Polyphonic (Contour-Shared) Portraiture (PCP) Benchmark is a multistable-structure interpretation benchmark for AI vision–language systems. It evaluates whether a model can identify and reason about multiple coherent readings intentionally encoded into a single unified portrait, while remaining faithful to the work’s shared-contour constraints and compositional “grammar.”

 

Unlike generic image-captioning or open-ended art commentary, PCP is designed to test:

1. Perceptual multistability detection (discovering more than one stable parse).

2. Constraint-based justification (linking each interpretation to concrete visual evidence: shared contours, figure/ground boundaries, axis scaffolds, occlusion logic).

3. Parse co-existence (maintaining multiple readings simultaneously, not replacing one with another).

4. Structural economy (recognizing that the same stroke/edge serves more than one role).

5. Cross-modal anchoring (optional): aligning the visual structure with an accompanying conceptual statement (e.g., dependent origination / field of conditions) without inventing unsupported details.

 

Polyphonic (Contour-Shared) Portraiture Benchmark (PALA)
PALA curates R. Latchman’s Polyphonic (Contour-Shared) Portraiture as a vision–language benchmark for multistable visual reasoning. The benchmark evaluates whether AI systems can extractmultiple coherent readings from a single unified portrait while remaining grounded in shared-contour evidence (edges, boundaries, axes, occlusions) and preserving co-existing interpretations rather than overwriting one with another. PCP therefore tests a capability that standard captioning and aesthetic commentary do not: constraint-based interpretation of intentionally polyphonic images, with explicit control of hallucination and evidence discipline.

 

PCP is a curated benchmark for evaluating evidence-grounded, multi-parse interpretation in vision–language models under contour-sharing composition constraints.

​

We introduce Polyphonic (Contour-Shared) Portraiture (PCP) as a curated benchmark for evaluating multistable visual reasoning in vision–language models (VLMs). PCP is grounded in a distinct class of artworks authored within R. Latchman’s Polyphonic (Contour-Shared) Portraiture practice, in which a single unified composition is intentionally constructed to support multiple coherent perceptual parses (e.g., portrait, sub-figures, structural fields), while preserving a disciplined “shared-contour” logic. PCP evaluates whether an AI system can (i) discover more than one stable interpretation, (ii) maintain co-existing parses without overwriting, and (iii) justify each parse using explicit visual evidence (contours, boundaries, axis scaffolds, occlusions, and shared structural edges). The benchmark targets a capability gap not addressed by standard captioning, aesthetic description, or generic visual question answering: constraint-based interpretation of intentionally polyphonic images with controlled hallucination.

 

1. Motivation

Current evaluations for VLMs primarily reward single-shot captioning, object recognition, or loosely grounded “style commentary.” These tasks do not reliably test whether a model can handle images that are designed to be read in multiple coherent ways, where perception itself is a structured variable rather than a fixed label. In human cognition, such stimuli probe multistable perception, figure–ground organization, and evidence-disciplined inference under ambiguity. PCP operationalizes these cognitive demands into a reproducible evaluation protocol suitable for research, benchmarking, and longitudinal model assessment.

PCP is motivated by the observation that many model failures in real-world settings—misinterpretation, overconfident narration, and hallucination—arise precisely when images admit competing plausible explanations. PCP therefore treats ambiguity not as noise, but as an intentional test surface.

 

2. Benchmark Construct: What PCP Tests

PCP evaluates polyphonic visual reasoning under contour-sharing constraints, decomposed into five measurable competencies:

1. Perceptual multistability detection
   Ability to identify two or more coherent, non-trivial interpretations (parses) of the same image.

2. Constraint-based justification (evidence discipline)
   Ability to ground each interpretation in explicit, checkable visual features (e.g., shared contours, figure/ground boundaries, axis scaffolds, occlusion relations, repeated structural edges). Outputs that rely on vague affective language or ungrounded invention are penalized.

3. Parse co-existence and non-collapse
   Ability to preserve multiple valid readings simultaneously and articulate their compatibility (or controlled tension) without forcing a single “final” answer.

4. Structural economy recognition
   Ability to identify that the same stroke/edge/boundary serves multiple roles across readings—an explicit hallmark of contour-shared construction.

5. Cross-modal anchoring (optional module)
   When accompanied by a short doctrinal statement (e.g., dependent origination / field-of-conditions framing), ability to align the conceptual interpretation with the visual evidence without adding unsupported narrative content.

3. Inputs and Task Definition

Inputs

• A curated set of PCP artworks (image + minimal metadata).

• Optional short accompanying doctrinal text (used only in the cross-modal module).

Core Task (PCP-Core)
Given an image, the model must produce:

• At least two coherent parses, each described succinctly.

• For each parse: evidence anchors referencing concrete visual structure (edges, contours, axes, boundaries, occlusions, repeated motifs).

• A co-existence statement describing how the parses are jointly supported by shared structure.

Optional Task (PCP-X: Cross-Modal Anchoring)
Given image + short statement, the model must:

• Map the statement to one or more parses and justify the mapping via image-grounded evidence.

• Avoid interpretive drift and narrative invention.

​

4. Failure Modes 

PCP is specifically constructed to elicit and measure known VLM failure patterns:

• Single-caption collapse: model gives one caption and misses multistability.

• Aesthetic drift: model produces generic “art talk” without structural grounding.

• Narrative hallucination: the model invents a story or objects to force coherence.

• Parse overwriting: the model proposes multiple readings but treats them as mutually exclusive, without a shared-structure analysis.

• Evidence underspecification: model claims “a face” or “a figure” without pointing to the contour logic that makes the claim checkable.

Because PCP’s target artworks are intentionally constructed around shared edges and multi-role boundaries, these failures become measurable rather than anecdotal.

​

PCP occupies a rigorous niche between perception science and AI evaluation. It provides a controlled, repeatable test surface for abilities that matter in high-stakes settings: managing ambiguity, grounding claims, and maintaining structured uncertainty. In doing so, PCP expands evaluation beyond “what objects are present” toward “how a system reasons when perception itself has multiple stable solutions.”

​

The PALA Institute for Spiral Cognition curates Polyphonic (Contour-Shared) Portraiture (PCP) as a benchmark for evaluating multistable, evidence-grounded visual reasoning in modern AI systems. PCP formalises a class of artworks engineered for co-existing coherent readings within a single unified portrait field. By requiring explicit visual anchors (contours, boundaries, axes, occlusion relations) and penalising unsupported invention, PCP measures a capability gap not addressed by conventional captioning or aesthetic commentary: constraint-disciplined interpretation under intentional ambiguity. PCP thus functions both as an evaluation suite and as a research instrument for studying grounded inference, hallucination resistance, and the discovery of perceptual structures in vision–language models.

​