Posts
All the articles I've posted.
- 7.0
Beyond Linear Probes: Dynamic Safety Monitoring for Language Models
动态安全监控：根据输入难度灵活调整计算成本的LLM激活监测方法
- 6.0
PREF-XAI: Preference-Based Personalized Rule Explanations of Black-Box Machine Learning Models
PREF-XAI：基于用户偏好/目标/认知约束的个性化规则解释方法
- 7.0
Calibrating Scientific Foundation Models with Inference-Time Stochastic Attention
Stochastic Attention：推理时随机化attention权重为科学基础模型提供校准预测不确定性的轻量级修改
- 8.0
Unsupervised Confidence Calibration for Reasoning LLMs from a Single Generation
仅需单次生成即可实现推理LLM无监督置信度校准的新方法，无需标签或重复采样
- 7.0
TEMPO: Scaling Test-time Training for Large Reasoning Models
TEMPO：扩展大推理模型的test-time training能力，解决TTT方法快速plateau和奖励漂移问题
- 7.0
SAW-INT4: System-Aware 4-Bit KV-Cache Quantization for Real-World LLM Serving
SAW-INT4：考虑真实LLM serving约束（paged memory、正则访存、融合attention）的4-bit KV缓存量化方案
- 8.0
LLMs Know They're Wrong and Agree Anyway: The Shared Sycophancy-Lying Circuit
跨12个模型发现同一小组attention head携带'此陈述错误'信号——沉默这些head即翻转谄媚行为，揭示sycophancy与lying共享神经回路
- 6.0
Guiding Distribution Matching Distillation with Gradient-Based Reinforcement Learning
将梯度RL融入Distribution Matching Distillation(DMD)，解决样本评分与蒸馏轨迹冲突问题
- 6.0
Rethinking Dataset Distillation: Hard Truths about Soft Labels
揭示soft labels导致大规模dataset distillation方法退化为random baseline的根本原因，对比coreset的hard label优势
- 6.0
Efficient Mixture-of-Experts LLM Inference with Apple Silicon NPUs
解决MoE LLM在Apple Neural Engine上推理的三大核心挑战：动态张量形状、不规则算子和显存碎片
- 6.5
- 7.3
Changes to GitHub Copilot Individual plans
· 04/22 18:32 采集 - 8.5
- 6.0
- 6.0
[Trending] thunderbird / thunderbolt
Thunderbird 推出 AI 客户端 Thunderbolt：用户自选模型、拥有数据、消除供应商锁定。
- 5.5
[Trending] sansan0 / TrendRadar
AI 驱动的舆情与趋势监控工具，支持多平台聚合、RSS 订阅和智能告警。
- 6.0
[Trending] koala73 / worldmonitor
实时全球情报仪表盘：AI 新闻聚合、地缘政治监测、基础设施追踪的统一态势感知界面。
- 7.0
[Trending] langfuse / langfuse
开源 LLM 工程平台：可观测性、评估、提示管理、Playground、数据集，集成 OpenTelemetry/LangChain/OpenAI SDK 等。
- 6.5
[Trending] KeygraphHQ / shannon
自主式白盒 AI 渗透测试工具：分析源代码识别攻击向量并执行真实利用验证漏洞。
- 5.5
[Trending] open-metadata / OpenMetadata
统一元数据平台：数据发现、可观测性、数据治理，中央元数据仓库 + 列级血缘。