TopoPerception: Diagnostic Benchmark for LVLMs

Updated 5 July 2026

TopoPerception is a diagnostic benchmark for LVLMs that measures global visual perception by isolating topological invariants from semantic signals.
It employs synthetic images with fixed text prompts and controlled partition granularity to neutralize shortcut strategies and local visual cues.
Empirical results indicate that current LVLMs struggle to preserve global image structure, suggesting a need for resolution-aware and topology-focused model improvements.

TopoPerception denotes, in its most specific contemporary usage, a diagnostic benchmark for evaluating the global visual perception of Large Vision-LLMs (LVLMs) through topology rather than semantics. Introduced in “TopoPerception: A Shortcut-Free Evaluation of Global Visual Perception in Large Vision-LLMs” (Zhou et al., 14 Nov 2025), it is designed to remove both semantic-level shortcuts and local visual shortcuts by fixing the text prompt across all samples and using synthetic images whose relevant attributes are topological invariants. In broader adjacent usage, the same term also appears as a label for topologically grounded accounts of perception in neuroscience, representation learning, autonomous driving, and medical imaging, where connectivity, holes, neighborhood preservation, and structural consistency are treated as primary perceptual primitives rather than secondary descriptors (Takahashi, 2018).

1. Definition and conceptual scope

TopoPerception, in the LVLM benchmark sense, targets global visual perception: the ability of a model to preserve and use properties that depend on the entire image structure and are invariant to local changes. The benchmark replaces semantically rich tasks such as VQA, captioning, and reasoning with topology-based classification, thereby isolating the visual signal from language priors and memorized world knowledge. Its central claim is methodological rather than taxonomic: topology provides a shortcut-free probe because connectivity, number of holes, and interior/exterior relations are insensitive to texture, color, and local semantics (Zhou et al., 14 Nov 2025).

The benchmark is motivated by a specific architectural diagnosis of current LVLMs. These systems typically pipeline a visual encoder into a LLM through a projection interface. Under fixed resolution, patching, resizing, and language-aligned objectives such as CLIP-like contrastive training, the visual pathway may discard or distort global structure before it reaches the LLM. Conventional multimodal evaluations then confound perception with reasoning and semantics, so strong benchmark performance can coexist with weak preservation of global visual structure.

A common misconception is that TopoPerception is a general multimodal reasoning benchmark. It is instead a diagnostic benchmark: the text modality is intentionally fixed, and the visual modality is intentionally stripped of natural-image semantics. Another misconception is that topology here requires persistent homology or filtration-based machinery. The benchmark does not require persistent homology, Vietoris–Rips, or Čech complexes; it constructs images whose topological class is unambiguous by design and labels them accordingly.

2. Topological foundations

TopoPerception is built around standard topological invariants for planar binary images. The benchmark uses three interrelated notions: connected components, holes or loops, and interior/exterior relations. In algebraic-topological terms, connected components are the zeroth Betti number $b_0$ , holes are the first Betti number $b_1$ , and the Euler characteristic is

$b_k = \operatorname{rank}(H_k),$

$\chi = \sum_{k=0}^{n} (-1)^k b_k.$

For planar binary images, this reduces effectively to

$\chi = b_0 - b_1.$

Because $\chi$ , $b_0$ , and $b_1$ are invariants under continuous deformations, they are intrinsically insensitive to local textures and stylistic changes. This is the formal basis for treating topology as a probe of global perception rather than local cue detection (Zhou et al., 14 Nov 2025).

The benchmark’s use of topology is deliberately classical and planar. It does not depend on persistent homology, and pixel connectivity conventions are not treated as the crux. Instead, the images are generated so that their global classes remain stable under standard interpretations, while difficulty is controlled by image partition granularity rather than by appearance. This makes the benchmark less a study of topological data analysis in the persistence-diagram sense than a study of whether visual encoders preserve global structure at all.

This focus distinguishes TopoPerception from shape-based and maze-based tests. Shape-based tests can often be solved from local cues, such as curvature at a point. Maze-based tests can reduce to finding any path when connected, which creates local-search shortcuts and uneven difficulty control. TopoPerception instead centers invariants that require global processing and can be scaled through granularity.

