arXiv subjects

Han Shen

Publications and source records attributed to Han Shen.

At least 19 recordsLinked to original sources

arXiv2026

Refusing Intent, Not Form: Wrapper-Based Intent-Group Supervision for LLM Safety

Safety tuning can improve harmful refusal, but models may learn surface-form shortcuts: wrapped harmful prompts bypass safety, while similarly wrapped benign prompts are over-refused. We propose Wrapper-Based Intent-Form Augmentation (WIFA), an automatic intent-group augmentation method that pairs wrapped harmful examples with structurally matched wrapped benign counterexamples, requiring no external teacher or manual per-wrapper intent labels. We use WIFA as a common data layer for two complementary fine-tuning routes: WIFA-Boost, a two-stage high-safety recipe, and Anchored Group-Consistent Refusal Training (A-GCRT), which regularizes refusal/compliance decision scores across same-intent wrappers and anchors harmful and benign groups on opposite sides of a margin. In the Qwen setting, WIFA-Boost reaches the strongest transformed-harmful refusal, while A-GCRT reduces OR-Bench over-refusal from 25.7\% for the base model to 17.4\%; reproduced baselines do not match these operating points. Llama results and ablations over data structure, two-stage order, and A-GCRT components support this intent-group interpretation without claiming universal below-base over-refusal.

cs.CL↗

arXiv2026

Divergent Evolution of Radial Metallicity Gradients in the Thin and Thick Disks of the Milky Way

Using 200,388 red clump stars from LAMOST and APOGEE, we investigate the radial metallicity gradients of the Galactic disk as a function of vertical height and stellar age. The thin disk displays a pronounced negative radial metallicity gradient near the Galactic mid-plane that progressively flattens with increasing $|Z|$, following $\Delta \mathrm{[Fe/H]}/\Delta R$ = $-$0.0784 $+$ 0.0776 (1 $-$ exp ($-$ $|Z|$/1.42)). The thin disk also exhibits a clear age dependence in radial metallicity gradients, evolving smoothly from a strong gradient regime for young stars to a weak gradient regime for old stars, following $\Delta \mathrm{[Fe/H]}/\Delta R$ = $-$0.0438 $+$ 0.0233 tanh (($\tau$ $-$ 11.29)/4.21). The thick disk shows weakly positive radial metallicity gradients that remain statistically invariant with respect to both vertical height and stellar age, following respectively, $\Delta \mathrm{[Fe/H]}/\Delta R$ = 0.0038 $+$ 0.0009 $|Z|$ and $\Delta \mathrm{[Fe/H]}/\Delta R$ = 0.0146 $-$ 0.0007 $\tau$. These results indicate that the thin disk retains radial metallicity gradients shaped by relatively ordered inside-out growth and long-term secular evolution processes. The thick disk exhibits spatially and temporally homogeneous radial metallicity gradients, which are consistent with a formation environment characterized by mergers of gas-rich systems and/or the turbulent ISM.

astro-ph.GA↗

arXiv2026

Normalizing Flow-Enhanced Message Passing for Multirobot Collaborative Localization

Accurate, robust, and adaptive localization is essential for various robotic operations. This paper proposes a new message passing (MP) algorithm for realizing collaborative localization in a distributed manner. The algorithm unifies Gaussian belief propagation (GBP) and mean-field (MF) approximation, where GBP preserves dependencies among robot states, and MF enables estimation of noise statistics. To effectively handle non-conjugate terms from nonlinear measurement models, the algorithm adopts a parametric formulation in which these terms are treated by gradient estimators. Beyond linearization and sampling, we further design a normalizing flow (NF)-based gradient estimator, enabling learnable sampling. End-to-end training tunes NF parameters according to the behavior of MP, improving the overall estimation performance. To support estimation of practical robotic states that involve rotations, the method is then extended to Lie group state spaces. Finally, the method is applied to multirobot localization task fusing odometry, global navigation satellite system (GNSS) measurements, and inter-robot ultra wideband (UWB) ranging. Simulations and experiments on autonomous surface vehicles (ASVs) demonstrate its improved accuracy, robustness, and adaptability.

cs.RO↗

arXiv2026

Revis: Sparse Latent Steering to Mitigate Object Hallucination in Large Vision-Language Models

