• Multi-agent errors can finally be computed, not guessed. GBC adds differentiable weights to the connections between agents, so loss flows back along the interaction chain and turns "which agent's prompt to fix" into token-level attribution — as long as the collaboration structure is differentiable, which discrete tool calls still test.

• Tokenizer-free pixel AR is closing the quality gap. PRA uses a low-dimensional intermediate and approximate rollout to suppress compounding error, hitting FID 2.58 on ImageNet 256 with 135M parameters — past the 3.60 of earlier billion-scale models, an order of magnitude smaller and better.

Read the full edition →