Research map / Architectures and efficiency / RNNs and normalization
Quiet Feature Learning in Algorithmic Tasks
Prudhviraj Naidu, Zixian Wang, Leon Bergen, Ramamohan Paturi
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 |
|---|---|---|---|
| keep carry feature (vs ablating it) (ablation) | Transformer++ on Addition · Addition (16) | test accuracy (%): 58.8 -> 100 | Improved |
| keep carry feature (vs ablating it) (ablation) | Transformer++ on Addition · Addition (64) | test accuracy (%): 24.9 -> 100 | Improved |
| keep first_operand feature (vs ablating it) (ablation) | Transformer++ on Addition · Addition (32) | test accuracy (%): 7.00 -> 100 | Improved |
| keep first_operand feature (vs ablating it) (ablation) | Transformer++ on Addition · Addition (16) | test accuracy (%): 100 -> 100 | No clear effect |
| keep first_operand feature (vs ablating it) (ablation) | Transformer++ on Multiplication · Multiplication (16) | test accuracy (%): 76.8 -> 76.8 | No clear effect |
| keep carry feature (vs ablating it) (ablation) | Transformer++ on Multiplication · Multiplication (16) | test accuracy (%): 76.8 -> 56.4 | Improved |
| keep queue feature (vs ablating it) (ablation) | Transformer++ on BFS · Breadth first search (11) | test accuracy (%): 99.7 -> 54.6 | Improved |
| keep is_prev_negative feature (vs ablating it) (ablation) | Transformer++ on Maximum Subarray · Maximum Subarray (64) | test accuracy (%): 95.6 -> 89.9 | Improved |
| ablate random direction instead of quiet feature (ablation) | Transformer++ · Addition (16) carry | test accuracy (%): 100 -> 100 | No clear effect |
| increase compute budget past slow phase | Transformer++ on algorithmic tasks · six of ten tasks | compute-optimal validation loss | Mixed |
Closest papers by meaning
- Tracing the Representation Geometry of Language Models from Pretraining to Post-training (NeurIPS 2025 · similarity 0.87)
- Transformers Converge to Invariant Algorithmic Cores (NeurIPS 2026 · similarity 0.87)
- What Happens During the Loss Plateau? Understanding Abrupt Learning in Transformers (NeurIPS 2025 · similarity 0.87)
- Slower Generalization, Faster Memorization: A Sweet Spot in Algorithmic Learning (NeurIPS 2026 · similarity 0.85)
- The Importance of Being Lazy: Scaling Limits of Continual Learning (ICML 2025 · similarity 0.85)
- Understanding Emergent Abilities of Language Models from the Loss Perspective (NeurIPS 2024 · similarity 0.85)
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.
