The stakes of this evaluation could hardly be higher. In the summer of 2020, record-breaking heavy precipitation swept across South Korea, claiming 46 lives and leaving people missing, while inflicting property damage of roughly 860 million dollars. The memory of that disaster, alongside the catastrophic July 2021 rainfall in China’s Henan Province that killed 398 people and caused about 19 billion dollars in direct economic losses, frames the urgency of the research. With climate change expected to increase the frequency of extreme precipitation events, the question of whether AI models can be trusted to warn of such hazards has become one of the most consequential in modern meteorology.

GraphCast represents a fundamentally different approach to forecasting than the numerical weather prediction models that have dominated the field for decades. Rather than solving the physical governing equations of the atmosphere, GraphCast uses graph neural networks arranged in an encoder-processor-decoder configuration to learn how atmospheric states evolve, drawing on more than four decades of historical reanalysis data from 1979 to 2017. Trained on the European Centre for Medium-Range Weather Forecasts’ ERA5 dataset, the model autoregressively predicts meteorological variables, including precipitation, at six-hour intervals with a horizontal resolution of about 25 kilometers. Its computational speed is extraordinary, producing forecasts in minutes on a single processor where traditional models require supercomputers running for hours.

To assess whether that speed comes at the cost of accuracy, the research team selected the ten days in the summer of 2020 with the largest accumulated precipitation over South Korea, spanning late June through mid-August. These cases featured daily maximum precipitation exceeding 130 millimeters and countrywide averages ranging from roughly 30 to 70 millimeters. Crucially, the team evaluated the model against observations from 549 rain gauges operated by the Korea Meteorological Administration, a ground-truth network widely regarded as the most reliable source of precipitation data. This choice matters, because most previous evaluations of AI weather models relied on reanalysis or analysis products, which can themselves carry biases that flatter models trained on similar data.

For comparison, the researchers ran the Weather Research and Forecasting model, a physics-based numerical system, at three horizontal resolutions of 27, 9, and 3 kilometers. For each heavy precipitation day, they performed 27 WRF simulations using different combinations of microphysics, boundary-layer, and radiation schemes, averaging the results to account for the model’s sensitivity to parameterization choices. Both the AI and numerical models were initialized with the same ERA5 data, creating as fair a head-to-head comparison as the profound architectural differences between the two systems allow.

The results on broad precipitation patterns were genuinely impressive for GraphCast. Using metrics that measure overall spatial structure, including normalized mean bias, normalized centered root-mean-square error, and spatial correlation coefficient, the AI model matched or beat the numerical simulations. Its normalized mean bias of about negative 14 percent was comparable to the 27-kilometer WRF’s roughly negative 13 percent, and its centered error was the smallest of all the models tested. Its spatial correlation of about 0.54 exceeded the 0.46 of the similarly resolved WRF, though the 3-kilometer WRF reached higher at about 0.62. GraphCast also proved remarkably adept at capturing the synoptic setup of the storms, reproducing the strong low-pressure systems, the monsoon-frontal pressure gradients, and the atmospheric-river-like moisture transports exceeding 1250 kilograms per meter per second that funnel water vapor toward the Korean Peninsula.

But the picture darkened when the team examined precipitation by intensity. Using the equitable threat score, neighborhood versions of that score, and contingency-table metrics across thresholds from 5 to 100 millimeters, GraphCast showed a systematic distortion of the rainfall distribution. It overpredicted light precipitation at or below 20 millimeters, painting wider areas of appreciable rain than actually occurred, while severely underpredicting heavy precipitation at or above 70 millimeters. Averaged across the 70 to 100 millimeter thresholds, its frequency bias was a mere 0.22, compared with 0.61 for the 27-kilometer WRF, and its probability of detection was just 0.16. In eight of the ten cases, GraphCast missed the observed precipitation peaks above 100 millimeters entirely, producing smoothed fields that washed out the most dangerous local maxima.

A clever resolution-matching experiment revealed that this weakness was not simply an artifact of the model’s 25-kilometer grid. When the higher-resolution WRF outputs were degraded to 27 kilometers, their advantage at heavy thresholds persisted, indicating that GraphCast’s deficit lies in its ability to represent precipitation peaks with spatial scales of roughly 25 to 100 kilometers, features that are resolvable even at coarse resolution. The researchers trace this limitation to two likely culprits. First, the ERA5 training data themselves struggle to capture heavy precipitation, meaning the model may have learned a systematically muted version of extreme rainfall. Second, the use of mean squared error as the training objective tends to blur small-scale variability through the well-known double penalty effect, in which a correctly predicted feature that is slightly misplaced is penalized twice, once as a false alarm and once as a miss.

