arXiv subjects

Yihang Peng

Publications and source records attributed to Yihang Peng.

7 recordsLinked to original sources

arXiv2026

CulMind: Benchmarking Multimodal Understanding and Reasoning in Chinese Cultural Heritage

Evaluating Multimodal Large Language Models (MLLMs) in Chinese Cultural Heritage (CCH) requires fine-grained reasoning over visual, textual, stylistic, and historical clues. However, existing CCH benchmarks mainly emphasize final-answer accuracy, while the accuracy and completeness of reasoning processes remain underexplored. To address this gap, we introduce CulMind and CulMind-R: a high-quality benchmark for multimodal CCH covering 50 tasks from collections of more than 100 museums, and a 24-task reasoning subset that adaptively defines task-specific dimensions for reasoning process evaluation. To evaluate reasoning quality, we propose ReaScore, a task-adaptive metric that evaluates reasoning by automatically weighting task-relevant dimensions. Experiments on 14 leading MLLMs reveal a substantial gap between answers and reasoning, especially on challenging tasks. Further analysis shows that task-adaptive dimension selection and weighting better align evaluation results with expert judgments. Overall, our benchmark and metric support a more expert-aligned assessment of CCH understanding and offer a transferable reference for broader evaluations of cultural heritage. We publicly release the data, code, and evaluation scripts at https://github.com/ZevTsao/CulMind to facilitate reproducible research.

cs.CL↗

arXiv2026

Thermal conductivity of seifertite and pyrite-type SiO$_2$: A comparative study

Thermal conductivity is a fundamental material property that plays a crucial role in understanding the dynamics and evolution of planetary interiors. Despite its importance, the thermal conductivity of seifertite and pyrite-type SiO$_2$ remains unknown. Here, we calculate the lattice thermal conductivities of seifertite and pyrite-type SiO$_2$ using the Green-Kubo method based on molecular dynamics (MD) simulations driven by two machine learning potentials (MLPs) constructed from the SCAN and PBEsol exchange-correlation functionals, with $\textit{ab initio}$-level accuracy. To demonstrate our methodology, we also compute thermal conductivities using the phonon quasiparticle approach for comparison. Overall, the Green-Kubo method predicts up to 119 % higher thermal conductivity with a temperature dependence close to $T^{-1}$, as it fully captures diffusion-like phonons at high temperatures that are missed by the phonon quasiparticle approach. The 19 % reduction in thermal conductivity across the phase transition from seifertite to the pyrite-type phase suggests the potential formation of a thermally insulating layer in the mantle of super-Earths.

cond-mat.mtrl-sci↗

arXiv2026

Echo-LoRA: Parameter-Efficient Fine-Tuning via Cross-Layer Representation Injection

Parameter-efficient fine-tuning (PEFT) has become a practical route for adapting large language models to downstream tasks, with LoRA-style methods being particularly attractive because they are inexpensive to train and easy to deploy. Most LoRA variants, however, revise the update rule within the weight space of each layer and leave the intermediate representations formed by deeper layers largely unused. We propose Echo-LoRA, a cross-layer representation injection method for parameter-efficient fine-tuning. During training, Echo-LoRA collects boundary hidden states from deeper source layers, aggregates them into a sample-level echo representation, and uses lightweight projection and gating networks to inject the resulting signal into shallow LoRA or DoRA modules. Answer-only masking, masked distillation, and stochastic routing are used to keep this auxiliary path stable and to reduce the gap between training and inference. On eight commonsense reasoning benchmarks, Echo-LoRA exceeds the reported LoRA baselines by 5.7 percentage points on average across LLaMA-7B, LLaMA2-7B, and LLaMA3-8B. Under reproduced LoRA baselines in our unified implementation, the average gain is 3.0 points; when combined with DoRA, the gain is 2.7 points. The Echo path is discarded after training, so the deployed model keeps the original low-rank LoRA/DoRA form and adds neither inference-time parameters nor inference computation.

