RevTogether: AI for Science Story Revision

• RevTogether is a web-based multi-agent system that scaffolds the revision of complex scientific narratives through interactive, human-like GPT-4o agents.
• It integrates three distinct agents—including critic-type commentators and a writing assistant—to provide structured feedback and affective cues via avatars.
• The system offers a graduated user agency model that blends free-form feedback with inline revision suggestions, enhancing both learning and narrative clarity.

RevTogether is a web-based multi-agent system (MAS) for supporting the revision of science stories through interaction with multiple human-like AI agents powered by GPT-4o. Designed to scaffold the transformation of complex scientific concepts into coherent, accessible narratives, RevTogether orchestrates three distinct AI agents around a central text-editing interface. These agents simulate affective reactions, provide feedback, and supply structured revision guidance at varying levels of user agency. It integrates affect simulation through avatar-based non-verbal cues and supports a graduated agency continuum, enabling writers to tailor the degree of AI-driven intervention in their work (Zhang et al., 3 Mar 2025).

1. System Architecture

RevTogether employs a single-human, multi-agent architecture, comprising two commentator agents (“Mad Scientist” and “Curious Girl”) and one writing assistant agent. The system is implemented as a modern web application:

• Front-end: Built in React, utilizing the react-quill editor, it enables text selection, comment display, avatars, technique tags, highlighted spans, and inline revision rendering.
• Back-end: Implemented with Python Flask, it orchestrates request routing, prompt template assembly, and communication with the LLM service.
• LLM Service: All agent reasoning is delegated to GPT-4o, which handles comment generation, sentiment analysis, writing technique recommendation, and auto-revision.
• Avatar Engine: The FLUX-pro 1.1 engine renders three affective avatars that visually represent agent sentiment and reactions.

Interaction is structured as a sequential flow—users select text, prompt a commentator agent, review and accept/reject feedback, and, if desired, invoke the writing assistant agent for technique-driven suggestions and candidate line edits. The front-end presents all feedback and workflow controls inline, augmenting user engagement with dynamically changing avatars.

2. Agent Roles and Feedback Logic

RevTogether distinguishes between critic-type commentator agents and a companion-type writing assistant agent, each embodying unique personas and operational logic.

Commentator Agents:

• Personas:
  • Mad Scientist: Prioritizes technical rigor and clarity; visually signals negative sentiment with an “angry” emoji.
  • Curious Girl: Focuses on accessibility and narrative engagement; uses a “disappointed” avatar to indicate negative tone.
• Workflow:
  • Users select a text span $T_{\text{sel}}$.
  • The system issues a persona-seeded prompt to GPT-4o, which returns a comment $C$ and sentiment label $$s \in \{\text{positive}, \text{neutral}, \text{negative}\}$$.
  • Feedback is parsed and presented above the respective avatar; hovering reveals the underlying affective cue.

Writing Assistant Agent:

• Function: Activated after a comment is accepted; proposes revision strategies grounded in a fixed set of four techniques (humor, analogy/metaphor, emotional arousal, suspense/surprise).
• Process:

1. LLM is prompted with the full story and the accepted comment.
2. GPT-4o returns a subset of applicable techniques and highlights candidate text spans.
3. User selects a technique, then a highlighted span, to preview/propose an inline revision; double-clicking confirms the edit.

• Affect: The writing assistant operates exclusively through text and does not engage in affect simulation.

3. Affect Simulation and Non-Verbal Cues

Affect simulation in RevTogether is instantiated at two system junctures:

• Comment Sentiment Display: Each feedback comment is tagged with a sentiment score $s \in \{+1, 0, -1\}$. The avatar function $A(s)$ maps sentiment to one of three pre-rendered avatars, dynamically updated on user interaction (hover).
• Reaction to Accept/Reject:
  • Accept: Triggers a temporary shift to a positive avatar ($A(+1)$) for $\Delta t = 1 \text{ s}$.
  • Reject: Triggers a negative avatar ($A(-1)$), again for 1 second.

This mechanism is formally described by:

$$A_t = \begin{cases} A(+1), & u = \text{accept},\; t \in [t_0,\,t_0+1\text{ s}] \ A(-1), & u = \text{reject},\; t \in [t_0,\,t_0+1\text{ s}] \ A(0), & \text{otherwise} \end{cases}$$

Affect displays are non-verbal and non-persistent, designed to cue user attention and elicit reflection or engagement without prescriptive intervention.

4. Levels of User Agency

RevTogether exposes a graduated continuum of human agency over the revision process, providing three structured levels:

Level	Control Modality	Writer Role
1	Human-like comments only (free-form feedback)	Full editorial autonomy
2	Writing technique suggestions on comment acceptance	Guided (choose path)
3	Inline auto-revision with selectable techniques	Primarily evaluative (accept/reject AI edits)

• High Agency: Writers receive only free-form feedback and retain full control over content.
• Medium Agency: When a comment is accepted, the system proposes one or more writing technique tags as structured but non-intrusive guidance.
• Low Agency: The system surfaces detailed revision suggestions; users preview and selectively accept inline edits, with the AI assuming a more active editorial role.

This layered framework enables writers to modulate their reliance on AI intervention and supports both learning and productivity optimizations across the continuum.

5. Preliminary User Study: Findings and Implications

A qualitative user study (N=3, PhD students aged 22–30, non-expert in science storytelling) was conducted to evaluate RevTogether’s usability and impact. Each participant drafted a science story (via GPT-4o), sought at least two comments per commentator, and accepted one from each for deeper revision. Data were collected through think-aloud protocols and post-hoc interviews.

Key findings:

• Transparency and Learning: Participants expressed a desire for greater visibility into the agents’ reasoning processes, viewing AI as a potential tutor. Exposing decision rationales or intermediate steps was perceived as supportive for skill acquisition and fostering trust.
• Emotional Engagement: Non-verbal affective cues—such as avatars displaying unhappiness upon comment rejection—influenced revision choices, even though participants understood these emotions were simulated. The avatar feedback elicited reflection and enhanced user engagement.

No statistical measures were derived, and all findings are exploratory. A plausible implication is that affect simulation and explainable reasoning mechanisms are individually salient for user motivation and metacognitive engagement.

6. Design Implications and Future Directions

The initial deployment of RevTogether yields several recommendations for MAS design in the scientific writing domain:

• Expose AI Reasoning: Incorporate “chain-of-thought” or stepwise annotation features to enhance transparency and support user learning objectives.
• Calibrate Affective Feedback: Systematically assess the influence of affect cues through ablation studies, tuning their intensity and timing for optimal engagement without distraction.
• Expand Agent Persona Repertoire: Broaden the spectrum of commentator personas (e.g., domain-expert critic, layperson skeptic) to replicate multiple reader viewpoints and enrich feedback diversity.
• Scale Evaluation Protocols: Enlist professional science writers and larger, more diverse participant samples to substantiate findings and generalize effectiveness.
• Enable Granular Agency Controls: Develop user-facing controls (potentially a continuum slider) to dynamically adjust the level of AI involvement and editorial autonomy in real time.

RevTogether exemplifies how a MAS integrating critic-type commentators, a revision-oriented companion agent, non-verbal affect simulation, and differentiated user agency can support iterative science story revision while facilitating transparent and emotionally engaged human-AI collaboration (Zhang et al., 3 Mar 2025).