arXiv subjects

Jie Cheng

Publications and source records attributed to Jie Cheng.

At least 37 records · Page 2Linked to original sources

arXiv2025

Stop Summation: Min-Form Credit Assignment Is All Process Reward Model Needs for Reasoning

Process reward models (PRMs) have proven effective for test-time scaling of Large Language Models (LLMs) on challenging reasoning tasks. However, reward hacking issues with PRMs limit their successful application in reinforcement fine-tuning. In this paper, we identify the main cause of PRM-induced reward hacking: the canonical summation-form credit assignment in reinforcement learning (RL), which defines the value as cumulative gamma-decayed future rewards, easily induces LLMs to hack steps with high rewards. To address this, we propose PURE: Process sUpervised Reinforcement lEarning. The key innovation of PURE is a min-form credit assignment that formulates the value function as the minimum of future rewards. This method significantly alleviates reward hacking by limiting the value function range and distributing advantages more reasonably. Through extensive experiments on 3 base models, we show that PRM-based approaches enabling min-form credit assignment achieve comparable reasoning performance to verifiable reward-based methods within only 30% steps. In contrast, the canonical sum-form credit assignment collapses training even at the beginning! Additionally, when we supplement PRM-based fine-tuning with just 10% verifiable rewards, we further alleviate reward hacking and produce the best fine-tuned model based on Qwen2.5-Math-7B in our experiments, achieving 82.5% accuracy on AMC23 and 53.3% average accuracy across 5 benchmarks. Moreover, we summarize the observed reward hacking cases and analyze the causes of training collapse. We release our code and model weights at https://github.com/CJReinforce/PURE.

cs.AI↗

arXiv2025

ZTree: A Subgroup Identification Based Decision Tree Learning Framework

Decision trees are a commonly used class of machine learning models valued for their interpretability and versatility, capable of both classification and regression. We propose ZTree, a novel decision tree learning framework that replaces CART's traditional purity based splitting with statistically principled subgroup identification. At each node, ZTree applies hypothesis testing (e.g., z-tests, t-tests, Mann-Whitney U, log-rank) to assess whether a candidate subgroup differs meaningfully from the complement. To adjust for the complication of multiple testing, we employ a cross-validation-based approach to determine if further node splitting is needed. This robust stopping criterion eliminates the need for post-pruning and makes the test threshold (z-threshold) the only parameter for controlling tree complexity. Because of the simplicity of the tree growing procedure, once a detailed tree is learned using the most lenient z-threshold, all simpler trees can be derived by simply removing nodes that do not meet the larger z-thresholds. This makes parameter tuning intuitive and efficient. Furthermore, this z-threshold is essentially a p-value, allowing users to easily plug in appropriate statistical tests into our framework without adjusting the range of parameter search. Empirical evaluation on five large-scale UCI datasets demonstrates that ZTree consistently delivers strong performance, especially at low data regimes. Compared to CART, ZTree also tends to grow simpler trees without sacrificing performance. ZTree introduces a statistically grounded alternative to traditional decision tree splitting by leveraging hypothesis testing and a cross-validation approach to multiple testing correction, resulting in an efficient and flexible framework.

cs.LG↗

arXiv2025

Robotic Fire Risk Detection based on Dynamic Knowledge Graph Reasoning: An LLM-Driven Approach with Graph Chain-of-Thought

Fire is a highly destructive disaster, but effective prevention can significantly reduce its likelihood of occurrence. When it happens, deploying emergency robots in fire-risk scenarios can help minimize the danger to human responders. However, current research on pre-disaster warnings and disaster-time rescue still faces significant challenges due to incomplete perception, inadequate fire situational awareness, and delayed response. To enhance intelligent perception and response planning for robots in fire scenarios, we first construct a knowledge graph (KG) by leveraging large language models (LLMs) to integrate fire domain knowledge derived from fire prevention guidelines and fire rescue task information from robotic emergency response documents. We then propose a new framework called Insights-on-Graph (IOG), which integrates the structured fire information of KG and Large Multimodal Models (LMMs). The framework generates perception-driven risk graphs from real-time scene imagery to enable early fire risk detection and provide interpretable emergency responses for task module and robot component configuration based on the evolving risk situation. Extensive simulations and real-world experiments show that IOG has good applicability and practical application value in fire risk detection and rescue decision-making.