3. Benchmark design and construction

TopoPerception neutralizes shortcut strategies by jointly controlling the text and image modalities. Every image is paired with a single fixed multiple-choice question with five options, A–E. Only B, C, and D correspond to valid topological categories; A and E are distractors. Because the text is fixed and does not leak image-specific information, the model must rely on vision. Because the images are synthetic and express only global topology, there are no natural-image semantics or local textures to exploit (Zhou et al., 14 Nov 2025).

The image-generation pipeline is combinatorial. A uniform spanning tree is constructed on a connected $n \times n$ graph. Each graph node maps to a $3 \times 3$ pixel block, and a connected graph with $b_1$ 0 nodes produces an image of resolution $b_1$ 1. The resulting sample space scales exponentially with $b_1$ 2:

$b_1$ 3

which is presented as making memorization infeasible and permitting arbitrarily high difficulty extensions.

Difficulty is defined through partition granularity. For an image partitioned to resolution $b_1$ 4, the difficulty index is $b_1$ 5. Lower granularity corresponds to coarser partitions and easier tasks; higher granularity corresponds to finer partitions and harder tasks. The benchmark links this directly to effective perceptual resolution: if a model’s perceptual granularity is below the image’s partition granularity, downsampling can distort topology; if it is equal or larger, topology is preserved.

Dataset construction is balanced. Each difficulty level contains 100 images per topological category across 3 categories, for 300 samples total. Level 0 has resolution $b_1$ 6, described as comparable to MNIST’s $b_1$ 7, and serves as a clean, low-bar reference for global perception. Evaluation uses accuracy, precision, recall, and weighted F1. Two chance levels are distinguished: 20% for random guessing over all five options, and 33.3% for perfect bias toward one of the three valid options while remaining random within that set.

4. Empirical results on large vision-LLMs

TopoPerception evaluates GPT-4o, o4-mini, o3, Claude-sonnet-4-0, Claude-opus-4-0, Gemini-2.5-flash, and Gemini-2.5-pro using standard API calls and default inference settings. At Level 0, no model exceeds the 33.3% “perfect bias” threshold, and most hover near the 20% random baseline. The paper states that even at the coarsest granularity, performance is statistically indistinguishable from guessing (Zhou et al., 14 Nov 2025).

Model	Level 0 accuracy
GPT-4o	22.00%
o4-mini	19.67%
o3	12.00%
Claude-sonnet-4-0	30.00%
Claude-opus-4-0	24.33%
Gemini-2.5-flash	33.33%
Gemini-2.5-pro	30.67%

The paper reports weighted precision, recall, and F1 per model, but the salient result is not metric ranking so much as the collective failure regime. Confusion matrices show strong, stable option biases, often concentrated on one valid option, and prediction distributions are nearly identical across ground-truth categories. This indicates that model outputs are not conditioned on the actual image topology; selections reflect intrinsic tendencies rather than visual recognition.

A further empirical pattern is family-internal degradation with stronger reasoning-oriented models. Within OpenAI, the ordering is GPT-4o > o4-mini > o3; within Anthropic, Claude-sonnet-4-0 > Claude-opus-4-0; within Google, Gemini-2.5-flash > Gemini-2.5-pro. The paper interprets this as evidence that scaling language reasoning does not improve global visual perception and may actively interfere with it, as stronger language priors can override fragile visual signals.

The paper does not report human baselines. It does, however, state that the simplicity of Level 0 for humans implies a large gap to human-level global perception. That implication is suggestive rather than directly measured.

5. Explanatory hypotheses, limitations, and proposed remedies

The central explanatory hypothesis is a visual encoder bottleneck. Resizing, patching, and token reduction may corrupt spatial relationships and discard subtle global cues before cross-modal fusion occurs. A second hypothesis concerns feature alignment: the projection interface and LLM embedding space may not faithfully embed global structure. A third is the lack of topological inductive bias: current architectures are not explicitly designed to encode homology-like invariants or preserve Euler characteristics under downsampling (Zhou et al., 14 Nov 2025).