Despite the advanced capabilities of Large Vision-Language Models (LVLMs), they frequently suffer from object hallucination. One reason is that visual features and pretrained textual representations often become intertwined in the deeper network layers. To address this, we propose REVIS, a training-free framework designed to explicitly re-activate this suppressed visual information. Rooted in latent space geometry, REVIS extracts the pure visual information vector via orthogonal projection and employs a calibrated strategy to perform sparse intervention only at the precise depth where suppression occurs. This surgical approach effectively restores visual information with minimal computational cost. Empirical evaluations on standard benchmarks demonstrate that REVIS reduces object hallucination rates by approximately 19% compared to state-of-the-art baselines, while preserving general reasoning capabilities.

cs.AI↗

arXiv2026

Light Alignment Improves LLM Safety via Model Self-Reflection with a Single Neuron

The safety of large language models (LLMs) has increasingly emerged as a fundamental aspect of their development. Existing safety alignment for LLMs is predominantly achieved through post-training methods, which are computationally expensive and often fail to generalize well across different models. A small number of lightweight alignment approaches either rely heavily on prior-computed safety injections or depend excessively on the model's own capabilities, resulting in limited generalization and degraded efficiency and usability during generation. In this work, we propose a safety-aware decoding method that requires only low-cost training of an expert model and employs a single neuron as a gating mechanism. By effectively balancing the model's intrinsic capabilities with external guidance, our approach simultaneously preserves utility and enhances output safety. It demonstrates clear advantages in training overhead and generalization across model scales, offering a new perspective on lightweight alignment for the safe and practical deployment of large language models. Code: https://github.com/Beijing-AISI/NGSD.

cs.AI↗

arXiv2025

On Entropy Control in LLM-RL Algorithms

For RL algorithms, appropriate entropy control is crucial to their effectiveness. To control the policy entropy, a commonly used method is entropy regularization, which is adopted in various popular RL algorithms including PPO, SAC and A3C. Although entropy regularization proves effective in robotic and games RL conventionally, studies found that it gives weak to no gains in LLM-RL training. In this work, we study the issues of entropy bonus in LLM-RL setting. Specifically, we first argue that the conventional entropy regularization suffers from the LLM's extremely large response space and the sparsity of the optimal outputs. As a remedy, we propose AEnt, an entropy control method that utilizes a new clamped entropy bonus with an automatically adjusted coefficient. The clamped entropy is evaluated with the re-normalized policy defined on certain smaller token space, which encourages exploration within a more compact response set. In addition, the algorithm automatically adjusts entropy coefficient according to the clamped entropy value, effectively controlling the entropy-induced bias while leveraging the entropy's benefits. AEnt is tested in math-reasoning tasks under different base models and datasets, and it is observed that AEnt outperforms the baselines consistently across multiple benchmarks.

cs.LG↗

arXiv2025

Kwai Keye-VL 1.5 Technical Report

In recent years, the development of Large Language Models (LLMs) has significantly advanced, extending their capabilities to multimodal tasks through Multimodal Large Language Models (MLLMs). However, video understanding remains a challenging area due to the dynamic and information-dense nature of videos. Existing models struggle with the trade-off between spatial resolution and temporal coverage when processing video content. We present Keye-VL-1.5, which addresses fundamental challenges in video comprehension through three key innovations. First, we introduce a novel Slow-Fast video encoding strategy that dynamically allocates computational resources based on inter-frame similarity, processing key frames with significant visual changes at higher resolution (Slow pathway) while handling relatively static frames with increased temporal coverage at lower resolution (Fast pathway). Second, we implement a progressive four-stage pre-training methodology that systematically extends the model's context length from 8K to 128K tokens, enabling processing of longer videos and more complex visual content. Third, we develop a comprehensive post-training pipeline focusing on reasoning enhancement and human preference alignment, incorporating a 5-step chain-of-thought data construction process, iterative GSPO-based reinforcement learning with progressive prompt hinting for difficult cases, and alignment training. Through extensive evaluation on public benchmarks and rigorous internal human assessment, Keye-VL-1.5 demonstrates significant improvements over existing models, particularly excelling in video understanding tasks while maintaining competitive performance on general multimodal benchmarks.

