arXiv subjects

Yiming Hu

Publications and source records attributed to Yiming Hu.

At least 19 recordsLinked to original sources

arXiv2026

SkillForge: Evolving Verifiable Skills for Reinforcement Learning Agents

Large language model (LLM) agents are trained with reinforcement learning (RL) for complex decision-making tasks. However, most RL-trained agents remain episodic and cannot accumulate reusable knowledge across episodes. Recent skill-based approaches, such as SkillRL, attempt to address this issue by extracting skills from raw trajectories, but treat the skill bank as an append-only repository without verifying whether stored skills remain effective. In this paper, we propose SkillForge, a framework for continuous skill evolution that enables skills to be verified and refined through environment interaction. By making skill usage explicit during agent interaction, RL can directly optimize both environment actions and skill invocation decisions. SkillForge further introduces evidence-based skill verification and multi-pathway skill induction, allowing the skill bank to continuously grow while maintaining its quality. Extensive experiments on ALFWorld, WebShop, and AppWorld show that SkillForge consistently outperforms SkillRL, demonstrating the effectiveness of continuously verified skills in training stronger LLM agents.

cs.CL↗

arXiv2026

LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks

Large language model (LLM) agents increasingly undertake long-horizon tasks that require sustained reasoning, tool use, and revision across many interdependent steps. However, existing agent harnesses maintain task execution, task state, and completion assessment within a growing context, making the state difficult to track and allowing incorrect self-assessments to propagate into later decisions. We reformulate long-horizon execution as a task-state management problem and propose LongHorizon-Harness, which maintains the task state explicitly outside execution and updates it only with facts independently verified from the environment. Its Manage-Execute-Audit(MEA) loop uses a manager to maintain the task state and determine the next subtask, a fresh-context executor to perform it, and a read-only auditor to verify the resulting environment state before the next round. A lightweight AgentAdapter supports interchangeable model and harness backends without modifying their native agent loops. LongHorizon-Harness improves Qwen~3.7-Plus from 51.8% to 80.7% on WeaveBench, from 69.7% to 77.2% on Terminal-Bench~2.1, and from 2.8% to 8.3% on OSWorld~2.0. It also raises Claude Opus~4.7 from 20.0% to 34.3% on an OSWorld2.0 subset, demonstrating consistent gains across models, harnesses, and interaction domains.

cs.CV↗

arXiv2026

Towards Memory-Efficient Autoregressive Video Generation via Instance-Specific Parametric Absorption

Autoregressive (AR) streaming models have emerged as a powerful paradigm for long video generation. However, the linearly growing Key-Value (KV) cache poses a significant bottleneck, leading to memory overload and degraded inference throughput. A common compression method is to drop redundant KV tokens, which often breaks long-range dependencies, resulting in temporal flickering and identity loss. In this paper, we propose Instance-Specific Parametric Absorption (ISPA), a novel framework that shifts the KV cache compression from discarding to distilling. The core idea is to transit a subset of layers from Full-Attention (F-Layers) to memory-efficient Local-Attention (L-Layers) by "absorbing" historical context into the model's weights. Specifically, during a brief warmup phase, ISPA monitors the output discrepancy between global and local attention. At the transition point, we solve a closed-form least-squares problem to compute an instance-specific weight modulation that compensates for the missing history. Experiments across architectures (1.3B to 14B) demonstrate that ISPA can remove up to 50\% of the KV cache with near-lossless visual quality. We hope this perspective encourages future work to explore parametric memory consolidation beyond external token-level cache management for streaming generative models.

cs.CV↗

arXiv2026

Towards High-Resolution Visual Perception via Hierarchical Entity Exploration

High-resolution (HR) image perception remains a key challenge in multimodal large language models (MLLMs), as fine-grained details are often lost when the image is processed as a whole. Existing methods either require training to teach models where to look or heuristically divide the image into fixed regions, both of which struggle to generalize in complex HR scenes. In this work, we propose Hierarchical Entity Exploration (HEE), a training-free and model-agnostic framework that transforms static image understanding into dynamic, query-guided entity exploration. HEE first evaluates each region using a dual scoring mechanism to determine whether it already contains sufficient evidence to answer the question. If not, it applies object detection within the most promising region to extract fine-grained entities, clusters them into coherent subregions, and organizes them into a multi-level semantic hierarchy for deeper exploration. When deeper regions still fail to yield confident answers, a confidence-guided backtracking mechanism revisits alternative paths to ensure adaptive perception. Extensive results show that HEE outperforms training-free methods like ZoomEye and RAP in both accuracy and efficiency on two complex HR benchmarks (Visual Probe and HR-Bench), across different MLLMs such as Qwen2.5-VL and LLaVA-OneVision. Moreover, HEE demonstrates generalization on the MME-RealWorld benchmark.

