Rohan Paul (@rohanpaul_ai)T2·78 pts
- PAN 2 splits world-model reasoning into conceptual planning via an LLM and perceptual rendering via a 29B diffusion model
- Removing the LLM drops multi-step performance to 8.7% correct steps with zero completed plans, showing action ordering depends on the language model
- The combined system reaches 75.1% correct steps and 34.0% fully completed plans, the best among models IFM tested