Researchers introduce Meta-Skill, a method where a Builder model learns principles to construct better execution environments for a Target model, improving performance by 8.95 percentage points. ·
Researchers propose RIDE, a method that extrapolates reinforcement learning induced representation changes to stabilize on-policy distillation training. ·
Researchers propose RIDE, a method that extrapolates reinforcement learning induced representation changes to stabilize on-policy distillation training. ·
Researchers find that direct-decision models compress ordinal scales, using only 67 to 76 percent of available support, a bias that targeted post-training can correct. ·
Researchers propose RIDE, a method that extrapolates reinforcement learning induced representation changes to stabilize on-policy distillation training. ·
Researchers find that direct-decision models compress ordinal scales, using only 67 to 76 percent of available support, a bias that targeted post-training can correct. ·
Researchers find that direct-decision models compress ordinal scales, using only 67 to 76 percent of available support, a bias that targeted post-training can correct. ·
Researchers propose DC-SAE, a decoupled compact semantic autoencoder that combines semantic and pixel-level encoders to enable high-compression image generation with faster diffusion model convergence. ·
Researchers introduce WorldAuditBench, a benchmark containing 213 anomaly tasks across 13 environments to evaluate multimodal agents' ability to audit interactive 3D worlds. ·
Researchers find that direct-decision models compress ordinal scales, using only 67 to 76 percent of available support, a bias that targeted post-training can correct. ·
Researchers propose DC-SAE, a decoupled compact semantic autoencoder that combines semantic and pixel-level encoders to enable high-compression image generation with faster diffusion model convergence. ·
Researchers introduce WorldAuditBench, a benchmark containing 213 anomaly tasks across 13 environments to evaluate multimodal agents' ability to audit interactive 3D worlds. ·
HF Daily Papers5h ago▲arxiv.org
X voices
5 stories
X posts other sources are linking to, then named voices on Bluesky and Mastodon
Magnitude, a YC S25 startup, releases an open-source inference engine that optimizes itself for specific hardware, making open models run up to two times faster than llama.cpp. ·
Magnitude, a YC S25 startup, releases an open-source inference engine that optimizes itself for specific hardware, making open models run up to two times faster than llama.cpp. ·
Data center operators increasingly use on-site power generation to bypass grid constraints, creating new environmental compliance challenges and bringing pollution risks closer to local communities. ·
Data Center Knowledge11h ago▲datacenterknowledge.com
Data center operators increasingly use on-site power generation to bypass grid constraints, creating new environmental compliance challenges and bringing pollution risks closer to local communities. ·
Data Center Knowledge11h ago▲datacenterknowledge.com