On February 14, ByteDance released Doubao-Seed 2.0, targeting Agent and large-scale production scenarios, enhancing efficient inference, multimodal understanding, and complex instruction execution. The release includes four versions: Pro, Lite, Mini, and Code. The Pro version emphasizes deep reasoning and long-chain tasks, achieving an HLE-text score of 54.2, and demonstrates leading performance on scientific benchmarks such as SuperGPQA and HealthBench, as well as competition metrics like IMO and ICPC. Officially, the token price is approximately one order of magnitude lower than top industry models. It has been launched on the Doubao app and is available for enterprise access via the Volcano Engine API.

Google and OpenAI have disclosed detection of 'distillation attacks' against their LLMs: adversaries programmatically collect model outputs using bulk prompts in an attempt to replicate reasoning logic. Google reported organized efforts using over 100,000 prompts to probe Gemini's non-English reasoning capabilities; related accounts have been suspended. OpenAI submitted a memo to the U.S. House Select Committee on the Chinese Communist Party, alleging that DeepSeek employees circumvented access restrictions and covertly scraped outputs via third-party routing for distillation. It called on the government to help close API routing vulnerabilities and restrict adversaries' access to U.S. cloud computing resources.

Tencent HunYuan has released GradLoc, a method for pinpointing gradient anomalies in reinforcement learning training processes like RLVR. The team traces global gradient spikes back to specific anomalous tokens in distributed training, using binary search and DFS-based greedy algorithms to identify primary contributors in logarithmic complexity. Based on micro-level observations, the paper identifies three key anomaly patterns: token-level train-inference inconsistency, sequence-level inconsistency, and inter-layer gradient heterogeneity. It proposes mitigation strategies including TokenClip, SeqClip, and LayerClip, aiming to reduce large model RL training debugging time from 'weekly' to 'hourly' and lower engineering barriers.

Ohio State University and Amazon have released MMDR-Bench, establishing a more verifiable evaluation framework for 'Multimodal Deep Research' agents. The benchmark shifts focus from 'whether the answer is correct' to 'whether the process is auditable': TRACE verifies consistency between claims and URL citations, while MOSAIC validates alignment between image evidence and textual statements sentence-by-sentence, heavily penalizing citation hallucinations and visual misinterpretations (e.g., entity recognition errors, incorrect number reading). Experiments show clear divergence among models in writing fluency versus evidential support capability, offering a reusable metric for evaluating, iterating, and deploying long-chain research agents.

ByteDance's technical team has introduced the 'Agent Bucket' storage paradigm, addressing permission control, billing difficulties, and neighbor noise issues caused by the 'single bucket with multiple prefixes' object storage model in massive multi-tenant Agent scenarios. The solution introduces an intermediate ObjectSet layer between buckets and objects, enabling native tenant awareness and supporting per-user rate limiting, quotas, independent domains, and metering. Set Slice partitions metadata by range to achieve linear scalability with 'logical unity, physical separation' and resource isolation; Set AccessPoint, combined with high-entropy domain names and STS temporary credentials, reduces lateral risk propagation after credential leaks.

ModelBest introduces the SALA hybrid attention architecture, using linear attention for local computation in most layers and sparse attention for global retrieval, striking a balance between speed and accuracy for million-token context inference. The article claims stable 1M-context inference on RTX 5090 at approximately 3.5 times the speed of Qwen3-8B. The accompanying HALO training paradigm converts existing full-attention models through layer transformation and continued training, reportedly reducing training costs by about 75%. HyPE hybrid positional encoding balances short-context performance with long-range information transmission, mitigating long-range signal decay.

A domestic team has used the reinforcement learning system PackingStar to set new records in the kissing number problem across multiple dimensions, including 25–31, achieving the best-known results historically. The method transforms high-dimensional spherical packing search into a multi-agent game of cosine matrix filling: the filling agent generates candidate structures, while the pruning agent performs geometric analysis and pruning, leveraging GPU parallelism to reduce search overhead. Reports indicate that AI-discovered configurations are mostly asymmetric, challenging previous research intuitions favoring symmetric constructions, and demonstrate a sustainable, engineering-driven iterative path in mathematical exploration where data and ground truths are lacking.

Following its Hong Kong listing, Zhipu AI is advancing its 'A+H' strategy. According to the China Securities Regulatory Commission's website, its STAR Market IPO filing was updated on February 13, adding Guotai Haigong Securities as a co-counseling institution alongside CICC. Reports suggest that due to the technical and operational complexity of AI firms—many of which are not yet profitable—and increasingly stringent regulatory scrutiny, underwriters must balance filing speed with documentation quality. This development reflects how leading large model companies are accelerating entry into A-share financing and compliance procedures to secure long-term capital for future R&D and infrastructure investment.

Researchers propose Spider-Sense, an agent protection framework aiming to balance security inspection with execution efficiency. Its IRS mechanism embeds risk awareness natively into model inference, eliminating additional checks for 99% of interactions deemed safe. HAS implements hierarchical screening: a vector database performs low-cost preliminary filtering, triggering LLM-based deep analysis only for suspicious samples. The authors report that compared to typical external protections causing 197%–381% extra latency, Spider-Sense reduces defense overhead to just 8.3%, covering critical stages including input, memory/planning, tool parameter auditing, and tool result verification.