Searcharxiv⌕ Search

arXiv subjects

Yi Ge

Publications and source records attributed to Yi Ge.

14 recordsLinked to original sources
arXiv2026

RoleFlow: Self-Evolving Multi-Role Agentic Collaboration via Structural Adaptation

Large language model (LLM) based multi-agent systems often rely on predefined collaboration structures, limiting their adaptability to evolving task requirements. Although recent approaches enable dynamic orchestration, trajectory-level rewards they apply assign one outcome to each decision, ofen struggling to identify the contributions of individual structural changes. We propose RoleFlow, a self-evolving multi-role agentic system that refines collaboration structures through feedback-driven atomic edits and on-demand role synthesis. To make structural adaptation learnable, we introduce Single-Edit Counterfactual Policy Optimization (SECPO), which evaluates each edit against its no-edit counterpart at the same chain state and assigns targeted credit to edit-specific tokens. Across seven benchmarks spanning four domains, RoleFlow achieves the optimal average performance, surpassing the strongest competitor by 4.7 points. It leads on all three multi-hop QA benchmarks by 3.2-7.5 F1 points and also shows supremacy on mathematical reasoning, web interaction, and code generation. Further analyses demonstrate cross-benchmark transferability, training-data efficiency, and cross-executor generalization.

cs.AI↗
arXiv2026

SAFE-Mem: Secure Agentic Memory with Fine-Grained Authority and Governance-Aware Retrieval

Large language model (LLM) agents increasingly rely on persistent memory to support long-term reasoning and personalized interactions. However, information collected from external sources can introduce security risks when untrusted claims (i.e., atomic statements) are later treated as authorized preferences, policies, or permissions. Existing defenses typically rely on content-based filtering or coarse-grained provenance tracking, which either fail to prevent unauthorized memory use or block legitimate information, sacrificing memory utility. To address these limitations, we propose SAFE-Mem, a fine-grained memory governance framework that improves security while preserving legitimate memory use. First, SAFE-Mem annotates each claim with three governance labels (i.e., kind, authority, and scope) to capture its provenance and permitted use. Second, governance-aware retrieval matches query requirements with claim permissions to block unauthorized influence while retaining legitimate information. Third, a governance-enabled memory graph connects source sentences, claims, and entities to support provenance tracking and efficient retrieval. Across four LLM backbones on MPBench and LoCoMo, SAFE-Mem attains the lowest attack success among defenses that retain legitimate memory, at least 77% below the strongest of them, and the highest answer accuracy of all evaluated defenses, improving it by up to 15 points. The code is available at https://github.com/LVenum/SAFE-Mem.

cs.CR↗
arXiv2026

FlexUMM: Flexible Representation Learning for Unified Multimodal Models

Unified multimodal models (UMMs) can integrate understanding and generation within a single backbone. However, it is commonly known that these two capabilities can cause conflicting representations, leading to a central question: how representation and computation should be shared across functionalities? Existing UMMs often rely on a largely fixed structure, where different inputs and functionalities follow the same depth and parameter-sharing pattern. We argue that this assumption is too restrictive: more effective unification requires sharing the computation that supports cross-functional communication while specializing the capacity that becomes task- or input-specific. We introduce FlexUMM, a flexible UMM framework that dynamically decides when to preserve shared computation, when to route specialized capacity, and when to stop further processing. Evidence shows that this flexible allocation improves UMMs along two complementary axes. It improves accuracy by reducing harmful competition between functionalities, and it improves efficiency by avoiding unnecessary computation once an input representation is sufficiently mature. Beyond simple architectural unification from the pre-trained models, FlexUMM works as a model-agnostic, functional plug-in to make existing UMMs more unified, stronger, and efficient. Code is available at: https://github.com/AIFrontierLab/TorchUMM/tree/main/src/umm/post_training/flexumm.

cs.AI↗
arXiv2026

ANTE: Anticipate, Then Execute -- Learning Internal World Models for Coding Agents from Verified Beliefs

