Actor, Judge, or Teacher? Understanding Decision Models in Long-Horizon Agents
A study of decision models finds that poor actors can still become effective teachers when they judge action progress with full-trajectory hindsight.
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 study of decision models finds that poor actors can still become effective teachers when they judge action progress with full-trajectory hindsight.
Open Vision Banana turns open image generators into multi-task perception models, showing how RGB targets, prompt design, and test-time ensembling shape accuracy.
PhysLoop combines LLM planning, physics simulation, and verification to generate more physically consistent videos and repair failures, while exposing trade-offs in visual quality and evaluation.
SO-Arena evaluates 21 System One models across capability, latency, difficulty, agentic tasks, and reliability—showing why accuracy alone is not enough for deployment.
G-STAR compresses autoregressive image generation by scaling per-token code capacity and using semi-autoregressive decoding, reaching 2.5-second 1024px synthesis in the authors’ tests.
Duo v1 combines autoregressive intent with causal audiovisual generation to support real-time, interruptible interaction. Its strongest results are in speech and visual fidelity, while semantic responsiveness remains challenging.
SimpleHC removes residual mixing from Hyper-Connections and groups memory access, matching mHC quality while cutting reported training overhead and improving stability.
BirdDreams combines camera video with a structured bird’s-eye view so a driving world model can control ego motion and simulate more plausible reactions from surrounding vehicles.
GTAGym turns GTA V-inspired missions into a verifiable benchmark and RL environment, revealing where visual agents still fail at dynamic control.
A new reinforcement-learning approach teaches language models to study textbooks while solving unfamiliar proofs, improving transfer beyond fixed retrieval.
ConceptTime uses predictive surprise to segment time series, discover inspectable concepts, classify with few labels, and flag inputs that fall outside its learned vocabulary.
DepthGraft treats cross-layer KV sharing as both a cache optimization and a training regularizer, reporting stronger benchmarks and faster prefill with little decoding overhead.