Direct-decision models tested for safety and forgery forensics

Same-day papers report a CPU-NPU prompt guardrail and faster direct-decision heads for multimodal forgery checks on Ascend 910C.

Two separate arXiv cs.CR papers dated 2026-09-27 study direct-decision models on Huawei Ascend 910C NPUs; they cover different tasks and do not report conflicting figures for the same system. COGNIT-Guard cascades a CPU fast gatekeeper to Laya-322M for prompt safety screening. On an unseen DUCS-Bench split of 607 examples it reports 98.85% accuracy, a 0.42% benign false-positive rate (1 of 238), 1.12% ECE, and a 0.0104 Brier score, while the live cascade averages 41.63 ms at 99.23% accuracy and 0.00% FPR versus 21.77 ms for pure NPU inference, and experience replay restores SafetyBench-ZH out-of-domain accuracy to 64.10-65.05%. MAD-Guard finds that, on matched Qwen3-VL-8B with 2,400 FakeClue samples and LoRA, a binary direct head is 2.60x-7.27x faster at 53.12 ms and better calibrated (ECE 0.0450 versus 0.0845) than autoregressive decoding, though accuracy is 93.10% versus 94.90% for AR-SFT. Its CLM-Head reaches 96.55% binary accuracy and 98.79% seven-class attribution, scoring 96.44% on GenImage and 97.73% on Chameleon across 5,000 images, but only 0.5913 ROC-AUC on FF++ compressed faces.

• COGNIT-Guard escalates uncertain prompts from a validation-calibrated CPU gatekeeper to NPU-resident Laya-322M, a 322M bidirectional direct-decision safety model, under an asymmetric false-positive penalty.
• On unseen DUCS-Bench (N=607), COGNIT-Guard reports 98.85% accuracy, 0.42% benign false-positive rate (1/238), 1.12% ECE, and a 0.0104 Brier score.
• On Huawei Ascend 910C, pure NPU inference averages 21.77 ms; the deployed CPU-NPU cascade averages 41.63 ms at 99.23% accuracy and 0.00% FPR.
• Experience replay restores SafetyBench-ZH out-of-domain accuracy to 64.10-65.05%.
• MAD-Guard compares autoregressive generation with direct decision heads on a matched Qwen3-VL-8B backbone using 2,400 FakeClue samples and LoRA on Ascend 910C.
• A binary direct head is 2.60x-7.27x faster at 53.12 ms and cuts ECE from 0.0845 to 0.0450, with accuracy 93.10% versus 94.90% for AR-SFT.
• CLM-Head reaches 96.55% binary accuracy and 98.79% seven-class attribution; on 5,000 out-of-sample images it scores 96.44% on GenImage and 97.73% on Chameleon, but FF++ compressed-face ROC-AUC is 0.5913.