Searcharxiv⌕ Search

arXiv subjects

Yijia Zhu

Publications and source records attributed to Yijia Zhu.

2 recordsLinked to original sources

Vortex solitons in spin-orbit-coupled atomic-molecular condensates

In our previous study of a three-dimensional (3D) atomic-molecular Bose-Einstein condensate (BEC) with quadratic nonlinearity and spin-orbit coupling (SOC), stable solitons are realized as semi-vortex (SV) or mixed-mode (MM) states, with SOC applied solely to the atomic component. In the SV state, the molecular part forms a pure vortex, while the atomic part remains a hybrid of fundamental and vortex modes. This prevents the overall self-bound state from achieving a fully pure vortex configuration, with the total vortex component always less than $100\%$. Here, we extend the study to two dimensions (2D) and uncover a qualitatively different behavior: purely quadratic 2D systems support not only ground-state (GS) SV solitons, but also higher-order excited-state (ES) solitons that are genuine full vortices (a feature absent in the 3D counterpart). For both GS and ES solitons, stable vortex states exist only when the SOC strength $λ$ exceeds a minimum threshold $λ_{\min}$, which increases monotonically with $S_1$. Conversely, the critical detuning $α_c$ decreases monotonically with $S_1$, indicating that higher-order vortices are more susceptible to destabilization. For $λ< λ_{\min}$, the system forms MM states instead of vortex solitons. Our results demonstrate that dimensional reduction to 2D enables the emergence of pure-vortex excited states, establishing a new pathway for stabilizing high-charge topological excitations in ultracold molecular gases.

cond-mat.quant-gas↗

Internalizing LLM Reasoning via Discovery and Replay of Latent Actions

The internalization of chain-of-thought processes into hidden states has emerged as a highly efficient paradigm for scaling test-time compute. However, existing activation steering methods rely on static control vectors that fail to adapt to the non-stationary evolution of complex reasoning tasks. To address this limitation, we propose STIR (Self-Distilled Tools for Internal Reasoning), a framework that reformulates reasoning enhancement as a dynamic latent trajectory control problem. STIR introduces a synergistic three-stage pipeline: (1) differential intrinsic action induction harvests latent reasoning successes to crystallize steering primitives; (2) sparse control basis construction curates a compact, geometrically diverse tool library; and (3) value-modulated trajectory intervention dynamically injects context-specific impulses via anchor-based gating. Extensive experiments on six arithmetic and logical benchmarks across four representative models demonstrate that STIR improves average accuracy by 1.9% to 7.5% while reducing average token consumption by up to 35% compared to vanilla decoding. These findings demonstrate that the benefits of explicit chain-of-thought can be realized through dynamic latent trajectory control, internalizing the reasoning process to bypass the explicit generation while achieving superior fidelity. Our code is available at https://github.com/sznnzs/LLM-Latent-Action.

cs.LG↗