DeepSeek AI research team made breakthrough progress in verifiable reasoning tasks such as mathematics and code by using rule verification rewards in coordination with group relative policy optimization (GRPO), but reliable reward models are difficult to build on tasks such as writing, and the risk of Reward Hacking still exists outside the verifiable field. Taking DeepSeek-R1 as an anchor, this paper summarizes the progress of detection and mitigation of Reward Hacking from four aspects: formal definition, data set and evaluation indicators, method level, limitations and response. This paper first combs the formal definition of Reward Hacking and the classification framework under the assumption of proxy compression; Then, typical data sets and evaluation indicators are introduced, and the level of task performance and detection mitigation achieved by current mainstream methods is summarized; Finally, the structure boundary of rule reward, the algorithm layer performance of GRPO deviation, and the direction of detection and mitigation are analyzed. The overview shows that Reward Hacking originates from information compression in agent reward optimization, which can verify that accurate rules in the domain can compress Reward Hacking space, while outside the domain, it needs to rely on a collaborative system consisting of KL monitoring, representation detection, algorithm correction and policy constraints.

Show more

With the accelerating intelligent transformation of global manufacturing, industrial robots have become critical equipment in modern smart factories. However, the growing prevalence of multi-variety, small-batch production has made the high programming threshold, tedious reconfiguration, and limited environmental adaptability of traditional industrial robots increasingly prominent. Over the past few years, large language models (LLMs) have proven to have strong semantic comprehension, knowledge reasoning, and task planning abilities, a new avenue in the intelligent upgrading of industrial robots. The systematic review of the literature on the topic of the industrial robot driven by LLM is presented in this paper, including the mainstream technical directions, and the representative studies in five areas of application: industrial inspection, assembly and collaboration, maintenance and fault recovery, manufacturing operation planning, and robot navigation. Based on this, the paper examines the key challenges to be addressed at present, including model output reliability and safety, the simulation-to-reality gap and real-time constraints, and data scarcity and distribution shift. It also discusses future research opportunities, including knowledge enrichment and reliable reasoning, lightweight deployment and edge computing, and life-long learning and domain adaptation. The purpose of this paper is to serve as a reference in future studies and implementation of LLMs in industrial robots.

Show more

Graph Neural Networks (GNNs) have become a dominant paradigm for learning from graph-structured data, yet their application to large-scale graphs remains constrained by three coupled system bottlenecks: memory, computation, and communication. This paper organizes the literature around a bottleneck-centric taxonomy that maps representative techniques, including Cluster-GCN, GraphSAINT, GraphSAGE, SGC, DistDGL, and AliGraph, to the specific constraint each primarily relieves. Three representative systems, GraphSAGE, Cluster-GCN, and DistDGL, are further analyzed as case studies and consolidated into a practical decision matrix linking graph scale, hardware budget, and structural properties to an appropriate strategy. The paper also identifies a largely unaddressed challenge: the shift toward high-dimensional, LLM-encoded node features, which is beginning to move the dominant bottleneck from computation toward memory bandwidth and I/O. The review concludes that current sampling and partitioning techniques remain vulnerable to long-tail bias, that reported near-linear distributed speedups depend on network conditions rarely disclosed in published work, and that the most promising direction for future systems lies not in a new algorithm but in adaptive schedulers that reconfigure strategy at runtime.

Show more

This study retrospectively evaluates one-day-ahead classification of daily extreme precipitation in Guangzhou. Seven consecutive days of ERA5 atmospheric fields over a 41×65 regional grid are used to predict whether precipitation on the following day exceeds the 95th percentile of training-period daily IMERG precipitation. A no-event baseline, logistic regression, random forest, and a convolutional long short-term memory network (ConvLSTM) are compared with a chronological 70%/15%/15% split. On the held-out test period, ConvLSTM attained the highest recall (0.586), F1 score (0.347), receiver operating characteristic area under the curve (0.888), and Critical Success Index (0.210), detecting 34 of 58 extreme days. It also generated 104 false alarms and had a precision of 0.246. The observed results therefore show a detection–false-alarm tradeoff rather than an operationally reliable warning system. Because the comparison changes both model architecture and feature representation, and because it uses one temporal split without uncertainty intervals, the results do not isolate the causal benefit of spatial learning or establish statistical superiority. The study provides a reproducible framework for further evaluation, calibration, and physically informed analysis of extreme-precipitation forecasts in the Pearl River Delta.

Show more