SCISENSE: Structured Sensemaking Framework Boosts Research Novelty
A new framework called SCISENSE operationalizes scientific ideation as eight cognitive stages, challenging the assumption that looser supervision yields more novel research. The framework, based on Pirolli and Card's 2005 sensemaking model, generates 100K-scale trajectories (SCISENSE-Traj) in two modes: Target (reconstructing paths to known papers) and Infer (proposing novel directions). Surprisingly, Target-trained models outperformed Infer-trained ones by 2.0% in trajectory quality, suggesting that constrained sensemaking produces more novel output. The distilled SCISENSE-LM family ranges from 3B to 70B parameters.
Key facts
- SCISENSE operationalizes ideation as eight cognitive stages based on Pirolli & Card 2005
- SCISENSE-Traj dataset contains 100K citation-conditioned research trajectories
- Two modes: Target (reconstructs known paper paths) and Infer (proposes novel directions)
- Target-trained models achieved 2.0% improvement in trajectory quality over Infer-trained models
- SCISENSE-LM family spans 3B to 70B parameters
- Contrary to assumption that looser supervision promotes exploration
Entities
—