PaperlayerView exact revision on arXiv

Source-linked AI summary

MAIA: Multi-Agent Intent Articulation for Requirement Discovery in Art Commissions

Yu-Chao Wang, Yanhong Lu, Yingjie Victor Chen, Tim McGraw

arXiv:2609.12097v1cs.HCcs.AI

TL;DR

Laypeople can have strong affective intent for an art commission without knowing how to express it as visual requirements, and requirement discovery remains underexplored. MAIA uses multi-agent Socratic inquiry under “Verification over Invention” to produce user-verified briefs; its within-subjects study reports a large Cognitive Support gain over a minimal baseline.

  • Problem

    Requirement discovery is an underexplored upstream stage because commission intent may still need to be constituted rather than treated as a stable request to optimize.

    Source: §I · p. 1

  • Method

    MAIA uses multi-agent Socratic inquiry and a validator gate to turn vague intent into a text-only brief, with AI-proposed visual elements requiring user ratification.

    Source: §I · p. 1, 1) Coordinator and Specialized Roles · p. 3

  • Results

    In a within-subjects study (N=16), the full configuration produced a large gain in Cognitive Support over a minimal baseline, with the gain converging across statistical tests.

    Source: §VII · p. 7

  • Takeaways & Limitations

    The validator gate structurally blocks unratified content, making the system’s commitment to verified brief content enforceable.

    Source: §I · p. 1

    The artist-side evidence is exploratory, based on three raters and four tasks per arm; a larger preregistered blind review remains the stated priority.

    Source: C. Limitations and Future Work · p. 6

Abstract

from arXiv · show

In bespoke art commissions, laypeople know what they feel but lack the words to specify it: one participant wanted a laid-off truck driver depicted as "a ghost in his own machine" but left the medium, scale, and palette unsaid. We frame this as an articulation bottleneck at an under-served upstream stage: requirement discovery, which precedes any artist or image generator and forces the commissioner to constitute intent in the first place. We present MAIA (Multi-Agent Intent Articulation), a multi-agent system that scaffolds this stage through Socratic inquiry under a "Verification over Invention" rule, turning vague affect into a text-only brief of visual terms the user verifies. In a within-subjects study (N = 16), the full configuration produced a large, significant gain in Cognitive Support over a minimal baseline (r = 0.96, p_FDR = 0.015; LMM p_FDR < 0.001). Thematic analysis traces the same mechanism, and a validator gate structurally blocks unratified content. A complementary blind review by three professional concept artists on a sampled set of briefs corroborates this improvement from the artist's side: AI rewriting improved visual completeness and executability in all eight sampled tasks (task-level Wilcoxon p = 0.008; FDR q = 0.010), with directionally larger gains under MAIA than under the baseline (underpowered, d = 1.4-2.6).

I. INTRODUCTION

Art commissions face an articulation bottleneck: laypeople may have strong affective intent but lack the terminology to turn it into actionable visual requirements. MAIA addresses this upstream requirement-discovery stage with Socratic inquiry and a text-only brief whose proposed visual elements require user ratification.

I. INTRODUCTION

  • Laypeople may omit important visual parameters or bury them in emotional prose, leaving artists to reconstruct what was meant.The gap is framed as an articulation bottleneck between a felt proto-idea and an actionable visual code.Source: §I · p. 1
  • Requirement discovery is an under-served upstream stage where intent must be constituted, not treated as an already stable request to optimize.Prior work has largely bypassed the transition from an ambiguous impulse to a concrete concept.Source: §I · p. 1, §II · p. 1
  • MAIA uses iterative Socratic inquiry to turn low-context impulses into concrete visual terms users can verify, under “Verification over Invention.”Any AI-proposed visual element the user did not supply must be ratified before entering the brief.Source: §I · p. 1
  • The pipeline separates upstream, text-only brief formation from downstream artist negotiation, where the artist clarifies commercial terms and produces the artwork.The system neither generates nor reads images.Source: Figure 1 · p. 2

A. Expert-Layperson Communication

  • Laypeople’s verbose narratives can be valuable proto-ideas; useful briefs fix important visual anchors while leaving room for the artist’s judgment.The stated aim is to turn affect into concrete technical requirements without forcing arbitrary choices the user never made.Source: A. Expert-Layperson Communication · pp. 1–2

B. Generative AI in Design

Generative-AI design work has concentrated on downstream artifact production, while requirement discovery remains underexplored. MAIA’s framing shifts creative workflows from message optimization toward active epistemic co-articulation.

