvement vs Vanilla-14B and +8.6% vs Vanilla-72B (despite using a smaller model).
• Ablation studies: Removing the mindscape summary reduces performance by 2.7% (absolute); omitting QA or claim verification supervision leads to further drops.

Graph-based experiments indicate that MiA-Emb selects nodes yielding higher judged Comprehensiveness (81.3% vs 18.7%).

7. Analytical Findings and Significance

Multiple analyses demonstrate the impact of the mindscape:

• Selective retrieval: t-SNE visualizations show MiA-Emb queries align more closely with relevant chunk subspaces (mean projection angle 37.1° vs 43.5% for vanilla), confirming bias toward document-relevant content.
• Enriched understanding: Layerwise attention analyses indicate that middle transformer layers exhibit increased attention from queries to summary tokens. Token-level inspection reveals focus on summary phrases aligning semantically with the query.
• Integrative reasoning: The Mindscape-Coherent Evidence Alignment (MCEA) metric quantifies attention correlation across query, chunk, and summary. MiA-Gen achieves higher MCEA, particularly in middle/late layers, and sharply loses coherence when the true summary is ablated.

These results indicate that hierarchical mindscape scaffolding enables human-like long-context retrieval (selective retrieval, enriched understanding) and coherent, evidence-grounded generation (integrative reasoning), with substantial gains across diverse benchmarks (Li et al., 19 Dec 2025).