Lucia Malíčková | Searcharxiv

arXiv2026

From Projected Subspaces to Full-Space Implementations: Representation-Equivalence Audits for Open-Shell ADAPT-VQE

A variational state optimized in a symmetry-projected space need not be the state prepared by exponentiating the corresponding unprojected parent generators. For an orthogonal target-space projector $P$, $Q=I-P$, and an anti-Hermitian generator $A$, exact equivalence for all amplitudes requires target-space invariance, $QAP=0$. We formulate this condition as a representation-equivalence audit for open-shell ADAPT-VQE. For a canonical CAS(15e,9o) type-1 Cu(II) Hamiltonian, 274 of 323 bare Hartree--Fock-origin single, double, and triple excitations violate this condition. Projected-doublet ADAPT reaches 1.5719 m$E_h$ with 121 operators, whereas the same amplitudes in the corresponding full-space sequence give 464.730 m$E_h$ error, 0.07343 fidelity with the projected state, and 54.45\% quartet weight. Reoptimizing the same 121 parent generators in the same order under the hard constraint $p_D\ge0.999$ reduces the best converged error to 8.482 m$E_h$, showing that the catastrophic same-angle failure is primarily a representation/parameter-transfer failure but is not completely repaired by full-space reoptimization of the fixed sequence. An independent 12-qubit NO benchmark exhibits the same mechanism more mildly. A spin-preserving generalized full-space construction removes the representation ambiguity and reaches 1.493 and 1.181 m$E_h$ at 18 and 24 qubits. For the 18-qubit positive control, selectively refining only four Suzuki-2 factors and re-transpiling with a fixed compiler seed passes the predeclared circuit-equivalence gate at 19,660 controlled-X (CX) gates and depth 23,016. Circuit synthesis and final-energy measurement therefore constitute separate validation layers. The contribution is a validation framework, not a new spin-adapted ADAPT algorithm.

quant-ph↗

arXiv2026

Behavioral Reprogramming of Open-Weights Models: Cognitive Plasticity and Alignment Bounds

Large language models (LLMs) are predominantly aligned to function as passive, sycophantic assistants. We challenge this default paradigm by empirically evaluating the cognitive plasticity of open-weight architectures when subjected to rigorous behavioral reprogramming. Our objective is to induce a proactive, Socratic conversational framework, characterized by high-frequency question generation under strictly constrained high-performance computing (HPC) conditions. Through a massively parallelized hyperparameter sweep comprising 405 HPC jobs, we define precise mathematical bounds for parameter-efficient fine-tuning (PEFT). We identify an architectural threshold at LoRA rank $r=16$ and demonstrate via extensive epoch ablation that generalization capacity strictly reaches its optimal convergence within an optimized training window of $e \in [2, 3]$ depending on dataset density (minimum validation loss of 0.919). Furthermore, scaling model capacity to 14B parameters yielded a lower localized evaluation perplexity (1.414). Subsequent Direct Preference Optimization (DPO) successfully decoupled the underlying assertive behavior from localized syntax, while rigorous cross-lingual stress testing reveals both the capabilities and the structural boundaries of zero-shot persona transfer, demonstrating robust alignment in closely related linguistic families alongside identifiable degradation pathways in morphologically distant targets. These findings establish a rigorous empirical framework for compute-efficient, cross-lingual behavioral modification.

cs.AI↗