Close reading

Turns a 33B text-to-video model (MiniMax-H3) into an interactive world model without adding any action module: each character/camera action is written as a short sentence, bound to one video-latent interval, restricted by an attention mask so it can only enter the video stream at its matched latent, and learned with LoRA on 0.199% of the parameters.

★★ · 35 min · 7 sections

Shows by human scoring that revealing the ground-truth future trajectory to a CoT teacher makes it justify the outcome instead of reasoning from the scene, then recasts planning as a choice among explicit candidate trajectories (AD-MCQ) and defers that choice until after the decision is committed, training it with verifiable rewards (DEFT-RLVR).

★★★ · 45 min · 7 sections

Introduces Declarative Attention (DA), a zero-shot protocol in which the model announces its own attention scope with <global>/<focus>/<local> tags inside its chain-of-thought, and the inference engine builds the KV-cache mask by parsing that text — no auxiliary scorer, no training, roughly half the attended tokens.

★★★ · 45 min · 8 sections

★★ · 45 min · 7 sections

Most looped-Transformer results compare at fixed model size, which quietly hands the looped model extra FLOPs. This paper closes three budgets at once — per-token FLOPs, total non-embedding parameters, and KV cache — using MoE as the currency, and still finds looping ahead: 6.8–18.0% of training compute saved on the compute-optimal frontier, across a scaling ladder up to 54B non-embedding parameters.

★★★ · 45 min · 8 sections

Proposes the Transformer, which drops recurrence and convolution entirely and does sequence transduction with attention alone, reaching state-of-the-art machine translation at a small fraction of the training cost.

★★ · 40 min · 8 sections