arXiv subjects

Han Lin

Publications and source records attributed to Han Lin.

At least 19 recordsLinked to original sources

arXiv2026

GRIP: Granular Reward-Guided Parameter Interpolation for Efficient Reasoning

Reasoning-oriented large language models often achieve strong problem-solving performance by generating long chains of thought, but this behavior substantially increases inference cost and latency. In contrast, instruction-tuned models tend to answer more concisely, yet often lack comparable reasoning ability. This accuracy-efficiency mismatch motivates a lightweight approach that combines the strengths of both models without full model retraining. In this paper, we propose GRIP (Granular Reward-guided Interpolation of Parameters), a reward-guided parameter interpolation framework for efficient reasoning. Given a reasoning model and an instruction model with identical architectures, GRIP assigns learnable interpolation ratios to individual modules and optimizes only these ratios while keeping both source models frozen. The interpolation ratios are trained with a reward signal that favors responses that are both correct and concise. Experiments show that GRIP achieves a better accuracy-efficiency trade-off than fixed or search-based merging baselines and further reveals module-wise fusion patterns associated with efficient reasoning.

cs.CL↗

arXiv2026

OPD+: Rethinking the Advantage Design for On-Policy Distillation

On-policy distillation (OPD) is a widely used technique to transfer capabilities from capable teacher language models to the base student models, and can be formulated in a reinforcement learning style objective using student generated rollouts. Yet, despite the divergence reward being dependent on student model likelihood, existing works usually adopt a stop gradient design primarily for stability, which makes the resulting advantage estimation questionable. In this work, we provide a generic optimization framework based on f-divergence between the student and teacher, and mathematically revisit whether such design space is valid. We prove that general stop-gradient operation would lead to biased estimates of the reward objective and corresponding gradient for general divergence functions. We propose OPD+, the corrected version of OPD that demonstrates improved performance over the baseline KL approach and also supports the choice of various f-divergence. We validate our findings on mathematical reasoning and tool-use benchmarks.

cs.LG↗

arXiv2026

Seeing Isn't Knowing: Do VLMs Know When Not to Answer Spatial Questions (and Why)?

Spatial reasoning is a fundamental capability for vision-language models (VLMs) deployed in real-world environments. However, visual observations are inherently limited representations of a 3D world: occlusion can render objects invisible, and perspective can make geometric properties misleading. Despite this, existing spatial reasoning benchmarks typically assume that observations are sufficient and reliable, focusing on whether models produce correct answers rather than whether they recognize when a question cannot be answered and what additional observations would be needed. In this work, we challenge this assumption by constructing a controlled evaluation framework, SpatialUncertain, and introducing two types of observation challenges: (1) occlusion, which hides target information, and (2) perspective ambiguity, which produces misleading visual cues. For each configuration, we design spatial questions that are answerable under clean observations but require abstention under the introduced challenges. We further evaluate whether models can identify which additional viewpoints would resolve perspective ambiguity. Our results across a diverse set of frontier open- and closed-source VLMs reveal two consistent failure modes. First, models are prone to overconfident answering, attempting to solve spatial reasoning tasks even when visual evidence is incomplete or misleading, with average accuracy around 30\% under occlusion and below 10\% under perspective ambiguity. Second, even when additional views are available, some models perform near random chance in identifying which would provide reliable evidence. Together, our findings call for moving beyond answer correctness toward evaluating whether models know when to abstain and how to seek reliable evidence.

cs.CV↗

arXiv2026

PhyMotion: Structured 3D Motion Reward for Physics-Grounded Human Video Generation

