Research map / Vision-language and multimodal / Human and embodied modeling
Neural Prediction Errors enable Analogical Visual Reasoning in Human Standard Intelligence Tests
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add prediction-error computation (PRB) to ResNet-4B encoder (ablation) | ResNet-4B · RAVEN-FAIR and I-RAVEN | accuracy (%): about +20 on RAVEN-FAIR, about +24 on I-RAVEN | Helped |
| remove prediction-error computation from PRB (Rmv.PC) (ablation) | PredRNet · I-RAVEN | accuracy (%): 96.5 to 92.1 | Hurt |
| replace subtraction in prediction error with ReLU(x)+ReLU(y) (ablation) | PredRNet · I-RAVEN | accuracy (%): 96.5 to 95.0 | Hurt |
| forward prediction errors without refining context features and errors (ablation) | PredRNet · I-RAVEN | accuracy (%): 96.5 to 93.1 | Hurt |
| replace PRBs with MLPs (ablation) | PredRNet · I-RAVEN | accuracy (%): 96.5 to 66.1 | Hurt |
| stack 2-3 PRBs vs 1 PRB on ResNet-4B (ablation) | ResNet-4B + PRBs · RAVEN, RAVEN-FAIR, I-RAVEN | accuracy (%): RAVEN 94.6 (K=1) to 95.8 (K=3); I-RAVEN 95.0 to 96.5 | Helped |
| increase stacked PRBs to 4 (ablation) | ResNet-4B + PRBs · RAVEN, I-RAVEN | accuracy (%): K=3 to K=4: RAVEN 95.8 to 96.0 (up), I-RAVEN 96.5 to 94.8 (down) | Hurt |
| add additional ResBlocks instead of PRBs to ResNet-4B (ablation) | ResNet-4B · RAVEN, RAVEN-FAIR, I-RAVEN | accuracy (%): ResNet-4B 68.1 avg to 56.8-65.8 avg | Hurt |
| use late fusion instead of early fusion in ResNet-18/50 encoders (ablation) | ResNet-18 and ResNet-50 image encoders · RAVEN-FAIR and I-RAVEN | accuracy (%): ResNet-18E 15.8 to ResNet-18L 77.6 on RAVEN-FAIR | Helped |
| PredRNet with PRB-based prediction-error reasoning (full model) vs prior SOTA | PredRNet · four RAVEN-style and PGM-N benchmarks | average accuracy (%): 96.7 vs 94.8 (STSN) | Helped |
Closest papers by meaning
- Learning Visual Abstract Reasoning through Dual-Stream Networks (AAAI 2024 · similarity 0.88)
- Concept-RuleNet: Grounded Multi-Agent Neurosymbolic Reasoning in Vision Language Models (AAAI 2026 · similarity 0.86)
- Towards Generative Abstract Reasoning: Completing Raven’s Progressive Matrix via Rule Abstraction and Selection (ICLR 2024 · similarity 0.86)
- SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs (ICML 2026 · similarity 0.86)
- Diversity Matters: Revisiting Test-Time Compute in Vision-Language Models (ICML 2026 · similarity 0.86)
- Monet: Reasoning in Latent Visual Space Beyond Image and Language (CVPR 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.