Traditional visual reasoning systems often rely on a "hidden" form of human intelligence: the manual translation of raw pixels into symbolic logic. Before a model even begins to reason, researchers typically pre-process images into structured predicates or functional programs. Davidson et al. identify this as a "symbolic translation bias," arguing that true fluid intelligence requires models to induce logical rules directly from raw, unstructured input without human-mediated symbolic pre-processing.
A growing body of research is now focused on end-to-end visual logic induction, where models learn to perceive and reason simultaneously, discovering their own internal symbolic representations from raw data.
The Induced Symbolic Space
The core challenge of reasoning from pixels is bridging the gap between continuous visual features and discrete logical operations. Rather than using human-defined labels for concepts like "red" or "behind," recent frameworks learn to "induce" these concepts. Wang et al. demonstrate this with the OCCAM framework, which maps raw visual features into an automatically discovered symbolic space using only image-question-answer pairs. This allows the model to navigate compositional logic without ever seeing a human-annotated program.
Similarly, Sun et al. show that models can learn implicit symbolic representations by reconstructing video frames from compressed "slot" tokens. By forcing the model to bottleneck visual information into these discrete slots, it naturally learns object-centric reasoning without requiring explicit object detectors or bounding box supervision.
Grounding Logic in Perception
For neurosymbolic systems, the difficulty lies in "grounding"—mapping a raw input (like a Sudoku grid) to a logical symbol (the number 4) without being told which is which. Topan et al. address this by combining unsupervised clustering with a "Symbol Grounding Loss," enabling architectures like SATNet to solve complex puzzles from raw images. This suggests that explicit intermediate labels are not a prerequisite for high-level logical performance.
This transition to unsupervised reasoning extends to abstract pattern matching. In the context of Raven’s Progressive Matrices, Zhuo et al. utilize "pseudo-targets" and negative sampling to learn abstract rules without ground truth labels. Their work indicates that models can achieve competitive generalization by treating reasoning as a self-supervised discovery process rather than a supervised mapping task.
Scaling and Self-Correction
As models scale, the ability to reason from pixels appears to emerge as a byproduct of large-scale sequential modeling. Bai et al. demonstrate that Large Vision Models (LVMs), trained purely on "visual sentences" (sequences of images and videos), can solve geometric reasoning tasks in-context. This suggests that complex logic can be developed through purely visual observation, independent of linguistic supervision.
To refine these reasoning capabilities without human intervention, researchers are turning to iterative feedback loops. Chen et al. propose a "reasoning-rendering-visual feedback" loop (RRVF) that allows models to self-correct by comparing their generated logic against visual outputs. Furthermore, the Unsupervised Visual Chain-of-Thought Reasoning via Preference Optimization framework leverages a model's own ability to generate and evaluate intermediate steps, eliminating the need for expensive, human-annotated reasoning chains.
By removing the crutch of human-defined symbols, this research direction moves AI closer to autonomous "fluid" intelligence—the ability to look at a novel visual problem and induce the underlying rules of the game from scratch.
Key Foundational Work
Towards Fluid Machine Intelligence: Can We Make a Gifted AI? (2019) — contributed 4 concept elements to this direction.
Global Fusion Attention for Vision and Language Understanding (Student Abstract) (2021) — contributed 2 concept elements to this direction.
Unsupervised Abstract Reasoning for Raven's Problem Matrices (2021) — contributed 2 concept elements to this direction.