Model merging for Large Language Models (LLMs) directly fuses the parameters of different models finetuned on various tasks, creating a unified model for multi-domain tasks.However, due to potential vulnerabilities in models available on open-source platforms, model merging is susceptible to backdoor attacks.In this paper, we propose Merge Hijacking, the first backdoor attack targeting model merging in LLMs.The attacker constructs a malicious upload model and releases it.Once a victim user merges it with any other models, the resulting merged model inherits the backdoor while maintaining utility across tasks.Merge Hijacking defines two main objectives-effectiveness and utility-and achieves them through four steps.Extensive experiments demonstrate the effectiveness of our attack across different models, merging algorithms, and tasks.Additionally, we show that the attack remains effective even when merging real-world models.Moreover, our attack demonstrates robustness against two inference-time defenses (Paraphrasing and CLEANGEN) and one training-time defense (Fine-pruning).

For complex dynamic systems, their dynamical behaviours often exhibit unique phenomena such as chaotic oscillations and multistability switching, one of the main reasons for which is the nonlinear characteristics. In addition, the time-varying parameters, strong coupling effects, and multiscale interactions of nonlinear systems further exacerbate the complexity of modelling and analysis. Therefore, in order to study nonlinear complex dynamic systems and to obtain their physical properties, this paper proposes a data-driven method, the nonlinear probing excitation method. Nonlinear complex dynamical systems are often regarded as black-box systems and cannot be explained directly. The method proposed in this paper constructs a data-driven model through inputs and outputs, which contains the essential properties of the system, and is able to extract the physical properties of the system from it, and then explains part of the black-box system, and by repeating this process, it is possible to construct multiple models, and obtain the physical properties of the system one by one, and ultimately complete the implementation of the interpretability of the whole black-box model. In this way, the correlation between the input and output of the system and the nonlinear nature of the system is established from the data-driven point of view, and the physical characteristics of the nonlinear system are obtained, in order to predict and control the behaviour of the nonlinear system.

This paper presents an autonomous method to calibrate joint torque sensors for advanced robots; as a by-product, the position sensors can also be calibrated. When one joint of a manipulator is rotated the gravity torque exerted on the joint varies sinusoidally with rotation angle. By means of sinusoidally curve fitting, the bias of the torque sensor and the gain of the position sensor can be found directly. The gravity vector expressed in the defined base coordinates, can also be found. We then determine the torque sensor gain as well as the rotation angle offset, by attaching a known load to robot in different end positions. Experimental results are shown for the first two joints of the Sarcos dextrous arm.>

Describes the use of a nonlinear proportional and derivative (NPD) controller for both non-contact transition force control (NCTFC) (zero velocity when contact occurs) and contact transition force control (CTFC) (non-zero velocity when contact occurs). The authors present a gain design method for NCTFC. Simulations and experiments on the Sarcos Dextrous Arm show that PD control becomes unstable for CTFC while NPD control is stable and reaches the steady state quickly. Experiments with human subjects on NCTFC and CTFC are also presented.>

The transmission system of the Utah/MIT Dextrous Hand (UMDH) is investigated theoretically and experimentally. It is shown that the friction of the routing pulleys is not negligible and should be considered in the force control of the UMDH. In the frequency range of the pneumatic actuators (80 Hz), tendons act like springs and the first mode of the tendons is above that frequency range.>

A recently developed kinematic calibration technique for six degree-of-freedom (DOF) arms, the Jacobian measurement method, is generalized to include underconstrained and overconstrained arms. Using joint torque sensing and a wrist six-axis force/torque sensor to measure the Jacobian, experimental results are presented for the first five DOFs of the Sarcos Dextrous Arm. Results are compared to two standard calibration methods, i.e., open loop calibration and circle point analysis. Measurements of endpoint position and orientation were obtained with the Optotruk and the Bird sensor, two commercial motion tracking systems.