Generating realistic human motion is a central yet unsolved challenge in video generation. While reinforcement learning (RL)-based post-training has driven recent gains in general video quality, extending it to human motion remains bottlenecked by a reward signal that cannot reliably score motion realism. Existing video rewards primarily rely on 2D perceptual signals, without explicitly modeling the 3D body state, contact, and dynamics underlying articulated human motion, and often assign high scores to videos with floating bodies or physically implausible movements. To address this, we propose PhyMotion, a structured, fine-grained motion reward that grounds recovered 3D human trajectories in a physics simulator and evaluates motion quality along multiple dimensions of physical feasibility. Concretely, we recover SMPL body meshes from generated videos, retarget them onto a humanoid in the MuJoCo physics simulator, and evaluate the resulting motion along three axes: kinematic plausibility, contact and balance consistency, and dynamic feasibility. Each component provides a continuous and interpretable signal tied to a specific aspect of motion quality, allowing the reward to capture which aspects of motion are physically correct or violated. Experiments show that PhyMotion achieves stronger correlation with human judgments than existing reward formulations. These gains carry over to RL-based post-training, where optimizing PhyMotion leads to larger and more consistent improvements than optimizing existing rewards, improving motion realism across both autoregressive and bidirectional video generators under both automatic metrics and blind human evaluation (+68 Elo gain). Ablations show that the three axes provide complementary supervision signals, while the reward preserves overall video generation quality with only modest training overhead.

cs.CV↗

arXiv2026

Unveiling contrasting impacts of heat mitigation and adaptation policies on U.S. internal migration

While climate-induced population migration has received rising attention, the role played by human climate endeavors remains underexplored. Here, we combine machine learning with attribution mapping to analyze the impacts of 4,713 heat-related policies (HPs) on 11,177 migration flows between U.S. counties. We find that heat adaptation policies (APs) and heat mitigation policies (MPs) have significant and opposing impacts on internal migration: APs reduce out-migration, while MPs increase it. These policies have heterogeneous effects on migration among policy types. Behavioral and cultural MPs at origins lead to a 0.24%-0.68% (95% confidence interval) increase in annual outflows per policy, whereas behavioral and cultural APs at destinations elevate outflows of origins by 0.11%-1.55% (95% confidence interval). Migration patterns are nonlinearly moderated by income, ageing, education, and racial diversity of both origin and destination counties. Ageing rates have the most noticeable U-shaped relationship in shaping migration responses to behavioral and cultural MPs at origins, and inverted U-shapes for institutional MPs at origins and nature-based MPs at destinations. These findings offer critical insights for policymakers on how HPs influence migration as global warming and policy interventions persist.

econ.GN↗

arXiv2026

JoyAI-LLM Flash: Advancing Mid-Scale LLMs with Token Efficiency

We introduce JoyAI-LLM Flash, an efficient Mixture-of-Experts (MoE) language model designed to redefine the trade-off between strong performance and token efficiency in the sub-50B parameter regime. JoyAI-LLM Flash is pretrained on a massive corpus of 20 trillion tokens and further optimized through a rigorous post-training pipeline, including supervised fine-tuning (SFT), Direct Preference Optimization (DPO), and large-scale reinforcement learning (RL) across diverse environments. To improve token efficiency, JoyAI-LLM Flash strategically balances \emph{thinking} and \emph{non-thinking} cognitive modes and introduces FiberPO, a novel RL algorithm inspired by fibration theory that decomposes trust-region maintenance into global and local components, providing unified multi-scale stability control for LLM policy optimization. To enhance architectural sparsity, the model comprises 48B total parameters while activating only 2.7B parameters per forward pass, achieving a substantially higher sparsity ratio than contemporary industry leading models of comparable scale. To further improve inference throughput, we adopt a joint training-inference co-design that incorporates dense Multi-Token Prediction (MTP) and Quantization-Aware Training (QAT). We release the checkpoints for both JoyAI-LLM-48B-A3B Base and its post-trained variants on Hugging Face to support the open-source community.

