Assessing regional carbon balance is crucial for developing effective greenhouse gas management strategies under global climate change. This study presents a CatBoost gradient boosting machine learning model to evaluate spatiotemporal variability of net ecosystem exchange of CO2 in terrestrial ecosystems. Applied to Japanese islands for 2024, the model effectively accounts for multiple factors influencing CO2 exchange and provides spatial distribution of CO2 fluxes at regional scale with monthly temporal resolution. The model demonstrated high prediction accuracy with an average coefficient of determination (R2) of 0.77 across all ecosystems. By integrating remote sensing data, ground-based measurements and environmental parameters through machine learning, this study provides a comprehensive framework for understanding regional CO2 exchange patterns. Results can be applied to CO2 fluxes assessments in other regions and contribute to developing climate mitigation measures. This modeling approach offers a valuable tool for monitoring ecosystem responses to changing climate conditions and informing regional carbon management policies.
_Latent-space_ monitoring techniques have shown promise as defenses against LLM attacks. These defenses act as scanners to detect harmful activations before they lead to undesirable actions. This prompts the question: can models execute harmful behavior _via inconspicuous latent states_? Here, we study such _obfuscated activations_. Our results are nuanced. We show that state-of-the-art latent-space defenses---such as activation probes and latent OOD detection---are vulnerable to obfuscated activations. For example, against probes trained to classify harmfulness, our obfuscation attacks can reduce monitor recall from 100% down to 0% while still achieving a 90% jailbreaking success rate. However, we also find that certain probe architectures are more robust than others, and we discover the existence of an _obfuscation tax_: on a complex task (writing SQL code), evading monitors reduces model performance. Together, our results demonstrate white-box monitors are not robust to adversarial attack, while also providing concrete suggestions to alleviate, but not completely fix, this weakness.
Recent LLMs like DeepSeek-R1 have demonstrated state-of-the-art performance by integrating deep thinking and complex reasoning during generation. However, the internal mechanisms behind these reasoning processes remain unexplored. We observe reasoning LLMs consistently use vocabulary associated with human reasoning processes. We hypothesize these words correspond to specific reasoning moments within the models' internal mechanisms. To test this hypothesis, we employ Sparse Autoencoders (SAEs), a technique for sparse decomposition of neural network activations into human-interpretable features. We introduce ReasonScore, an automatic metric to identify active SAE features during these reasoning moments. We perform manual and automatic interpretation of the features detected by our metric, and find those with activation patterns matching uncertainty, exploratory thinking, and reflection. Through steering experiments, we demonstrate that amplifying these features increases performance on reasoning-intensive benchmarks (+2.2%) while producing longer reasoning traces (+20.5%). Using the model diffing technique, we provide evidence that these features are present only in models with reasoning capabilities. Our work provides the first step towards a mechanistic understanding of reasoning in LLMs.
The quest for advanced hard materials with enhanced ductility drives the exploration of novel ternary borides. In this study, we use evolutionary algorithm USPEX combined with ab initio calculations to systematically investigate the ternary W–Nb–B system. Our results successfully reproduce all known stable binary compounds and reveal a new structural motif for NbB2 (space group R 3̅ m). Building on this motif, we predict a series of thermodynamically stable and metastable ternary compounds forming a continuous WxNb1 – xB24 solid solution, with the 50
The ability to tune spontaneous emission dynamics is pivotal for advancing quantum photonic technologies, such as micro-lasers and quantum sensors. Colloidal quantum dots (QDs) are promising candidates due to their high quantum yield and spectral tunability. However, their radiative decay rates are fundamentally limited by the local photonic density of states in conventional environments. Here, we demonstrate a tailored dielectric metasurface that resonantly couples to environmentally benign AgAuSe QDs. The metasurface is designed to support a magnetic dipole resonance and an electric dipole mode, the latter arises from the broken symmetry of a protected BIC state. The synergistic interplay between these modes creates a highly localized photonic environment, leading to a pronounced 15-fold amplification of NIR-II photoluminescence at room temperature. Optical characterization confirms that the enhancement stems from resonant coupling to quasi-bound states in the continuum (quasi-BIC) mode, drastically increases radiative decay rate. The use of CMOS-compatible, all-dielectric silicon nanostructures enables scalable fabrication. We demonstrate that the precise control of the resonance wavelength via structural symmetry breaking allows for selective fluorescence enhancement. This work establishes a quasi-BIC-based light-matter interaction scheme for sustainable quantum emitters and provides a design blueprint for developing non-toxic, high-efficiency quantum light sources and lasers.
