Meituan has open-sourced its 1.6 trillion-parameter MoE model LongCat-2.0, trained using a peak of 50,000 domestic AI chips, completely脱离NVIDIA GPUs. Leveraging the ScMoE architecture, communication latency is embedded into computation, reducing inference latency by approximately 50%. A dynamic activation mechanism allocates compute based on token complexity—activating 33B experts for simple tokens and 56B experts for complex ones. The model achieves a steady-state daily throughput of 1T tokens and compresses memory usage below 60GB. Weights are released under the MIT license, with Agent benchmark performance matching Gemini 3.1 Pro, ranking second globally. The team also invested 135B parameters in N-gram Embedding, arguing that MoE sparsity has passed its sweet spot and stacking more experts offers diminishing returns.

NYU and LeCun's team propose AdaJEPA, a framework enabling world models to continuously adapt during deployment via test-time self-adaptation. Traditional frozen models suffer prediction inaccuracies under environmental distribution shifts, where MPC single-step errors recursively amplify, leading to planning failure. AdaJEPA establishes a closed loop of plan-execute-observe-update-replan, lightly updating the encoder and last few layers of the predictor using real observations as self-supervised signals after each action. This requires only one gradient step, adding just 0.01–0.03 seconds of latency. On PushObj and PointMaze benchmarks, planning success rates for unseen object shapes and maze layouts increase from ~50% to 70–80%, nearly doubling in out-of-distribution scenarios.

Due to the implementation of the 「Interim Measures for the Management of Anthropomorphic AI Interaction Services」, Doubao and Tongyi Qwen will shut down their agent functionalities on July 15, setting red lines for virtual companionship and minor protection. Concurrently, major tech firms are restricting external AI models internally: Alibaba bans Claude Code and promotes its in-house Qoder; Meituan limits Doubao Qwen in favor of self-developed LongCat; Meta restricts engineers from using Claude and Codex due to model distillation risks. Competition among domestic large models intensifies: Kling AI raises about $2.795 billion, Kimi reaches a valuation of $31.5 billion with ARR exceeding $300 million, and DeepSeek V4 introduces peak-off-peak pricing, accelerating commercialization.

NVIDIA GEAR and Li Fei-Fei's team introduce SimFoundry, a system that automatically generates interactive robotic simulation environments from just one real-world video. Through depth estimation, VLM-based segmentation, and 2D-to-3D generation, it reconstructs digital twins, significantly reducing manual modeling costs. Its core innovation, Digital Cousins, automatically alters object appearances, layouts, and tasks within the digital twin to generate massive new data while preserving functional semantics, effectively expanding a single video into an almost infinite data generation space. SimFoundry supports zero-shot sim-to-real transfer—the policy trained in simulation can be directly deployed on real robots, and simulation evaluations highly correlate with real-world performance, replacing costly physical testing.

After slow progress in developing proprietary AI models, Meta plans to monetize its vast AI infrastructure by opening access to external customers and selling GPU compute. While the Llama series is open-sourced, direct commercialization remains challenging; Muse Spark and Watermelon have failed to gain leadership, yet capital expenditures have exceeded $100 billion, pushing Meta to treat GPUs as priceable leased assets. In H1 2026, Meta signed contracts for over 5GW capacity, with 2.5GW of data center campuses under construction, and nearly 10GW in related deals since early 2024. Meta is negotiating with Anthropic to obtain private instances of Claude, potentially offering Claude-as-a-service externally, similar to Amazon Bedrock. Following the announcement, Meta’s stock rose nearly 9%.

Huawei released the second edition of its Tao's Law paper, adding extensive engineering details, empirical data, and product roadmaps. Tao's Law uses response time constant τ instead of transistor density as the core metric, composed of four hierarchical levels from transistor to system spanning 12 orders of magnitude; compressing τ becomes the new path to chip performance enhancement in the post-Moore era. LogicFolding 3D stacking technology vertically stacks circuits to shorten interconnects—on next-gen mobile chips at the same process node, transistor density increases from 155 million per mm² to 238 million, equivalent to a three-year process advancement. The AI data center trifecta (unified bus, Hi-ONE, 3D folding) targets data movement bottlenecks, reducing cross-node latency from microsecond-level to around 100 nanoseconds. Huawei acknowledges for the first time that thermal dissipation remains an unsolved challenge.

Citing a new ICML 2026 study, Stanford AI Lab highlights that even after deduplication, residual repeated data in language model pretraining may still waste up to 33% of compute. The research further finds that worst-case repeating structures can be predicted from model scale, providing theoretical grounding for optimizing pretraining data pipelines. This reveals limitations in current deduplication methods and suggests the industry needs finer-grained data processing strategies to improve compute efficiency.

An MIT paper proves from a statistical identifiability perspective that no matter how much pairwise comparison data is collected in RLHF, it cannot learn the correlation structure among preferences—the covariance matrix is lost during data collection, creating a statistical information bottleneck. The study proposes Best-of-Three, which observes full ranking structures (six permutations) from three-way comparisons containing covariance information, enabling complete recovery of correlated preference models. It constructs a polynomial-time estimator, experimentally outperforming traditional methods. The authors link this work to OpenAI's hallucination paper, noting that binary preference feedback loses dimensions such as caution, honesty, and risk expression, sharing a common root with hallucination issues.

Lin Junyang, former technical lead of Qwen, argues that hybrid reasoning models are inherently unstable due to conflicting optimization objectives—instruction models reward concise, low-latency outputs, while reasoning models reward spending more tokens on difficult problems. Merging them hastily harms both. He believes the field is shifting from reasoning-centric to agent-centric thinking: the former focuses on internal deliberation before a single answer, judged by verifiable rewards; the latter defines intelligence through closed-loop interaction with the environment, measuring success by continuous progress. Agent-based reinforcement learning introduces an infrastructure bottleneck—training and inference must be decoupled—where agents waiting for real-time tool servers cause inference stalls. He emphasizes that reward hacking is a critical risk, as tool access greatly expands the attack surface, making optimizing environment quality as crucial as optimizing data quality in the Agent era.

An ACL 2026 paper introduces T*, a progressive block expansion curriculum that solves the problem of deteriorating inference quality in block-based diffusion language models as block size increases. Larger blocks offer better parallelism but introduce more undetermined tokens and weaker conditional signals, causing training collapse in reinforcement learning stages. T* starts training with smaller blocks (e.g., B=4) to stabilize denoising trajectories, then gradually increases block size. Experiments show a 4B model improves from 60.73% to 76.00% on MATH500, with gains also seen on GSM8K and AIME24, maintaining advantages even at large block sizes like B=32. The core of T* is not adding complex model modules, but reordering the difficulty sequence in reinforcement learning, without sacrificing parallelism for accuracy.