cs.CV↗

arXiv2025

The Kinematic Signature of the Warp and Waves in the Milky Way Disk

Using over 170,000 red clump (RC) stars selected from LAMOST and APOGEE, we conduct a detailed analysis of the stellar $V_{Z}$ as a function of $L_{Z}$ (or $R_{g}$) across different $\phi$ bins for various disk populations. The $V_{Z}$ of the whole RC sample stars exhibits a wave-like pattern superimposed on an exponentially increasing trend, indicating the contribution from disk warp, disk flare and disk waves. Our results across various populations suggest that the thin disk is similar to the whole RC sample behavior, while the thick disk displays a wave-like pattern superimposed on a linearly increasing trend, meaning that the features of disk warp and waves are present in both thin and thick disks, and the disk flare feature is only present in the thin disk. These results indicate that the disk warp is potentially driven by secular processes like disk perturbations from intergalactic magnetic fields and a misaligned dark halo. The line-of-node of the disk warp of various populations displays a slight difference, with $\phi_{0}$ = 5.68 $\pm$ 2.91 degree for the whole RC sample stars, $\phi_{0}$ = 5.78 $\pm$ 2.89 degree for the thin disk stars, and $\phi_{0}$ = 4.10 $\pm$ 3.43 degree for the thick disk stars.

astro-ph.GA↗

arXiv2025

Fundamental Safety-Capability Trade-offs in Fine-tuning Large Language Models

Fine-tuning Large Language Models (LLMs) on some task-specific datasets has been a primary use of LLMs. However, it has been empirically observed that this approach to enhancing capability inevitably compromises safety, a phenomenon also known as the safety-capability trade-off in LLM fine-tuning. This paper presents a theoretical framework for understanding the interplay between safety and capability in two primary safety-aware LLM fine-tuning strategies, providing new insights into the effects of data similarity, context overlap, and alignment loss landscape. Our theoretical results characterize the fundamental limits of the safety-capability trade-off in LLM fine-tuning, which are also validated by numerical experiments.

stat.ML↗

arXiv2024

The Galactic Disk North-south Asymmetry in Metallicity May Be A New Tracer for the Disk Warp

Galactic disk warp has been widely characterized by stellar distributions and stellar kinematics but has not been traced by stellar chemistry. Here, we use a sample with over 170,000 red clump (RC) stars selected from LAMOST and APOGEE first to establish a correlation between the north-south asymmetry in metallicity ([Fe/H]) and the disk warp. Our results indicate that the height of the [Fe/H] mid-plane for the whole RC sample stars is accurately described as $Z_{w}$ = 0.017 ($R$ $-$ 7.112)$^{2}$ sin($\phi$ $-$ 9.218). This morphology aligns closely with the warp traced by Cepheids, suggesting that the disk north-south asymmetry in [Fe/H] may serve as a new tracer for the Galactic warp. Our detailed analysis of the young/thin disk stars of this RC sample suggests that its warp is well-modeled as $Z_{w}$ = 0.016 ($R$ $-$ 6.507)$^{2}$ sin($\phi$ $-$ 4.240), indicating that the line of node (LON) of the Galactic warp is oriented at 4.240$_{-1.747}^{+1.641}$ degree.

astro-ph.GA↗

arXiv2024

The Age-velocity Dispersion Relations of the Galactic Disk as Revealed by the LAMOST-Gaia Red Clump Stars

Using nearly 230,000 red clump (RC) stars selected from LAMOST and Gaia, we conduct a comprehensive analysis of the stellar age-velocity dispersion relations (AVRs) for various disk populations, within 5.0 $\leq$ $R$ $\leq$ 15.0 kpc and $|Z|$ $\leq$ 3.0 kpc. The AVRs of the whole RC sample stars are accurately described as $\sigma_{v}$ = $\sigma_{v,0}$ ($\tau$ + 0.1)$^{\beta_{v}}$, with $\beta_{R}$, $\beta_{\phi}$ and $\beta_{Z}$ displaying a global exponential decreasing trend with $R$, which may point to the difference in spatial distributions of various disk heating mechanisms. The measurements of $\beta$ $-$ $R$ for various disks suggest that the thin disk exhibits a radial dependence, with a global exponential decreasing trend in $\beta_{R}$ $-$ $R$ and $\beta_{Z}$ $-$ $R$, while $\beta_{\phi}$ remains a nearly constant value (around 0.20$\sim$0.25) within 8.5 $\leq$ $R$ $\leq$ 11.5 kpc. The thick disk displays a global increasing trend in $\beta_{R}$ $-$ $R$, $\beta_{\phi}$ $-$ $R$ and $\beta_{Z}$ $-$ $R$. These results indicate that the thin disk stars are likely heated by long-term heating from GMCs and spiral arms, while thick disk stars are likely heated by some violent heating process from merger and accretion, and/or formed by the inside-out and upside-down star formation scenarios, and/or born in the chaotic mergers of gas-rich systems and/or turbulent ISM. Our results also suggest that the disk perturbation by a recent minor merger from Sagittarius may have occurred within 3.0 Gyr.

astro-ph.GA↗

arXiv2024

Understanding Forgetting in LLM Supervised Fine-Tuning and Preference Learning -- A Convex Optimization Perspective

