Perspectra: Choosing Your Experts Enhances Critical Thinking in Multi-Agent Research Ideation
Perspectra’s contribution is strongest as a system-level reframing: it turns multi-agent ideation from a flat chat into a user-steered deliberation space with explicit expert selection and visible argument structure. The study suggests this can improve critical-thinking behaviors, but the evidence is still bounded to one task and a modest lab sample.
Axes Lens
Rare contribution shape, typical evidence profile. The point here is not a score. It is to show what kind of claim the paper makes, and whether the evidence pattern is unusual or baseline in this 268 -review set.
Contribution shape
- Knowledge form
- technical knowledge typical · 50/268
- Novelty type
- system architecture typical · 35/268
- Abstraction level
- system typical · 61/268
- Generalization target
- task class typical · 63/268
- Validation mode
- controlled experiment typical · 47/268
Evidence profile
- Evidence strength
- moderate typical · 105/268
- Claim alignment
- medium typical · 32/268
- Overclaim risk
- medium typical · 210/268
Review Summary
Perspectra is a credible CHI contribution because it does more than add another LLM agent wrapper: it proposes a specific interaction architecture for multi-agent ideation in which the user actively chooses which experts participate, can branch discussion by subtopic, and can inspect a structured representation of arguments and rationales. That combination is the paper’s main novelty, and it is well aligned with the claim that critical thinking in ideation is not just a property of model quality but of how deliberation is organized and made legible to the user. The evaluation is also appropriately scoped for a CHI system paper: a within-subjects study with 18 participants comparing Perspectra to a group-chat baseline on interdisciplinary proposal development. The reported outcomes—more critical-thinking behaviors, more interdisciplinary replies, and more proposal revisions—support the central design claim, but they do not justify broad generalization beyond this task class. The paper’s own limitations matter: it acknowledges possible bias from LLM-as-a-judge evaluation, the need for more expert assessment, and open questions about stability, cognitive load, and longer-term use. So the strongest reading is that Perspectra offers a useful design pattern for steerable multi-agent deliberation, with promising but still task-bounded evidence rather than a universal solution for multi-agent collaboration.
What Changed
Canon before
Prior CHI work on multi-agent ideation and persona-based LLM systems largely emphasized single-stream chat, agent role assignment, or conversational assistance; this paper positions user-steered expert selection and structured deliberation as the missing control layer for critical thinking.
Departure from common sense
The paper argues against the intuitive “one chat stream with multiple personas” model by showing that users need explicit control over which experts join which subtopic, plus branching threads and visible argument structure, to better support critical thinking during ideation.
Actual novelty
Perspectra combines user-invoked expert selection (@-mention/reply), threaded parallel exploration, and a mind-map view grounded in argumentation acts to externalize deliberation structure and rationales. The novelty is not just more agents, but a control-and-visualization layer for steering adversarial multi-agent discourse.
Evidence
The paper presents a new multi-agent ideation system and evaluates it in a within-subjects study with 18 participants on interdisciplinary research proposal development. The reported evidence includes increased critical-thinking behaviors, more interdisciplinary replies, and more proposal revisions versus a group-chat baseline, alongside explicit discussion of evaluation and deployment limits.
“ The interface design combines 1) a threaded canvas for parallel topic exploration with visualization of agent discourse dynamics informed by argumentation theory to aid sensemaking; and 2) feature that enables users to invite multiple self-chosen agents into an ongoing discussion”
actual novelty · Abstract + Introduction + 3.2.2 argumentation-act visualization · confidence 0.70
“ In order to obtain a more systematic understanding of how participants engage in critical thinking activities during system use, We analyzed users’ reply messages in Perspectra and compared them with their chat messages in the group-based condition, by annotating the messages with the critical thinking codebook described in section ”
departure from common sense · Introduction / system design description · confidence 0.62
“ Moreover, the use of LLM-as-a-judge during the evaluation of user-written proposals can bring in potential biases due to embedded LLM-generated snippets or content [ 71 , 98 ], which should be further examined in future work through more comprehensive expert evaluation”
limitation · Limitations and Future Work · confidence 0.80
“ We conducted a within-subjects study with 18 participants, who were asked to use both Perspectra and the baseline group-chat design to develop a brief proposal on an interdisciplinary research topi”
validation scope · Abstract + Methods (within-subjects study description) · confidence 0.78
Limits
Method limits
The evaluation is a within-subjects comparison on a specific ideation task with 18 participants, so the evidence supports task-level effects rather than broad claims about all multi-agent collaboration settings. The paper also notes concerns about LLM-as-a-judge bias and the need for more comprehensive expert evaluation.
Deployment limits
The system is framed for research ideation and interdisciplinary proposal development, not as a general-purpose replacement for all multi-agent chat workflows. Practical deployment depends on stable agent behavior, user willingness to manage threads and expert selection, and careful handling of evaluation bias.
Boundary conditions
Findings are most applicable when users are doing open-ended ideation that benefits from explicit expert selection, branching discussion, and visible rationales. The paper’s own limitations suggest caution for domains with different collaboration norms, unfamiliar topics, or settings where cognitive load from richer controls may outweigh benefits.
Position in field
This work sits at the intersection of multi-agent LLM interfaces, collaborative ideation tools, and critical-thinking support. Its contribution is a CHI-style interaction/system paper that reframes multi-agent assistance as a steerable deliberation environment rather than a single conversational agent.