Research map / Large language models / Retrieval and knowledge
Explainable Neural Computation via Stack Neural Module Networks
Ronghang Hu, Jacob Andreas, Trevor Darrell, Kate Saenko
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 expert layout supervision (ablation) | Stack-NMN · CLEVR val (VQA) | accuracy (%): 93.0 -> 96.6 | Improved |
| add expert layout supervision (ablation) | Stack-NMN · CLEVR-Ref val (REF) | accuracy (%): 93.4 -> 96.0 | Improved |
| add joint VQA+REF training (no expert layout) (ablation) | Stack-NMN · CLEVR val (VQA) | accuracy (%): 93.0 -> 93.9 | Improved |
| add joint VQA+REF training (no expert layout) (ablation) | Stack-NMN · CLEVR-Ref val (REF) | accuracy (%): 93.4 -> 95.4 | Improved |
| add joint VQA+REF training (with expert layout) (ablation) | Stack-NMN · CLEVR val (VQA) | accuracy (%): 96.6 -> 96.5 | No clear effect |
| use soft differentiable layout instead of discrete layout, no expert layout | Stack-NMN vs N2NMN · CLEVR | accuracy (%): 69.0 -> 93.0 | Improved |
| discretize module weights to one-hot at test time (ablation) | Stack-NMN (no expert layout) · CLEVR val (VQA) | accuracy (%): 93.0 -> 90.0 | Hurt |
| discretize module weights to one-hot at test time (ablation) | Stack-NMN (expert layout) · CLEVR val (VQA) | accuracy (%): 96.5 -> 94.8 | Hurt |
| add expert layout supervision (ablation) | Stack-NMN · human forward prediction (failure detection) | human accuracy: 0.625 -> 0.545 | Hurt |
| add expert layout supervision (ablation) | Stack-NMN · human subjective understanding | mean Likert score: 3.33 -> 3.47 | Improved |
| use modular Stack-NMN instead of non-modular MAC | interpretability · human subjective understanding | mean Likert score: 2.46 -> 3.33 | Improved |
| use Stack-NMN instead of MAC | VQA model · CLEVR | accuracy (%): 98.9 -> 93.0 | Hurt |
Closest papers by meaning
- Compositional Reasoning with Transformers, RNNs, and Chain of Thought (NeurIPS 2025 · similarity 0.86)
- Compositional Reasoning in Language Models under Reinforcement Learning Post-Training (NeurIPS 2026 · similarity 0.85)
- Are Transformers Able to Reason by Connecting Separated Knowledge in Training Data? (ICLR 2025 · similarity 0.85)
- Think in Latent, Explain in Language: Self-Explainable Latent Reasoning (ICML 2026 · similarity 0.85)
- NeSyCoCo: A Neuro-Symbolic Concept Composer for Compositional Generalization (AAAI 2025 · similarity 0.85)
- Attention as a Hypernetwork (ICLR 2025 · similarity 0.84)
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.
