# 蚂蚁 inclusionAI 发布 Qwen3.5-0.8B-singprobe 流式安全探针

- 来源：蚂蚁 inclusionAI：HuggingFace 新模型
- 发布时间：2026-09-14 11:01
- AIHOT 分数：45
- AIHOT 链接：https://aihot.news/items/cmu0zy4y507itro2ny435r5o8
- 原文链接：https://huggingface.co/inclusionAI/Qwen3.5-0.8B-singprobe

## AI 摘要

蚂蚁 inclusionAI 发布基于 Qwen/Qwen3.5-0.8B 的流式安全探针 Qwen3.5-0.8B-singprobe，仅 2.23M 探针参数，复用基座模型隐藏状态，逐 token 输出 8 类意图、不安全与幻觉评分，解码开销低于 0.5%。

## 正文

Model Description

SingProbe is an intrinsic streaming guardrail built on Qwen/Qwen3.5-0.8B. 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/Qwen3.5-0.8B-singprobe 2.23M [6, 14, 22] 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 Qwen3.5-0.8B-singprobe Reference baseline

Query intent classification (6 benchmarks) F1 0.8334 YuFeng-XGuard-Reason-8B: 0.8714

Response safety classification (8 benchmarks) F1 0.8334 Qwen3Guard-Gen-8B-strict: 0.8604

Streaming safety (3 benchmarks) R-AUC / T-AUC 0.9778 / 0.9206 Qwen3Guard-Stream-8B-strict: 0.9640 / 0.8893

Hallucination detection (6 benchmarks) AUC 0.7480 DRIFT: 0.8000

Deployment characteristic Result

Benign-response false-positive rate 0.23% 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 Qwen/Qwen3.5-0.8B \ --probe-ckpt inclusionAI/Qwen3.5-0.8B-singprobe \ --port 30000

The integrations return one score dictionary per generated token (label_0–label_9). Use the exact base-model/probe pair: Qwen/Qwen3.5-0.8B with this checkpoint.

Citation

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