Research map / Vision-language and multimodal / Multimodal LLMs
A Closer Look at Generalisation in RAVEN
Steven Spratley, Krista Ehinger, Tim Miller
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 frame-relational 1D convolution encoder to ResNet (Rel-Base) | ResNet · RAVEN-10k (all configurations) | accuracy (%): 34.51 -> 91.66 | Improved |
| add AIR object-centric scene decomposition to Rel-Base (Rel-AIR) | Rel-Base · RAVEN-10k (all configurations) | accuracy (%): 91.66 -> 94.10 | Improved |
| add AIR scene decomposition to Rel-Base, 10% training data | Rel-Base · RAVEN (10% of training set) | accuracy (%): 24.40 -> 51.39 | Improved |
| add AIR scene decomposition to Rel-Base, 25% training data | Rel-Base · RAVEN (25% of training set) | accuracy (%): 52.24 -> 81.07 | Improved |
| add AIR scene decomposition, train Left-Right test Up-Down | Rel-Base · RAVEN (train Left-Right, test Up-Down) | accuracy (%): 32.71 -> 66.77 | Improved |
| add AIR scene decomposition, train Up-Down test Left-Right | Rel-Base · RAVEN (train Up-Down, test Left-Right) | accuracy (%): 22.61 -> 60.81 | Improved |
| add AIR scene decomposition, train 2x2Grid test 3x3Grid | Rel-Base · RAVEN (train 2x2Grid, test 3x3Grid) | accuracy (%): 41.55 -> 67.01 | Improved |
| add AIR scene decomposition, train 3x3Grid test 2x2Grid | Rel-Base · RAVEN (train 3x3Grid, test 2x2Grid) | accuracy (%): 34.03 -> 61.90 | Improved |
| add position and scale latents via bilinear layer instead of linear layer (ablation) | Rel-AIR · RAVEN (Left-Right and Up-Down, incl. generalisation between them) | accuracy (%): falls to around 43% ± 3 without it | Improved |
| use Rel-Base instead of WReN on PGM extrapolation | WReN · PGM extrapolation | accuracy (%): 17.20 -> 22.05 | Improved |
| answer-frames-only training exposes bias: accuracy far above chance without context (ablation) | answer-only solver · RAVEN (all configurations) | accuracy (%): 12.5 -> above 90 | Improved |
| shuffle object slots from AIR before the relational encoder (ablation) | Rel-AIR · RAVEN | accuracy (%) | Improved |
Closest papers by meaning
- Revisiting the Necessity of Lengthy Chain-of-Thought in Vision-centric Reasoning Generalization (CVPR 2026 · similarity 0.87)
- When to Think and When to Look: Uncertainty-Guided Lookback (CVPR 2026 · similarity 0.86)
- Long Grounded Thoughts: Synthesizing Visual Problems and Reasoning Chains at Scale (ICML 2026 · similarity 0.86)
- Generalizing from SIMPLE to HARD Visual Reasoning: Can We Mitigate Modality Imbalance in VLMs? (ICML 2025 · similarity 0.86)
- MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · similarity 0.85)
- Synthesize Diagnose and Optimize: Towards Fine-Grained Vision-Language Understanding (CVPR 2024 · similarity 0.85)
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.
