arXiv subjects

Jinghan He

Publications and source records attributed to Jinghan He.

At least 19 recordsLinked to original sources

arXiv2026

Ace-Skill: Bootstrapping Multimodal Agents with Prioritized and Clustered Evolution

Self-evolving agents present a promising path toward continual adaptation by distilling task interactions into reusable knowledge artifacts. In practice, this paradigm remains hindered by two coupled bottlenecks: data inefficiency, where costly rollout effort is disproportionately spent on low-value samples rather than informative ones, and knowledge interference, where heterogeneous knowledge stored in shared repositories leads to noisy retrieval and task-misaligned guidance. Together, these issues form a self-reinforcing failure loop in which uninformative rollouts yield noisy knowledge, which in turn degrades subsequent rollouts. In this work, we introduce Ace-Skill, a co-evolutionary framework that jointly optimizes rollout allocation and knowledge organization for self-evolving multimodal agents. Specifically, Ace-Skill combines aprioritized sampler with lazy-decay proficiency tracking to focus rollouts on informative and insufficiently mastered samples, and a clustered organizer that semantically clusters knowledge for cleaner retrieval and more reliable adaptation. By improving sampling and organization together, Ace-Skill turns self-evolution into a virtuous cycle in which more informative rollouts produce higher-quality knowledge that supports stronger subsequent rollouts. Across four multimodal tool-use benchmarks, Ace-Skill delivers strong gains (e.g., +35.46% relative improvement in Avg@4 accuracy), enabling an opensource 35B MoE model to match or surpass proprietary models. The acquired knowledge also transfers effectively in a zero-shot manner to smaller 9B and 4B models, allowing resource-constrained agents to inherit advanced capabilities without additional training. The code has been publicly available at https://github.com/AMAP-ML/Ace-Skill.

cs.AI↗

arXiv2026

R-Diverse: Mitigating Diversity Illusion in Self-Play LLM Training

Self-play bootstraps LLM reasoning through an iterative Challenger-Solver loop: the Challenger is trained to generate questions that target the Solver's capabilities, and the Solver is optimized on the generated data to expand its reasoning skills. However, existing frameworks like R-Zero often exhibit non-sustained improvement, where early gains degrade as self-play continues. We identify a key failure mode, Diversity Illusion, where the Solver's training signals appear diverse yet collapse into recurring underlying patterns. It manifests as (1) Local Diversity Illusion, where diversity is enforced only within-batch, inducing cross-iteration mode cycling; and (2) Surface Diversity Illusion, where questions vary superficially but require near-identical reasoning skills. To mitigate them, we propose R-Diverse with two aligned innovations: Memory-Augmented Penalty (MAP), which uses a persistent memory bank to discourage recycling across iterations, and Skill-Aware Measurement (SAM), which evaluates diversity by the reasoning skills exercised rather than surface variation of questions. Across 10 math and general reasoning benchmarks, R-Diverse sustains gains over more iterations and consistently outperforms prior self-play methods. Code is available at https://github.com/Gengsheng-Li/R-Diverse.

cs.LG↗

arXiv2026

Active Zero: Self-Evolving Vision-Language Models through Active Environment Exploration

Self-play has enabled large language models to autonomously improve through self-generated challenges. However, existing self-play methods for vision-language models rely on passive interaction with static image collections, resulting in strong dependence on initial datasets and inefficient learning. Without the ability to actively seek visual data tailored to their evolving capabilities, agents waste computational effort on samples that are either trivial or beyond their current skill level. To address these limitations, we propose Active-Zero, a framework that shifts from passive interaction to active exploration of visual environments. Active-Zero employs three co-evolving agents: a Searcher that retrieves images from open-world repositories based on the model's capability frontier, a Questioner that synthesizes calibrated reasoning tasks, and a Solver refined through accuracy rewards. This closed loop enables self-scaffolding auto-curricula where the model autonomously constructs its learning trajectory. On Qwen2.5-VL-7B-Instruct across 12 benchmarks, Active-Zero achieves 53.97 average accuracy on reasoning tasks (5.7% improvement) and 59.77 on general understanding (3.9% improvement), consistently outperforming existing self-play baselines. These results highlight active exploration as a key ingredient for scalable and adaptive self-evolving vision-language systems.

