PaperBridge: AI-Enhanced Research Narratives

• PaperBridge is a human–AI co-exploration system that synthesizes research narratives by dynamically grouping publications using bi-directional LLM analysis.
• It integrates automatic paper clustering, narrative spark generation, and editable slide deck creation to support iterative narrative refinement.
• User studies indicate high usability and low cognitive overhead, demonstrating its effectiveness in interdisciplinary academic communication.

PaperBridge refers to a human–AI co-exploration system that assists researchers in exploring and synthesizing their publication corpus into coherent, context-adaptive research narratives. The system is specifically architected to support both top-down and bottom-up narrative construction via a bi-directional analysis engine powered by LLMs. PaperBridge was developed to address the challenge of assembling scholarly work into compelling narratives, particularly for interdisciplinary researchers whose output spans diverse domains and publication venues (Zhang et al., 19 Jul 2025).

1. System Architecture and Bi-Directional Analysis Engine

The PaperBridge system consists of three major components:

• Paper Management Module (Panels A and B): Ingests a user’s publication list (retrieved automatically via a Google Scholar URL) and generates an initial topic-based and user-aligned categorization.
• Narrative Exploration Workspace (Panels C, D, E): Supports multi-perspective, iterative exploration of narrative structures. PaperBridge injects narrative “sparks” (candidate groupings and contribution statements) using LLM-driven structured prompts and synthesizes thematic clusters based on high-level frameworks (parallel, linear, circular, coordinate).
• Slide Draft Preview (Panel F): Automatically transforms the curated narrative into a presentation-ready slide sequence, which users can further edit.

At its core, PaperBridge’s bi-directional engine enables incremental, iterative refinement. In the top-down mode, users select a global narrative framework, and the system proposes candidate groups and abstractions. In bottom-up mode, users directly rearrange or assign papers among themes, after which the LLM refines contribution statements and thematic structure. PaperBridge maintains all intermediate representations in a structured JSON schema, bridging both organizational flows for seamless revision.

2. User Interaction Model

The user workflow in PaperBridge begins with corpus retrieval and narrative intent specification. Paper clustering and initial grouping are auto-generated but are fully editable—users can relabel, merge, or reorganize paper themes. In the exploration area, users select a narrative framework then interact with candidates or “sparks,” which consist of:

• A synthesized contribution statement
• Multiple thematic clusters (each cluster showing characteristic keywords and assigned papers)

Users adjust groupings through drag-and-drop; all updates invoke an LLM-driven revision of contribution statements and cluster descriptors. A Rationale Mode enables refinement of justification strategies (e.g., ethos, pathos, logos) by dragging rhetorical structures into the workspace. Once a candidate narrative is adopted, the system generates an editable slide deck.

3. Narrative Construction and Evaluation

PaperBridge operationalizes narrative construction via:

• Structured Narrative Frameworks: Top-down selection (parallel, linear, circular, coordinate) leads to systematic “sparks.” For example, in the linear model, clusters are temporally or thematically sequenced, illustrating developmental progression.
• Cluster Synthesis and Statement Alignment: The system uses LLMs to abstract publication metadata and to validate statement-cluster coherence.
• Unified JSON Narrative Schema: All candidate perspectives, contribution statements, and thematic clusters are encoded, serving as an intermediate language for both the system and user.
• Quantitative Evaluation Metrics: Multiple semantic and structural metrics are computed for each candidate. The comprehensive quality score for each narrative perspective is given by

$$\text{FinalScore} = 0.2~\mathrm{SCA} + 0.2~\mathrm{SC} + 0.2~\mathrm{ARI} + 0.2~\mathrm{PCS} + 0.2~\mathrm{ICC}$$

where SCA is statement-cluster alignment, SC is structural consistency, ARI is adjusted Rand index, PCS is paper–cluster similarity, and ICC is intra-cluster cohesion. Each is normalized to $[0,1]$.

4. Empirical Evidence and Usability Findings

A user paper with 12 seasoned HCI researchers (N=12) demonstrated exceptionally high system usability (mean System Usability Scale >84) and low cognitive overhead. Participants efficiently generated and compared alternative narratives, lowering the barrier to narrative ideation. The “spark” mechanism often validated users’ preconceptions yet also prompted new narrative framings that had not previously been considered. The interface’s affordances (direct manipulation, immediate LLM synthesis, rationale annotation) fostered a reflective, exploratory process, frequently surfacing underexplored facets of a researcher’s portfolio. Some users expressed a desire for more flexible framework composition and granular control, suggesting extensions for future iterations.

5. Thematic Clustering and Justification

Thematic clustering leverages LLM-based analysis of paper titles, abstracts, and keywords to identify distinguishing features. Initial clusters are automatically surfaced but are fully user-editable. PaperBridge recalculates and re-synthesizes key cluster descriptors and contribution statements on-the-fly. The process is scaffolded by:

• Semantic Cohesion Metrics: Paper–cluster similarity and intra-cluster cohesion are quantified using measures such as cosine similarity or average pairwise keyword overlap.
• Structural Alignment: Adjusted Rand index (ARI) assesses agreement between system and user groupings relative to baseline clusterings or ground-truth user-provided themes.
• Narrative Consistency Checking: Statement-cluster alignment (SCA) quantifies logical and semantic fit between high-level claims and the underlying cluster content.

Justification strategies (ethos, pathos, logos) can be injected as “rationale blocks,” allowing researchers to align narrative logic with communication goals.

6. Implications for Academic Communication

PaperBridge’s design supports flexible, multi-perspective reframing of scholarly portfolios for diverse contexts: job talks, grant applications, or public engagement. By exposing high-level frameworks and supporting low-level curation, it is particularly valuable for interdisciplinary researchers, whose work resists traditional linear or mono-thematic organization. The underlying narrative structuring pattern—bi-directional, LLM-driven, and schema-grounded—offers a roadmap for future interactive systems that scaffold reflective academic communication.

Potential future research directions include:

• Integration of richer input sources (full papers, slides) beyond abstracts and titles.
• Expansion of narrative frameworks and customizability.
• Comparative user studies with general-purpose LLM interfaces.
• Domain transfer to educational portfolios, course syllabi, or interdisciplinary grant consortia.

7. Summary and Prospective Impact

PaperBridge demonstrates the efficacy of human-AI co-exploration in the construction and communication of coherent research narratives. Core innovations include a bi-directional, LLM-driven analysis engine; tightly integrated narrative frameworks; and quantifiable evaluation metrics for narrative perspective selection. The system’s design has immediate utility for academic storytelling and long-term implications for mixed-initiative tools supporting creative, interdisciplinary scholarly communication (Zhang et al., 19 Jul 2025).