The post-training of LLMs, which typically consists of the supervised fine-tuning (SFT) stage and the preference learning stage (RLHF or DPO), is crucial to effective and safe LLM applications. The widely adopted approach in post-training popular open-source LLMs is to sequentially perform SFT and RLHF/DPO. However, this is suboptimal in terms of SFT and RLHF/DPO trade-off: the LLM gradually forgets about the first stage's training when undergoing the second stage's training. This sequential paradigm persists largely due to its simplicity and modularity, which make it easier to implement and manage at scale despite its limitations. We theoretically prove the sub-optimality of sequential post-training and propose a practical joint post-training framework which has theoretical convergence guarantees and empirically outperforms sequential post-training framework, with up to 23% overall performance improvement across multiple LLM evaluation benchmarks, while having minimal computational overhead. Our code is available at https://github.com/heshandevaka/XRIGHT.

cs.LG↗

arXiv2024

SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selection

Fine-tuning on task-specific data to boost downstream performance is a crucial step for leveraging Large Language Models (LLMs). However, previous studies have demonstrated that fine-tuning the models on several adversarial samples or even benign data can greatly comprise the model's pre-equipped alignment and safety capabilities. In this work, we propose SEAL, a novel framework to enhance safety in LLM fine-tuning. SEAL learns a data ranker based on the bilevel optimization to up rank the safe and high-quality fine-tuning data and down rank the unsafe or low-quality ones. Models trained with SEAL demonstrate superior quality over multiple baselines, with 8.5% and 9.7% win rate increase compared to random selection respectively on Llama-3-8b-Instruct and Merlinite-7b models. Our code is available on github https://github.com/hanshen95/SEAL.

cs.LG↗

arXiv2024

On the chemical and kinematic signatures of the resonances of the Galactic bar as revealed by the LAMOST-APOGEE red clump stars

The Milky Way is widely considered to exhibit features of a rotational bar or quadrupole bar. In either case, the feature of the resonance of the Galactic bar should be present in the properties of the chemistry and kinematics, over a large area of the disk. With a sample of over 170,000 red clump (RC) stars from LAMOST-APOGEE data, we attempt to detect the chemical and kinematic signatures of the resonances of the Galactic bar, within 4.0 $\leq$ $R$ $\leq$ 15.0 kpc and $|Z|$ $\leq$ 3.0 kpc. The measurement of the $\Delta$[Fe/H]/$\Delta|Z|$ $-$ $R$ with subtracted the global profiles trends, shows that the thin and thick disks values are Cor_$\Delta$[Fe/H]/$\Delta|Z|$ = 0.010 $\mathrm{sin}$ (1.598 $R$ + 2.551) and Cor_$\Delta$[Fe/H]/$\Delta|Z|$ = 0.006 $\mathrm{sin}$ (1.258 $R$ $-$ 0.019), respectively. The analysis of the tilt angle of the velocity ellipsoid indicates that the thin and thick disks are accurately described as $\alpha$ = $\alpha_{0}$ arctan (Z/R), with $\alpha_{0}$ = 0.198 $\mathrm{sin}$ (0.853 $R$ + 1.982) + 0.630 and $\alpha_{0}$ = 0.220 $\mathrm{sin}$ (0.884 $R$ + 2.012) + 0.679 for thin and thick disks, respectively. These periodic oscillations in Cor_$\Delta$[Fe/H]/$\Delta|Z|$ and $\alpha_{0}$ with $R$ appear in both thin and thick disks, are the most likely chemical and kinematic signatures of the resonance of the Galactic bar. The difference in the phase of the functions of the fitted periodic oscillations for the thin and thick disks may be related to the presence of a second Galactic bar.

astro-ph.GA↗

arXiv2024

Mapping the Chemo-dynamics of the Galactic disk using the LAMOST and APOGEE red clump stars