cs.CV↗

arXiv2025

MLLM-CTBench: A Benchmark for Continual Instruction Tuning with Reasoning Process Diagnosis

Continual instruction tuning(CIT) during the post-training phase is crucial for adapting multimodal large language models (MLLMs) to evolving real-world demands. However, the progress is hampered by the lack of benchmarks with rigorous, protocol-consistent evaluation. To bridge this gap, we introduce MLLM-CTBench, a comprehensive benchmark for CIT of MLLMs, covering seven challenging tasks across six diverse domains. MLLM-CTBench makes three key contributions. First, we establish a multidimensional evaluation framework that jointly assesses final-answer accuracy and process-level reasoning quality, where Chain-of-Thought (CoT) traces serve as an observable signal to diagnose catastrophic forgetting beyond answer-only evaluation. Second, we conduct a large-scale evaluation of continual learning methods by systematically assessing eight representative algorithms from four major families under a unified protocol across task orders, providing actionable insights for algorithm design. Third, we expand the scope from Supervised Fine-Tuning (SFT) to Reinforcement Fine-Tuning (RFT) in CIT. By investigating GRPO, an on-policy RL algorithm that stabilizes updates through explicit KL-divergence control to a prior policy, we aim to analyze how this mechanism affects cross-task knowledge retention. Our experiments yield several findings:(1) Process-level reasoning quality is often more resilient to catastrophic forgetting than final-answer accuracy, and forgetting is primarily driven by degradation in domain knowledge. (2) Model capability is critical factor influencing continual learning outcomes, with stronger baseline models exhibiting greater resistance to catastrophic forgetting. (3) On-policy RFT (GRPO), with its inherent KL control, achieves more stable cross-task retention than SFT. While removing KL control can amplify forgetting despite potential gains on new ones.

cs.CL↗

arXiv2025

Steering LVLMs via Sparse Autoencoder for Hallucination Mitigation

Large vision-language models (LVLMs) have achieved remarkable performance on multimodal tasks. However, they still suffer from hallucinations, generating text inconsistent with visual input, posing significant risks in real-world applications. Existing approaches to address this issue focus on incorporating external knowledge bases, alignment training, or decoding strategies, all of which require substantial computational cost and time. Recent works try to explore more efficient alternatives by adjusting LVLMs' internal representations. Although promising, these methods may cause hallucinations to be insufficiently suppressed or lead to excessive interventions that negatively affect normal semantics. In this work, we leverage sparse autoencoders (SAEs) to identify semantic directions closely associated with faithfulness or hallucination, extracting more precise and disentangled hallucination-related representations. Our analysis demonstrates that interventions along the identified faithful direction can mitigate hallucinations, while those along the hallucinatory direction can exacerbate them. Building on these insights, we propose Steering LVLMs via SAE Latent Directions (SSL), a plug-and-play method based on SAE-derived latent directions to mitigate hallucinations in LVLMs. Extensive experiments demonstrate that SSL significantly outperforms existing decoding approaches in mitigating hallucinations, while maintaining transferability across different model architectures with negligible additional time overhead. The code is available at https://github.com/huazhenglin2003/SSL.

cs.CV↗

arXiv2024

Cracking the Code of Hallucination in LVLMs with Vision-aware Head Divergence

Large vision-language models (LVLMs) have made substantial progress in integrating large language models (LLMs) with visual inputs, enabling advanced multimodal reasoning. Despite their success, a persistent challenge is hallucination-where generated text fails to accurately reflect visual content-undermining both accuracy and reliability. Existing methods focus on alignment training or decoding refinements but primarily address symptoms at the generation stage without probing the underlying causes. In this work, we investigate the internal mechanisms driving hallucination in LVLMs, with an emphasis on the multi-head attention module. Specifically, we introduce Vision-aware Head Divergence (VHD), a metric that quantifies the sensitivity of attention head outputs to visual context. Based on this, our findings reveal the presence of vision-aware attention heads that are more attuned to visual information; however, the model's overreliance on its prior language patterns is closely related to hallucinations. Building on these insights, we propose Vision-aware Head Reinforcement (VHR), a training-free approach to mitigate hallucination by enhancing the role of vision-aware attention heads. Extensive experiments demonstrate that our method achieves superior performance compared to state-of-the-art approaches in mitigating hallucinations, while maintaining high efficiency with negligible additional time overhead.

