Research map / Large language models / Retrieval and knowledge
UNICORN on RAINBOW: A Universal Commonsense Reasoning Model on a New Multitask Benchmark
Nicholas Lourie, Ronan Le Bras, Chandra Bhagavatula, Yejin Choi
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 |
|---|---|---|---|
| Train on the other Rainbow tasks, then fine-tune on the target | T5-Large · Rainbow tasks | dev accuracy: ANLI 77.8 -> 79.5; CosmosQA 81.9 -> 83.2; HellaSwag 82.8 -> 83.0; PhysicalIQA 80.2 -> 82.2; SocialIQA 73.8 -> 75.5; Wino | Improved |
| Train on multiple tasks simultaneously, including the target | T5-Large · Rainbow tasks | dev accuracy: WinoGrande 77.0 -> 72.1 | Mixed |
| Use Rainbow rather than GLUE or SuperGLUE for sequential transfer | T5-Large · Rainbow tasks | data efficiency | Improved |
| Add commonsense knowledge-graph generation as a multitask source | T5-Large · Rainbow tasks | dev accuracy: ATOMIC: ANLI 77.8 -> 78.3; CosmosQA 81.9 -> 81.8; HellaSwag 82.8 -> 82.8; PhysicalIQA 80.2 -> 79.9; SocialIQA 73.8 -> 75 | No clear effect |
| Scale the pretrained model up when using transfer | T5 models transferred from Rainbow to CommonsenseQA · CommonsenseQA | data efficiency | Improved |
| Use sequential training instead of multitask training or multitask fine-tuning | T5 models · Rainbow tasks | data efficiency | Improved |
| Initialize from T5-11B multitasked on Rainbow, then fine-tune on each target task | T5-11B · ANLI | accuracy: 86.8% -> 87.3% | Improved |
| Initialize from T5-11B multitasked on Rainbow, then fine-tune on each target task | T5-11B · CosmosQA | accuracy: 90.3% -> 91.8% | Improved |
Closest papers by meaning
- MDK12-Bench: A Multi-Discipline Benchmark for Evaluating Reasoning in Multimodal Large Language Models (AAAI 2026 · similarity 0.88)
- The Generalization and Robustness of Transformer-Based Language Models on Commonsense Reasoning (AAAI 2024 · similarity 0.88)
- CodePlan: Unlocking Reasoning Potential in Large Language Models by Scaling Code-form Planning (ICLR 2025 · similarity 0.87)
- CORECODE: A Common Sense Annotated Dialogue Dataset with Benchmark Tasks for Chinese Large Language Models (AAAI 2024 · similarity 0.87)
- Global PIQA: Evaluating Commonsense Reasoning Across 100+ Languages and Cultures (NeurIPS 2026 · similarity 0.87)
- Buffer of Thoughts: Thought-Augmented Reasoning with Large Language Models (NeurIPS 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.