cs.CL↗

arXiv2026

V-Co: A Closer Look at Visual Representation Alignment via Co-Denoising

Pixel-space diffusion has recently re-emerged as a strong alternative to latent diffusion, enabling high-quality generation without pretrained autoencoders. However, standard pixel-space diffusion models receive relatively weak semantic supervision and are not explicitly designed to capture high-level visual structure. Recent representation-alignment methods (e.g., REPA) suggest that pretrained visual features can substantially improve diffusion training, and visual co-denoising has emerged as a promising direction for incorporating such features into the generative process. However, existing co-denoising approaches often entangle multiple design choices, making it unclear which are truly essential. We therefore present V-Co, a systematic study of visual co-denoising in a unified JiT-based framework. This controlled setting allows us to isolate the ingredients that make visual co-denoising effective. Our study reveals two main ingredients. First, co-denoising benefits from preserving feature-specific computation while enabling flexible cross-stream interaction, which leads to a fully dual-stream architecture together with a structurally defined unconditional prediction for classifier-free guidance. Second, it requires both stronger semantic supervision and proper cross-stream calibration, which we realize through a perceptual-drifting hybrid loss and RMS-based feature rescaling. Together, these findings yield a simple recipe for visual co-denoising. Experiments on ImageNet-256 show that, at comparable model sizes, V-Co outperforms the underlying pixel-space diffusion baseline and strong prior pixel-diffusion methods while using fewer training epochs, offering practical guidance for future representation-aligned generative models.

cs.CV↗

arXiv2026

Spectral Dataset of Stripped-Envelope Supernovae from the Tsinghua Supernova Group

The extent of envelope stripping in the progenitor stars is directly reflected in the diversity of spectral features observed in stripped-envelope supernovae (SESNe). Through extensive spectral observation and analysis, we aim to clarify the statistical differences between the subclasses of SESNe. The Tsinghua Supernova group obtained 249 optical spectra of 62 SESNe during the years from 2010 to 2020, covering phases from $-$16 to over 190 days relative to maximum light. Most spectra were obtained during the photospheric phases after the supernova explosion. For each spectrum, the pseudo-equivalent widths (pEWs) and blueshift velocities of principal lines were measured. We further investigated the common spectral features by analysing their velocity and strength correlations across all subtypes. We identify the feature near 6200~\AA\ in SNe Ib as H$\mathrm{\alpha}$ through comparison with SNe IIb and Ic, which resolves inconsistent literature interpretations. Our finding reveals prevalent residual hydrogen in SNe Ib, further supporting a continuous stripping sequence from SNe IIb to Ib. We observe a trend in increasing velocity among different subtypes of stripped-envelope SNe, with SNe IIb exhibiting the lowest line velocities, followed by Ib, Ic, and Ic-BL. Typically, the O~I lines in SNe Ic/Ic-BL are stronger than those seen in SNe IIb/Ib. In nebular phases, the [Ca II] emission dominates over [O I] in SNe IIb/Ib while [O I] is stronger in SNe Ic, including the He-rich SN 2016coi. This spectral dichotomy implies that progenitors of SNe Ic (BL) have more massive CO cores and hence higher initial masses.

astro-ph.HE↗

arXiv2026

RUMAD: Reinforcement-Unifying Multi-Agent Debate

Multi-agent debate (MAD) systems leverage collective intelligence to enhance reasoning capabilities, yet existing approaches struggle to simultaneously optimize accuracy, consensus formation, and computational efficiency. Static topology methods lack adaptability to task complexity variations, while external LLM-based coordination risks introducing privileged knowledge that compromises debate neutrality. This work presents RUMAD (Reinforcement-Unifying Multi-Agent Debate), a novel framework that formulates dynamic communication topology control in MAD as a reinforcement learning (RL) problem. RUMAD employs a content-agnostic observation scheme that captures high-level debate dynamics avoiding access to raw agent reasoning content. RUMAD uses a multi-objective reward to model solution quality, cohesion and efficiency. A PPO-trained controller dynamically adjusts edge weights in the communication graph, while a dual-threshold mechanism enables fine-grained control over both agent activation and information visibility. Experimental evaluation across MMLU, GSM8K, and GPQA benchmarks demonstrates that RUMAD achieves substantial efficiency gains, reducing token costs by over 80\%, while still improving reasoning accuracy compared to single LLM model and multiple MAD baselines. Notably, RUMAD trained exclusively on MMLU exhibits robust zero-shot generalization to out-of-domain (OOD) tasks, indicating that the learned communication strategies capture task-independent principles of effective multi-agent coordination. These results establish RUMAD as a efficient and robust approach for deploying multi-agent reasoning application with practical resource constraints.

