List your company

Interactive Reasoning on ScienceWorld 30 tasks

84.7ScoreSwiftSage
34.46847.50960.5573.591Mar 18, 2026

Evaluation results

MethodLinks
SwiftSageBackbone=SWIFT + SAGE (GPT-4), Regime=Hybrid (BC + prompting)2026.0384.7
CLINBackbone=GPT-4, Regime=Continual memory update2026.0359.8
RetrospexBackbone=Flan-T5-large + RL, Regime=Trained (IL+RL)2026.0356
Retrospex: IL-T5Backbone=Flan-T5-large, Regime=Trained (IL)2026.0348.8
SwiftSage: ReflexionBackbone=GPT-4, Regime=Prompt-only (iterative reflection)2026.0345.3
SwiftSage: ReActBackbone=GPT-4, Regime=Prompt-only (few-shot)2026.0336.4