Perhaps the most novel contribution of the study is its examination of physical consistency, an aspect of AI weather models that had not previously been tested for precipitation. The team focused on the Sobaek Mountains, where topographic lifting enhances rainfall during warm-season events. Observations showed the terrain-adjacent region received 20 millimeters more precipitation than nearby lowland areas on average. GraphCast reproduced only 42 percent of that enhancement, or an 8-millimeter difference, while the WRF models captured 63 to 90 percent. Strikingly, although both GraphCast and the 27-kilometer WRF simulated nearly identical synoptic conditions and moisture transport intensities, the upward motions over the mountains were considerably weaker in GraphCast, failing to exceed the 0.5 pascals per second seen in the numerical model, despite nearly identical mountain heights in both simulations. This suggests the AI model, unconstrained by physical equations, fails to fully translate moisture-laden flow over terrain into the orographic lifting that drives enhanced precipitation.

The authors are careful to frame these findings not as a verdict on AI versus physics in general, but as a measurement of where the current generation of AI models stands. Their results echo earlier reports that machine learning models struggle with mesoscale variability in temperature, wind, and intense precipitation, and they point toward promising remedies: better objective functions than mean squared error, high-resolution open-access training datasets from convection-permitting models, and hybrid strategies that pair AI models for locating precipitation features with numerical models for quantifying their intensity. Given GraphCast’s enormous computational advantages and its demonstrated skill with large-scale patterns, the study suggests the future of high-impact weather forecasting may lie not in choosing between silicon and physics, but in designing systems that combine the strengths of both, ensuring that the deadliest details of a storm are no longer smoothed away.

Subject of Research: Skill assessment of the AI-based weather prediction model GraphCast in forecasting heavy precipitation over South Korea

Article Title: Skill assessment of AI-based weather model in heavy precipitation prediction over South Korea

Article References: Hong, S.-H., Park, K., Kim, D.-H., Baik, J.-J., & Jin, H.-G. (2026). Skill assessment of AI-based weather model in heavy precipitation prediction over South Korea. Theoretical and Applied Climatology, 157(9), Article 606. https://doi.org/10.1007/s00704-026-06536-w

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06536-w

Keywords: GraphCast, AI weather prediction, heavy precipitation, South Korea, WRF model, forecast verification, topographic precipitation, extreme weather, machine learning, ERA5, equitable threat score, physical consistency

The core problem the survey identifies is deceptively simple. Modern generative models—diffusion models that iteratively denoise random patterns into images, and autoregressive models that predict visual tokens one after another—are trained with surrogate objectives such as maximum likelihood estimation or reconstruction loss. These mathematical proxies measure how well a model reproduces its training data, but they say little about whether a generated video moves convincingly, whether a synthesized face looks beautiful, or whether a prompt asking for “a cat juggling on a bicycle” is actually satisfied. The consequences are familiar to anyone who has played with text-to-video systems: limbs that morph mid-stride, objects that float when they should fall, and scenes that drift semantically from the original request. Likelihood training simply was never designed to reward physical plausibility or aesthetic judgment.

Reinforcement learning offers a principled escape from this trap. Originally formulated to solve Markov decision processes—sequential decision problems in which an agent learns through trial and error to maximize cumulative reward—the framework can optimize objectives that are non-differentiable, preference-driven, or temporally structured. A human preference, a physics violation, or an aesthetic score can all be packaged as reward signals, even when no gradient can flow through them directly. The survey traces reinforcement learning’s conceptual evolution through four phases: first as a solver of well-defined decision problems using value-based methods like Q-learning and policy-based methods like REINFORCE; then as a family of specialized subfields including offline reinforcement learning, multi-agent systems, risk-sensitive methods, and safe learning; then as a tool for learning environment dynamics and aligning with human intent; and finally as a general-purpose substrate for decision-making embedded within larger systems that combine planning, simulation, and feedback.

That final phase matters most for generative modeling. The landmark demonstration came from reinforcement learning with human feedback, the technique that turned a 1.3-billion-parameter language model fine-tuned on human preference rankings into a system that outperformed the original 175-billion-parameter GPT-3 at following instructions. The lesson generalized: instead of hand-crafting reward functions, researchers collect comparisons between outputs, train a reward model to predict those preferences, and then use reinforcement learning to push the generator toward highly rewarded behavior. For visual generation, this reframes vague goals like “make it look better” into concrete optimization problems. The same logic underpins world-model approaches such as Dreamer and MuZero, which learn internal simulators of their environments and plan within them—a strategy the survey links directly to the modern idea of generative models as learned simulators of visual reality.

In image generation, the survey organizes the methodological landscape into three families. Policy-based methods treat the denoising process of a diffusion model as a multi-step decision problem. Denoising Diffusion Policy Optimization, or DDPO, and its cousin DPOK were early exemplars, using policy gradients with Kullback–Leibler regularization to improve both image quality and text-image alignment. More recently, Group Relative Policy Optimization, or GRPO—an algorithm introduced with DeepSeekMath—has been adapted with striking breadth: DanceGRPO unifies diffusion models and rectified flows under a single framework applicable to text-to-image, text-to-video, and image-to-video tasks, while Flow-GRPO reformulates flow-matching generation as a stochastic differential equation to enable effective exploration. A parallel family, Direct Preference Optimization or DPO, sidesteps explicit reward modeling entirely, treating alignment as a classification problem over ranked output pairs. Variants now address patch-level detail, personalization, safety through unlearning, curriculum learning, and even AI-generated preference labels that reduce dependence on costly human annotation.