cs.RO↗

arXiv2025

SafeWork-R1: Coevolving Safety and Intelligence under the AI-45$^{\circ}$ Law

We introduce SafeWork-R1, a cutting-edge multimodal reasoning model that demonstrates the coevolution of capabilities and safety. It is developed by our proposed SafeLadder framework, which incorporates large-scale, progressive, safety-oriented reinforcement learning post-training, supported by a suite of multi-principled verifiers. Unlike previous alignment methods such as RLHF that simply learn human preferences, SafeLadder enables SafeWork-R1 to develop intrinsic safety reasoning and self-reflection abilities, giving rise to safety `aha' moments. Notably, SafeWork-R1 achieves an average improvement of $46.54\%$ over its base model Qwen2.5-VL-72B on safety-related benchmarks without compromising general capabilities, and delivers state-of-the-art safety performance compared to leading proprietary models such as GPT-4.1 and Claude Opus 4. To further bolster its reliability, we implement two distinct inference-time intervention methods and a deliberative search mechanism, enforcing step-level verification. Finally, we further develop SafeWork-R1-InternVL3-78B, SafeWork-R1-DeepSeek-70B, and SafeWork-R1-Qwen2.5VL-7B. All resulting models demonstrate that safety and capability can co-evolve synergistically, highlighting the generalizability of our framework in building robust, reliable, and trustworthy general-purpose AI.

cs.AI↗

arXiv2025

Step-3 is Large yet Affordable: Model-system Co-design for Cost-effective Decoding

Large language models (LLMs) face low hardware efficiency during decoding, especially for long-context reasoning tasks. This paper introduces Step-3, a 321B-parameter VLM with hardware-aware model-system co-design optimized for minimizing decoding costs. Step-3 innovates in two key dimensions: (1) A novel Multi-Matrix Factorization Attention (MFA) mechanism that significantly reduces both KV cache size and computation while maintaining high attention expressiveness, and (2) Attention-FFN Disaggregation (AFD), a distributed inference system that decouples attention and Feed-Forward Network (FFN) layers into specialized subsystems. This co-design achieves unprecedented cost efficiency: Step-3 significantly reduces theoretical decoding costs compared with models like DeepSeek-V3 and Qwen3 MoE 235B, with the gains widening at longer context. Step-3 achieves low cost while activating 38B parameters per token (more than DeepSeek-V3 and Qwen3 MoE 235B), demonstrating that hardware-aligned attention arithmetic intensity, MoE sparsity, and AFD are critical to cost-effectiveness. We perform a head-to-head comparison with DeepSeek-V3 in its favorable scenarios. Our implementation on Hopper GPUs achieves a decoding throughput of up to 4,039 tokens per second per GPU under 50ms TPOT SLA (4K context, FP8, no MTP). It is higher than DeepSeek-V3's 2,324 in the same setup and sets a new Pareto frontier for LLM decoding.

cs.LG↗

arXiv2025

HAMF: A Hybrid Attention-Mamba Framework for Joint Scene Context Understanding and Future Motion Representation Learning

Motion forecasting represents a critical challenge in autonomous driving systems, requiring accurate prediction of surrounding agents' future trajectories. While existing approaches predict future motion states with the extracted scene context feature from historical agent trajectories and road layouts, they suffer from the information degradation during the scene feature encoding. To address the limitation, we propose HAMF, a novel motion forecasting framework that learns future motion representations with the scene context encoding jointly, to coherently combine the scene understanding and future motion state prediction. We first embed the observed agent states and map information into 1D token sequences, together with the target multi-modal future motion features as a set of learnable tokens. Then we design a unified Attention-based encoder, which synergistically combines self-attention and cross-attention mechanisms to model the scene context information and aggregate future motion features jointly. Complementing the encoder, we implement the Mamba module in the decoding stage to further preserve the consistency and correlations among the learned future motion representations, to generate the accurate and diverse final trajectories. Extensive experiments on Argoverse 2 benchmark demonstrate that our hybrid Attention-Mamba model achieves state-of-the-art motion forecasting performance with the simple and lightweight architecture.

cs.CV↗

arXiv2025

Simulation of the Background from $^{13}$C$(α, n)^{16}$O Reaction in the JUNO Scintillator

