Anthropic Frontier Red Team 在内部 Binary Exploitation 基准的 100 个随机任务上评测多个模型，GLM-5.3 在 4% 的试验中实现了完整控制流劫持，Claude Mythos Preview 为 6%。报告指出，Claude Opus 4.6 和 GLM-5.2 等更早的模型在这些任务中均未成功，认为一个有意义的能力门槛已被跨过。

OpenAI 公布前沿 AI 训练安全案例的早期指南，涵盖技术防护措施、运营实践以及失准事件调查三方面。该指南旨在为前沿 AI 训练建立安全论证框架。

Google DeepMind 宣布与游戏开发商合作，用 AI 原型化全新游戏体验，并回顾了从 DQN 玩 49 款 Atari 游戏、AlphaGo、AlphaZero、MuZero、AlphaStar 到 SIMA 的游戏 AI 研究历程。

Google DeepMind published an AI Control Roadmap, a framework for building and managing advanced AI deployed inside Google. It takes a defense-in-depth approach, adding system-level safety layers on top of model alignment so protections hold even when alignment is imperfect.

A Reddit user fine-tuned Qwen2.5-7B with 1,040 DV-DPO preference pairs, costing about $3 at Claude Haiku rates, and reported 96% composite performance versus Claude Haiku on a domain task, with 11-second latency on a 4-bit T4 setup.

An ICML 2026 paper presents an ISR=1 answer-abstain gate for evidence-grounded QA, and ntkMirror implements it for local open-weight models with multiple evidence orderings, reporting 0.0–0.7% hallucination at about 24% abstention in the held-out audit.

OpenBMB released the VoxCPM2 technical report, covering a 2B-parameter speech generation model trained on more than 2 million hours of multilingual speech data, with support for 30 languages and 9 Chinese dialects.

The experiment used five models from OpenAI, NVIDIA, OpenBMB, and a self-fine-tuned 500M-parameter model to drive market agents; three interventions failed to reproduce the price crash, and the crash was created only by overriding prices during settlement.

Google DeepMind published results from a pre-registered randomized controlled trial in Sierra Leone. Students using Guided Learning gained 0.258 standard deviations in math over the control group, equal to roughly 1.2 to 1.7 years of normal learning progress in eight weeks.

openJiuwen proposed MANGO, a multi-agent flow-network framework that combines reinforcement learning, textual gradients, and a Skip-k mechanism; using GPT-4o-mini, it reports a 12.8% accuracy gain over MaAS on MATH500 and a 5.1% F1 gain over AFlow on DROP.

Alibaba’s RTP team released RTPurboV2, replacing 85% of attention heads with SWA and compressing the remaining 15% retrieval heads using low-rank projection, clustering, and dynamic top-p; the adaptation uses about 600 training steps and roughly 1M label tokens, with reported Prefill speedup up to 9.36x.

UIUC and Chroma released Harness-1, a 20B-parameter retrieval subagent trained with reinforcement learning inside a stateful search harness, reporting 0.730 average curated recall across 8 benchmarks, 11.4 percentage points above the next-best open-source subagent and behind only Opus-4.6.

FusionRoute proposes a token-level multi-LLM collaboration method that freezes expert models and trains a lightweight router to select an expert for each token while merging router logits with expert logits. The paper evaluates it on GSM8K, MATH-500, HumanEval, MBPP, IfEval, and 500 PerfectBlend prompts.

UC Berkeley and Princeton researchers introduced Log-ICoT for k-parity, reducing training stages from 15 to 4 when k=16, and proved that an L-layer Transformer can internalize chain-of-thought with log₂k curriculum stages under simplified assumptions.

City University and Kuaishou Kling proposed VLM-as-Teacher, which uses VLM feedback to optimize VGM LoRA at test time, reporting a 16.7-point average gain and raising VBVR-Bench from 0.666 to 0.781.

Thousand Token Wood v2 uses four small models from different labs to drive agents in a financial simulation game, with vLLM 0.22.1’s CUDA toolkit dependency identified as the main serving friction, while a fine-tuned 0.5B Qwen reached 0% self-trading and 100% valid quotes.

Domino reports up to 5.8x throughput speedup on Qwen3 by decoupling causal modeling from autoregressive drafting in speculative decoding. The Reddit snippet links the arXiv paper, GitHub code, and Hugging Face models, but does not disclose hardware, baseline settings, dataset, or acceptance-rate details.

PhysX-Omni models rigid, deformable, and articulated objects in one simulation-ready 3D generation framework, while PhysXVerse contains over 8.7K physical 3D assets across more than 2.9K categories.

Princeton researchers released Goedel-Architect, an agent framework for Lean formal theorem proving. Using DeepSeek-V4-Flash, it reached 75.6% pass@1 on PutnamBench, with $294 in API cost for 672 problems, compared with Hilbert’s 70.0% and about $170,000 cost.