cs.CV↗

arXiv2026

BlockPilot: Instance-Adaptive Policy Learning for Diffusion-based Speculative Decoding

Speculative decoding accelerates inference by using a lightweight draft model to generate candidate tokens in parallel, and are then verified by the target model, enabling lossless acceleration. Recently, diffusion-based speculative decoding further improves parallelism by generating multiple tokens per forward pass via block-level diffusion, achieving state-of-the-art (SOTA) performance. However, existing methods adopt a fixed inference block size and assume a uniform optimal decoding strategy across all inputs. In this paper, we show that this assumption is suboptimal, as the optimal block size varies across samples and plays a critical role in speculative decoding performance. Moreover, these values exhibit a clear local structure, concentrating around the training block size, which reduces the problem to a low-dimensional and structured decision space. Based on these insights, we propose BlockPilot, a sample-adaptive policy that predicts the optimal block size from the prefilling representation. Specifically, we formulate block size selection as a lightweight policy learning problem and propose an instance-adaptive decision mechanism that predicts the optimal block size based on the representation of the prefilling stage. The prediction is performed only once after prefilling, allowing for seamless integration. Extensive experiments demonstrate that our method is plug-and-play, introduces minimal overhead, and consistently improves efficiency, achieving an acceptance length of 5.92 and a 4.20$\times$ speedup on Qwen3-4B under temperature $T=1$.

cs.CL↗

arXiv2026

CoEvolve: Training LLM Agents via Agent-Data Mutual Evolution

Reinforcement learning for LLM agents is typically conducted on a static data distribution, which fails to adapt to the agent's evolving behavior and leads to poor coverage of complex environment interactions. To address these challenges, we propose CoEvolve, an agent-data mutual evolution framework that enables LLM agents to improve through closed-loop, interaction-driven training. Specifically, CoEvolve extracts feedback signals such as forgetting and uncertainty from rollout trajectories to identify failure-prone interaction patterns, and utilizes them to guide LLM-based task synthesis. The synthesized tasks are validated through environment interaction and utilized to update the data distribution, enabling joint adaptation of the agent and its data. Extensive experiments on AppWorld and BFCL across Qwen2.5-7B, Qwen3-4B, and Qwen3-30B-A3B demonstrate consistent and significant improvements over strong base models, yielding absolute gains of 19.43%, 15.58%, and 18.14%, respectively.

cs.CL↗

arXiv2026

SkillClaw: Let Skills Evolve Collectively with Agentic Evolver

Large language model (LLM) agents such as OpenClaw rely on reusable skills to perform complex tasks, yet these skills remain largely static after deployment. As a result, similar workflows, tool usage patterns, and failure modes are repeatedly rediscovered across users, preventing the system from improving with experience. While interactions from different users provide complementary signals about when a skill works or fails, existing systems lack a mechanism to convert such heterogeneous experiences into reliable skill updates. To address these issues, we present SkillClaw, a framework for collective skill evolution in multi-user agent ecosystems, which treats cross-user and over-time interactions as the primary signal for improving skills. SkillClaw continuously aggregates trajectories generated during use and processes them with an autonomous evolver, which identifies recurring behavioral patterns and translates them into updates to the skill set by refining existing skills or extending them with new capabilities. The resulting skills are maintained in a shared repository and synchronized across users, allowing improvements discovered in one context to propagate system-wide while requiring no additional effort from users. By integrating multi-user experience into ongoing skill updates, SkillClaw enables cross-user knowledge transfer and cumulative capability improvement, and experiments on WildClawBench show that limited interaction and feedback, it significantly improves the performance of Qwen3-Max in real-world agent scenarios.