A detailed measurement is made of the metallicity distributions, kinematics and dynamics of the thin and thick disks, across a large disk volume (5.0 $\leq$ $R$ $\leq$ 15.0 kpc and $|Z|$ $\leq$3.0 kpc), by using the LAMOST-APOGEE red clump stars. The metallicity distributions results show that the radial metallicity gradient $\Delta$[Fe/H]/$\Delta$R of the thin disk weakens with $|Z|$ from $-$0.06 dex kpc$^{-1}$ at around $|Z|$ $<$ 0.25 kpc to $-$0.02 dex kpc$^{-1}$ at around $|Z|$ $>$ 2.75 kpc, while the thick disk displays a global weak positive $\Delta$[Fe/H]/$\Delta$R, generally weaker than 0.01 dex kpc$^{-1}$. The vertical metallicity gradient $\Delta$[Fe/H]/$\Delta|Z|$ weakened steadily from $-$0.36 dex kpc$^{-1}$ at $R$ $\sim$ 5.5 kpc to $-$0.05 dex kpc$^{-1}$ at around R $>$ 11.5 kpc for the thin disk, while the thick disk presents an almost constant value (nearly $-$0.06 $\sim$ $-$0.08 dex kpc$^{-1}$) for all the $R$ bins. These results indicate the contribution of the radial migration to the disk evolution, and the obvious north-south asymmetry in [Fe/H] may be linked to the disk warp and/or the disk perturbation events. The oscillations of the corrected $\Delta$[Fe/H]/$\Delta|Z|$ with $R$ are likely because of the resonances with the Galactic Bar. Our detailed measurements of $\Delta$V$_{\phi}$/$\Delta$[Fe/H] indicate an "inside-out" and "upside-down" star formation scenario for the thick disk. The results of eccentricity distributions and [$\alpha$/Fe]--velocity dispersion relations are likely to suggest that the thick disk stars require an obvious contribution from other heating mechanisms such as merger and accretion, or born in the chaotic mergers of gas-rich systems and/or turbulent interstellar medium.

astro-ph.GA↗

arXiv2024

Principled Penalty-based Methods for Bilevel Reinforcement Learning and RLHF

Bilevel optimization has been recently applied to many machine learning tasks. However, their applications have been restricted to the supervised learning setting, where static objective functions with benign structures are considered. But bilevel problems such as incentive design, inverse reinforcement learning (RL), and RL from human feedback (RLHF) are often modeled as dynamic objective functions that go beyond the simple static objective structures, which pose significant challenges of using existing bilevel solutions. To tackle this new class of bilevel problems, we introduce the first principled algorithmic framework for solving bilevel RL problems through the lens of penalty formulation. We provide theoretical studies of the problem landscape and its penalty-based (policy) gradient algorithms. We demonstrate the effectiveness of our algorithms via simulations in the Stackelberg Markov game, RL from human feedback and incentive design.

cs.LG↗

arXiv2024

Joint Unsupervised and Supervised Training for Automatic Speech Recognition via Bilevel Optimization

In this paper, we present a novel bilevel optimization-based training approach to training acoustic models for automatic speech recognition (ASR) tasks that we term {bi-level joint unsupervised and supervised training (BL-JUST)}. {BL-JUST employs a lower and upper level optimization with an unsupervised loss and a supervised loss respectively, leveraging recent advances in penalty-based bilevel optimization to solve this challenging ASR problem with affordable complexity and rigorous convergence guarantees.} To evaluate BL-JUST, extensive experiments on the LibriSpeech and TED-LIUM v2 datasets have been conducted. BL-JUST achieves superior performance over the commonly used pre-training followed by fine-tuning strategy.

cs.CL↗

arXiv2023

Mapping the Galactic disk with the LAMOST and Gaia Red clump sample: VIII: Mapping the kinematics of the Galactic disk using mono-age and mono-abundance stellar populations

We present a comprehensive study of the kinematic properties of the different Galactic disk populations, as defined by the chemical abundance ratios and stellar ages, across a large disk volume (4.5 $\leq$ R $\leq$ 15.0 kpc and $|Z|$ $\leq$ 3.0 kpc), by using the LAMOST-Gaia red clump sample stars. We determine the median velocities for various spatial and population bins, finding large-scale bulk motions, such as the wave-like behavior in radial velocity, the north-south discrepancy in azimuthal velocity and the warp signal in vertical velocity, and the amplitudes and spatial-dependences of those bulk motions show significant variations for different mono-age and mono-abundance populations. The global spatial behaviors of the velocity dispersions clearly show a signal of spiral arms and, a signal of the disk perturbation event within 4 Gyr, as well as the disk flaring in the outer region (i.e., $R \ge 12$ kpc) mostly for young or alpha-poor stellar populations. Our detailed measurements of age/[$\alpha$/Fe]-velocity dispersion relations for different disk volumes indicate that young/$\alpha$-poor populations are likely originated from dynamically heated by both giant molecular clouds and spiral arms, while old/$\alpha$-enhanced populations require an obvious contribution from other heating mechanisms such as merger and accretion, or born in the chaotic mergers of gas-rich systems and/or turbulent interstellar medium.

astro-ph.GA↗