Large-scale organic liquid scintillator detectors are highly efficient in the detection of MeV-scale electron antineutrinos. These signal events can be detected through inverse beta decay on protons, which produce a positron accompanied by a neutron. A noteworthy background for antineutrinos coming from nuclear power reactors and from the depths of the Earth (geoneutrinos) is generated by ($α, n$) reactions. In organic liquid scintillator detectors, $α$ particles emitted from intrinsic contaminants such as $^{238}$U, $^{232}$Th, and $^{210}$Pb/$^{210}$Po, can be captured on $^{13}$C nuclei, followed by the emission of a MeV-scale neutron. Three distinct interaction mechanisms can produce prompt energy depositions preceding the delayed neutron capture, leading to a pair of events correlated in space and time within the detector. Thus, ($α, n$) reactions represent an indistinguishable background in liquid scintillator-based antineutrino detectors, where their expected rate and energy spectrum are typically evaluated via Monte Carlo simulations. This work presents results from the open-source SaG4n software, used to calculate the expected energy depositions from the neutron and any associated de-excitation products. Also simulated is a detailed detector response to these interactions, using a dedicated Geant4-based simulation software from the JUNO experiment. An expected measurable $^{13}$C$(α, n)^{16}$O event rate and reconstructed prompt energy spectrum with associated uncertainties, are presented in the context of JUNO, however, the methods and results are applicable and relevant to other organic liquid scintillator neutrino detectors.

physics.ins-det↗

arXiv2025

ElimPCL: Eliminating Noise Accumulation with Progressive Curriculum Labeling for Source-Free Domain Adaptation

Source-Free Domain Adaptation (SFDA) aims to train a target model without source data, and the key is to generate pseudo-labels using a pre-trained source model. However, we observe that the source model often produces highly uncertain pseudo-labels for hard samples, particularly those heavily affected by domain shifts, leading to these noisy pseudo-labels being introduced even before adaptation and further reinforced through parameter updates. Additionally, they continuously influence neighbor samples through propagation in the feature space.To eliminate the issue of noise accumulation, we propose a novel Progressive Curriculum Labeling (ElimPCL) method, which iteratively filters trustworthy pseudo-labeled samples based on prototype consistency to exclude high-noise samples from training. Furthermore, a Dual MixUP technique is designed in the feature space to enhance the separability of hard samples, thereby mitigating the interference of noisy samples on their neighbors.Extensive experiments validate the effectiveness of ElimPCL, achieving up to a 3.4% improvement on challenging tasks compared to state-of-the-art methods.

cs.CV↗

arXiv2025

Offline Reinforcement Learning with Discrete Diffusion Skills

Skills have been introduced to offline reinforcement learning (RL) as temporal abstractions to tackle complex, long-horizon tasks, promoting consistent behavior and enabling meaningful exploration. While skills in offline RL are predominantly modeled within a continuous latent space, the potential of discrete skill spaces remains largely underexplored. In this paper, we propose a compact discrete skill space for offline RL tasks supported by state-of-the-art transformer-based encoder and diffusion-based decoder. Coupled with a high-level policy trained via offline RL techniques, our method establishes a hierarchical RL framework where the trained diffusion decoder plays a pivotal role. Empirical evaluations show that the proposed algorithm, Discrete Diffusion Skill (DDS), is a powerful offline RL method. DDS performs competitively on Locomotion and Kitchen tasks and excels on long-horizon tasks, achieving at least a 12 percent improvement on AntMaze-v2 benchmarks compared to existing offline RL approaches. Furthermore, DDS offers improved interpretability, training stability, and online exploration compared to previous skill-based methods.

cs.LG↗

arXiv2025

Neutral-current background induced by atmospheric neutrinos at large liquid-scintillator detectors: III. Comprehensive prediction for low energy neutrinos

Atmospheric neutrinos play a vital role in generating irreducible backgrounds in liquid-scintillator (LS) detectors via their neutral-current (NC) interactions with $^{12}$C nuclei. These interactions may affect a wide range of research areas from the MeV to GeV energy range, such as the reactor and geo neutrinos, diffuse supernova neutrino background (DSNB), dark matter, and nucleon decay searches. In this work, we extend our preceding paper, by conducting a first-time systematic exploration of NC backgrounds as low as the MeV region of reactor and geo neutrinos. We utilize up-to-date neutrino generator models from GENIE and NuWro, a TALYS-based nuclear deexcitation package and a GEANT4-based detector simulation toolkit for our complete calculation. Our primary focus is to predict the NC background for experimental searches of inverse-beta-decay signals below the 100 MeV visible energy. In order to have deeper understanding of the characteristics of atmospheric neutrino NC interactions in LS, we investigate the model dependence of NC background predictions by using various data-driven models, including the initial neutrino-nucleon interactions, nuclear ground-state structure, final-state interactions, nuclear deexcitation processes, and secondary interactions of final-state particles.