The granularity analysis strengthens the first hypothesis. A nearest-neighbor downsampling thought experiment is used to argue that when effective perceptual resolution falls below image granularity, topological invariants distort; when it is equal or higher, topology is preserved. This localizes failure to information retention in the visual pathway rather than to language reasoning alone, although the family-level trends suggest that language-heavy reasoning can compound the problem.

Several remedies are proposed. Architectural directions include resolution-aware encoders, iterative re-examination of the image during reasoning, and explicit vision-to-decision heads for topology. Training directions include topology-focused curricula, loss functions that penalize topology changes under augmentations, and supervision on $b_1$ 8, $b_1$ 9, and $b_k = \operatorname{rank}(H_k),$ 0 for binary masks or regions. The paper also mentions multi-view or 3D cues and differentiable topology estimators or invariant-preserving constraints such as stability of $b_k = \operatorname{rank}(H_k),$ 1 under transformations.

The benchmark has clear scope limits. It is not a measure of general multimodal competence, not a substitute for semantic reasoning evaluations, and not an empirical study of human perception. A plausible implication is that TopoPerception should be read as a bottleneck diagnostic for current LVLM pipelines rather than as a global ranking of multimodal intelligence.

Outside the LVLM benchmark, the term also appears in broader topological accounts of perception. In neuroscience, “Topographic maps in the brain are fundamental to processing of causality” (Takahashi, 2018) uses TopoPerception to describe a framework in which neural topographic maps instantiate partially ordered sets defined by molecular coordinates, preserving neighborhood relations and supporting causal representation. There, the emphasis is not on benchmark design but on adjacency-preserving, monotone mappings across neural manifolds.

In model architecture, “Topoformer: brain-like topographic organization in Transformer LLMs through spatial querying and reweighting” (Binhuraib et al., 21 Oct 2025) advances a topographic inductive bias inside Transformers through spatial querying and spatial reweighting, yielding interpretable topographic organization in linguistic representations. In computational neuroscience, “Topological decoding of grid cell activity via path lifting to covering spaces” (Yao et al., 17 Oct 2025) uses persistent cohomology, toroidal coordinates, and path lifting to reconstruct trajectories from co-modular grid-cell population activity.

In autonomous driving, the term is used in topology-aware perception systems for lane and road-network understanding. “TopoSD: Topology-Enhanced Lane Segment Perception with SDMap Prior” (Yang et al., 2024), “TopoPoint: Enhance Topology Reasoning via Endpoint Detection in Autonomous Driving” (Fu et al., 23 May 2025), and “TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes” (Fu et al., 2024) all treat perception and topology reasoning as coupled problems, emphasizing lane geometry, endpoint consistency, adjacency, and graph construction.

In medical imaging and VLM evaluation, “TopoAgent: An Agentic Framework for Automated Topology Learning in Medical Imaging” (Meng et al., 29 Jun 2026) frames topologically grounded perception as automated descriptor selection over persistent homology pipelines, while “Topo-R1: Detecting Topological Anomalies via Vision-LLMs” (Xu et al., 13 Mar 2026) treats topological anomaly detection in tubular structures as a structured visual reasoning problem. Related but distinct lines include topological activation-space analysis in “TopoAct: Visually Exploring the Shape of Activations in Deep Learning” (Rathore et al., 2019), geometric specification by “Topological Parallax: A Geometric Specification for Deep Perception Models” (Smith et al., 2023), and human-perception studies of topology-based visualizations in “A Comparative Study of the Perceptual Sensitivity of Topological Visualizations to Feature Variations” (Athawale et al., 2023).

Taken together, these usages suggest two persistent meanings of TopoPerception. The narrow meaning is the shortcut-free LVLM benchmark grounded in $b_k = \operatorname{rank}(H_k),$ 2, $b_k = \operatorname{rank}(H_k),$ 3, and $b_k = \operatorname{rank}(H_k),$ 4. The broader meaning is a research program in which perception is evaluated, structured, or regularized through topological and topographic invariants. The benchmark of (Zhou et al., 14 Nov 2025) is distinctive within that landscape because it turns topology into a controlled diagnostic instrument for exposing a profound inability of current LVLMs to perceive global visual features.