Research papers, read to you in five minutes.
Research papers, read to you in five minutes.

Catching a thrown box on a humanoid from onboard RGB-D, without mocap or a fixed base: train a privileged teacher in sim, distill its motor skill through an 8-dim object-dynamics latent, then ground that latent in vision. Hits 76% of human catch rate at 2.5m on a real Unitree G1.

Diffusion · Added Oct 13

DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models

DiffusionART attacks the per-task EMA-tuning tax that makes RL post-training of diffusion models fragile at scale. By replacing forward-velocity targets with one prompt-level reward direction plus a two-update restoration split, it hits prior video-RL rewards with 12× fewer GPU hours on a 13B model, no retuning.

Agents · Added Oct 13

SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments

SoftwareGym attacks the computer-use agent data bottleneck by generating 182 verified desktop apps from scratch, using GUI-driven test-and-repair loops. Fine-tuning a 9B model purely on this synthetic data lifts OSWorld Pass@3 from 41% to 57% with zero OSWorld apps seen—source-only self-critique actually regressed functionality.

Agents · Added Oct 13

EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents

EmbodiedCLBench asks whether household LLM agents actually reuse skills across tasks, using a state-transition graph over a closed simulator for reproducible eval. The kicker: one prior basic task lifts advanced success from ~10% to ~35%, but scaling to six examples adds nothing — mostly by cutting invalid actions.

Agents · Added Oct 11

CheckerBench: Can Long-Horizon Agents Synthesize Static-Analysis Checkers?

CheckerBench asks coding agents to synthesize reusable static-analysis checkers for 300 real CVEs across five languages, with independent rebuild-and-rescan verification. Best model-harness pair hits only 45% Pass@1; one-shot generation collapses to 0.33%, so skill libraries and live analyzer feedback dominate prompt tweaks.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Reinforcement Learning · Added Oct 13

Reflex: Visually Grounded Reactive Humanoid Control

Catching a thrown box on a humanoid from onboard RGB-D, without mocap or a fixed base: train a privileged teacher in sim, distill its motor skill through an 8-dim object-dynamics latent, then ground that latent in vision. Hits 76% of human catch rate at 2.5m on a real Unitree G1.

Diffusion · Added Oct 13

DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models

DiffusionART attacks the per-task EMA-tuning tax that makes RL post-training of diffusion models fragile at scale. By replacing forward-velocity targets with one prompt-level reward direction plus a two-update restoration split, it hits prior video-RL rewards with 12× fewer GPU hours on a 13B model, no retuning.

Agents · Added Oct 13

SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments

SoftwareGym attacks the computer-use agent data bottleneck by generating 182 verified desktop apps from scratch, using GUI-driven test-and-repair loops. Fine-tuning a 9B model purely on this synthetic data lifts OSWorld Pass@3 from 41% to 57% with zero OSWorld apps seen—source-only self-critique actually regressed functionality.

Agents · Added Oct 13

EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents

EmbodiedCLBench asks whether household LLM agents actually reuse skills across tasks, using a state-transition graph over a closed simulator for reproducible eval. The kicker: one prior basic task lifts advanced success from ~10% to ~35%, but scaling to six examples adds nothing — mostly by cutting invalid actions.

Agents · Added Oct 11

CheckerBench: Can Long-Horizon Agents Synthesize Static-Analysis Checkers?

CheckerBench asks coding agents to synthesize reusable static-analysis checkers for 300 real CVEs across five languages, with independent rebuild-and-rescan verification. Best model-harness pair hits only 45% Pass@1; one-shot generation collapses to 0.33%, so skill libraries and live analyzer feedback dominate prompt tweaks.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Browse topics

Research papers, read to you in five minutes.Research papers, read to you in five minutes.

Research papers, read to you in five minutes.
Research papers, read to you in five minutes.