Research map / Architectures and efficiency / Attention and transformers
Emergent Discrete Controller Modules for Symbolic Planning in Transformers
S M Rafiuddin, Muntaha Nujat Khan
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 discrete Gumbel-Softmax controller modules | Transformer · Sorting (n=320) | accuracy: 38.7 -> 91.6 | Improved |
| add discrete controller modules | Transformer · Sorting (n=80, in-distribution) | accuracy: 85.4 -> 97.4 | Improved |
| add discrete controller modules | Transformer · DROP numeric subset | F1: 78.3 -> 85.1 | Improved |
| add discrete controller modules | Transformer · RobustFill | exact match: 67.8 -> 73.9 | Improved |
| use discrete Gumbel-Softmax selection instead of soft mixture (ablation) | Ctrl-Transformer · Sorting (n=320) | accuracy: 66.3 -> 91.6 | Improved |
| use discrete selection instead of soft mixture (ablation) | Ctrl-Transformer · BFS (n=320) | accuracy: 74.2 -> 83.5 | Improved |
| insert controller in every block instead of every 2nd (ablation) | Ctrl-Transformer · Sorting (n=320) | accuracy: 91.6 -> 92.7 | Improved |
| insert controller every 4th block instead of every 2nd (ablation) | Ctrl-Transformer · Sorting (n=320) | accuracy: 91.6 -> 84.9 | Hurt |
| add semantic consistency loss (ablation) | Ctrl-Transformer · Sorting (n=320) | accuracy: -4.5 points when removed | Improved |
| keep controller active at inference (vs knockout) (ablation) | Ctrl-Transformer · Sorting (n=320) | accuracy: 44.8 -> 91.6 | Improved |
| add discrete controller modules | Transformer · BFS length extrapolation | accuracy: 57.9 -> 83.5 | Improved |
Closest papers by meaning
- Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping (NeurIPS 2026 · similarity 0.85)
- From Shortcut to Induction Head: How Data Diversity Shapes Algorithm Selection in Transformers (NeurIPS 2025 · similarity 0.84)
- MambaPEFT: Exploring Parameter-Efficient Fine-Tuning for Mamba (ICLR 2025 · similarity 0.84)
- A Formal Framework for Understanding Length Generalization in Transformers (ICLR 2025 · similarity 0.84)
- Do Efficient Transformers Really Save Computation? (ICML 2024 · similarity 0.84)
- Optimal Control for Transformer Architectures: Enhancing Generalization, Robustness and Efficiency (NeurIPS 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.
