Research map / Large language models / Language, safety and interpretability
Leveraging the Inductive Bias of Large Language Models for Abstract Textual Reasoning
Christopher Michael Rytting, David Wingate
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 |
|---|---|---|---|
| Train on the full, nonspecific noun set rather than a 2k common-noun subset | T5-3B container reasoner · Containers, validation words (all) | exact: 0.60 -> 0.78 | Improved |
| Train on the full, nonspecific noun set rather than a 2k concrete-noun subset | T5-3B container reasoner · Containers, validation words (all) | exact: 0.59 -> 0.78 | Improved |
| Train on the full, nonspecific noun set rather than 2k randomly sampled nouns | T5-3B container reasoner · Containers, validation words (all) | exact: 0.55 -> 0.78 | Improved |
| Use a larger, nonspecific noun training distribution | T5-3B container reasoner · Containers, new POS (random strings) | exact: 0.37 -> 0.44; 0.31 -> 0.44; 0.40 -> 0.44 | Mixed |
| Train first on elemental navigation and container tasks before the composite task | T5-3B hard object reasoner · Hard object composite task | performance and sample efficiency | Improved |
| Train on a path-length-balanced navigation distribution instead of randomly sampling maps and endpoints | T5-3B navigation reasoner · Navigation route, four- and five-step plans | exact string accuracy | Improved |
| Use a path-length-balanced navigation distribution | T5-3B navigation reasoner · Navigation route, two-step plans | exact string accuracy | Hurt |
| Replace natural English template words with one-to-one gibberish equivalents (ablation) | T5-3B container reasoner · Containers, gibberish-template comparison | learning performance | Hurt |
| Increase the number of objects and containers beyond the training range | T5-3B container reasoner · Containers, cardinality extrapolation | exact string accuracy | Mixed |
| Require longer route outputs than seen during training | T5-3B navigation route reasoner · Navigation route, longer plans | exact string accuracy | Hurt |
Closest papers by meaning
- On the Inductive Bias of Stacking Towards Improving Reasoning (NeurIPS 2024 · similarity 0.90)
- Hypothesis Search: Inductive Reasoning with Language Models (ICLR 2024 · similarity 0.89)
- Meaningful Learning: Enhancing Abstract Reasoning in Large Language Models via Generic Fact Guidance (NeurIPS 2024 · similarity 0.89)
- How do Transformers Learn Implicit Reasoning? (NeurIPS 2025 · similarity 0.89)
- Phenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinement (ICLR 2024 · similarity 0.89)
- Grokking of Implicit Reasoning in Transformers: A Mechanistic Journey to the Edge of Generalization (NeurIPS 2024 · similarity 0.89)
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.