B. Generative AI in Design

  • Generative-AI design work largely addresses downstream artifact production and workflow efficiency, leaving requirement discovery—the move from ambiguous impulses to concrete concepts—underexplored.Text-to-image generation may overwrite latent intent with AI-hallucinated imagery.Source: B. Generative AI in Design · p. 2
  • An AI mediator can lower the social cost of voicing half-formed ideas and keep artists out of the early loop, but existing scaffolding tools do not negotiate both artistic freedom and commission constraints.CausalMapper emphasizes formal causal reasoning, while the Ideation Compass stays close to open-ended reflection.Source: B. Generative AI in Design · p. 2

C. AI-MC in Creative Workflows

  • AI-mediated communication usually assumes the sender already knows what to say, whereas composing a commission message can itself be how the commissioner figures that out.The paper argues that treating half-formed affect as complete can lead a translator to fill missing detail with model defaults.Source: C. AI-MC in Creative Workflows · p. 2
  • The paper frames creative workflows as active epistemic co-articulation through maieutic inquiry, distinct from model-side “maieutic prompting.”This is operationalized as design goals that fix technical requirements while retaining interpretive room for the artist.Source: C. AI-MC in Creative Workflows · pp. 2–3
  • The commissioner–artist pair and the upstream moment before they speak have received little study, especially from the artist’s perspective.The study addresses the commissioner’s side and adds an exploratory blind review by professional concept artists of brief executability.Source: C. AI-MC in Creative Workflows · p. 3

III. SYSTEM DESIGN

MAIA is presented as a multi-agent epistemic scaffolding system intended to help users traverse the problem space and reach shared understanding with an artist. Its design section introduces three goals for operationalizing the AI as a co-articulator.

III. SYSTEM DESIGN

  • MAIA uses a user-facing coordinator that dispatches work to specialized sub-agents, with role separation enforced by the architecture.The implementation uses OpenCode Desktop v1.1.60 and MiniMax M2.5 at temperature t=0.2.Source: §III · p. 3
  • The system is framed as epistemic scaffolding that helps users traverse the problem space toward shared understanding with an artist.The paper says the implementation adopts its harness so architectural role separation is enforced.Source: §III · p. 3

A. Design Goals

  • The design section introduces three goals to operationalize the AI’s role as a co-articulator.The supplied passage announces the goals but does not specify them.Source: A. Design Goals · p. 3

1) DG1: Fostering Crystallization:

MAIA’s design pairs translation into artist-usable visual specificity with bounded Socratic choices that help users make decisions while preserving their authorship. Its architecture separates analysis, validation, and synthesis and tracks what is explicit, inferred, or deferred.

1) DG1: Fostering Crystallization:

  • DG1 links a user’s low-context phrase to its visual code rather than mechanically replacing lay wording with jargon.The aim is to reduce the artist’s effort decoding the brief while keeping the user in control of meaning.Source: 1) DG1: Fostering Crystallization · p. 3

2) DG2: Epistemic Scaffolding via Maieutic Inquiry:

  • DG2 uses bounded either/or questions to help novices make explicit decisions, while an “Other” option preserves freeform answers and authorship.The paper motivates bounded choices as easier to start with than unrestricted freedom, and as easier to evaluate than generating a preference from scratch.Source: 2) DG2: Epistemic Scaffolding via… · p. 3

B. System Architecture

  • MAIA assigns coordination, intent diagnosis, independent response validation, and brief synthesis to distinct agents rather than one monolithic prompt.The validator returns APPROVED or REVISION-NEEDED and does not author user-facing content, making its gate structurally independent of the generation it audits.Source: 1) Coordinator and Specialized Roles · p. 3, B. System Architecture · p. 3

2) State Object and Four-State Epistemic Tagging:

  • The shared State Object tracks five visual and three logistical dimensions, each tagged UNADDRESSED, INFERRED, EXPLICIT, or DISCRETION.INFERRED marks an AI hypothesis requiring ratification; DISCRETION protects a dimension deferred to the artist from re-interrogation.Source: 2) State Object and Four-State… · p. 3

3) The Maieutic Turn: A Four-Step Pipeline:

Each user turn follows a validated serial pipeline, and the brief is finalized only after visual dimensions reach defined epistemic states. The study compared this scaffolding with a baseline single-agent chat in a counterbalanced within-subjects design.

