Research map / Large language models / Agents, code and math
Programming with a Differentiable Forth Interpreter
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 |
|---|---|---|---|
| Use a Forth sketch with a length-generalizing outer loop instead of Seq2Seq | Sorting · Bubble sort, test length 64, train length 2 | accuracy (Hamming distance): 13.3 -> 100.0 | Helped |
| Provide comparison behavior in the sketch, leaving only the comparison to learn | ∂4 sorting sketches · Bubble sort, test length 64, train length 4 | accuracy (Hamming distance): 7.81 -> 20.65 | Helped |
| Train the sorting sketches on length-4 sequences rather than shorter sequences | ∂4 sorting sketches · Bubble sort, test length 64 | accuracy (Hamming distance): Permute: 100.0 -> 7.81; Compare: 100.0 -> 20.65 | Hurt |
| Use a recursive addition sketch rather than Seq2Seq | Addition · Addition, test length 64, train length 2 | accuracy (Hamming distance): 15.0 -> 100.0 | Helped |
| Specify carry and output-digit choices in the sketch rather than directly manipulating state | ∂4 addition sketches · Addition, test length 64, train length 2 | accuracy (Hamming distance): 99.49 -> 100.0 | Helped |
| Train the Manipulate sketch with more examples | ∂4 Manipulate · Addition, test length 64, train length 2 versus 4 | accuracy (Hamming distance): 99.49 -> 100.0 | Helped |
| Train a joint text reader and differentiable interpreter end-to-end from question-answer pairs | Common Core (CC) · Common Core (CC) | accuracy (%): 95.0 -> 96.0 | Helped |
| Use ∂4 without data augmentation instead of GeNeRe with additional data | Common Core (CC) · Common Core (CC) | accuracy (%): 98.5 -> 96.0 | Hurt |
Closest papers by meaning
- Learning Differentiable Programs with Admissible Neural Heuristics (NeurIPS 2020 · similarity 0.90)
- Data-Efficient Learning with Neural Programs (NeurIPS 2024 · similarity 0.86)
- HYSYNTH: Context-Free LLM Approximation for Guiding Program Synthesis (NeurIPS 2024 · similarity 0.85)
- Program-as-Weights: A Programming Paradigm for Fuzzy Functions (NeurIPS 2026 · similarity 0.85)
- Large Language Models are Interpretable Learners (ICLR 2025 · similarity 0.84)
- Skip a Layer or Loop It? Learning Program-of-Layers in LLMs (ICML 2026 · 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.