Scaling Laws for Sparse Mixture-of-Experts at Inference
New routing strategies that reshape how sparse MoE models scale when deployed at inference time.
Authors: J. Park et al. · Stanford / DeepMind
← Back to home
Papers & News
Curated research from top-tier venues and the reporting we actually read.
🔬 Featured Papers
Long-Horizon Reasoning via Verifier-Guided Search
Combining process reward models with Monte-Carlo search to dramatically improve multi-step reasoning accuracy.
Authors: L. Chen et al. · MIT CSAIL
Continual Pretraining with Minimal Forgetting
A low-rank replay mechanism that lets foundation models absorb new domains without losing prior knowledge.
Authors: Anthropic Research
Unified Video-Language Models for Embodied Agents
A single representation spanning vision, language, and action for long-horizon robot task planning.
Authors: UC Berkeley · Covariant
📰 Industry News
OpenAI launches next-gen GPT with native long context
The company claims a 40% reduction in inference cost compared to the previous generation; enterprise API preview opens today.
Source: The Information
Anthropic partners with sovereign funds on AI safety research
A joint $1B long-term program focused on alignment and interpretability research.
Source: Financial Times
DeepSeek tops multiple open benchmarks with fully open weights
On math and code tasks the new release rivals closed frontier models — the community response has been enthusiastic.
Source: Hugging Face Blog