Advancing Generalization Across a Variety of Abstract Visual Reasoning Tasks

TL;DR

This work addresses the challenge of out-of-distribution generalization in Abstract Visual Reasoning (AVR) by proposing PoNG, a two-stage neural architecture that combines a panel encoder with a Reasoner and a Novel Pathways block utilizing normalized group convolution. PoNG employs three prediction heads to jointly learn answer selection and rule-structure representations, enabling strong performance across RPM benchmarks (PGM, I-RAVEN, I-RAVEN-Mesh, A-I-RAVEN) and visual analogy datasets (VAP, VASR) in both i.i.d. and o.o.d settings. Comprehensive ablations reveal that each component—panel encoding, the P1–P4 pathways, and the Temporal Convolutional Network (TCN)—contributes critically to generalization, particularly for held-out attribute regimes. The results show PoNG achieving state-of-the-art or near-top performance across diverse AVR tasks, illustrating its versatility and potential for broad applicability in reasoning over structured visual data and beyond.

Abstract

The abstract visual reasoning (AVR) domain presents a diverse suite of analogy-based tasks devoted to studying model generalization. Recent years have brought dynamic progress in the field, particularly in i.i.d. scenarios, in which models are trained and evaluated on the same data distributions. Nevertheless, o.o.d. setups that assess model generalization to new test distributions remain challenging even for the most recent models. To advance generalization in AVR tasks, we present the Pathways of Normalized Group Convolution model (PoNG), a novel neural architecture that features group convolution, normalization, and a parallel design. We consider a wide set of AVR benchmarks, including Raven's Progressive Matrices and visual analogy problems with both synthetic and real-world images. The experiments demonstrate strong generalization capabilities of the proposed model, which in several settings outperforms the existing literature methods.

Figures (7)

• Figure 1: Raven's Progressive Matrices (RPMs).
• Figure 2: Visual analogies.
• Figure 3: A-I-RAVEN malkinski2025airaven.
• Figure 4: I-RAVEN-Mesh malkinski2025airaven.
• Figure 5: PoNG. (a) The panel encoder embeds each input image $x_i$ independently, producing $h_i$. Context panel embeddings $\{h_i\}_{i=1}^8$ together with the embedding of $k$'th answer $h_k$ are stacked and processed with the reasoner, leading to $z_k$. (b) The pathways block, a key component of PoNG, comprises four parallel pathways P1 -- P4. (c) P3 and (d) P4 employ novel normalized group convolution operators. PosEmb denotes position embedding, G-C the group convolution module used in P3, and GP-C the group-pair convolution module used in P4. The red dashed line marks the point after which G-C and GP-C perform analogous computation.