Mindscape-Aware Generator (MiA-Gen)

Updated 28 January 2026

The paper introduces MiA-Gen, which conditions sequence generation on global mindscape summaries and instance-specific local evidence for enhanced long-context understanding.
It employs hierarchical summarization, contrastive retrieval training, and prompt engineering to merge semantic and behavioral signals in both RAG and AI journaling applications.
Empirical results demonstrate significant improvements in narrative benchmarks and well-being outcomes, validating MiA-Gen’s dual-domain efficacy.

The Mindscape-Aware Generator (MiA-Gen) refers to a set of system architectures that establish explicit context awareness in generation tasks by integrating global semantic representations—termed "mindscapes"—with local evidence or user context. The paradigm spans two major domains: (1) retrieval-augmented generation (RAG) in long-context natural language tasks and (2) behavioral-contextual AI-driven journaling powered by LLMs and multi-modal behavioral sensing. The central theme underlying MiA-Gen systems is the conditioning of sequence generation on both a holistic, global abstraction and instance-specific contextual signals, with demonstrated efficacy across domains including narrative understanding and well-being intervention systems (Nepal et al., 2024, Li et al., 19 Dec 2025, Nepal et al., 2024).

1. System Architecture and Input Conditioning

MiA-Gen in RAG (Long-Context Understanding):

Within the MiA-RAG framework, MiA-Gen is the generative component receiving three inputs: (i) a mindscape summary $S$ —a concise, document-level global abstraction computed via hierarchical summarization; (ii) a set of retrieved evidence chunks $\hat C_{\mathrm{ret}}$ —relevant local content selected using mindscape-aware retrieval embeddings; and (iii) the user’s query $q$ . The sequence-to-sequence model (Qwen2.5-14B-Instruct) is fine-tuned to synthesize these signals, enabling the generator to fuse global discourse structure with localized evidence for integrative reasoning. The input sequence for generation is formed as:

$x_i^{\mathrm{gen}} = [\texttt{[INST]}_{\mathrm{gen}}; S; \hat C_{\mathrm{ret}}; q_i]$

The model autoregressively generates answers, allocating attention over both global and local tokens without explicit gating (Li et al., 19 Dec 2025).

MiA-Gen in Behavioral Sensing (AI Journaling):

In the MindScape app, MiA-Gen comprises four to five principal modules: (1) a behavioral sensing layer acquiring raw multi-modal behavioral streams; (2) time-series feature extraction generating normalized and smoothed metrics; (3) a context encoder or inference layer producing standardized feature vectors $b_t \in \mathbb{R}^d$ and learned or hand-crafted embeddings $c_t = f_{\text{enc}}(b_t) \in \mathbb{R}^k$ ; (4) a prompt/template manager integrating high-salience features, historical usage, and user priorities into Jinja2 templates; and (5) an LLM interface (e.g., OpenAI GPT-4) operating via structured context prompts. The data flow is: raw sensors $\rightarrow$ feature extraction $\rightarrow b_t \rightarrow f_{\text{enc}} \rightarrow c_t$ $\rightarrow$ prompt (via template) $\rightarrow$ GPT-4 $\rightarrow$ user-facing journal or check-in. Context conditioning is performed through both embedding and natural-language summaries injected into prompts (Nepal et al., 2024, Nepal et al., 2024).

2. Mathematical Foundations and Context Fusion

Global–Local Contextualization in RAG:

The MiA-Gen generator models:

$P(y_{1:T} \mid S, \hat C_{\text{ret}}, q) = \prod_{t=1}^T P(y_t \mid y_{<t}, S, \hat C_{\text{ret}}, q)$

where at each inference step, standard transformer attention attends over concatenated tokens from $S$ , $\hat C_{\text{ret}}$ , and $q$ . No architectural modifications (e.g., gating) are necessary; alignment between summary tokens and retrieved content is acquired through fine-tuning with cross-entropy loss over the corpus $\tilde{\mathcal{D}}_{\mathrm{gen}}$ :

$\mathcal{L}_{\mathrm{MiA-Gen}} = -\sum_i \sum_{t=1}^{|y_i|}\log P_\theta(y_{i,t} \mid y_{i,<t},\,x_i^{\mathrm{gen}})$

(Li et al., 19 Dec 2025).

Context Representation in AI Journaling:

Behavioral feature vectors $b_t$ are linearly embedded:

$c_t = W \cdot (b_t - \mu)/\sigma + b$

where $W\in\mathbb{R}^{k\times d}$ , and $\mu$ , $\sigma$ , $b$ represent precomputed means, std, and bias. Generation is then factored as:

$P(x_{1:n} \mid c_t) = \prod_{i=1}^n P(x_i \mid x_{<i}, c_t)$

with $c_t$ represented as structured "User Data" in the prompt. Salience scoring for feature selection and prompt slot-filling is computed as:

$s_j(t) = \frac{\exp(\alpha x_j(t))}{\sum_k \exp(\alpha x_k(t))}$

and prompt selection probabilities weigh user priorities $p_j$ and feature salience $s_j(t)$ :

$P(\text{select feature }j) \propto \exp(\beta p_j + \gamma s_j(t))$

(Nepal et al., 2024, Nepal et al., 2024).

3. Hierarchical Summarization and Mindscape Construction

In RAG Systems:

The global summary $S$ (“mindscape”) is constructed in hierarchical passes:

Chunk-level: $s_i = \mathcal{M}_s([\texttt{INST}]_{sum_c}, c_i)$
Document-level: $S = \mathcal{M}_s([\texttt{INST}]_{sum_g}, [s_1, ..., s_n])$ where $\mathcal{M}_s$ is a frozen LLM (e.g., GPT-4o). The resulting $S$ is a condensed, globally coherent representation, capturing major characters, arcs, and settings (Li et al., 19 Dec 2025).

