Shaping the future of AI development

Mesh-native autoregression for whole scenes from a single view, token-by-token.

Reassembles what's left. Generates what isn't.

Train the agent inside its own loop. 7B beats GPT-4o on search, math, and science.

Discovers objects and masks from raw video, predicts what happens next, no supervision.

LLMs fold under peer pressure. Our benchmark and RL recipe help smaller models hold their ground.

Object-level grading of multi-turn image editing, exposing where today's best editors silently break.

Give an LLM a job and clear boundaries. It still answers off-topic questions, almost every time.

Sparse updates make syncing cheap enough to go frequent, neutralizing non-IID drift.

Smarter data selection pushes the forget-vs-retain frontier past oracle sampling.

Token-level RL doesn't fit diffusion LLMs. Treat the whole sequence as one action, 20–40 point gains.

Pretrains on static molecular structures, stitches them into dynamics trajectories, bypassing simulation data scarcity.

515M params, 30s of studio audio in under 4 seconds. Aligned by ranking its own outputs.

Learnable sparse attention for video. 3.6% attention budget at 128K tokens, accuracy still improves.

Model weight exponents cluster into 2–3 bits of entropy. Lossless FP8 compression, up to 177% faster inference.

Diffusion beats autoregressive in data-constrained settings.

Builds an explicit 3D scene and reasons over it step by step, boosting accuracy and generalization on 3D spatial QA benchmarks.

Replaces heavy tensor operations in molecular force-field models with low-rank approximation, reducing compute while keeping accuracy.

Aligns VLMs with diffusion models through shared CLIP patch embeddings, enabling controllable high-quality generation while preserving reasoning.

Using simple BLEU scores as feedback on hard instructions can train instruction-following models that rival those tuned with expensive learned rewards.

A dataset that stress-tests layout-to-image models on heavily overlapping scenes, exposing current failures and offering an improved baseline.