Research map / Large language models / Language, safety and interpretability
LogiGAN: Learning Logical Reasoning via Adversarial Pre-training
Xinyu Pi, Wanjun Zhong, Yan Gao, Nan Duan, Jian-Guang Lou
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 |
|---|---|---|---|
| pre-train T5 base with LogiGAN (ss+es) | T5 base · 12 downstream tasks (classification avg) | accuracy: 49.52 -> 54.62 | Improved |
| pre-train T5 base with LogiGAN (ss+es) | T5 base · generation tasks avg | score: 36.65 -> 37.77 | Improved |
| pre-train T5 large with LogiGAN | T5 large · classification tasks avg | accuracy: 63.56 -> 66.47 | Improved |
| pre-train T5 large with LogiGAN | T5 large · generation tasks avg | score: 38.87 -> 40.28 | Improved |
| pre-train T5 large with LogiGAN | T5 large · ReClor | accuracy: 50.40 -> 54.80 | Improved |
| add Elastic Search retrieved pseudo-statements to self-sampling | LogiGAN base · 12 tasks classification avg | accuracy: 54.06 -> 54.62 | Improved |
| add Elastic Search retrieved pseudo-statements to self-sampling | LogiGAN base · LogiQA | accuracy: 34.72 -> 37.02 | Improved |
| add Elastic Search retrieved pseudo-statements to self-sampling | LogiGAN base · ReClor | accuracy: 40.20 -> 40.00 | No clear effect |
| replace logical statements with random masked sentences (I) (ablation) | LogiGAN base · 7-dataset average | accuracy: 49.80 -> 50.10 | No clear effect |
| MLE-only logic pre-training without adversarial training (II) (ablation) | T5 base · 7-dataset average | accuracy: 49.80 -> 52.71 | Improved |
| add adversarial Generator-Verifier training on top of MLE-only (full LogiGAN vs II) (ablation) | MLE logic pre-train · 7-dataset average | accuracy: 52.71 -> 54.59 | Improved |
| iterative multi-task generation plus verification in one model (III) vs MLE-only (II) (ablation) | MLE logic pre-train · 7-dataset average | accuracy: 52.71 -> 52.08 | Hurt |
Closest papers by meaning
- Generative Adversarial Reasoner: Enhancing LLM Reasoning with Adversarial Reinforcement Learning (ICLR 2026 · similarity 0.89)
- AdvChain: Adversarial Chain-of-Thought Tuning for Robust Safety Alignment of Large Reasoning Models (ICLR 2026 · similarity 0.86)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.86)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.86)
- Selection, Reflection and Self-Refinement: Revisit Reasoning Tasks via a Causal Lens (ICLR 2026 · similarity 0.86)
- I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders (AAAI 2026 · 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.