cs.CL↗

arXiv2024

SEEKR: Selective Attention-Guided Knowledge Retention for Continual Learning of Large Language Models

Continual learning (CL) is crucial for language models to dynamically adapt to the evolving real-world demands. To mitigate the catastrophic forgetting problem in CL, data replay has been proven a simple and effective strategy, and the subsequent data-replay-based distillation can further enhance the performance. However, existing methods fail to fully exploit the knowledge embedded in models from previous tasks, resulting in the need for a relatively large number of replay samples to achieve good results. In this work, we first explore and emphasize the importance of attention weights in knowledge retention, and then propose a SElective attEntion-guided Knowledge Retention method (SEEKR) for data-efficient replay-based continual learning of large language models (LLMs). Specifically, SEEKR performs attention distillation on the selected attention heads for finer-grained knowledge retention, where the proposed forgettability-based and task-sensitivity-based measures are used to identify the most valuable attention heads. Experimental results on two continual learning benchmarks for LLMs demonstrate the superiority of SEEKR over the existing methods on both performance and efficiency. Explicitly, SEEKR achieves comparable or even better performance with only 1/10 of the replayed data used by other methods, and reduces the proportion of replayed data to 1%.

cs.CL↗

arXiv2023

Continual Instruction Tuning for Large Multimodal Models

Instruction tuning is now a widely adopted approach to aligning large multimodal models (LMMs) to follow human intent. It unifies the data format of vision-language tasks, enabling multi-task joint training. However, vision-language tasks are constantly being created in practice. Instead of always re-training LMMs when new tasks arrive, continual learning offers flexibility for models to continually and efficiently exploit the evolving data. This work aims to explore the following two questions: 1) Do LMMs still suffer from catastrophic forgetting in continual instruction tuning? 2) Are the existing three classes of continual learning methods still applicable to the continual instruction tuning of LMMs? An extensive study is conducted to address the above questions. First, we establish the first benchmark in this setting and reveal that catastrophic forgetting is still observed when continually instruction-tuning LMMs. However, the multi-task joint instruction tuning can facilitate the model's continual learning ability and mitigate forgetting. Second, we integrate and adapt classic continual learning methods to our context, demonstrating the efficacy of data replay and model expansion strategies across diverse scenarios. In contrast, regularization-based methods only perform well on models that have been jointly instruction-tuned on multiple tasks. Third, we delve into the correlation and forgetting dynamics between vision-language task pairs and propose task-similarity-informed regularization and model expansion methods for continual instruction tuning of LMMs. Experimental results show that our approach consistently boosts the model's performance.

cs.LG↗

arXiv2022

Resilience-Motivated Distribution System Restoration Considering Electricity-Water-Gas Interdependency

A major outage in the electricity distribution system may affect the operation of water and natural gas supply systems, leading to an interruption of multiple services to critical customers. Therefore, enhancing resilience of critical infrastructures requires joint efforts of multiple sectors. In this paper, a distribution system service restoration method considering the electricity-water-gas interdependency is proposed. The objective is to provide electricity, water, and natural gas supplies to critical customers in the desired ratio according to their needs after an extreme event. The operational constraints of electricity, water, and natural gas networks are considered. The characteristics of electricity-driven coupling components, including water pumps and gas compressors, are also modeled. Relaxation techniques are applied to nonconvex constraints posed by physical laws of those networks. Consequently, the restoration problem is formulated as a mixed-integer second-order cone program, which can readily be solved by the off-the-shelf solvers. The proposed method is validated by numerical simulations on electricity-water-gas integrated systems, developed based on benchmark models of the subsystems. The results indicate that considering the interdependency refines the allocation of limited generation resources and demonstrate the exactness of the proposed convex relaxation.

eess.SY↗

arXiv2021

Multifunctional photoresponsive organic molecule for electric field sensing and modulation