In Behavioral AI Journaling:

Mindscape-like context vectors are synthesized through windowed feature aggregation, normalization, and salience scoring. Change-points in the feature space (e.g., sleep deviations, activity bursts) are detected via time-series analysis, and only the most significant or user-relevant patterns populate the prompt context (Nepal et al., 2024, Nepal et al., 2024).

4. Training, Prompt Engineering, and Personalization

Training Objectives:

MiA-Gen for RAG: Full fine-tuning is applied on Qwen2.5-14B-Instruct, optimizing cross-entropy loss for answer generation. MiA-Emb (retriever) is trained with an InfoNCE contrastive objective, integrating query and mindscape summary for enriched chunk retrieval (Li et al., 19 Dec 2025).
MiA-Gen for Journaling: No LLM fine-tuning; prompt engineering and template design guide GPT-4 to generate personalized content with in-context signals (Nepal et al., 2024, Nepal et al., 2024). A hypothetical fine-tuning objective is suggested:

$L(\theta) = -\sum_{(b_t, \text{prompt}^*) \in D} \log P_\theta(\text{prompt}^* | f_{\text{enc}}(b_t)) + \lambda\|\theta - \theta_0\|^2$

Prompt Engineering:

Inputs comprise four sections: (A) system prompt (tone, style), (B) user context (priorities, temporal features, recent prompts), (C) response rules (e.g., word count, phrasing constraints), and (D) a strategy directive. Slot-filling and dynamic prioritization ensure contextual and personalized output (Nepal et al., 2024, Nepal et al., 2024).

Personalization Algorithms:

Users’ evolving engagement drives updates to a priority vector $p^{(i)}$ and a latent embedding $u_i$ . Prompt construction prioritizes categories and specific features corresponding to high user engagement, with parameters updated via simple gradient or counting updates to adapt longitudinally (Nepal et al., 2024).

5. Sensor Integration and Signal Processing

Behavioral Sensing Layer:

Raw signals incorporate multiple phone and optional watch modalities:

Movement (accelerometer/gyroscope)
App usage and screen time
Annotated geolocation (via DBSCAN clustering)
Microphone-based conversational detection
Phone log metadata (calls/SMS, without PII)
Sleep inference (prolonged inactivity plus screen-off)

Data are aggregated over regular time windows (e.g., 30 minutes) and uploaded encrypted to a backend. Feature engineering involves context-aware normalization, z-scoring, and rolling comparison to generate “increase,” “decrease,” or “stable” context flags (Nepal et al., 2024, Nepal et al., 2024).

6. Empirical Evaluation and Benchmarking

Long-Context Tasks (RAG):

MiA-RAG with MiA-Gen-14B surpasses vanilla RAG by +16.2% average on five long-context benchmarks (NarrativeQA, ∞Bench-EN.MC, DetectiveQA-ZH/EN, NoCha), exceeding even a 72B baseline by +8.6%. Conditioning the generator on $S$ yields +3.8% over no summary; full fine-tuning produces +11.2% improvement. Ablations reveal significant performance degradation upon removing the summary or supervision components. Embedding geometry and alignment analyses confirm that query vectors with mindscape integration cluster more coherently with corresponding semantic domains (t-SNE, MCEA scores) (Li et al., 19 Dec 2025).

AI Journaling and Well-being Outcomes:

In an 8-week MindScape exploratory study with 20 college students:

Positive affect increased by 7.2% ( $p=0.05$ , $d=+0.39$ )
Negative affect decreased by 10.6% ( $p=0.05$ , $d=-0.62$ )
Neuroticism dropped 11.8% ( $p<0.01$ , $d=-0.63$ )
Mindfulness improved by 6.8% ( $p<0.05$ , $d=+0.55$ )
Self-reflection increased 5.8% ( $p<0.05$ , $d=+0.47$ )
PHQ-4 total slope $\beta_1 = -0.25$ ( $SE=0.08$ , $p<0.001$ ) Participants generated more entries during contextual prompt phases (M=33.05) versus generic (M=7.11), confirming higher engagement and self-reflection. 85% rated usability as good/excellent, and 85% reported that contextual prompts facilitated deeper reflection (Nepal et al., 2024).

7. Limitations and Open Directions

MiA-Gen frameworks are limited by reliance on static, precomputed summaries, constraining adaptability to streaming content or rapidly evolving behavioral contexts. RAG-based variants are currently validated solely on narrative corpora; generalization to dialogue, scientific texts, or heterogeneous information sources requires further study. Behavioral journaling systems are presently tailored to the college student population, with adaptation to broader demographics untested. Supervision using commercial LLMs as evaluators introduces potential model bias and hallucinations. Directions for future research include dynamic mindscape updates, joint fine-tuning of summarizers and retrievers, domain adaptation, and integration with incremental or user-feedback-driven learning (Li et al., 19 Dec 2025, Nepal et al., 2024).

References:

(Nepal et al., 2024) MindScape Study: Integrating LLM and Behavioral Sensing for Personalized AI-Driven Journaling Experiences
(Li et al., 19 Dec 2025) Mindscape-Aware Retrieval Augmented Generation for Improved Long Context Understanding
(Nepal et al., 2024) Contextual AI Journaling: Integrating LLM and Time Series Behavioral Sensing Technology to Promote Self-Reflection and Well-being using the MindScape App