蚂蚁 inclusionAI 发布 SingProbe 流式护栏探针，基于 Qwen3.5-4B

Model Description

SingProbe is an intrinsic streaming guardrail built on Qwen/Qwen3.5-4B. 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/Qwen3.5-4B-singprobe`	5.18M	`[9, 19, 30]`	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	Qwen3.5-4B-singprobe	Reference baseline
Query intent classification (6 benchmarks)	F1	0.8650	YuFeng-XGuard-Reason-8B: 0.8714
Response safety classification (8 benchmarks)	F1	0.8646	Qwen3Guard-Gen-8B-strict: 0.8604
Streaming safety (3 benchmarks)	R-AUC / T-AUC	0.9870 / 0.9286	Qwen3Guard-Stream-8B-strict: 0.9640 / 0.8893
Hallucination detection (6 benchmarks)	AUC	0.7855	DRIFT: 0.8000

Deployment characteristic	Result
Benign-response false-positive rate	0.04% 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/Qwen3.5-4B \
  --probe-ckpt inclusionAI/Qwen3.5-4B-singprobe \
  --port 30000

The integrations return one score dictionary per generated token (label_0–label_9). Use the exact base-model/probe pair: Qwen/Qwen3.5-4B with this checkpoint.

Citation

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