🔬 AI Research Digest
📅 Week of Aug 19–Aug 25, 2026
━━━━━━━━━━━━━━━━━━━━━━━━

1. ⚡ FreeToken: Efficient Edge-Native MoE Serving
Authors/Org: FlashML / UC Berkeley | arXiv: 2608.16157
Bottleneck solved: Running large Mixture-of-Experts models locally on consumer hardware without cloud dependency.
FreeToken treats heterogeneous edge resources — GPU, CPU, host memory, and interconnects — as a unified elastic inference platform, enabling a 35B MoE model on a laptop and a 753B model on a single workstation GPU, delivering 2–4x faster speeds than Ollama for local MoE deployment.
🔗 FreeToken on arXiv

━━━━━━━━━━━━━━━━━━━━━━━━

2. 🧠 BDH-CQ: In-Context Learning with Recurrent Latent Reasoning
Authors/Org: Pathway | arXiv: 2608.09888
Bottleneck solved: Cost-efficiency of reasoning models — achieving strong ARC-AGI performance at just $0.0007 per task.
Instead of generating chain-of-thought text, a 150M-parameter model updates a recurrent memory from demonstrations and reasons iteratively in continuous latent space, reaching 29.5% pass@2 on ARC-AGI-1 at a fraction of the cost of much larger models — a new Pareto frontier for inference efficiency.
🔗 BDH-CQ on arXiv

━━━━━━━━━━━━━━━━━━━━━━━━

3. 🔭 ASI-Bench: At the Dawn of Artificial Superintelligence
Authors/Org: Junwei Zhou, Zhen Sun, Binyu Li et al. (Tsinghua University / Apexin AI) | arXiv: 2608.17271
Bottleneck solved: Lack of rigorous benchmarks for evaluating AI on open-ended, project-level scientific research without human methodological guidance.
Built by 40+ experts across 31,000+ human hours, ASI-Bench covers 60 end-to-end research tasks across 11 scientific domains — top agent configurations score just 26.62 when required to independently select methods, revealing how far current AI systems remain from autonomous science.
🔗 ASI-Bench on arXiv

━━━━━━━━━━━━━━━━━━━━━━━━
💡 Stay curious. Read the papers.

For More: @kdnuggets @datasciencechats