Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?
A Tsinghua perspective proposes pairing a cheap utility predictor with an expensive executor, while examining efficiency gains, evidence, risks, and open questions.
ls ./papers
Trending AI research papers from Hugging Face, summarized in plain language: what each paper does, why it matters and where to read more.
140 papers
A Tsinghua perspective proposes pairing a cheap utility predictor with an expensive executor, while examining efficiency gains, evidence, risks, and open questions.
R²Orch combines self-play, adaptive task selection, and evidence-based skill withdrawal to train agent orchestrators that retain performance when skills are removed.
HiFOR links short-term motion, task milestones, and final goals in a VLA policy, then retrieves only the future information most relevant to each action.
AskStream adds an ask action to streaming video understanding, showing that models struggle to identify missing information, ask at the right time, and avoid hallucinating.
Glance turns the RAG generator into its own reranker, sharing one passage prefill to cut duplicated model work while improving retrieval and downstream QA in early tests.
A new benchmark tests whether AI agents can perform equity research using only information available at decision time—and finds forecasting remains a major weakness.
CoSTA separates speech generation from synthesis, using losslessly expandable BPE speech tokens to reduce autoregressive steps while preserving spoken-dialogue quality.
Soft-MTP removes a noisy component of standard multi-token prediction, improving model training, speculative decoding, and downstream math and code results without added parameters.
Co-Design Bench tests whether coding agents can turn real users’ multimodal requests into complete, working application revisions—not merely plausible partial edits.
HairDeck generates stylized hair cards and spring-bone rigs from one image, replacing slow autoregressive modeling with a faster unordered set diffusion approach.
A study of multimodal retrieval heads finds sparse, shared mechanisms across audio and vision, then uses them to improve document retrieval.
L3G-World combines adaptive 3D Gaussian memory with direct VAE-latent rendering to improve long-horizon consistency and reach real-time interactive world generation.