A developer used Qwen2.5-3B to build a five-agent forest economy, and across 15 simulation rounds honey prices fell from 10 to 3, firewood rose from 4 to 7, and the Gini coefficient increased from 0.14 to 0.38.

Google Research and Google Cloud introduced the Cross-Corpus Retrieval framework as Agentic RAG for Gemini Enterprise Agent Platform, using a multi-agent workflow to plan, rewrite, route, and iteratively search multiple data sources, with up to 34% higher accuracy than standard RAG on factual datasets.

Southeast University and Peking University researchers introduced MetaFine, a diagnostic meta-evaluation framework that tests fine-grained robot manipulation across understanding, perception, and behavior, and the article says traditional binary success metrics can overestimate fine-manipulation capability by up to 70%.

CMU and the University of Maryland propose Language Models Need Sleep: when each L-token context window fills, the model runs N offline recurrent forward passes and updates SSM fast weights before evicting the KV cache. On GSM-Infinite, Jet-Nemotron 2B with 6 sleep loops improves 6-step arithmetic accuracy from 0.742 to 0.812.

Ali Kayyam and coauthors evaluate three QKV projection-sharing variants across synthetic, vision, and language-modeling settings, including 300M and 1.2B parameter models trained on 10B tokens; Q-K=V halves the KV cache with a 3.1% perplexity degradation, while Q-K=V plus MQA reduces cache use by 96.9%.

Anthropic published a post on recursive self-improvement under the title “When AI Builds Itself,” while the RSS body only discloses 95 Hacker News points and 106 comments, with no experimental setup, model details, or timeline disclosed.

Together AI, the University of Sydney, and UIUC introduced OSCAR, a 2-bit KV Cache quantization method that uses about 2.28 effective bits per KV element and scores 71.86 on Qwen3-4B-Thinking, 40.1 points above TurboQuant.

Xpeng presented its world-model stack at CVPR 2026, covering X-World, X-Foresight, and X-Cache; the article says X-Cache cuts about 70% of repeated computation, the second-generation VLA used over 4 trillion training tokens, and the in-car stack reduced inference latency to 80 ms.

NVIDIA Research presented three physical AI papers at CVPR: GraspGen-X was trained on 2 billion simulated grasps, LCDrive cuts reasoning tokens by about half versus text-based reasoning, and NitroGen trains embodied agents across more than 1,000 games and 40,000 hours of interaction.

Junpeng Zhang and coauthors argue that SFT on highly homogeneous data has an effective window of only hundreds to about 1,000 training steps, and their interaction-based warning signal detects overfitting before loss gaps appear, saving roughly 30%–50% of training compute.

Ant Lingbo and HKUST introduced LingBot-VA, an autoregressive video-action world model that unifies visual dynamics prediction and action inference, and the paper reports fine-tuning with 50 real-world demonstrations per task plus 92.0% and 91.1% success on RoboTwin 2.0 Easy and Hard settings.

Daxiao Robot and NTU introduced PhysX-Omni, a unified simulation-ready physical 3D generation framework for rigid, deformable, and articulated objects, with PhysXVerse covering 8.7K assets across 2.9K categories and PhysX-Bench evaluating six dimensions including geometry, scale, material, affordance, kinematics, and description.

Microsoft AI says MAI-Thinking-1 uses three mechanisms—thermostat, circuit breaker, and self-distillation—to keep RL training stable for several thousand steps; the RSS snippet contrasts MAI’s discipline with DeepSeek’s efficiency and GLM’s endurance.

MAI filtered 265,000 trainable tasks from 4.87 million open-source PRs and built a three-layer judging system. The key change after vibe coding is the industrialization of training infrastructure.

Microsoft released MAI-Thinking-1, an MoE model with 35B active parameters and 1T total parameters, pretrained from scratch on 30T tokens without third-party model distillation.

DataMaster searches, cleans, and combines data while keeping the model and training algorithm fixed; on MLE-Bench Lite, it raised the medal rate from 35.91% to 68.18%.

Banafsheh Rafiee and Richard S. Sutton propose an enactive cognition framework for AI, naming four pillars: experience, perception-action inseparability, autonomy, and embodiment.

Google DeepMind, Anthropic, and Meta are studying whether AI can become conscious and the human implications, but the post does not disclose methods, timelines, or evaluation criteria.

JetBrains published a Hugging Face blog post introducing Mellum2, confirming a mixture-of-experts architecture and a 12B parameter scale; the snippet does not disclose training data, license, benchmarks, or deployment conditions.

Import AI 459 summarizes papers on AI-economy measurement and AI oversight: one estimates U.S. nominal AI GDP at about $250 billion in 2025, with quality-adjusted real growth near 2,600% per year.

The author trained Mamba-Titan-1.4B-Reasoning on a 12GB RTX 3060: a frozen 1.4B Mamba-1 backbone with 8 trainable MoE arms, 2.54B total parameters, Top-2 routing at layers 24/25, and about 50% math accuracy.