Research map / Large language models / Language, safety and interpretability
Abductive Commonsense Reasoning
Chandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya, Keisuke Sakaguchi, Ari Holtzman, Hannah Rashkin, Doug Downey, Wen-tau Yih, Yejin Choi
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 |
|---|---|---|---|
| Jointly model both observations rather than impose linear-chain independence (ablation) | BERT for alphaNLI · ART test | accuracy: 68.9 (0.5) -> 68.6 (0.5) | No clear effect |
| Restrict input to the second observation and hypothesis instead of using the linear-chain model (ablation) | BERT for alphaNLI · ART test | accuracy: 68.9 (0.5) -> 66.6 (0.2) | Hurt |
| Restrict input to the first observation and hypothesis instead of using the linear-chain model (ablation) | BERT for alphaNLI · ART test | accuracy: 68.9 (0.5) -> 63.5 (0.7) | Hurt |
| Use only the hypothesis instead of the linear-chain model (ablation) | BERT for alphaNLI · ART test | accuracy: 68.9 (0.5) -> 59.5 (0.2) | Hurt |
| Add COMeT embeddings to GPT2 conditioned on both observations (ablation) | GPT2 alphaNLG · ART test | BLEU: 2.23 -> 3.03 | Improved |
| Add COMeT embeddings to GPT2 conditioned on both observations (ablation) | GPT2 alphaNLG · ART test | Human evaluation: 42.26 -> 44.56 | Improved |
| Provide COMeT in text form rather than as embeddings (ablation) | GPT2 alphaNLG · ART test | BLEU: 2.23 -> 2.29 | No clear effect |
| Pre-train on ART before fine-tuning on WinoGrande (ablation) | BERT · WinoGrande | accuracy: 65.8% -> 67.2% | Improved |
| Pre-train on ART before fine-tuning on WSC (ablation) | BERT · WSC | accuracy: 70.0% -> 74.0% | Improved |
| Pre-train on ART before fine-tuning on DPR (ablation) | BERT · DPR | accuracy: 72.5% -> 86.0% | Improved |
| Pre-train on ART before fine-tuning on HellaSwag (ablation) | BERT · HellaSwag | accuracy: 46.7% -> 46.1% | Hurt |
Closest papers by meaning
- Abductive Reasoning with Probabilistic Commonsense (ICML 2026 · similarity 0.91)
- Graph of States: Solving Abductive Tasks with Large Language Models (ICML 2026 · similarity 0.88)
- AbductiveMLLM: Boosting Visual Abductive Reasoning Within MLLMs (AAAI 2026 · similarity 0.88)
- NL-Eye: Abductive NLI For Images (ICLR 2025 · similarity 0.87)
- Hypothesis Search: Inductive Reasoning with Language Models (ICLR 2024 · similarity 0.87)
- The Generalization and Robustness of Transformer-Based Language Models on Commonsense Reasoning (AAAI 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.
