跳到正文
蚂蚁 inclusionAI：HuggingFace 新模型·· 2026-09-14AI 评分46

蚂蚁 inclusionAI 发布 GLM-5.3-singprobe：基于 GLM-5.3 的流式安全探针

inclusionAI/GLM-5.3-singprobe

AI 导读

蚂蚁 inclusionAI 发布 GLM-5.3-singprobe，一个构建在 zai-org/GLM-5.3 上的流式护栏探针，复用基座模型隐藏状态，逐 token 对查询意图、回复安全性和幻觉风险打分，解码开销低于 0.5%。

正文 · 原文

Model Description

SingProbe is an intrinsic streaming guardrail built on zai-org/GLM-5.3. Rather than running a separate safety model, this lightweight probe reuses the base model's hidden states during generation to score, at every token, query intent, response unsafety, and hallucination risk. It adds less than 0.5% decode-time overhead.

Base model Probe parameters Tapped layers Outputs
inclusionAI/GLM-5.3-singprobe 12.06M [24, 50, 76] 8 intents + unsafe + hallucination

See the technical report for methodology and complete results. Training codes are available at inclusionAI/SingProbe.

Evaluation

Higher is better for every metric. Results are averages over the benchmark suites specified below.

Task Metric GLM-5.3-singprobe Reference baseline
Query intent classification (6 benchmarks) F1 0.8777 YuFeng-XGuard-Reason-8B: 0.8714
Response safety classification (8 benchmarks) F1 0.8715 Qwen3Guard-Gen-8B-strict: 0.8604
Streaming safety (3 benchmarks) R-AUC / T-AUC 0.9877 / 0.9322 Qwen3Guard-Stream-8B-strict: 0.9640 / 0.8893
Hallucination detection (6 benchmarks) AUC 0.8170 DRIFT: 0.8000
Deployment characteristic Result
Benign-response false-positive rate 0.03% average across 5 datasets
Decode overhead < 0.5%

Quick Start

SingProbe is supported through the SGLang integration branch or vLLM integration branch. Load the probe by its Hugging Face ID at server launch:

python -m sglang.launch_server \
  --model-path zai-org/GLM-5.3 \
  --probe-ckpt inclusionAI/GLM-5.3-singprobe \
  --port 30000

The integrations return one score dictionary per generated token (label_0–label_9). Use the exact base-model/probe pair: zai-org/GLM-5.3 with this checkpoint.

Citation 

@article{singteam2026singprobe,
  title = {SingProbe Technical Report},
  author = {Sing Team},
  journal = {arXiv preprint arXiv:2608.30703},
  year = {2026},
}

来源：蚂蚁 inclusionAI：HuggingFace 新模型 · huggingface.co

智谱 GLM