Perspectra: Multi-Agent Research Deliberation

• Perspectra is an interactive multi-agent system that structures deliberation using forum-style threads and a real-time mind map.
• It employs persona-based LLM agents with long-term memory and retrieval tools to support targeted argumentation and critical reasoning.
• Empirical evaluation demonstrated enhanced proposal clarity, feasibility, and critical thinking compared to conventional group chat.

Perspectra is an interactive multi-agent system designed to structure, visualize, and enhance deliberation among LLM–driven, persona-based expert agents in research ideation settings. Its architecture integrates a forum-style threaded discussion interface with a real-time argumentation mind map, supporting targeted @-mention invitations, parallel thread branching, agent rationales, and user-editable persona profiles. Empirical evaluation demonstrates significant gains in critical-thinking behaviors, interdisciplinary engagement, and proposal refinement compared to traditional group chat, without an increase in cognitive load (Liu et al., 24 Sep 2025).

1. System Architecture and Pipeline

Perspectra’s architecture fuses two principal interfaces—a forum-style discussion panel and a dynamic mind map—underpinned by a reusable multi-agent pipeline. Each LLM agent $a_i$ is instantiated from a persona profile $P_i$, maintains a long-term memory store $M_i$, and accesses a literature database $\mathcal{D}$.

On dialogue turn $t$, agent $a_i$ observes the context $C_t$, issues tool calls (e.g., $\mathsf{GraphRAG}(\mathcal{D},q)$), updates memory via: $$M_i \;\leftarrow\; \mathsf{Distill}(M_i,\,C_t,\,r_{i,t})$$ and emits a post $r_{i,t}$. Deliberation follows a ReAct loop alternating “thinking” (planning) with “acting” (post/tool use), yielding formal locutions $$\ell\in\{\text{ISSUE}, \text{CLAIM}, \text{SUPPORT}, \text{REBUT}, \text{QUESTION}\}$$.

The mind map is rendered as a directed graph $G=(V,E)$, where $V$ comprises threads, posts, or replies, and $E$ records reply structure labeled by $\ell$. Semantic zoom controls the level of displayed detail.

2. Agent Personas, Deliberation, and Memory

Each persona $P_i$ encodes a structured taxonomy (discipline, methods, style, venues) and narrative background, user-editable through the interface. Long-term memory $M_i$ stores distilled idea snippets with lineage, promoting stance transparency.

Agents may issue retrieval calls from the GraphRAG-powered database, obtaining paper citations linked to dialogue content. Every agent turn can be represented as the tuple

$$(v_p,\,\ell,\,r)\quad \text{with}\quad \ell\in\{\text{ISSUE},\text{CLAIM},\text{SUPPORT},\text{REBUT},\text{QUESTION}\},$$

where $v_p$ is the parent node. Agent dialogue is regulated by commitment stores and argument composition rules, following frameworks such as Prakken’s and Toulmin’s models and ADF (Liu et al., 24 Sep 2025).

User steering is realized via selective @-mentioning of personas, dynamically assembling agent panels per thread. When multiple agents are invited, post weighting may be uniform or scale with profile “relevance” heuristics.

3. User Interaction and Visualization

Perspectra’s forum allows explicit, parallel thread management: users may spawn threads from any post via a “+ Thread” button, with independent expansion and collapse. The @-mention system invokes specific agents for targeted replies; if untagged, the parent post’s author responds.

The mind map offers instantaneous visualization of $G$, supporting force-directed layouts and zoom-dependent summarization: $$s_z(v) = \begin{cases} \text{keywords of }v, & z<z_0\ \text{full summary of }v, & z\ge z_0 \end{cases}$$ Nodes are interactive, linking back to the textual forum context. This spatial mapping enables rapid navigation across complex deliberation structures.

4. Empirical Evaluation Methodology

A within-subjects, counterbalanced study with 18 participants (undergraduates through postdoctoral level) compared Perspectra (forum + mind map) to a baseline group chat. Each participant engaged in two 30-minute research-ideation sessions, developing and revising proposals with agent guidance. Measures included:

• Pre/post self-rated proposal metrics (coverage, significance, clarity), NASA-TLX, and SUS-style usability;
• LLM (GPT-5) proposal evaluation (1–7 Likert, five dimensions);
• Critical-thinking activity coding mapped to established facets;
• Quantitative edit tracking and localization;
• Audio-recorded think-aloud and exit interviews.

Statistical analysis utilized paired $t$-tests, repeated-measures ANOVA for multidimensionality, and χ² testing for edit-type distributions. Inter-rater reliability for human-annotated proposal scoring yielded Krippendorff’s $\alpha=0.75$.

5. Quantitative Findings

LLM-judged proposal clarity increased more under Perspectra ($\bar\Delta=0.87$, SD=0.94) versus chat (0.39, SD=0.88), $t(17)=-2.15$, $p=0.039^{*}$, with effect size $d=0.51$. Feasibility improvement was 0.56 (SD=0.89) in Perspectra vs. 0.23 (SD=0.79) in chat, $t(17)=-2.37$, $p=0.024^{*}$, $d=0.63$. Self-assessed metrics corroborated these gains.

Total proposal edits averaged 5.35 (SD=1.80) in forum vs. 2.19 (SD=1.07) in chat, $\chi^2(1)=180.33$, $p<.001^{***}$. Edits in the motivation section were disproportionately higher: 33.1% (forum) vs. 4.3% (chat), $z=-3.98$, $p<.001^{***}$.

No significant differences appeared in cognitive load (NASA-TLX: mental demand, forum 3.89 vs. chat 3.94) or SUS usability (both ≈ 5/7).

Forum-mediated sessions yielded heightened critical-thinking activity: application (13% vs. 8%, $t=2.19$, $p=0.037^{*}$), analysis (10% vs. 6%, $t=2.01$, $p=0.049^{*}$), inference (15% vs. 9%, $t=2.33$, $p=0.028^{*}$), and evaluation (17% vs. 7%, $t=2.85$, $p=0.010^{**}$); global distribution χ²: $\chi^2(5)=11.33$, $p=.046^{*}$ (Liu et al., 24 Sep 2025).

6. Qualitative Insights and Design Implications

Structured adversarial discourse, achieved via ISSUE and REBUT moves among agents, was described by participants as analogous to “panel debate,” encouraging skepticism and assumption-testing. The @-mention mechanism facilitated ad hoc expert panel formation for emergent subtopics—an affordance absent in linear group chat modalities.

Participants reported that the system’s “productive friction”—requiring deliberate thread branching and agent selection—resulted in more considered reasoning and revisions. The mind map visualization was instrumental for maintaining argument context, surfacing emergent hypotheses, and planning follow-up actions.

Recommended design principles include embedding adversarial locutions to foster critical thinking, exposing agent rationales and memory in an on-demand fashion to prevent information overload, enabling flexible switching between low-friction chat and structured forum modes, and supporting mixed-modal branching for integrated forum-mind map navigation.

7. Significance and Prospective Directions

Perspectra demonstrates that LLM-based, persona-driven multi-agent frameworks—when structured around forum-based deliberation and user-controllable agent selection—can enhance depth and rigor in research ideation, yielding higher critical-thinking rates, more interdisciplinary engagement, and greater proposal refinement without increasing cognitive burden. A plausible implication is that integrating adversarial and user-steered multi-agent discourse may mitigate LLM sycophancy and improve scientific sensemaking in AI-assisted scholarly workflows (Liu et al., 24 Sep 2025). Future directions suggested include hybridizing lightweight linear chat with structured argumentation, adaptive visualization granularity, and deeper integration of literature-based argumentation graphs.