hep-ph↗

arXiv2025

Magnetic phases and electron-phonon coupling in La$_3$Ni$_2$O$_7$ under pressure

Motivated by recent reports of pressure-induced superconductivity in bilayer nickelate La$_3$Ni$_2$O$_7$, we present a comprehensive investigation into the structural, electronic, magnetic, and phonon properties of this compound across a pressure range of 0 to 29.5 GPa. DFT+U calculations reveal that the A-type antiferromagnetic ground state of La$_3$Ni$_2$O$_7$ persists throughout the studied pressure range. Electronic structure analysis shows that the Ni-$d_{xy}$ and Ni-$d_{z^2}$ orbitals dominate near the Fermi level in both the $Fmmm$ and $Amam$ phases of La$_3$Ni$_2$O$_7$. Phonon dispersion calculations for the $Fmmm$ phase reveal no imaginary modes from 12 to 29.5 GPa, confirming its dynamical stability in this pressure range. The vibrational frequencies of O atoms are substantially higher than those of Ni and La atoms, primarily due to the lower mass of oxygen. At 29.5 GPa, the electron-phonon coupling constant $λ$ for the $Fmmm$ phase is calculated to be 0.13. This small value suggests that conventional electron-phonon coupling is insufficient to explain the reported superconductivity in La$_3$Ni$_2$O$_7$, indicating a potentially unconventional mechanism. The study offers nuanced, actionable insights that can strategically inform and direct subsequent experimental investigations into the design and optimization of nickel-based superconducting materials.

cond-mat.supr-con↗

arXiv2025

Prediction of high-temperature ambient-pressure superconductivity in hexagonal boron-rich clathrates

Inspired by recent predictions of superconductivity in B-C framework clathrates, we employ density functional theory to explore potential superconductors among hexagonal hydride-substituted compounds with compositions XB$_8$C, XB$_7$C$_2$, XB$_6$C$_3$, XB$_3$C$_6$, XB$_2$C$_7$, and XBC$_8$. Our high-throughput calculations on 96 compounds reveal several dynamically stable candidates exhibiting superconductivity at ambient pressure. Analysis of electronic structures and electron-phonon coupling demonstrates that CaB$_8$C, SrB$_8$C, and BaB$_8$C possess superconducting transition temperatures ($T_c$) exceeding 50 K, with CaB$_8$C exhibiting the highest predicted $T_c$ of 77.1 K among all stable compounds studied. These findings expand the family of B-C clathrate superconductors and provide valuable insights for experimental efforts aimed at discovering novel superconducting materials.

cond-mat.supr-con↗

arXiv2025

JUNO Sensitivity to Invisible Decay Modes of Neutrons

We explore the decay of bound neutrons into invisible particles (e.g., $n\rightarrow 3 ν$ or $nn \rightarrow 2 ν$) in the JUNO liquid scintillator detector, which do not produce an observable signal. The invisible decay includes two decay modes: $ n \rightarrow { inv} $ and $ nn \rightarrow { inv} $. The invisible decays of $s$-shell neutrons in $^{12}{\rm C}$ will leave a highly excited residual nucleus. Subsequently, some de-excitation modes of the excited residual nuclei can produce a time- and space-correlated triple coincidence signal in the JUNO detector. Based on a full Monte Carlo simulation informed with the latest available data, we estimate all backgrounds, including inverse beta decay events of the reactor antineutrino $\barν_e$, natural radioactivity, cosmogenic isotopes and neutral current interactions of atmospheric neutrinos. Pulse shape discrimination and multivariate analysis techniques are employed to further suppress backgrounds. With two years of exposure, JUNO is expected to give an order of magnitude improvement compared to the current best limits. After 10 years of data taking, the JUNO expected sensitivities at a 90% confidence level are $τ/B( n \rightarrow { inv} ) > 5.0 \times 10^{31} \, {\rm yr}$ and $τ/B( nn \rightarrow { inv} ) > 1.4 \times 10^{32} \, {\rm yr}$.

hep-ex↗

arXiv2025

Potential to identify neutrino mass ordering with reactor antineutrinos at JUNO

The Jiangmen Underground Neutrino Observatory (JUNO) is a multi-purpose neutrino experiment under construction in South China. This paper presents an updated estimate of JUNO's sensitivity to neutrino mass ordering using the reactor antineutrinos emitted from eight nuclear reactor cores in the Taishan and Yangjiang nuclear power plants. This measurement is planned by studying the fine interference pattern caused by quasi-vacuum oscillations in the oscillated antineutrino spectrum at a baseline of 52.5~km and is completely independent of the CP violating phase and neutrino mixing angle $θ_{23}$. The sensitivity is obtained through a joint analysis of JUNO and Taishan Antineutrino Observatory (TAO) detectors utilizing the best available knowledge to date about the location and overburden of the JUNO experimental site, local and global nuclear reactors, JUNO and TAO detector responses, expected event rates and spectra of signals and backgrounds, and systematic uncertainties of analysis inputs. We find that a 3$σ$ median sensitivity to reject the wrong mass ordering hypothesis can be reached with an exposure to approximately 6.5 years $\times$ 26.6 GW thermal power.

