Research map / Architectures and efficiency / RNNs and normalization
Weights to Code: Extracting Interpretable Algorithms from the Discrete Transformer
Yifan Zhang, Wei Bi, Kechi Zhang, Dongming Jin, Jie Fu, Zhi Jin
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 |
|---|---|---|---|
| increase operand arity k from 1 to 2 (ablation) | sub-MLP symbolic regression · add_mod_3 | RMSE: 6.46 x 10^-2 -> 3.10 x 10^-4 | Improved |
| increase operand arity k from 1 to 2 (ablation) | sub-MLP symbolic regression · bit_dot_prod_mod_2 | RMSE: 2.36 x 10^-1 -> 3.00 x 10^-4 | Improved |
| increase operand arity k from 2 to 4 (ablation) | sub-MLP symbolic regression · add_mod_3 | RMSE: 3.10 x 10^-4 -> 2.98 x 10^-3 | Hurt |
| increase operand arity k from 2 to 4 (ablation) | sub-MLP symbolic regression · bit_dot_prod_mod_2 | RMSE: 3.00 x 10^-4 -> 8.44 x 10^-4 | Hurt |
| use k=3 instead of k=2 (ablation) | sub-MLP symbolic regression · parity_last3 | RMSE: 4.69 x 10^-1 -> 3.33 x 10^-6 | Improved |
| add magnitude pruning of output head (ablation) | Discrete Transformer · diff_last2 | program length (lines): 33 -> 22 | Improved |
| add magnitude pruning of output head (ablation) | Discrete Transformer · sum_last2 | program length (lines): 34 -> 28 | Improved |
| add magnitude pruning of output head (ablation) | Discrete Transformer · parity_last2 | program length (lines): 34 -> 34 | No clear effect |
| add l1 regularization on output head (coefficient 0.001) (ablation) | over-parameterized Discrete Transformer · sum_last2 | program length (lines): 78 -> 22 | Improved |
| add l1 regularization on output head (coefficient 0.001) (ablation) | over-parameterized Discrete Transformer · spring | program length (lines): 96 -> 34 | Improved |
| slow annealing horizon from 25 to 100 epochs (ablation) | over-parameterized Discrete Transformer · spring | training MSE: 3.69 x 10^-3 -> 3.94 x 10^-7 | Improved |
| set number of sub-MLPs to zero (ablation) | Discrete Transformer · maximum_prev2 | mechanism | Mixed |
Closest papers by meaning
- Discovering Interpretable Algorithms by Decompiling Transformers to RASP (ICML 2026 · similarity 0.88)
- Transformers Provably Solve Parity Efficiently with Chain of Thought (ICLR 2025 · similarity 0.85)
- SMILE: Bridging Continuous Optimization and Discrete Symbolic Recovery (NeurIPS 2026 · similarity 0.84)
- HYSYNTH: Context-Free LLM Approximation for Guiding Program Synthesis (NeurIPS 2024 · similarity 0.84)
- From I/O to Code with Discovery Agent (NeurIPS 2026 · similarity 0.84)
- BoolXAI: Explainable AI Using Expressive Boolean Formulas (AAAI 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.