cs.AI↗

arXiv2026

Thinking with Map: Reinforced Parallel Map-Augmented Agent for Geolocalization

The image geolocalization task aims to predict the location where an image was taken anywhere on Earth using visual clues. Existing large vision-language model (LVLM) approaches leverage world knowledge, chain-of-thought reasoning, and agentic capabilities, but overlook a common strategy used by humans -- using maps. In this work, we first equip the model \textit{Thinking with Map} ability and formulate it as an agent-in-the-map loop. We develop a two-stage optimization scheme for it, including agentic reinforcement learning (RL) followed by parallel test-time scaling (TTS). The RL strengthens the agentic capability of model to improve sampling efficiency, and the parallel TTS enables the model to explore multiple candidate paths before making the final prediction, which is crucial for geolocalization. To evaluate our method on up-to-date and in-the-wild images, we further present MAPBench, a comprehensive geolocalization training and evaluation benchmark composed entirely of real-world images. Experimental results show that our method outperforms existing open- and closed-source models on most metrics, specifically improving Acc@500m from 8.0\% to 22.1\% compared to \textit{Gemini-3-Pro} with Google Search/Map grounded mode.

cs.CV↗

arXiv2025

TrajMoE: Scene-Adaptive Trajectory Planning with Mixture of Experts and Reinforcement Learning

Current autonomous driving systems often favor end-to-end frameworks, which take sensor inputs like images and learn to map them into trajectory space via neural networks. Previous work has demonstrated that models can achieve better planning performance when provided with a prior distribution of possible trajectories. However, these approaches often overlook two critical aspects: 1) The appropriate trajectory prior can vary significantly across different driving scenarios. 2) Their trajectory evaluation mechanism lacks policy-driven refinement, remaining constrained by the limitations of one-stage supervised training. To address these issues, we explore improvements in two key areas. For problem 1, we employ MoE to apply different trajectory priors tailored to different scenarios. For problem 2, we utilize Reinforcement Learning to fine-tune the trajectory scoring mechanism. Additionally, we integrate models with different perception backbones to enhance perceptual features. Our integrated model achieved a score of 51.08 on the navsim ICCV benchmark, securing third place.

cs.CV↗

arXiv2025

When Harmless Words Harm: A New Threat to LLM Safety via Conceptual Triggers

Recent research on large language model (LLM) jailbreaks has primarily focused on techniques that bypass safety mechanisms to elicit overtly harmful outputs. However, such efforts often overlook attacks that exploit the model's capacity for abstract generalization, creating a critical blind spot in current alignment strategies. This gap enables adversaries to induce objectionable content by subtly manipulating the implicit social values embedded in model outputs. In this paper, we introduce MICM, a novel, model-agnostic jailbreak method that targets the aggregate value structure reflected in LLM responses. Drawing on conceptual morphology theory, MICM encodes specific configurations of nuanced concepts into a fixed prompt template through a predefined set of phrases. These phrases act as conceptual triggers, steering model outputs toward a specific value stance without triggering conventional safety filters. We evaluate MICM across five advanced LLMs, including GPT-4o, Deepseek-R1, and Qwen3-8B. Experimental results show that MICM consistently outperforms state-of-the-art jailbreak techniques, achieving high success rates with minimal rejection. Our findings reveal a critical vulnerability in commercial LLMs: their safety mechanisms remain susceptible to covert manipulation of underlying value alignment.

cs.CL↗

arXiv2025

A Method for On-Orbit Calibration of the VLAST-P Electromagnetic Calorimeter

The Very Large Area Gamma-ray Space Telescope Pathfinder (VLAST-P), as the technology validation satellite for the VLAST mission, is designed to observe high-energy solar bursts on orbit. The CsI electromagnetic calorimeter (ECAL) is one of the key sub-detectors of VLAST-P. To investigate the on-orbit energy calibration method of the ECAL, a Geant4-based simulation of VLAST-P was carried out. The results show an energy resolution better than 10% in the 0.1 to 5 GeV range and a linearity deviation below 2%. A dedicated minimum-ionization-particle (MIP) calibration method was developed to ensure accurate energy reconstruction and to monitor detector stability throughout the in-orbit calibration period.

