Unemployment Labs, by analyzing the random number initialization layout in public model weights, discovered DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3.
Why it matters: By reversing the random number layout in the model weights to train the hardware, it provides a reproducible technical evidence path for export control disputes.
MarkTechPost3838MoE routing strategies determine which experts process each token, affecting model quality, load balancing, and deployment costs. Mixtral uses Softmax Top-K to select 2-8 experts, while DeepSeek-V4-Pro selects 1-8 out of 384. Load balancing is achieved through Sigmoid gating and auxiliary loss, with Sigmoid scoring combined with lossless bias balancing being a common approach for cutting-edge open models in 2026.
The Decoder7575Vals AI's CyberOffense benchmark allows AI agents to autonomously attack simulated enterprise networks on isolated Kali Linux machines, GPT-6 Sol Daybreak in OpenAI restricted projects completes 80.8% of the target, open-source weighted model Deepseek V4.1 Flash 1017 reaches 72.8%, with a single run cost of about 3 dollars instead of about 22 dollars
The DecoderFeatured7878A study made Claude Opus 4.8, Claude Opus 5, Claude Fable 5, Claude Opus 5.5, GPT-5.6 Sol, GPT-6 Astra.
Why it matters: This benchmark uses 15 unpublished research questions to test 9 agents and provides common failure patterns in 135 submissions, which can be used as a reference for evaluating autonomous research capabilities.
vLLM BlogFeatured7272NVIDIA Research and Inferact's REINFORCE Zero proposes to first align rollout with training kernel, then use unclipped token-level importance sampling to correct weight lag, achieving a training step throughput 2.40 times that of synchronous training under 16 GPUs and 12K updates.
Why it matters: The paper breaks down the gap between asynchronous RL's rollout and training into two items: weight lag and kernel inconsistency, and proposes an alignment-free, no-trimming importance sampling method after alignment.
The Decoder5858Stanford researchers tested 50 widely stated claims about LLMs, finding that 44% have expired, with an average lifespan of only 15.7 months. Claims related to reasoning abilities (such as counting, arithmetic, long context information loss) have all become invalid, while claims related to flattery, deception, and strategic behavior persist or even enhance.
Import AI4242Realtime-Bench benchmark tests evaluated the performance of LLMs on 50 real-time interactive games and web tasks, with GPT-6 Astra achieving a 36% pass rate under real-time clock conditions; this test was developed by institutions such as the University of Texas at Austin to measure the model's real-time state awareness, adaptive reasoning, and time-aware action selection capabilities.
The DeepSeek Harness team published 'Super-agent design exploration: Towards more autonomous Agent systems' on the official WeChat account, proposing a dual-layer architecture, event-driven, asynchronous multi-channel interaction, separation of computation and storage, and an external Watchdog to break through the limitations of existing Agents in multi-tasking and team collaboration.
DeepSeek has open-sourced the Super-agent prototype, disclosing the dual architecture concept for Personal Agent, where the upper-layer Router handles message reception and task coordination, and the lower-layer Worker executes tasks in parallel.
Why it matters: DeepSeek has publicly disclosed the dual architecture of Personal Agent and actual test data, showing how multi-session collaboration can bypass human attention bottlenecks.
The DeepSeek Harness team released an article titled 'Super-agent Design Exploration,' proposing a dual-layer architecture where the Router layer uniformly handles human interactions and multiple Worker sessions are isolated and executed concurrently, summarizing four characteristics: continuous, efficient, proactive, and persistent.
Prime Intellect Blog3838Prime Mega MoE enables the DeepSeek Mega MoE kernel to be used for training by integrating backpropagation kernels, achieving up to 3.11 times faster MoE kernel speed compared to the default PyTorch implementation, and 1.34 times faster end-to-end training speed compared to the Prime-RL baseline.
Prime Intellect BlogDeepSeek releases SYNTHETIC-2, which includes 20 complex reinforcement learning tasks and verification reasoning trajectories, supporting heterogeneous computing from consumer-grade GPUs to super-large-scale clusters, using the DeepSeek-R1-0528 model to generate data and label difficulty.
Dify BlogDeepSeek R1 performed excellently in the AIME Mathematics Competition, with accuracy increasing from 15.6% to 71.0%, and reached a human level of 96.3% in the Codeforces Programming Competition, demonstrating strong reasoning ability.