cs.AI↗

arXiv2026

AnchorWeave: World-Consistent Video Generation with Retrieved Local Spatial Memories

Maintaining spatial world consistency over long horizons remains a central challenge for camera-controllable video generation. Existing memory-based approaches often condition generation on globally reconstructed 3D scenes by rendering anchor videos from the reconstructed geometry in the history. However, reconstructing a global 3D scene from multiple views inevitably introduces cross-view misalignment, as pose and depth estimation errors cause the same surfaces to be reconstructed at slightly different 3D locations across views. When fused, these inconsistencies accumulate into noisy geometry that contaminates the conditioning signals and degrades generation quality. We introduce AnchorWeave, a memory-augmented video generation framework that replaces a single misaligned global memory with multiple clean local geometric memories and learns to reconcile their cross-view inconsistencies. To this end, AnchorWeave performs coverage-driven local memory retrieval aligned with the target trajectory and integrates the selected local memories through a multi-anchor weaving controller during generation. Extensive experiments demonstrate that AnchorWeave significantly improves long-term scene consistency while maintaining strong visual quality, with ablation and analysis studies further validating the effectiveness of local geometric conditioning, multi-anchor control, and coverage-driven retrieval.

cs.CV↗

arXiv2026

Hinge Regression Tree: A Newton Method for Oblique Regression Tree Splitting

Oblique decision trees combine the transparency of trees with the power of multivariate decision boundaries, but learning high-quality oblique splits is NP-hard, and practical methods still rely on slow search or theory-free heuristics. We present the Hinge Regression Tree (HRT), which reframes each split as a non-linear least-squares problem over two linear predictors whose max/min envelope induces ReLU-like expressive power. The resulting alternating fitting procedure is exactly equivalent to a damped Newton (Gauss-Newton) method within fixed partitions. We analyze this node-level optimization and, for a backtracking line-search variant, prove that the local objective decreases monotonically and converges; in practice, both fixed and adaptive damping yield fast, stable convergence and can be combined with optional ridge regularization. We further prove that HRT's model class is a universal approximator with an explicit $O(\delta^2)$ approximation rate, and show on synthetic and real-world benchmarks that it matches or outperforms single-tree baselines with more compact structures.

cs.LG↗

arXiv2026

Remember Me, Not Save Me: A Collective Memory System for Evolving Virtual Identities in Augmented Reality

This paper presents "Remember Me, Not Save Me," an AR & AI system enabling virtual citizens to develop personality through collective dialogue. Core innovations include: Dynamic Collective Memory (DCM) model with narrative tension mechanisms for handling contradictory memories; State-Reflective Avatar for ambient explainability; and Geo-Cultural Context Anchoring for local identity. Deployed at the 2024 Jinan Biennale, the system demonstrated stable personality emergence (ISTP type via Apply Magic Sauce analysis) from over 2,500 public interactions. We provide a framework for designing evolving digital entities that transform collective memory into coherent identity.

cs.HC↗

arXiv2026

Pre-Supernova Eruptions Triggered by Sudden Energy Deposition in Low-Mass Core-Collapse Supernova Progenitors

