Authors: Shaheer U. Saeed, Zion Tse, Gregory Slabaugh, Daniel C. Alexander, Yipeng Hu
Organizations: Digital Environment Research Institute, Queen Mary University of London · Centre for Bioengineering, Queen Mary University of London · UCL Hawkes Institute, University College London
Modern machines can solve equations, follow logic and reason through language, yet they still struggle with problems that require visual non-verbal reasoning. This is evident even in advanced systems such as DeepSeek, ChatGPT and Gemini, which remain limited on visual tasks such as completing abstract patterns or delineating ambiguous regions-of-interest. Unlike explicit intermediate steps in verbal reasoning which can be verified through logic or automated tests, visual reasoning involves implicit intermediate steps with no clear methods for verification. Non-verbal reasoning is therefore challenging to model and, as a result, remains a fundamental limitation of current machine intelligence. Here we introduce GenVerOpt, a general framework that unifies generation, verification and optimisation into a reasoning pipeline for machines. By generating candidate solutions, verifying their correctness, and iteratively refining them, GenVerOpt enables per-sample optimisation rather than relying solely on dataset-level trends captured during model training. This process mirrors human problem-solving where multiple possibilities are imagined, evaluated, and refined until a correct solution emerges. Across diverse domains, GenVerOpt demonstrates capabilities beyond existing systems: it surpasses humans on abstract IQ puzzles, matches clinicians in prostate cancer segmentation, exceeds experts in liver vessel reconstruction and outperforms state-of-the-art approaches in tumour boundary alignment. These results show that machine reasoning can extend beyond language, establishing visual non-verbal reasoning as a foundation for more flexible and generalisable intelligence.
Figures & tables
Figure 1: IQ test puzzles solved by recent state-of-the-art reasoning AI models.
Figure 2: Overview of GenVerOpt (top: conceptual summary; bottom: details of one implementation of GenVerOpt for the dummy task of human segmentation on natural images).
Figure 3: IQ test puzzles solved by GenVerOpt and a comparison with humans and commonly used AI models. Open-source models LLaVA and Deepseek were fine-tuned using low-rank adaptation on the dataset before inference and closed-source models were primed using four-shot prompting with examples randomly selected from the datasets and performance averaged across 8 runs. Task-specific models were trained from scratch on the same data as GenVerOpt.
Figure 4: Examples of medical imaging problems solved by GenVerOpt and compared with human experts and recent state-of-the-art AI models.
Figure 5: Performance comparison of GenVerOpt against expert humans and recent state-of-the-art AI systems for medical imaging tasks.
Dataset
Samples
Role
Logical Operation
395
Training
Spatial Relationship
102
Training
Geometric Relationship
355
Training
Temporal Movement
32 + 379
Fine-tuning + Evaluation
Table 1: IQ test pattern puzzels dataset
Dataset
Modality
Structure
Cohort Country
Samples
Role
Antonelli et al. (2021)
CT
Spleen
United States
41
Training
Synapse (2013)
CT
Gallbladder
United States
30
Training
Synapse (2013)
CT
Adrenal Gland
United States
30
Training
Synapse (2013)
CT
Major Vessels
United States
30
Training
Synapse (2013)
CT
Stomach
United States
30
Training
Rister et al. (2020)
CT
Kidneys
> 3 countries
119
Training
Table 2: A summary of the data sets used for training and evaluation for the prostate cancer and liver vessel segmentation tasks.
Generator + Verifier
Optimiser
Joint Training
Reptile
Cross-task training
Self-play
RL
Performance
✓
✓
✓
✓
✓
62.4 ± 4.9
✓
✗
✗
✓
✓
51.4 ± 5.5
✗
✓
✓
✓
✓
54.8 ± 4.6
✓
✗
✓
✓
✓
60.6 ± 5.4
✗
✗
✗
✓
✓
49.5 ± 4.7
Table 3: Ablation study - removing components from the framework to test efficacy of different approaches. The standard deviation is across different runs rather than patient cases.
Samples
Performance
Reasoning Time / h
16
56.1 ± 4.7
40
32
59.6 ± 5.4
35
48
60.1 ± 4.9
31
64
62.4 ± 4.9
26
80
61.5 ± 5.1
24
96
61.7 ± 5.5
25
Table 4: Measuring performance of GenVerOpt, given different amounts of labelled task-specific examples.
Reasoning Time / h
Performance
0
38.2 ± 4.9
4
42.4 ± 4.9
8
47.6 ± 5.2
12
52.9 ± 4.3
16
55.7 ± 5.3
20
58.3 ± 5.7
Table 5: Performance vriation with reasoning time.