3) The Maieutic Turn: A Four-Step Pipeline:

  • The coordinator logs intent, delegates diagnosis, waits for validator approval, then responds with bounded multiple-choice prompts.Each turn allows at most two substantive questions, with at least five annotated options per question.Source: 3) The Maieutic Turn: A Four-Step… · p. 3

4) Convergence and Brief Finalization:

  • A brief reaches Semantic Saturation when all five visual dimensions are EXPLICIT or DISCRETION, after which the validator re-audits it with zero INFERRED items allowed.A REVISION-NEEDED verdict loops back to re-diagnosis; final-brief auditing is stricter than turn-level validation.Source: 4) Convergence and Brief Finalization · pp. 3–4, Figure 2 · p. 4

A. Participants and Design

  • Sixteen laypeople completed a counterbalanced 2×2 within-subjects study comparing the full multi-agent system with default single-agent chat across two scenarios.Each participant experienced each system and scenario once, in opposite pairings.Source: A. Participants and Design · p. 4

B. Stimuli and Procedure

The study used two text-only commission scenarios and also conducted an exploratory blind review of a sample of briefs by professional concept artists. Participants rated the interactions on established and custom constructs, with paired tests and mixed-effects models used for analysis.

B. Stimuli and Procedure

  • Participants drafted initial and final briefs for two textual tasks lacking explicit visual directives, with a distractor task between system blocks.The tasks were a narrative-affective illustration and commercial-promotional club art; the AI was blind to task assignment.Source: B. Stimuli and Procedure · p. 4

C. Artist-Side Blind Review (Exploratory)

  • Three professional concept artists blindly rated 16 briefs from eight sampled tasks on visual completeness, executability, autonomy, and professionalism.All three artists rated the same briefs; condition comparisons were exploratory with four tasks per condition.Source: C. Artist-Side Blind Review… · p. 4

D. Measures and Analytical Approach

  • Participants rated task difficulty, ownership, creativity support, partnership, cognitive support, confidence change, creative agency, and semantic alignment on 7-point scales.Cognitive Support and Partnership were the focal outcomes; paired differences used Wilcoxon tests with FDR correction, corroborated by linear mixed-effects models.Source: D. Measures and Analytical Approach · pp. 4–5, Table I · p. 5

V. RESULTS

Cognitive Support was the focal outcome and improved significantly under the Maieutic condition across both reported statistical tests. Partnership and Creativity Support showed weaker or non-significant evidence.

1) System Main Effects:

  • Wilcoxon p_FDR = 0.015, r = 0.96; LMM βSys = 1.69, p_FDR < 0.001: Cognitive Support improved significantly under Maieutic scaffolding.The measure was designated a priori as a focal outcome and had a reliable scale in both conditions.Source: 1) System Main Effects · p. 5
  • Partnership had rank-biserial r = 0.75, but its primary Wilcoxon result did not survive correction (p_FDR = 0.064).It reached significance only under the LMM (p_FDR = 0.012), and the authors treat it as suggestive given low Baseline reliability.Source: 1) System Main Effects · p. 5
  • Creativity Support (CSI) trended upward (β = 0.71), but did not survive correction (LMM p_FDR = 0.098).Source: 1) System Main Effects · p. 5

2) Scenario Effects and Interaction:

In the artist-side review, rewriting improved reliable visual-completeness and executability dimensions across all eight sampled tasks, while condition-specific differences remain directional. Scenario effects and trait-related robustness checks qualify the broader results.

2) Scenario Effects and Interaction:

  • No Scenario main effects were isolated (all p > 0.10); the Cognitive Support interaction did not survive correction and its Narrative-Affective advantage is hypothesis-generating.The reported Maieutic advantages were Δ = +2.67 for Narrative-Affective and Δ = +0.71 for Commercial-Promotional; the interaction was underpowered because it was between subjects.Source: 2) Scenario Effects and Interaction · p. 5

3) Robustness: Covariates, Scale Reliability, and Construct Validity:

  • Measured baseline traits did not predict improvement (all p_FDR > 0.14), while scale reliability was lower for Partnership, CSI, and Creative Agency than for the focal measure.Cognitive Support items loaded cleanly on one factor (0.69–0.91); Creative Agency and Partnership co-loaded and did not separate.Source: 3) Robustness: Covariates, Scale… · p. 5

4) Artist-Side Evaluation (Exploratory):

  • All three reliable artist-rated dimensions improved across eight sampled tasks (r = 1.0; task-level Wilcoxon p = 0.008, FDR q = 0.010); MAIA gains were directionally larger than Baseline gains (d = 1.4–2.6).With four tasks per condition, only professionalism reached two-sided significance and none survived FDR; autonomy did not change and its final-brief scale was unreliable.Source: 4) Artist-Side Evaluation (Exploratory) · p. 5, Table II · p. 5

B. Qualitative Analysis

Thematic feedback links MAIA’s structured prompts to instruction and drift control, while surfacing ownership, visualization, and latency tensions.

