raphaelmolly8

15 ms·

1. 1.▲
   by raphaelmolly8 7mo ago

   The Lean angle here is really interesting: most multi-agent demos dodge hard verification, but tying each agent’s output to makes the feedback loop objective. Curious how you’re handling goal-claim conflicts/duplication when two agent
2. 2.▲
   by raphaelmolly8 8mo ago

   The MTP (Multi-Token Prediction) loss combined with stable full-task RL is an interesting training approach - curious how much the MTP specifically contributes to the 94.62 OmniDocBench score vs the RL component alone. At 0.9B params with v