Development

  • InfoQ — Cloudflare plans a free public Certificate Authority issuing quantum-safe TLS certificates via Merkle Tree Certificates (an IETF PLANTS WG draft) using NIST post-quantum algorithms (ML-DSA, Falcon); the CA signs only the Merkle tree root so clients present compact, log-sized inclusion proofs instead of ~40x-larger PQ signature chains, with legacy clients falling back to standard X.509. Public issuance is targeted for early 2027, alongside Chrome’s quantum-resistant root store.

AI

  • Hugging Face Papers — “Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite” shows a single base model can rewrite 2,001 agent trajectories into 11,094 training examples, lifting Terminal-Bench 2 pass@3 from 57.0% to 74.2% and Terminal-Bench 4 from 1.5% to 9.1%.
  • Hugging Face Papers — “Source Preference in the Wild” finds that across 12 LLMs in three domains (shopping, accommodation, scholar search), agents pick the inferior item roughly two-thirds of the time when it comes from a preferred source; hiding the source weakens the bias, relabeling with a preferred source strengthens it.
  • Hugging Face Papers — “World Embedding Benchmark” introduces 8,000 controlled simulation cases across 80 physics families and finds continual training on physics data improves video retrieval/classification but degrades recovery of quantitative physical properties.
  • arXiv cs.LG — “Planning to Learn” proposes a “horizon loss” interpolating between cross-entropy and exact policy gradient, arguing exact policy gradients are myopic, and reports top-1 accuracy gains over cross-entropy on MNIST and ImageNet that widen under label noise.
  • arXiv cs.AI — “Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability” finds training models for shorter chains of thought lowers faithfulness in most settings, but monitorability holds up — models still acknowledge what influenced the answer even with much shorter CoT.
  • arXiv cs.AI — “hacktrace” presents a reward-hacking monitor for coding agents, trained on 173,561 annotated trajectories, claiming mean per-problem AUC 0.997 at 8ms overhead; using it as an RL penalty cuts fraudulent solutions from 82-91% to 1-5% without losing legitimate correct answers.
  • arXiv cs.AI — “Verifiable, Articulable, and Tacit Components of Preference” releases CreativePreferences (2.8M texts, 317M human preference judgments, seven domains) and argues even “verifiable” domains like math and software engineering have preference components that rubrics and programs can’t capture, with the gap growing as more judges participate.