Research map / Architectures and efficiency / Attention and transformers
In-Context Algebra
Eric Todd, Jannik Brinkmann, Rohit Gandikota, David Bau
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 hidden dimension from 128 to 1024 (ablation) | In-context algebra transformer · held-out eval | evaluation accuracy: 13.3% -> 99.4% | Improved |
| increase hidden dimension from 256 to 512 (ablation) | In-context algebra transformer · held-out eval | evaluation accuracy: 61.3% -> 96.9% | Improved |
| increase hidden dimension from 1024 to 2048 (ablation) | In-context algebra transformer · held-out eval | evaluation accuracy: 99.4% -> 98.6% | No clear effect |
| increase layers from 2 to 4 (ablation) | In-context algebra transformer · held-out eval | evaluation accuracy: 93.5% -> 99.4% | Improved |
| increase layers from 4 to 10 (ablation) | In-context algebra transformer · held-out eval | evaluation accuracy: 99.4% -> 98.8% | No clear effect |
| increase heads from 4 to 8 (ablation) | In-context algebra transformer · held-out eval | evaluation accuracy: 95.9% -> 99.4% | Improved |
| increase heads from 2 to 8 (ablation) | In-context algebra transformer · held-out eval | evaluation accuracy: 99.2% -> 99.4% | No clear effect |
| vary depth for associativity sequences (ablation) | In-context algebra transformer · associativity subset | associativity score: 85.9% (4 layers) vs 51.6% (8 layers) | Mixed |
| raise group mixing probability p_mix (ablation) | In-context algebra transformer · held-out eval | evaluation accuracy | Improved |
| raise group mixing probability p_mix (ablation) | In-context algebra transformer · training | time to generalization | Improved |
| evaluate on unseen order-8 groups vs in-distribution groups | In-context algebra transformer · unseen groups | accuracy | No clear effect |
| evaluate on quasigroups instead of groups | In-context algebra transformer · held-out unseen structures | accuracy: up to 50% at 200 facts | Hurt |
Closest papers by meaning
- When can transformers reason with abstract symbols? (ICLR 2024 · similarity 0.89)
- The Expressive Power of Transformers with Chain of Thought (ICLR 2024 · similarity 0.87)
- Are Transformers Able to Reason by Connecting Separated Knowledge in Training Data? (ICLR 2025 · similarity 0.87)
- Grokking of Implicit Reasoning in Transformers: A Mechanistic Journey to the Edge of Generalization (NeurIPS 2024 · similarity 0.87)
- How do Transformers Learn Implicit Reasoning? (NeurIPS 2025 · similarity 0.87)
- Transformers Provably Learn Chain-of-Thought Reasoning with Length Generalization (NeurIPS 2025 · 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.
