Research map / Label-efficient and robust learning / Recognition and long-tail
DeepMath - Deep Sequence Models for Premise Selection
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 |
|---|---|---|---|
| add hard negative mining during training (ablation) | CNN premise model · Mizar premise selection top-16 | theorems proved | Helped |
| add hard negative mining during training (ablation) | best neural model · Mizar premise selection union over cutoffs | % proved: 61.3 -> 66.4 | Helped |
| use definition-based embeddings instead of naive word embeddings (ablation) | word-level CNN · Mizar premise selection top-16 | % proved: 25.9 -> 26.8 | Helped |
| use definition-based embeddings instead of naive word embeddings (ablation) | word-level CNN · Mizar premise selection top-1024 | % proved: 64 -> 66.4 | Helped |
| replace CNN with LSTM on top of conv layers (ablation) | def-CNN · Mizar premise selection top-1024 | % proved: 66.4 -> 60.1 | Hurt |
| ensemble def-CNN with char-CNN scores | def-CNN · Mizar premise selection top-16 | % proved: 26.8 -> 30.5 | Helped |
| ensemble def-CNN with char-CNN scores | def-CNN · Mizar premise selection top-1024 | % proved: 66.4 -> 67.9 | Helped |
| replace k-NN handcrafted features with def-CNN | k-NN baseline · Mizar premise selection top-1024 | % proved: 65.1 -> 66.4 | Helped |
| replace k-NN with character-level CNN | k-NN baseline · Mizar premise selection top-32 | % proved: 39.4 -> 37.5 | Hurt |
| union with k-NN predictions instead of another neural model | def-CNN · Mizar premise selection union | % proved: 69.8 -> 74.25 | Helped |
| iterate definition-embedding stage additional times | def-CNN · Mizar premise selection | % proved | No clear effect |
| train RNNs on shorter sequences (ablation) | character-level RNN · Mizar premise selection | performance | Helped |
Closest papers by meaning
- REFACTOR: Learning to Extract Theorems from Proofs (ICLR 2024 · similarity 0.88)
- ProofAug: Efficient Neural Theorem Proving via Fine-grained Proof Structure Analysis (ICML 2025 · similarity 0.87)
- Magnushammer: A Transformer-Based Approach to Premise Selection (ICLR 2024 · similarity 0.87)
- Subgoal-based Demonstration Learning for Formal Theorem Proving (ICML 2024 · similarity 0.86)
- Reviving DSP for Advanced Theorem Proving in the Era of Reasoning Models (NeurIPS 2025 · similarity 0.86)
- PutnamBench: Evaluating Neural Theorem-Provers on the Putnam Mathematical Competition (NeurIPS 2024 · similarity 0.86)
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.