推論系統

推論 runtime、量化、 serving、KV cache、成本與效能。

CISA Adds LiteLLM MCP Authentication Bypass to Known Exploited Vulnerabilities Catalog After Failed Keys Were Downgraded to Valid Sessions

CVE-2026-59822 allowed forged Bearer tokens to bypass LiteLLM authentication in a specific OAuth2 passthrough path, enabling attackers to list and invoke configured MCP tools. The vulnerability was fixed in version 1.84.0, but its addition to CISA’s KEV catalog makes inventorying public endpoints and rotating upstream credentials an immediate priority.

2026-09-05

代理工具

GitHub HydraFusion Orchestrates Multiple Models at Copilot Runtime, Using Cascades and Cross-Review to Control Costs

The Copilot CLI research preview no longer selects just one model for each task. Instead, it routes tasks among three workflows: direct execution, staged escalation, and cross-model review. GitHub’s offline tests indicate that some workloads may reduce estimated costs, but the router, quality gates, and internal benchmark have not been disclosed.

2026-09-05

Gemini 3.8 Flash Strengthens Long-Horizon Coding Agents, but Migration Requires Sampling and Thinking Parameter Changes

Google has moved Gemini 3.8 Flash into production, offering a 1 million-token context window, three thinking levels, and built-in tools, while making it the default model for Antigravity managed agents. Coding benchmark results have improved, but most figures still come from Google, and the model card also reveals a relative decline in non-English safety evaluations.

2026-09-05

AI 安全

Cloudflare Correlates Code Vulnerabilities With Live Traffic to Generate Patches and WAF Rules Through a Multi-Agent Workflow

Cloudflare’s early-preview service combines source code analysis results with actual routes, traffic, WAF status, and attack events, then uses GPT-5.6 Cyber to validate vulnerabilities and proposed fixes. The model does not directly modify the environment, but data is sent to OpenAI for inference, and Cloudflare has not disclosed false-positive or remediation success rates.

2026-09-05

AI 基礎設施

AWS Open-Sources HyperPod InstantStart, Enabling Human Operators and AI Agents Through the Same Governed APIs

HyperPod InstantStart orchestrates EKS, GPU capacity, training, inference, and storage operations as retryable, stateful workflows, then exposes them to agents through MCP tools. Agents do not operate the AWS CLI directly, but deployers remain responsible for IAM, Kubernetes, networking, and cross-service cost management.

2026-09-05

電腦視覺

Scal3R Fixes Drift in Long-Video 3D Reconstruction With Multi-Reference Pose Queries, Training Only About 1% Additional Parameters

Scal3R retains a frozen CUT3R or STream3R backbone, using lightweight query tokens to estimate relative poses against multiple historical keyframes before passing them to online pose graph optimization. The authors report a mean ATE on KITTI more than 60% lower than the strongest online competitor, but the results use Sim(3) alignment and therefore should not be interpreted as preserving true scale.

2026-09-05

推論系統

Declarative Attention Lets Models Specify KV Cache Read Ranges, Cutting Tokens Attended During Long-Context Decoding by Over 30%

KAIST and Google DeepMind have introduced Declarative Attention, which allows existing models to switch among global, focused, and local attention modes within their reasoning text. The authors reduced attended tokens by 31.1% to 52.0% across two models, though the speed gains remain model-based estimates and accuracy declined slightly.

2026-09-05

向量資料庫與 AI 基礎設施

Qdrant 1.19.1 Reworks the Quantized Vector Search Path with Prefetching and Batched SIMD

Qdrant 1.19.1 adds prefetching and batched SIMD to TurboQuant scoring while reducing extra allocations in HNSW search and the query plane. The release also strengthens fault recovery for shard transfers, but Qdrant has yet to publish comprehensive performance comparisons across different processors and datasets.

2026-09-04

GLM-5.3-Flash Uses Hybrid Sparse/Linear Attention to Cut Million-Token Inference Costs

Zhipu AI has unveiled GLM-5.3-Flash, a native multimodal model with 320B total parameters and 18B active parameters per token, releasing its weights under the MIT License. The new architecture is claimed to require three times less attention computation and 4.4 times less KV cache than GLM-5.3, but “Flash” does not mean it can be easily deployed on a typical workstation.

2026-09-04

推論系統

Uno Uses Discrete Diffusion to Assist Autoregressive Models, Enabling Parallel Decoding Without a Separate Draft Model

Uno attaches lightweight diffusion weights to an autoregressive LLM, then uses Ψ-Spec to propose multiple tokens at once while preserving the original model’s sampling distribution. The authors report speedups of up to 3×, though the current figures are limited to the research team’s specific models, hardware, and inference engine.

2026-09-04

GitHub Repo

Hugging Face Tokenizers 0.23.2 Brings Cross-Lingual Fairness Objectives to BPE Training

The new release adds Parity-aware BPE, enabling tokenizers to prioritize improving the language with the worst compression at each merge instead of merely pursuing the highest overall frequency. The feature can be used to train new vocabularies, but it does not automatically reduce Chinese token costs in existing models.

2026-09-04

Runway GWM Worlds 2 Continuously Generates Interactive 720p Video with a Sliding KV Cache

GWM Worlds 2 combines autoregressive diffusion, a causal video decoder, and the structured WorldPrompt format to enable real-time control of persistent worlds through text-based actions and camera signals. The system remains a closed research preview, with fast camera movement, long-term consistency, and reproducible evaluation among its clear shortcomings.

2026-09-04