Research map / Large language models / Agents, code and math
Using natural language and program abstractions to instill human inductive biases in machines
Sreejan Kumar, Carlos G Correa, Ishita Dasgupta, Raja Marjieh, Michael Hu, Robert D. Hawkins, Jonathan Cohen, Nathaniel Daw, Karthik R Narasimhan, Thomas L. Griffiths
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 |
|---|---|---|---|
| Ground the agent with human-generated language descriptions | PPO meta-RL agent · GSP boards | white tiles revealed: p<0.0001 | Improved |
| Ground the agent with human-generated language descriptions | PPO meta-RL agent · Control boards | white tiles revealed: p=0.021 | Hurt |
| Ground the agent with a flattened-board autoencoder target | PPO meta-RL agent · GSP boards | white tiles revealed: p<0.001 | Improved |
| Ground the agent with a flattened-board autoencoder target | PPO meta-RL agent · GSP boards vs control boards | performance difference: p=0.252 | No clear effect |
| Ground the agent with synthetic descriptions | PPO meta-RL agent · GSP boards and control boards | white tiles revealed: p=0.0016 | Mixed |
| Use library-learning program embeddings instead of program embeddings without library learning (ablation) | PPO meta-RL agent · GSP boards | white tiles revealed: p<0.0001 | Improved |
| Ground the agent with program embeddings without library learning (ablation) | PPO meta-RL agent · GSP boards vs control boards | performance difference: p=0.11 | No clear effect |
| Add library-learned abstractions to program representations (ablation) | DreamCoder program representations · GSP boards | program description length | Improved |
| Use human-generated rather than synthetic language descriptions (ablation) | Language representations · GSP boards | description length | Improved |
| Use library-learning program description lengths rather than program description lengths without library learning (ablation) | DreamCoder program representations · GSP boards | correlation with human-language description length: p=0.0034 | Improved |
| Ground the agent with a DreamCoder recognition-model representation rather than a generic DNN representation (ablation) | PPO meta-RL agent · GSP boards vs control boards | selective performance favoring GSP boards | Improved |
| Train a generic meta-RL agent without grounding | PPO meta-RL agent · GSP boards vs control boards | white tiles revealed: p<0.0001 | Hurt |
Closest papers by meaning
- Human-like Category Learning by Injecting Ecological Priors from Large Language Models into Neural Networks (ICML 2024 · similarity 0.88)
- Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data (NeurIPS 2025 · similarity 0.86)
- In-Context Learning Agents Are Asymmetric Belief Updaters (ICML 2024 · similarity 0.85)
- Beyond Induction Heads: In-Context Meta Learning Induces Multi-Phase Circuit Emergence (ICML 2025 · similarity 0.85)
- Text-to-Decision Agent: Offline Meta-Reinforcement Learning from Natural Language Supervision (NeurIPS 2025 · similarity 0.85)
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (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.