hep-ex↗

arXiv2025

Prediction of Energy Resolution in the JUNO Experiment

This paper presents an energy resolution study of the JUNO experiment, incorporating the latest knowledge acquired during the detector construction phase. The determination of neutrino mass ordering in JUNO requires an exceptional energy resolution better than 3\% at 1~MeV. To achieve this ambitious goal, significant efforts have been undertaken in the design and production of the key components of the JUNO detector. Various factors affecting the detection of inverse beta decay signals have an impact on the energy resolution, extending beyond the statistical fluctuations of the detected number of photons, such as the properties of the liquid scintillator, performance of photomultiplier tubes, and the energy reconstruction algorithm. To account for these effects, a full JUNO simulation and reconstruction approach is employed. This enables the modeling of all relevant effects and the evaluation of associated inputs to accurately estimate the energy resolution. The results of study reveal an energy resolution of 2.95\% at 1~MeV. Furthermore, this study assesses the contribution of major effects to the overall energy resolution budget. This analysis serves as a reference for interpreting future measurements of energy resolution during JUNO data collection. Moreover, it provides a guideline for comprehending the energy resolution characteristics of liquid scintillator-based detectors.

hep-ex↗

arXiv2024

Theoretical Prediction of High-Temperature Superconductivity in SrAuH$_3$ at Ambient Pressure

We present a comprehensive computational investigation of electron-phonon interactions in MXH$_3$ hydride compounds, where $M$ represents alkali and post-transition metals, and $X$ denotes 3$d$, 4$d$, and 5$d$ transition metals. Our density functional theory calculations identify 17 dynamically stable compounds. Notably, SrAuH$_3$ and SrZnH$_3$ emerge as theoretical ambient-pressure superconductors with predicted critical temperatures ($T_c$) exceeding 100 K. Analysis of the electronic structure reveals that the $X$ component dominates the density of states at the Fermi level, playing a crucial role in determining electron-phonon coupling strength and superconducting properties. We elucidate the underlying mechanisms governing these properties through detailed examination of the electronic and vibrational spectra. Our findings may challenge the prevailing notion that high-$T_c$ superconductivity in hydrides requires extreme pressures, potentially paving the way for practical applications. This study also provides valuable insights to guide future experimental efforts in the synthesis of ambient-pressure hydride superconductors.

cond-mat.supr-con↗

arXiv2024

LHPF: Look back the History and Plan for the Future in Autonomous Driving

Decision-making and planning in autonomous driving critically reflect the safety of the system, making effective planning imperative. Current imitation learning-based planning algorithms often merge historical trajectories with present observations to predict future candidate paths. However, these algorithms typically assess the current and historical plans independently, leading to discontinuities in driving intentions and an accumulation of errors with each step in a discontinuous plan. To tackle this challenge, this paper introduces LHPF, an imitation learning planner that integrates historical planning information. Our approach employs a historical intention aggregation module that pools historical planning intentions, which are then combined with a spatial query vector to decode the final planning trajectory. Furthermore, we incorporate a comfort auxiliary task to enhance the human-like quality of the driving behavior. Extensive experiments using both real-world and synthetic data demonstrate that LHPF not only surpasses existing advanced learning-based planners in planning performance but also marks the first instance of a purely learning-based planner outperforming the expert. Additionally, the application of the historical intention aggregation module across various backbones highlights the considerable potential of the proposed method. The code will be made publicly available.

cs.RO↗

arXiv2024

RIME: Robust Preference-based Reinforcement Learning with Noisy Preferences

Preference-based Reinforcement Learning (PbRL) circumvents the need for reward engineering by harnessing human preferences as the reward signal. However, current PbRL methods excessively depend on high-quality feedback from domain experts, which results in a lack of robustness. In this paper, we present RIME, a robust PbRL algorithm for effective reward learning from noisy preferences. Our method utilizes a sample selection-based discriminator to dynamically filter out noise and ensure robust training. To counteract the cumulative error stemming from incorrect selection, we suggest a warm start for the reward model, which additionally bridges the performance gap during the transition from pre-training to online training in PbRL. Our experiments on robotic manipulation and locomotion tasks demonstrate that RIME significantly enhances the robustness of the state-of-the-art PbRL method. Code is available at https://github.com/CJReinforce/RIME_ICML2024.

cs.LG↗