Research map / Large language models / Language, safety and interpretability
Can Large Language Models Generalize Procedures Across Representations?
Fangru Lin, Valentin Hofmann, Xingchen Wan, Weixing Wang, Zifeng Ding, Anthony G Cohn, Janet B. Pierrehumbert
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 |
|---|---|---|---|
| add Graph stage-1 GRPO training before NL training | Qwen2.5-1.5B-Instruct GRPO · Asynchronous planning NL test | accuracy: 0.698 -> 0.782 | Improved |
| reverse curriculum order to NL then Graph (ablation) | Graph->NL curriculum · Asynchronous planning NL test | accuracy: 0.698 -> 0.431 | Hurt |
| replace stage 2 GRPO with NL distillation (ablation) | Graph->NL curriculum · Asynchronous planning NL test | accuracy: 0.782 -> 0.462 | Hurt |
| apply the curriculum with standard SFT (ablation) | Qwen2.5-1.5B-Instruct SFT · Asynchronous planning NL test | accuracy: 0.236 -> 0.249 | No clear effect |
| interleave Graph and NL data instead of sequential curriculum (ablation) | Graph->NL curriculum · Asynchronous planning NL test | accuracy: 0.782 -> 0.382 | Hurt |
| use Code instead of Graph as stage 1 (ablation) | symbolic-then-NL curriculum · Asynchronous planning NL test | accuracy: 0.782 -> 0.382 | Hurt |
| mix Graph and Code in stage 1 (ablation) | Graph->NL curriculum · Asynchronous planning NL test | accuracy: 0.782 -> 0.533 | Hurt |
| add Graph stage-1 curriculum before NL training | Qwen2.5-1.5B-Instruct GRPO · NL-AAVE dialect test | accuracy: 0.507 -> 0.573 | Improved |
| add Code stage-1 curriculum before NL training on math | Qwen2.5-1.5B-Instruct GRPO · SciBench physics test (trained on math) | accuracy: 0.135 -> 0.230 | Improved |
| add Code stage-1 curriculum before NL training on physics | Qwen2.5-1.5B-Instruct GRPO · SciBench physics test (in-domain) | accuracy: 0.555 -> 0.550 | No clear effect |
| scale model size from 1.5B to 7B under NL-only training | Qwen2.5 GRPO NL training · Asynchronous planning NL test | accuracy: 0.698 -> 0.698 | No clear effect |
| train on Graph or Code only, test on NL | SFT and GRPO post-training · Asynchronous planning NL test | accuracy | No clear effect |
Closest papers by meaning
- Beyond English-Centric Training: How Reinforcement Learning Improves Cross-Lingual Reasoning in LLMs (ICLR 2026 · similarity 0.89)
- Enhancing Decision-Making of Large Language Models via Actor-Critic (ICML 2025 · similarity 0.88)
- Can large language models explore in-context? (NeurIPS 2024 · similarity 0.88)
- Do LLMs Build World Representations? Probing Through the Lens of State Abstraction (NeurIPS 2024 · similarity 0.88)
- LLM-Empowered State Representation for Reinforcement Learning (ICML 2024 · similarity 0.88)
- Can Graph Learning Improve Planning in LLM-based Agents? (NeurIPS 2024 · 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.