Rapid advances in large language models have made software engineering agents capable of tackling complex coding tasks. Within a containerized environment, a coding agent receives tool responses or execution results, and determines its later actions by those earlier results. However, most existing coding agents never look ahead: their policies do not anticipate the consequences of actions, and the terminal reward barely attributes a failure to certain earlier steps, so an ill-judged edit can corrupt the repository state irreversibly before any signal arrives. Recent code world models aim to address these challenges, but they are either trained and then frozen before reinforcement learning (RL), or reduced to an auxiliary next-token loss on the environment's outputs. We present ANTE (Anticipate, Then Execute), an agentic RL framework that verifies the policy's pre-action predictions by real execution and turns them into per-step credit for its actions and trust-or-abstain decisions. With an internal belief stated before each action, Belief-Anchored Policy Optimization (BAPO) turns the settled predictions into per-step advantages by anchoring credit on the policy's own forecasts. To turn the trained forecasts into better actions at inference, thereby facilitating effective test-time scaling, ANTE-TTO (test-time optimization) selects among candidate actions without an external scorer. ANTE raises Qwen3-32B from 25.0% to 57.2% pass@1 on SWE-bench Verified, 6.2% above a matched GRPO baseline trained without the belief state; its world model's accuracy climbs from 44% to 69% during RL, and the learned foresight capability transfers across harnesses and out-of-distribution domains. In summary, we aim to build a code world model that knows what has happened, what holds now, and what the next action will achieve.

cs.LG↗
arXiv2026

TokenRouter: Efficient Serving System for Token-Level LLM Routing

Large language model (LLM) routing distributes inference work across different models, advancing the cost-quality Pareto frontier of LLM serving. While coarse-grained routing at the session or query level has been widely adopted in production systems, recent algorithmic work shows that fine-grained token-level routing can yield substantial efficiency and quality gains. However, efficiently serving token-level routed inference poses significant challenges to existing systems. Built on single-LLM assumptions, current systems suffer from severe step desynchronization and frequent batch admission delays under token-level routing, and they also impose high implementation complexity on developers. To address these challenges, we design TokenRouter, an efficient and developer-friendly serving system for token-level routed LLM inference. TokenRouter follows the principle of request-centric programming, model-centric execution: developers describe routing logic from the perspective of a single request, while the runtime launches a subserver for each LLM and dispatches requests asynchronously. Each subserver employs a delayed-batching scheduler, whose optimal hyperparameters are derived from a mathematical throughput model of the system. Across diverse routing algorithms, workloads, and model pairs, TokenRouter achieves 2.01-64.15x higher decoding throughput than existing systems, substantially advancing the serving efficiency of token-level LLM routing. Our code is available at https://github.com/thu-nics/TokenRouter.

cs.CL↗
arXiv2026

Multi-Crit: Benchmarking Multimodal Judges on Pluralistic Criteria-Following

Large multimodal models (LMMs) are increasingly adopted as judges in multimodal evaluation systems due to their strong instruction following and consistency with human preferences. However, their ability to follow diverse, fine-grained evaluation criteria remains underexplored. We develop Multi-Crit, a benchmark for evaluating multimodal judges on their capacity to follow pluralistic criteria and produce reliable criterion-level judgments. Covering both open-ended generation and verifiable reasoning tasks, Multi-Crit is built through a rigorous data curation pipeline that gathers challenging response pairs with multi-criterion human annotations. It further introduces three novel metrics for systematically assessing pluralistic adherence, criterion-switching flexibility, and the ability to recognize criterion-level preference conflicts. Comprehensive analysis of 25 LMMs reveals that 1) proprietary models still struggle to maintain consistent adherence to pluralistic criteria--especially in open-ended evaluation; 2) open-source models lag further behind in flexibly following diverse criteria; and 3) critic fine-tuning with holistic judgment signals enhances visual grounding but fails to generalize to pluralistic criterion-level judgment. Additional analyses on reasoning fine-tuning, test-time scaling, and boundary consistency between open-source and proprietary models further probe the limits of current multimodal judges. As a pioneering study, Multi-Crit lays the foundation for building reliable and steerable multimodal AI evaluation.