Video generation poses harder challenges because time introduces motion inconsistency, semantic drift, and physical implausibility. Here the survey catalogs reinforcement learning deployed at every stage of the pipeline. At the sampling stage, AdaDiff learns an adaptive policy for choosing denoising step sizes, trading coarse updates for fine ones to accelerate generation without sacrificing fidelity. At the planning stage, systems like FLIP use actor-critic frameworks with dense feedback from vision-language models to select video clips that fulfill textual instructions, while RLAVE applies reinforcement learning to automatic editing, rewarding narrative coherence, pacing, and aesthetics. For alignment, VideoDPO, HuViDPO, and DenseDPO extend preference optimization with multi-dimensional rewards, patch-level feedback, and segment-level annotations. Perhaps most intriguingly, RDPO generates preference pairs automatically from real videos using physics-based heuristics—a ball that falls is preferred over one that floats—encoding physical plausibility without any human labeling. Phys-AR goes further, converting frames into symbolic tokens and rewarding trajectories that obey velocity consistency and mass-informed motion, producing parabolic arcs and realistic collisions.

The survey also highlights a subtle but consequential innovation: reinforcement learning applied at inference time rather than during training. The InfLVG system samples candidate continuations at each generation step, scores them with a composite reward balancing face identity consistency, prompt relevance, and artifact suppression, and updates its sampling policy on the fly. This lets the model extend generated videos to nine times their baseline length while maintaining coherence—an achievement that would be prohibitively expensive with conventional autoregressive sampling alone. Alongside these methods, reward fine-tuning approaches like InstructVideo and VADER, which supervise generators directly with differentiable reward gradients from expert models such as CLIP and object detectors, blur the boundary between classical reinforcement learning and gradient-based alignment, though the survey is careful to distinguish the two.

In three-dimensional content generation, reinforcement learning proves equally versatile. Early work voxelized shapes and rewarded topologically valid growth; recent systems operate on meshes, point clouds, neural radiance fields, and 3D Gaussian splatting. DeepMesh applies preference optimization to autoregressive mesh creation, while Mesh-RFT introduces topology-aware scoring metrics to refine flawed geometric regions automatically. DreamReward built a preference dataset of over twenty-five thousand prompt–asset pairs and trained a reward model that guides text-to-3D sampling; DreamDPO eliminates the reward model by exploiting large vision-language models as zero-shot judges. Addressing the notorious Janus problem, in which multi-view reconstructions show conflicting geometry, Carve3D fine-tunes diffusion models with a multi-view reconstruction consistency reward, and Nabla-R2D3 transforms two-dimensional reward signals into structured three-dimensional rewards through a probabilistic refinement mechanism. Domain-specific applications extend to point cloud completion, sequential indoor scene synthesis with physically constrained layouts, scene-aware human motion generation, and music-synchronized 3D dance through actor-critic GPT architectures.

The survey’s synthesis is that reinforcement learning has outgrown its role as a post-training trick and become a structural component of generative system design. It enables optimization of non-differentiable objectives, fine-grained sequential control, incorporation of temporal and physical feedback, and principled alignment with subjective human goals. The authors argue that preference-based paradigms like DPO have redefined the relationship between learning and generation, shifting the field from exploration-heavy training toward stable, sample-efficient alignment, and they anticipate a future of multi-objective, potentially multi-agent generation in which models must balance quality, diversity, safety, and efficiency simultaneously. Open challenges remain—reward hacking, annotation cost, generalization, and the scalability of preference data among them—but the trajectory is unmistakable. Generation is no longer conceived as a static mapping from input to output; it is an interactive, iterative, goal-driven process. As generative systems become more autonomous and user-facing, the capacity to learn from feedback and adapt to diverse human preferences, the authors conclude, will be indispensable—and reinforcement learning supplies the theoretical and algorithmic machinery to deliver it.

Subject of Research: Integration of reinforcement learning with visual generative models across image, video, and 3D content generation

Article Title: Integrating reinforcement learning with visual generative models: foundations and advances

Article References: Liang, Y., Fang, Y., Li, R., Ni, Z., Su, R., & Zhang, C. (2026). Integrating reinforcement learning with visual generative models: foundations and advances. Vicinagearth, 3(1), Article 2. https://doi.org/10.1007/s44336-025-00030-z

Image Credits: AI Generated

DOI: 10.1007/s44336-025-00030-z

Keywords: reinforcement learning, generative models, diffusion models, direct preference optimization, video generation, 3D generation, human feedback, policy optimization, text-to-image, world models, physical consistency, multimodal learning