Large language models (LLMs) can often be made to behave in undesirable ways that they are explicitly fine-tuned not to. For example, the LLM red-teaming literature has produced a wide variety of'jailbreaking'techniques to elicit harmful text from models that were fine-tuned to be harmless. Recent work on red-teaming, model editing, and interpretability suggests that this challenge stems from how (adversarial) fine-tuning largely serves to suppress rather than remove undesirable capabilities from LLMs. Prior work has introduced latent adversarial training (LAT) as a way to improve robustness to broad classes of failures. These prior works have considered untargeted latent space attacks where the adversary perturbs latent activations to maximize loss on examples of desirable behavior. Untargeted LAT can provide a generic type of robustness but does not leverage information about specific failure modes. Here, we experiment with targeted LAT where the adversary seeks to minimize loss on a specific competing task. We find that it can augment a wide variety of state-of-the-art methods. First, we use targeted LAT to improve robustness to jailbreaks, outperforming a strong R2D2 baseline with orders of magnitude less compute. Second, we use it to more effectively remove backdoors with no knowledge of the trigger. Finally, we use it to more effectively unlearn knowledge for specific undesirable tasks in a way that is also more robust to re-learning. Overall, our results suggest that targeted LAT can be an effective tool for defending against harmful behaviors from LLMs.
The critical behavior of the two-dimensional XY model has been explored in the literature using various methods. They include the high-temperature expansion (HTE) method, Monte Carlo (MC) approach, strong coupling expansion method, and tensor network (TN) methods. This model undergoes a Berezinskii-Kosterlitz-Thouless (BKT) type of phase transition. This model can be modified by adding spin-nematic interaction terms with a period to give rise to the generalized XY model. The modified model contains excitations of integer and half-integer vortices. These vortices govern the critical behavior of the theory and produce rich physics. With the help of tensor networks, we investigate the transition behavior between the integer vortex binding and half-integer vortex binding phases of the model and how this transition line merges into two BKT transition lines.
Navdeep Singh Dhindsa, Anosh Joseph, Abhishek Samlodia
et al.
We present our analysis of the deconfinement phase transition in the bosonic BMN matrix model. The model is investigated using a non-perturbative lattice framework. We used the Polyakov loop as the order parameter to monitor the phase transition, and the results were verified using the separatrix ratio. The calculations are performed using a large number of colors and a broad range of temperatures for all couplings. Our results indicate a first-order phase transition in this theory for all the coupling values that connect the perturbative and non-perturbative regimes of the theory.
We summarize the results obtained for the quark masses (u,d,s,c, and b) in Refs.~\cite{AlamKhan:2023ili,AlamKhan:2023kgs} and strong coupling ($α_s$) using renormalization group (RG) improvement of the theoretical expressions and experimental inputs that enter in the QCD sum rules. We obtain $m_{u}(2 GeV)=2.00_{-0.40}^{+0.33}$ MeV, $m_{d}(2 GeV)=4.21_{-0.45}^{+0.48}$ MeV, and $m_{s}(2 GeV)=104.34_{-4.24}^{+4.32}$ MeV using Borel Laplace sum rules for the divergence of the axial vector currents. The relativistic sum rules for the moments of the heavy quark currents lead to the determination of $α_s(M_Z)=0.1171(7)$, $\overline{m}_{c}=1281.1(3.8)$ MeV and $\overline{m}_{b}=4174.3(9.5)$ MeV.
Claudio Bonanno, Francesco D'Angelo, Massimo D'Elia
et al.
We compute the sphaleron rate on the lattice from the inversion of the Euclidean time correlators of the topological charge density, performing also controlled continuum and zero-smoothing extrapolations. The correlator inversion is performed by means of a recently-proposed modification of the Backus-Gilbert method.