Organic molecules with nonlinear optical behavior have advanced a wide range of fields spanning from integrated photonics to biological imaging. With advances in microscopy, an emerging application is multifunctional nonlinear organic imaging agents. Unlike conventional imaging probes which simply emit light through single or multi photon processes, multifunctional materials allow systems to be simultaneously imaged and controlled. In this work, we report a multifunctional molecular probe for modulating and reporting electric fields. The probe molecule consists of two distinct functional modules which are connected by a long alkyl chain. The electric field detector module relies on the two-photon (2p) imaging agent and photo-induced electron transfer (PeT) dye, TPE. Two-photon imaging agents have demonstrated less damage and larger penetration depths in cells and live tissue imaging. The electric field modulator module relies on the organic photoconductor, NAI. To reduce cross-talk and optimize absorption and emission wavelengths, the molecular structure is first studied using density functional theory modeling, and then the multi-functional molecular probe is synthesized. The photophysical, photoconductivity, and biotoxicity of the probe molecule are studied in a range of solvents and solid state, and the results agree with the theoretical predictions. Specifically, 2p excitation in a biocompatible solvent is demonstrated, the photoconductivity is rapid and reversible, and the material has low cytotoxicity. Additionally, the entire system is optically controlled, including signal read-out, and the two modules can be operated simultaneously or individually. This work sets the stage for modulation and detection of bioelectric fields in a range of cell and tissue types.

physics.chem-ph↗

arXiv2021

Stretchable optical diffraction grating from poly(acrylic acid)/polyethylene oxide stereocomplex

Optical gratings are a key component in many spectroscopy, communications, and imaging systems. While initially static elements, advances in optical materials have enabled dynamically tunable gratings to be designed. One common tuning strategy is relying on mechanical deformation of the grating pitch to modify the diffraction pattern. To date, most mechanically adaptive optical gratings consist of a hybrid system where rigid moieties are patterned on an elastomeric substrate. In the present work, we demonstrate an all-polymer tunable grating that is fabricated using replica molding from the poly(acrylic acid) (PAA)/polyethylene oxide (PEO) polymer stereocomplex. PAA/PEO pristine films exhibit excellent optical transmittance at or above 80% from 500 nm to 1400 nm and stretchability over 800% strain. The experimental studies on the changes of diffraction mode distances with respect to the applied strains agree well with the finite-difference time-domain (FDTD) theoretical modeling.

physics.optics↗

arXiv2020

Cascaded Stokes and anti-Stokes laser based on an optical resonator with a self-assembled organic monolayer

Due to their high circulating intensities, ultra-high quality factor dielectric whispering-gallery mode resonators have enabled the development of low threshold Raman microlasers. Subsequently, other Raman-related phenomena, such as cascaded stimulated Raman scattering (CSRS) and stimulated anti-Stokes Raman scattering (SARS), were observed. While low threshold frequency conversion and generation have clear applications, CSRS and SARS have been limited by the low Raman gain. In this work, the surface of a silica resonator is modified with an organic monolayer, increasing the Raman gain. Up to four orders of CSRS is observed with sub-mW input power, and the SARS efficiency is improved by three orders of magnitude compared to previous studies with hybrid resonators.

physics.optics↗

arXiv2020

Nonlinear nanophotonic devices in the Ultraviolet to Visible wavelength range

Although the first lasers invented operated in the visible, the first on-chip devices were optimized for near-infrared (IR) performance driven by demand in telecommunications. However, as the applications of integrated photonics has broadened, the wavelength demand has as well, and we are now returning to the visible (Vis) and pushing into the ultraviolet (UV). This shift has required innovations in device design and in materials as well as leveraging nonlinear behavior to reach these wavelengths. This review discusses the key nonlinear phenomena that can be used as well as presents several emerging material systems and devices that have reached the UV-Vis wavelength range.