cs.LG↗

arXiv2026

HistLens: Mapping Idea Change across Concepts and Corpora

Language change both reflects and shapes social processes, and the semantic evolution of foundational concepts provides a measurable trace of historical and social transformation. Despite recent advances in diachronic semantics and discourse analysis, existing computational approaches often (i) concentrate on a single concept or a single corpus, making findings difficult to compare across heterogeneous sources, and (ii) remain confined to surface lexical evidence, offering insufficient computational and interpretive granularity when concepts are expressed implicitly. We propose HistLens, a unified, SAE-based framework for multi-concept, multi-corpus conceptual-history analysis. The framework decomposes concept representations into interpretable features and tracks their activation dynamics over time and across sources, yielding comparable conceptual trajectories within a shared coordinate system. Experiments on long-span press corpora show that HistLens supports cross-concept, cross-corpus computation of patterns of idea evolution and enables implicit concept computation. By bridging conceptual modeling with interpretive needs, HistLens broadens the analytical perspectives and methodological repertoire available to social science and the humanities for diachronic text analysis.

cs.CL↗

arXiv2026

Ab initio electronic conductivity of Fe-bearing post-perovskite

The electrical conductivity of high-pressure silicates profoundly influences the interior dynamics of rocky planets. Employing the Kubo-Greenwood formalism, we perform ab initio calculations of electronic conductivity in Fe-bearing post-perovskite under super-Earth mantle conditions, up to 4000 K and 500 GPa. Electronic structures are obtained via many-body perturbation theory, incorporating dynamical screening and correlations among localized Fe-3d orbitals. In contrast to (Fe,Mg)O, for which metallization has been reported at comparable conditions, our results indicate that post-perovskite with Earth-like Fe contents is unlikely to metallize in super-Earth mantles via band-gap closure, yielding negligible low-frequency conductivity. Any substantial conductivity would require non-electronic mechanisms, such as thermally activated small-polaron hopping, which fall beyond the scope of band conduction.

cond-mat.mtrl-sci↗

arXiv2025

Induction Heating in Super-Earths: A Thermochemical Perspective

Electromagnetic induction heating has recently been proposed as an important internal heat source in the mantles of rocky exoplanets. However, its dependence on planetary interior properties remains poorly constrained. Here we construct electrical conductivity profiles for super-Earth mantles considering different temperatures and compositions, and evaluate induction heating in super-Earth mantles in both solid and partially molten states. We find that high mantle temperature, iron content, and melt fraction all suppress the overall induction heating efficiency due to increased mantle conductivity and magnetic shielding. In GJ 486b, induction heating likely exceeds both radiogenic heating and tidal heating, driving persistent surface volcanism and early volatile depletion, whereas HD 3167b and GJ 357b experience insignificant induction heating due to weak stellar magnetic fields. Our findings highlight induction heating as a critical factor in the thermal and atmospheric evolution of close-in super-Earths around magnetically active stars.

astro-ph.EP↗

arXiv2023

Hydrogen diffusion in the lower mantle revealed by machine learning potentials

Hydrogen may be incorporated into nominally anhydrous minerals including bridgmanite and post-perovskite as defects, making the Earth's deep mantle a potentially significant water reservoir. The diffusion of hydrogen and its contribution to the electrical conductivity in the lower mantle are rarely explored and remain largely unconstrained. Here we calculate hydrogen diffusivity in hydrous bridgmanite and post-perovskite, using molecular dynamics simulations driven by machine learning potentials of ab initio quality. Our findings reveal that hydrogen diffusivity significantly increases with increasing temperature and decreasing pressure, and is considerably sensitive to hydrogen incorporation mechanism. Among the four defect mechanisms examined, (Mg + 2H)$_{\rm Si}$ and (Al + H)$_{\rm Si}$ show similar patterns and yield the highest hydrogen diffusivity. Hydrogen diffusion is generally faster in post-perovskite than in bridgmanite, and these two phases exhibit distinct diffusion anisotropies. Overall, hydrogen diffusion is slow on geological time scales and may result in heterogeneous water distribution in the lower mantle. Additionally, the proton conductivity of bridgmanite for (Mg + 2H)$_{\rm Si}$ and (Al + H)$_{\rm Si}$ defects aligns with the same order of magnitude of lower mantle conductivity, suggesting that the water distribution in the lower mantle may be inferred by examining the heterogeneity of electrical conductivity.

cond-mat.mtrl-sci↗