Research map / Large language models / Agents, code and math
NL2CA: Auto-formalizing Cognitive Decision-Making from Natural Language Using an Unsupervised CriticNL2LTL Framework
Zihao Deng, Yijia Li, Renrui Zhang, Peijun Ye
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 Critic Tree refinement on top of fine-tuned LLM | fine-tuned Qwen3-0.6B · NL-to-LTL Dataset1 | accuracy: 22.2 -> 80.6 | Improved |
| add Critic Tree refinement on top of fine-tuned LLM | fine-tuned Qwen3-0.6B · NL-to-LTL Dataset2 | accuracy: 97.5 -> 90.2 | Hurt |
| add Self-Refine to end-to-end DeepSeek R1 | DeepSeek R1 · NL-to-LTL Dataset1 | accuracy: 63.9 -> 63.9 | No clear effect |
| add Self-Refine to end-to-end DeepSeek R1 | DeepSeek R1 · NL-to-LTL Dataset1 | BLEU: 0.546 -> 0.490 | Hurt |
| use heterogeneous critic Tree instead of Self-Refine | DeepSeek R1 · NL-to-LTL Dataset1 | accuracy: 63.9 -> 80.6 | Improved |
| interactive human-in-loop GPT-4 vs automated Critic Tree | CriticNL2LTL · NL-to-LTL Dataset2 | accuracy: 78.3 -> 90.2 | Improved |
| let the LLM infer missing environment information when building rules (Msup) (ablation) | Mlit · Highway driving simulation | success rate: 58.70 -> 12.00 | Hurt |
| let the LLM infer missing information (Msup) (ablation) | Mlit · Lane Changing Interference | success rate: 91.30 -> 63.82 | Hurt |
| use NL2CA literal rules (Mlit) instead of manual expert rules | manual baseline · Driving simulation, average over 3 scenarios | success rate: 65.30 -> 78.57 | Improved |
| use NL2CA literal rules instead of manual rules | manual baseline · Driving simulation | RSR: 99.47 -> 95.69 | Hurt |
| add cognitive reinforcement learning on human data | Mlit vs manual · Driving simulation | JS divergence | Improved |
Closest papers by meaning
- Eliciting Reasoning in Language Models with Cognitive Tools (NeurIPS 2025 · similarity 0.88)
- Turning large language models into cognitive models (ICLR 2024 · similarity 0.86)
- Text-to-Decision Agent: Offline Meta-Reinforcement Learning from Natural Language Supervision (NeurIPS 2025 · similarity 0.86)
- Reflection-Bench: Evaluating Epistemic Agency in Large Language Models (ICML 2025 · similarity 0.86)
- Generating Computational Cognitive models using Large Language Models (NeurIPS 2025 · similarity 0.86)
- CogBench: a large language model walks into a psychology lab (ICML 2024 · 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.