physics.optics↗

arXiv2020

Optically tunable microresonator using an azobenzene monolayer

Photoswitchable organic molecules can undergo reversible structural changes with an external light stimulus. These optically controlled molecules have been used in the development of smart polymers, optical writing of grating films, and even controllable in-vivo drug release. Being the simplest class of photoswitches in terms of structure, azobenzenes have become the most ubiquitous, well-characterized, and implemented organic molecular switch. Given their predictable response, they are ideally suited to create an all-optically controlled switch. However, fabricating a monolithic optical device comprised solely from azobenzene while maintaining the photoswitching functionality is challenging. In this work, we combine integrated photonics with optically switchable organic molecules to create an optically controlled integrated device. A silica toroidal resonant cavity is functionalized with a monolayer of an azobenzene derivative. After functionalization, the loaded cavity Q is above 100,000. When 450 nm light is coupled into cavity resonance, the azobenzene isomerizes from trans-isomer to cis-isomer, inducing a refractive index change. Because the resonant wavelength of the cavity is governed by the index, the resonant wavelength changes in parallel. At the probe wavelength of 1300 nm, the wavelength shift is determined by the duration and intensity of the 450 nm light and the density of azobenzene functional groups on the device surface, providing multiple control mechanisms. Using this photoswitchable device, resonance frequency tuning as far as sixty percent of the cavity free spectral range in the near-IR is demonstrated. The kinetics of the tuning are in agreement with spectroscopic and ellipsometry measurements coupled with finite element method calculations.

physics.optics↗

arXiv2020

On the Radiality Constraints for Distribution System Restoration and Reconfiguration Problems

Radiality constraints are involved in both distribution system restoration and reconfiguration problems. However, a set of widely used radiality constraints, i.e., the spanning tree (ST) constraints, has its limitations which have not been well recognized. In this letter, the limitation of the ST constraints is analyzed and an effective set of constraints, referred to as the single-commodity flow constraints, is presented. Furthermore, a combined set of constraints is proposed and case studies indicate that the combined constraints can gain computational efficiency in the reconfiguration problem. Recommendations on the use of radiality constraints are also provided.

eess.SY↗

arXiv2020

All-optical reversible controls of integrated photonics by self-assembled azobenzene

The next frontier in photonics will rely on the synergistic combination of disparate material systems. One unique organic molecule is azobenzene. This molecule can reversibly change conformations when optically excited in the blue (trans-to-cis) or mid-IR (cis-to-trans). Here, we demonstrate SiO2 optical resonators modified with a monolayer of azobenzene-containing 4-(4-diethylaminophenylazo)pyridine (Aazo) with quality factors over 106. Using a pair of lasers, the molecule is reversibly flipped between molecular conformations, inducing resonant wavelength shifts, and multiple switching cycles are demonstrated. The magnitude of the shift scales with the relative surface density of Aazo. The experimental data agrees with theoretical modeling.

physics.optics↗

arXiv2020

Emerging material systems for integrated optical Kerr frequency combs

The experimental realization of a Kerr frequency comb represented the convergence of research in materials, physics, and engineering, and this symbiotic relationship continues to underpin efforts in comb innovation today. While the initial focus developing cavity-based frequency combs relied on existing microresonator architectures and classic optical materials, in recent years, this trend has been disrupted. This paper reviews the latest achievements in frequency comb generation using resonant cavities, placing them within the broader historical context of the field. After presenting well-established material systems and device designs, the emerging materials and device architectures are examined. Specifically, the unconventional material systems as well as atypical device designs that have enabled tailored dispersion profiles and improved comb performance are compared to the current state of art. The remaining challenges and future outlook for the field of cavity-based frequency combs is evaluated.

physics.optics↗

arXiv2019

A Two-Layer Distributed Control Method for Islanded Networked Microgrid Systems

This paper presents a two-layer, four-level distributed control method for networked microgrid (NMG) systems, taking into account the proprietary nature of microgrid (MG) owners. The proposed control architecture consists of a MG-control layer and a NMG-control layer. In the MG layer, the primary and distrib-uted secondary control realize accurate power sharing among distributed generators (DGs) and the frequency/voltage reference following within each MG. In the NMG layer, the tertiary control enables regulation of the power flowing through the point of common coupling (PCC) of each MG in a decentralized manner. Furthermore, the distributed quaternary control restores system frequency and critical bus voltage to their nominal values and ensures accurate power sharing among MGs. A small-signal dynamic model is developed to evaluate dynamic performance of NMG systems with the proposed control method. Time-domain simulations as well as experiments on NMG test systems are performed to validate the effectiveness of the proposed method.

math.OC↗