September 12, 2026

The paper identifies environment execution (exclusive sandboxes, real machine time) as the scaling bottleneck in RL post-training of automatic research agents, and replaces it with a learned world model (WMRL). Two corrections — Online Debiasing and Inverse-Variance Denoising — fix the biased, noisy rewards the world model introduces, with proofs that both strictly improve convergence. The authors report a 3-4x training speedup at matched or better final performance; 4B and 9B post-trained agents beat 48B and 120B open-weight agents on held-out benchmarks, and the method transfers to embodied VLA policies.

The Intern-NCP team scaled a latent-space language model to 8.9B parameters trained on 5.73T tokens, the largest such demonstration to date. On top of next-token prediction it adds Next Concept Prediction: a product-quantized concept vocabulary built from hidden states, a Concept Module predicting future multi-token concepts, and feedback of those predictions to guide token-level generation, trained end-to-end. It matches OLMo-3-7B's final pretraining loss with 51.3% of the tokens and beats it by 2.45 points macro-average (+5.99 GSM8K) after full pretraining; the 17M-parameter VQ module alone enables cheap domain adaptation and improves speculative-drafting accepted length by 4.17%.

SenseTime's SenseNova-U1.5 is an 8B mixture-of-transformers model doing visual understanding, reasoning, generation and editing in one encoder-free, VAE-free architecture — no separate vision encoder, no diffusion VAE. It uses spatially coherent patch reconstruction as the visual interface, native resolution up to 4K, and post-training with specialized experts (aesthetics, bilingual text rendering, infographics, editing) merged via multi-expert on-policy distillation. The report shows gains in fidelity, text rendering, multi-reference editing and identity-preserving edits; training code for SFT, RL and on-policy distillation is planned for open release.