The central bottleneck in computational geothermal science is not simulator fidelity or data scarcity—it is the abstraction itself. Geothermal energy is increasingly important to the clean energy transition, yet its computational core still follows a legacy simulate-then-optimize paradigm: a deterministic simulator is calibrated to sparse observations and then used to optimize decisions within a fixed model. Hidden inside this pipeline are three commitments—one predicted future, one mostly static operating strategy, and one fitted model per site. We argue that, for next-generation enhanced geothermal systems, the subsurface is partially observed, heterogeneous, and intervention-sensitive, and the information available to characterize it is limited. As a result, forecasting and decision-making must reason over multiple physically plausible futures under uncertainty. Our central claim is that geothermal should be reframed as an adaptive problem of inference, intervention, and discovery. Under this view, simulation becomes conditional generation over plausible reservoir futures rather than point prediction of one trajectory. Operation becomes adaptive decision making over belief states rather than offline scheduling under a presumed known state. Calibration becomes the separation of transferable physical structure from site-specific corrections rather than repeated fitting within a fixed equation class. These are not three independent engineering problems; they are three phases of a single inference cycle. This reframing matters because, in geothermal, uncertainty is not merely something to quantify; it is something operations act upon and reshape. Likewise, persistent model mismatch is not merely an engineering nuisance to suppress; it is the primary scientific signal from which missing or site-modulated physics can be discovered. We therefore organize the paper around three consequences of this reframing: generative world models of reservoir evolution, belief-state policy learning for sustainable operation, and data-to-equation discovery for transferable geophysics. Taken together, these directions define a new agenda for geothermal AI beyond faster surrogate prediction toward adaptive subsurface intelligence where inference, intervention, and discovery are intrinsically coupled.
Multimodal deep learning has achieved remarkable progress by leveraging complementary information across heterogeneous data sources such as texts, images, audios, and structured signals. While increasingly powerful encoders and fusion mechanisms have improved predictive performance, the reliability of multimodal systems remains a critical challenge. In particular, modality disagreement, distribution shifts, and noisy inputs can lead to overconfident yet incorrect predictions. Distinct from existing surveys on uncertainty in deep learning [45] or on multimodal learning [8; 135], this survey jointly covers three aspects: (i) the structural foundations of multimodal classification examined through the lens of the un- certainty challenges each design choice introduces; (ii) un- certainty quantification in both unimodal and multimodal settings; and (iii) set-valued classification as a decision-level strategy for cautious multimodal prediction. We first review foundational aspects of multimodal representation learning and fusion strategies, highlighting their structural limitations in modeling inter-modal dependence and the uncertainty challenges each stage introduces. We then examine uncertainty quantification methods in deep learning, including both probabilistic and evidence-theoretic approaches, and analyze how these techniques extend to multimodal settings. Special attention is given to conflict-aware fusion mechanisms and to decision-level strategies such as set- valued classification, which enable more cautious and informative predictions. Beyond reviewing existing methods, we identify key open challenges, including the modeling of partial dependence be- tween modalities, the need for systematic benchmarking of multimodal uncertainty, and the integration of uncertainty into decision-making pipelines. Finally, we discuss how these reliability challenges extend to emerging multimodal agentic systems. By synthesizing advances across multimodal learning and uncertainty modeling, this survey aims to provide a unified perspective and to outline recent research directions toward more reliable multimodal AI systems.
Large language models (LLMs) and large vision-language models (LVLMs) have demonstrated impressive generative capabilities, yet ensuring their outputs align with user intent is still challenging. While most existing approaches address this issue at the training stage, inference-time approaches like decoding methods offer a more efficient and scalable solution. Decoding methods control model generation by guiding token-level selection, performing sequencelevel generation, or generating tokens in parallel to accelerate the process. In this survey, we identify three emerging paradigms from recent works on decoding methods for LLMs and LVLMs, provide a systematic review of these methods, highlight ongoing challenges, and discuss potential future research directions. Our goal is to underscore the efficiency and effectiveness of decoding methods and offer a practical view of their applications. Paper lists and more resources on decoding methods for LLMs and LVLMs can be found at https://github.com/wang2226/Awesome-LLM-Decoding.
A challenge in mitigating social bias in fine-tuned language models (LMs) is the potential reduction in language modeling capability, which can harm downstream performance. Counterfactual data augmentation (CDA), a widely used method for fine-tuning, highlights this issue by generating synthetic data that may align poorly with real-world distributions or creating overly simplistic counterfactuals that ignore the social context of altered sensitive attributes (e.g., gender) in the pretraining corpus. To address these limitations, we propose a simple yet effective context-augmented CDA method, Context-CDA, which uses large LMs to enhance the diversity and contextual relevance of the debiasing corpus. By minimizing discrepancies between the debiasing corpus and pretraining data through augmented context, this approach ensures better alignment, enhancing language modeling capability. We then employ uncertainty-based filtering to exclude generated counterfactuals considered low-quality by the target smaller LMs (i.e., LMs to be debiased), further improving the fine-tuning corpus quality. Experimental results on gender bias benchmarks demonstrate that Context-CDA effectively mitigates bias without sacrificing language modeling performance while offering insights into social biases by analyzing distribution shifts in next-token generation probabilities.
Large language models (LLMs) are trained on massive corpora that may contain sensitive information, creating privacy risks under membership inference attacks (MIAs). Knowledge distillation is widely used to compress LLMs into smaller student models, but its privacy implications are poorly understood. We systematically evaluate how distillation affects MIA vulnerability across six teacher-student model pairs and six attack methods. We find that distilled student models do not consistently exhibit lower MIA success than their teacher models, and in some cases demonstrate substantially higher member-specific attack success, challenging the assumption that knowledge distillation inherently improves privacy. We attribute this to mixed supervision in distillation: for vulnerable training data points, teacher predictions often align with ground-truth labels, causing student models to learn overly confident predictions that amplify the separability between members and non-members; conversely, for non-vulnerable points, teacher predictions and ground truth frequently diverge, providing inconsistent learning signals. To mitigate this, we propose three practical interventions – restricting distillation to non-vulnerable points, adding a low-dimensional Bottleneck Projection, and a normalization variant (NoNorm). Experiments show these methods reduce both aggregate and member-specific MIA success while preserving model utility, improving privacy-utility trade-offs for distilled LLMs.