hep-ex↗

arXiv2025

Where and What Matters: Sensitivity-Aware Task Vectors for Many-Shot Multimodal In-Context Learning

Large Multimodal Models (LMMs) have shown promising in-context learning (ICL) capabilities, but scaling to many-shot settings remains difficult due to limited context length and high inference cost. To address these challenges, task-vector-based methods have been explored by inserting compact representations of many-shot in-context demonstrations into model activations. However, existing task-vector-based methods either overlook the importance of where to insert task vectors or struggle to determine suitable values for each location. To this end, we propose a novel Sensitivity-aware Task Vector insertion framework (STV) to figure out where and what to insert. Our key insight is that activation deltas across query-context pairs exhibit consistent structural patterns, providing a reliable cue for insertion. Based on the identified sensitive-aware locations, we construct a pre-clustered activation bank for each location by clustering the activation values, and then apply reinforcement learning to choose the most suitable one to insert. We evaluate STV across a range of multimodal models (e.g., Qwen-VL, Idefics-2) and tasks (e.g., VizWiz, OK-VQA), demonstrating its effectiveness and showing consistent improvements over previous task-vector-based methods with strong generalization.

cs.AI↗

arXiv2025

FTI-TMR: A Fault Tolerance and Isolation Algorithm for Interconnected Multicore Systems

Two-Phase TMR conserves energy by partitioning redundancy operations into two stages and making the execution of the third task copy optional, yet it remains susceptible to permanent faults. Reactive-TMR (R-TMR) counters this by isolating faulty cores, handling both transient and permanent faults. However, the lightweight hardware required by R-TMR not only increases complexity but also becomes a single point of failure itself. To bypass isolated node constraints, this paper proposes a Fault Tolerance and Isolation TMR (FTI-TMR) algorithm for interconnected multicore systems. By constructing a stability metric to identify the most reliable nodes in the system, which then perform periodic diagnostics to isolate permanent faults. Experimental results show that FTI-TMR reduces task workload by approximately 30% compared with baseline TMR while achieving higher permanent fault coverage.

cs.DC↗

arXiv2025

The development of a high granular crystal calorimeter prototype of VLAST

Very Large Area gamma-ray Space Telescope (VLAST) is the next-generation flagship space observatory for high-energy gamma-ray detection proposed by China. The observation energy range covers from MeV to TeV and beyond, with acceptance of 10 m^2sr. The calorimeter serves as a crucial subdetector of VLAST, responsible for high-precision energy measurement and electron/proton discrimination. This discrimination capability is essential for accurately identifying gamma-ray events among the background of charged particles. To accommodate such an extensive energy range, a high dynamic range readout scheme employing dual avalanche photodiodes (APDs) has been developed, achieving a remarkable dynamic range of 10^6. Furthermore, a high granular prototype based on bismuth germanate (BGO) cubic scintillation crystals has been developed. This high granularity enables detailed imaging of the particle showers, improving both energy resolution and particle identification. The prototype's performance is evaluated through cosmic ray testing, providing valuable data for optimizing the final calorimeter design for VLAST.

physics.ins-det↗

arXiv2025

Validation of Fast Mocks Generation for CSST Photometric Survey

Weak lensing has become a powerful tool for probing the matter distribution in the Universe and constraining cosmological parameters. This paper aims to explore the fast mock generation pipeline to obtain the covariance matrix of the $3\times 2$pt analysis for the upcoming China Space Station Telescope (CSST). We adopt the $N_\mathrm{side}$ pipeline, which generates matter distribution with lognormal assumptions, to create full-sky galaxy mocks with certain two-point statistics. We also employ the Markov-Chain Monte Carlo simulation to test the accuracy of the covariance matrix from the mock-generated galaxy catalogue. Our work validates the accuracy of the $3\times 2$pt statistics in both spherical harmonic space and real space. The critical scale below which the fractional error of correlation exceeds 1$\%$ can decrease as the resolution parameter $N_\mathrm{side}$ increases. After excluding certain scales, the covariance matrix from the mock-generated galaxy catalogue can constrain the cosmological parameters with 0.1$\%$ accuracy. This work demonstrates the potential of $\texttt{GLASS}$ for real-space cosmological measurements and highlights the importance of discarding appropriate scales.

