Focused on frontier AI technologies, empowering AI researchers' progress.
Anthropic's new Claude Haiku 5.5 surpasses GPT-6 Luna and Chinese rivals on benchmarks while matching Luna's pricing and undercutting DeepSeek-V4.1 Flash, though tokenizer changes increase token usage and API breaking changes require migration effort.
OpenAI released 722 mathematical manuscripts from an unreleased internal model, claiming progress on millennium problems like the Riemann hypothesis, BSD conjecture, and Hodge conjecture, but Fields medalists and other mathematicians warn the volume exceeds human verification capacity and do not endorse the results.
Tsinghua researchers open-source VeriLoop E2, a post-trained model using verifiable recurrence to achieve 67.35% zero proportion lower bound for Riemann hypothesis, while setting new benchmarks in code, math, and physics reasoning with external verification.
Google DeepMind's Dream-RSI introduces a recursive self-improvement framework that treats historical exploration data as an offline simulator to evolve search strategies, reducing compute costs by 1-2 orders of magnitude across algorithm engineering, mathematical optimization, and GPU kernel generation tasks.
This paper introduces Diffusion Reward Models (DRM) that replace the scalar value head with a Diffusion Transformer to model full reward distributions, capturing multimodal human disagreement, enabling uncertainty-aware decisions, test-time scaling via repeated reward sampling, and improving downstream RLHF performance.
This article explains how to effectively respond to reviewer concerns about hyperparameter sensitivity by providing hyperparameter sweeps, justifying default values via validation sets or efficiency trade-offs, and avoiding test-set tuning, with concrete rebuttal templates and examples.
Jiajun Wu's IROS 2026 talk presents structured representations as scaffolding for robot reasoning, covering learning predicates and actions from demonstrations, differentiable planning with foundation models, post-training for unseen objects, and generating interaction sequences via video diffusion models without human demonstrations.
DeepSeek's DSec infrastructure supports millions of agent sandboxes through layered environments, on-demand image loading via 3FS, memory sharing with virtio-pmem/DAX, CPU scheduling, and trajectory forking, achieving 50x resource overcommit while addressing security challenges like agent-discovered vulnerabilities.
The article examines how embodied intelligence data competition is shifting from sheer volume to measurable capability gains, detailing Real2Sim2Real closed loops, world models, and commercial validation showing simulation success rates jumping from 9.7% to 79.8% and real-world task success improving over 50%.
Einsia AI's PPTBench evaluates coding agents on reconstructing 500 scientific flowcharts into editable PPTX slides using a three-stage Agentic Judge; GPT-6 Astra High scores 77.34 with 80.8% passing semantic and rendering checks, yet semantic understanding remains the primary bottleneck across all models.
