Shanghai University of Engineering Science

With the development of large language models (LLMs), numerous studies have demonstrated their vulnerability to carefully crafted jailbreak attacks. However, existing mitigation measures rarely balance model usability with significant protective effects, raising concerns about model abuse. To address this, we introduce CoT Defender. It preemptively occupies the model's first few generated tokens with a chain-of-thought analysis that hinders attackers from steering the output towards harmful content. We designed a two-stage training framework that strengthens security while preserving usability. Stage 1 fine-tunes the model to follow a structured chain-of-thought format before answering. Stage 2 employs reinforcement learning to refine this reasoning. An auxiliary attacker model continuously synthesizes new jailbreak prompts, and a lightweight evaluator-Probabilistic Structured Output Evaluation (PSOE)-supplies fine-grained rewards by scoring both sentence-level intent capture and token-level format fidelity. We conducted a series of experiments on four models and six attack methods. Across all models, we successfully reduced the average attack success rate to below 8.0 %, with no more than a 7.0 % impact on the response rate for benign requests. Code is available here. Warning: This paper contains red-team data that may be offensive!

Aqueous sodium-ion batteries (ASIBs) have surfaced as viable solutions for grid-scale applications characterized by exceptional safety, cost efficiency and eco-friendliness. However, the high freezing point severely restricts low-temperature viability. To overcome this issue, the dimethylacetamide (DMAC) is employed as a co-solvent for the inorganic and cheap 2 m NaCl (m: mol kg^-1) electrolyte, achieving a freezing point below -45 °C with remarkable ionic conductivity (2.93 mS cm^-1 under -30 °C). Theoretical calculations and experimental measurements reveal that carbonyl group in DMAC engages hydroxyl group from H₂O molecules to primarily form 1H₂O-DMAC conformation, disrupting the intrinsic hydrogen bonds interaction of H₂O molecules, effectively lowering freezing point of hybrid system. Using optimized electrolyte, the assembled Na₂CoFe(CN)₆//activated carbon (AC) batteries deliver 70.7 mAh g^-1 at 1C (1 C = 150 mA g^-1)，with 95 % capacity retention over 10,000 cycles at 10C at -30 °C. Notably, ASIBs successfully power light emitting diodes (1.8 V) at -40 °C. The electrolyte engineering strategy not only significantly enhances the performance of aqueous batteries in cold environments but also underscores their substantial potential for energy storage.

Fiber reinforced polymer (FRP) composites play a critical role in key industries including aerospace and automotive, and transportation. However, internal wrinkles compromise their structural performance. Nondestructive testing techniques play a vital role in the detection and evaluation of structural failures in FRP composites. The total focusing method (TFM), which utilizes ultrasonic data acquired in full matrix capture (FMC) mode, offers superior resolution compared with conventional phased array methods. However, it performs poorly for internal wrinkles detection, as the weak echoes caused by wrinkles are ofen overlapped with interlayer reflections, mechanical noises and multiple reflection waves. In this paper, a maximum eigenvalue vector coherence factor (MEVCF) weighted TFM is proposed to quantify the wrinkle geometry. Wrinkle angle information is extracted using the structural tensor-based image processing method. The signal phase concentration and the eigenvalue size are jointly applied to measure the phase information validity, thereby enhancing the continuity of the fiber ply. On this basis, a beam focusing deflection angle threshold is processed to effectively suppress the background noises and improve the imaging quality. Experimental results show that the texture clarity and image contrast are improved by approximately 18% and 43% compared with VCF, respectively. Meanwhile, the structural tensor method is less sensitive to noises compared with the gradient operator method. The fiber ply orientation can be accurately captured, enabling the quantitative assessment of wrinkle severity.