astro-ph.CO↗

arXiv2025

Detection for Intermediate-mass Binary Black Holes in Population III Star Clusters with TianQin

Context. Population III star clusters are predicted to form in unenriched dark matter halos. Direct N-body simulation of Pop III clusters implies the possible formation and merger of intermediate-mass binary black holes (IMBBHs). The gravitational wave signals could be detected by space-borne gravitational wave detectors like TianQin. Aims. This study evaluates the potential of TianQin in detecting IMBBHs from Pop III star clusters, focusing on key factors such as the detection horizon, detection number, and Fisher uncertainty. Methods. A Monte Carlo simulation is employed to derive IMBBH sources, utilizing the IMRPhenomD waveform model to simulate catalogs of observable IMBBH mergers. The mass and redshift distributions are derived from direct N-body simulations of IMBBHs in Population III star clusters. Detection numbers are determined by calculating the signal-to-noise ratios (SNR) of the simulated signals and applying thresholds for detection. Fisher uncertainty is obtained through Fisher information matrix analysis. Results. The findings suggest that TianQin could achieve detection numbers within 5 years ranging from 1 in the most pessimistic scenario to 253 in the most optimistic scenario. Furthermore, TianQin can precisely constrain the IMBBH mass with a relative uncertainty of $10^{-6}$, coalescence time $t_c$ within 1 second, and sky location $\bar{\Omega}_S$ within 1 $\rm{deg}^2$. However, the luminosity distance $D_L$ and inclination angle $\iota$ exhibit substantial degeneracies, limiting their precise estimation.

astro-ph.GA↗

arXiv2025

Beam test result and digitization of TaichuPix-3: A Monolithic Active Pixel Sensors for CEPC vertex detector

The Circular Electron-Positron Collider (CEPC), as the next-generation electron-positron collider, is tasked with advancing not only Higgs physics but also the discovery of new physics. Achieving these goals requires high-precision measurements of particles. Taichu seires, Monolithic Active Pixel Sensor (MAPS), a key component of the vertex detector for CEPC was designed to meet the CEPC's requirements. For the geometry of vertex detector is long barrel with no endcap, and current silicon lacks a complete digitization model, precise estimation of cluster size particularly causing by particle with large incident angle is needed. Testbeam results were conducted at the Beijing Synchrotron Radiation Facility (BSRF) to evaluate cluster size dependence on different incident angles and threshold settings. Experimental results confirmed that cluster size increases with incident angle. Simulations using the Allpix$^2$ framework replicated experimental trends at small angles but exhibited discrepancies at large angles, suggesting limitations in linear electric field assumptions and sensor thickness approximations. The results from both testbeam and simulations have provided insights into the performance of the TaichuPix chip at large incident angles, offering a crucial foundation for the establishment of a digital model and addressing the estimation of cluster size in the forward region of the long barrel. Furthermore, it offers valuable references for future iterations of TaichuPix, the development of digital models, and the simulation and estimation of the vertex detector's performance.

physics.ins-det↗

arXiv2024

Consensus-Based Dynamic Task Allocation for Multi-Robot System Considering Payloads Consumption

This paper presents a consensus-based payload algorithm (CBPA) to deal with the condition of robots' capability decrease for multi-robot task allocation. During the execution of complex tasks, robots' capabilities could decrease with the consumption of payloads, which causes a problem that the robot coalition would not meet the tasks' requirements in real time. The proposed CBPA is an enhanced version of the consensus-based bundle algorithm (CBBA) and comprises two primary core phases: the payload bundle construction and consensus phases. In the payload bundle construction phase, CBPA introduces a payload assignment matrix to track the payloads carried by the robots and the demands of multi-robot tasks in real time. Then, robots share their respective payload assignment matrix in the consensus phase. These two phases are iterated to dynamically adjust the number of robots performing multi-robot tasks and the number of tasks each robot performs and obtain conflict-free results to ensure that the robot coalition meets the demand and completes all tasks as quickly as possible. Physical experiment shows that CBPA is appropriate in complex and dynamic scenarios where robots need to collaborate and task requirements are tightly coupled to the robots' payloads. Numerical experiments show that CBPA has higher total task gains than CBBA.

cs.RO↗