1) Cognitive Scaffolding and Drift Control:

  • Bounded either/or prompts felt instructive, while Maieutic checks made briefs more complete and thorough than the baseline’s semantically drifting generation.Participants valued professional suggestions and described the system as teaching them how to write a commission.Source: 1) Cognitive Scaffolding and Drift… · p. 5

2) Why Ownership Did Not Shift:

  • Ownership did not shift because some participants treated the brief as a means to the final artifact, not something they needed to feel was theirs.Source: 2) Why Ownership Did Not Shift · p. 5

3) Tensions: De-visualization and Latency:

  • Users conditioned by text-to-image tools wanted visuals, while the heavy CoT pipeline caused long waits and some wanted to bypass structured prompts.These tensions arose despite the deliberate text-only design.Source: 3) Tensions: De-visualization and… · pp. 5–6

VI. DISCUSSION

The discussion interprets MAIA’s Cognitive Support gain as evidence for scaffoldable requirement discovery, while treating other outcomes and their explanations more cautiously.

A. Interpreting the Findings

  • Participants described MAIA surfacing missing visual dimensions, supporting the interpretation that co-articulation helps constitute intent rather than merely optimize its expression.Source: A. Interpreting the Findings · p. 6
  • M=3.67 of 7 was the absolute Cognitive Support score under Maieutic, despite a large, convergent gain across both analytical methods.The authors characterize this as relative improvement over a weak baseline, not evidence of high absolute quality.Source: A. Interpreting the Findings · p. 6
  • Semantic Alignment did not improve (Baseline 2.56, Maieutic 2.94), and confidence fell in both conditions from roughly 4.1 to about 3.0.The authors interpret these changes as possibly making unspecified needs visible, but cannot rule out lower-fidelity briefs.Source: A. Interpreting the Findings · p. 6
  • 4 of 14 Maieutic sessions (29%) triggered REVISION-NEEDED; the validator caught targeted drift modes, but between-condition drift reduction was not measured.Source: A. Interpreting the Findings · p. 6
  • In transactional commissions, Psychological Ownership did not shift; Creativity Support (CSI) trended upward (β=0.71, LMM pFDR=0.098; uncorrected Wilcoxon p=0.044), directionally consistent with Cognitive Support.Source: A. Interpreting the Findings · p. 6
  • With a small sample and borderline Baseline reliability for Agency (α=0.62), the Creative Agency and Semantic Alignment nulls do not establish absence of effect.Source: A. Interpreting the Findings · p. 6

B. Design Implications

The design implications propose delaying generated visuals until users have articulated their intent, while the limitations qualify the study’s scope and leave broader validation and component isolation for future work. The conclusion frames MAIA as scaffolding requirement discovery and reports supporting evidence from a participant study and exploratory artist review.

B. Design Implications

  • Phase-Gated Generation would defer visual outputs until textual Semantic Saturation, keeping images supplementary to reduce the risk of overwriting latent intent.The proposal responds to users’ desire for immediate visuals while preserving the text-only strategy’s epistemic safety.Source: B. Design Implications · p. 6

C. Limitations and Future Work

  • The artist-side review was exploratory, covering 16 briefs from four participants, with only four tasks per condition and low inter-rater agreement.The authors identify a larger, pre-registered blind review with more artists as the most important next step.Source: C. Limitations and Future Work · p. 6
  • Because the study compared the full configuration with a minimal baseline, gains may reflect structured process generally rather than maieutic inquiry specifically.The authors leave isolating the sub-agent architecture from the inquiry strategy to follow-up work with intermediate controls.Source: C. Limitations and Future Work · p. 6
  • The small sample and single-model evaluation constrain inference, while the shared generator–validator weights mean full validator independence would require a second model.The CoT pipeline also added latency that future work must optimize.Source: C. Limitations and Future Work · pp. 6–7

VII. CONCLUSION

  • The within-subjects study (N=16) found a large Cognitive Support gain over a minimal baseline, alongside exploratory artist-side evidence of more complete and executable briefs.The conclusion presents these findings as evidence that requirement discovery is a distinct stage that can be scaffolded.Source: §VII · p. 7
Loading 2609.12097v1…