In low-mass core-collapse supernova (CCSN) progenitors, nuclear burning beyond oxygen can become explosive under degenerate conditions, triggering eruptive mass loss before the final explosion. We investigate such pre-SN eruptions using \texttt{SNEC} hydrodynamic simulations and realistic stellar models, parameterizing the nuclear energy deposition as a fraction of the binding energy of the combined He layer and H-rich envelope. For the lowest-mass model (9 $M_\odot$), the ejecta mass ($M_{\rm ej}$) scales with the energy gained by the H-rich envelope via a power law (index$\sim$3.5). Across 9-10 $M_\odot$, this relation shows limited scatter within a factor of $\sim$2.6, enabling an estimation of the gained energy from $M_{\rm ej}$. The shock passage also flattens the bound envelope, which can affect the SN light curve morphology and provide another diagnostic for the eruption. Then, we compute the associated precursor light curves for the 9 $M_\odot$ model with the multi-group radiative-transfer code \texttt{STELLA}. These signals are typically faint, with bolometric luminosities of $\sim10^{39}$ erg s$^{-1}$ lasting hundreds of days. Their cool black-body spectra make them brighter in the infrared, yet several magnitudes fainter than observed pre-SN precursors at the threshold for full envelope ejection. To aid future studies, we make our post-eruption stellar profiles and precursor light curves publicly available.

astro-ph.HE↗

arXiv2026

E^2-LLM: Bridging Neural Signals and Interpretable Affective Analysis

Emotion recognition from electroencephalography (EEG) signals remains challenging due to high inter-subject variability, limited labeled data, and the lack of interpretable reasoning in existing approaches. While recent multimodal large language models (MLLMs) have advanced emotion analysis, they have not been adapted to handle the unique spatiotemporal characteristics of neural signals. We present E^2-LLM (EEG-to-Emotion Large Language Model), the first MLLM framework for interpretable emotion analysis from EEG. E^2-LLM integrates a pretrained EEG encoder with Qwen-based LLMs through learnable projection layers, employing a multi-stage training pipeline that encompasses emotion-discriminative pretraining, cross-modal alignment, and instruction tuning with chain-of-thought reasoning. We design a comprehensive evaluation protocol covering basic emotion prediction, multi-task reasoning, and zero-shot scenario understanding. Experiments on the dataset across seven emotion categories demonstrate that E^2-LLM achieves excellent performance on emotion classification, with larger variants showing enhanced reliability and superior zero-shot generalization to complex reasoning scenarios. Our work establishes a new paradigm combining physiological signals with LLM reasoning capabilities, showing that model scaling improves both recognition accuracy and interpretable emotional understanding in affective computing.

cs.LG↗

arXiv2025

Deep-learning-enabled inverse design of large-scale metasurfaces with full-wave accuracy

Recent advances in meta-optics have enabled diverse functionalities in compact optical devices; however, conventional forward design approaches become inadequate as device complexity and scale grow. Inverse design offers a powerful alternative but often requires massive computational resources and neglects mutual coupling effects. Here, we propose and experimentally validate a deep-learning-enabled framework for rapid inverse design of large-scale, aperiodic metasurfaces with full-wave accuracy.The framework integrates an inverse design network responsible that maps target near-field responses to metasurface geometries in a non-iterative and scalable manner. A lightweight forward prediction network, integrated as a full-wave solver surrogate within the framework, enables efficient end-to-end training of the inverse design network while capturing mutual coupling effects by considering both local and neighboring geometries.The framework's effectiveness is experimentally verified through a multi-foci metalens and a holographic metasurface. This framework enables the inverse design from micrometer to centimeter scales (> 20k{\lambda}), with near-field responses discrepancies less than 3% compared to full-wave solvers at subwavelength (< {\lambda}/10) resolution.Moreover, it is generalizable to metasurfaces of arbitrary size and operates efficiently without high-performance resources, overcoming the computational bottlenecks of previous inverse design methods.

