Research map / Large language models / Agents, code and math
Faith and Fate: Limits of Transformers on Compositionality
Nouha Dziri, Ximing Lu, Melanie Sclar, Xiang Lorraine Li, Liwei Jiang, Bill Yuchen Lin, Sean Welleck, Peter West, Chandra Bhagavatula, Ronan Le Bras, Jena D. Hwang, Soumya Sanyal and 4 more
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 task complexity (problem size / graph depth) in zero-shot prompting | GPT-3, ChatGPT, GPT-4 · multiplication, puzzle, DP | accuracy | Hurt |
| add few-shot question-answer examples | ChatGPT/GPT-4 · multiplication and DP | accuracy | No clear effect |
| add few-shot examples | LLMs · Einstein puzzle | accuracy | Hurt |
| finetune GPT-3 on question-answer pairs up to a size, test out-of-domain | GPT-3 · multiplication, puzzle, DP | accuracy | Mixed |
| add scratchpad (question-scratchpad finetuning) | GPT-3 · multiplication, puzzle, DP | accuracy | Mixed |
| train far beyond overfitting point to seek grokking | GPT-3 · multiplication OOD | accuracy | No clear effect |
| train GPT2-XL from scratch with digit-level tokens on up to 4x4 (90M points) | GPT2-XL · 3x3 multiplication | accuracy | No clear effect |
| score of subgraph frequency in training data: correct vs wrong predictions | GPT-3 finetuned · multiplication and DP | training-subgraph frequency | Improved |
Closest papers by meaning
- Do Efficient Transformers Really Save Computation? (ICML 2024 · similarity 0.89)
- The Expressive Power of Transformers with Chain of Thought (ICLR 2024 · similarity 0.89)
- Limits of Transformer Language Models on Learning to Compose Algorithms (NeurIPS 2024 · similarity 0.88)
- Grokking of Implicit Reasoning in Transformers: A Mechanistic Journey to the Edge of Generalization (NeurIPS 2024 · similarity 0.88)
- Born a Transformer -- Always a Transformer? On the Effect of Pretraining on Architectural Abilities (NeurIPS 2025 · similarity 0.88)
- Assign and Add: A Mechanistic Study of Compositional Arithmetic (NeurIPS 2026 · similarity 0.88)
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.