physics.optics↗

arXiv2025

Exploring MLLM-Diffusion Information Transfer with MetaCanvas

Multimodal learning has rapidly advanced visual understanding, largely via multimodal large language models (MLLMs) that use powerful LLMs as cognitive cores. In visual generation, however, these powerful core models are typically reduced to global text encoders for diffusion models, leaving most of their reasoning and planning ability unused. This creates a gap: current multimodal LLMs can parse complex layouts, attributes, and knowledge-intensive scenes, yet struggle to generate images or videos with equally precise and structured control. We propose MetaCanvas, a lightweight framework that lets MLLMs reason and plan directly in spatial and spatiotemporal latent spaces and interface tightly with diffusion generators. We empirically implement MetaCanvas on three different diffusion backbones and evaluate it across six tasks, including text-to-image generation, text/image-to-video generation, image/video editing, and in-context video generation, each requiring precise layouts, robust attribute binding, and reasoning-intensive control. MetaCanvas consistently outperforms global-conditioning baselines, suggesting that treating MLLMs as latent-space planners is a promising direction for narrowing the gap between multimodal understanding and generation.

cs.CV↗

arXiv2025

Planning with Sketch-Guided Verification for Physics-Aware Video Generation

Recent video generation approaches increasingly rely on planning intermediate control signals such as object trajectories to improve temporal coherence and motion fidelity. However, these methods mostly employ single-shot plans that are typically limited to simple motions, or iterative refinement which requires multiple calls to the video generator, incuring high computational cost. To overcome these limitations, we propose SketchVerify, a training-free, sketch-verification-based planning framework that improves motion planning quality with more dynamically coherent trajectories (i.e., physically plausible and instruction-consistent motions) prior to full video generation by introducing a test-time sampling and verification loop. Given a prompt and a reference image, our method predicts multiple candidate motion plans and ranks them using a vision-language verifier that jointly evaluates semantic alignment with the instruction and physical plausibility. To efficiently score candidate motion plans, we render each trajectory as a lightweight video sketch by compositing objects over a static background, which bypasses the need for expensive, repeated diffusion-based synthesis while achieving comparable performance. We iteratively refine the motion plan until a satisfactory one is identified, which is then passed to the trajectory-conditioned generator for final synthesis. Experiments on WorldModelBench and PhyWorldBench demonstrate that our method significantly improves motion quality, physical realism, and long-term consistency compared to competitive baselines while being substantially more efficient. Our ablation study further shows that scaling up the number of trajectory candidates consistently enhances overall performance.

cs.CV↗

arXiv2025

Error-Driven Scene Editing for 3D Grounding in Large Language Models

Despite recent progress in 3D-LLMs, they remain limited in accurately grounding language to visual and spatial elements in 3D environments. This limitation stems in part from training data that focuses on language reasoning rather than spatial understanding due to scarce 3D resources, leaving inherent grounding biases unresolved. To address this, we propose 3D scene editing as a key mechanism to generate visual counterfactuals that mitigate these biases through fine-grained spatial manipulation, without requiring costly scene reconstruction or large-scale 3D data collection. Furthermore, to make these edits targeted and directly address the specific weaknesses of the model, we introduce DEER-3D, an error-driven framework that diagnoses grounding failures and generates targeted counterfactual training supervision via a structured "Decompose, Diagnose, Edit, and Retrain" loop. Specifically, given a grounding failure, DEER-3D first identifies the predicate-level error (e.g., attribute or spatial relation). It then performs minimal predicate-aligned scene edits, such as recoloring or repositioning, and constructs aligned question-answer pairs that explicitly target the failed predicate, forming targeted counterfactual training examples. We evaluate our editing pipeline across multiple benchmarks for 3D grounding and scene understanding tasks, consistently demonstrating improvements across all grounding datasets through iterative refinement (4-6% gains). DEER-3D underscores the effectiveness of targeted, error-driven scene editing in bridging linguistic reasoning with spatial grounding in 3D LLMs.

cs.CV↗