cs.CV↗
arXiv2025

Study on Dynamic Matching and Dynamic Characteristics of Hydrostatic Transmission System of Forklift Truck

In the fields of agricultural machinery, construction equipment, and special-purpose vehicles, hydrostatic transmission (HST) drive systems have witnessed a significant increase in market penetration. With the rapid development of intelligent and environmentally friendly trends, higher requirements have been imposed on HST control systems regarding operational efficiency, control accuracy, and adaptability to working conditions. This study focuses on a forklift truck drive system equipped with a dual variable-displacement pump-motor assembly. A mathematical model of the HST system was constructed, with detailed analysis of each stage during vehicle acceleration. A power matching strategy was proposed to enhance dynamic performance, whose core concept consisted of three elements: 1) correlating the driver's pedal signal with the target engine speed of the diesel engine, 2) dynamically adjusting motor displacement to maximize acceleration capability, and 3) coordinating pump displacement variations. Furthermore, a dynamic compensation mechanism for control variables was introduced to improve system responsiveness. The controller was evaluated under multiple operating conditions, with results demonstrating superior acceleration performance, rapid response, and precise control characteristics. These results validate the effectiveness of the proposed control strategy and its robust load adaptability in hydrostatic drive systems.

hep-ex↗
arXiv2025

R2R: Efficiently Navigating Divergent Reasoning Paths with Small-Large Model Token Routing

Large Language Models (LLMs) achieve impressive reasoning capabilities at the cost of substantial inference overhead, posing substantial deployment challenges. Although distilled Small Language Models (SLMs) significantly enhance efficiency, their performance suffers as they fail to follow LLMs' reasoning paths. Luckily, we reveal that only a small fraction of tokens genuinely diverge reasoning paths between LLMs and SLMs. Most generated tokens are either identical or exhibit neutral differences, such as minor variations in abbreviations or expressions. Leveraging this insight, we introduce **Roads to Rome (R2R)**, a neural token routing method that selectively utilizes LLMs only for these critical, path-divergent tokens, while leaving the majority of token generation to the SLM. We also develop an automatic data generation pipeline that identifies divergent tokens and generates token-level routing labels to train the lightweight router. We apply R2R to combine R1-1.5B and R1-32B models from the DeepSeek family, and evaluate on challenging math, coding, and QA benchmarks. With an average activated parameter size of 5.6B, R2R surpasses the average accuracy of R1-7B by 1.6x, outperforming even the R1-14B model. Compared to R1-32B, it delivers a 2.8x wall-clock speedup with comparable performance, advancing the Pareto frontier of test-time scaling efficiency. Our code is available at https://github.com/thu-nics/R2R.

cs.CL↗
arXiv2025

QeRL: Beyond Efficiency -- Quantization-enhanced Reinforcement Learning for LLMs

We propose QeRL, a Quantization-enhanced Reinforcement Learning framework for large language models (LLMs). While RL is essential for LLMs' reasoning capabilities, it is resource-intensive, requiring substantial GPU memory and long rollout durations. QeRL addresses these issues by combining NVFP4 quantization with Low-Rank Adaptation (LoRA), accelerating rollout phase of RL while reducing memory overhead. Beyond efficiency, our findings show that quantization noise increases policy entropy, enhancing exploration, and enabling the discovery of better strategies during RL. To further optimize exploration, QeRL introduces an Adaptive Quantization Noise (AQN) mechanism, which dynamically adjusts noise during training. Experiments demonstrate that QeRL delivers over 1.5 times speedup in the rollout phase. Moreover, this is the first framework to enable RL training of a 32B LLM on a single H100 80GB GPU, while delivering overall speedups for RL training. It also achieves faster reward growth and higher final accuracy than 16-bit LoRA and QLoRA, while matching the performance of full-parameter fine-tuning on mathematical benchmarks such as GSM8K (90.8%) and MATH 500 (77.4%) in the 7B model. These results establish QeRL as an efficient and effective framework for RL training in LLMs.

cs.LG↗
arXiv2023

Weighted Bayesian Gaussian Mixture Model for Roadside LiDAR Object Detection

Background modeling is widely used for intelligent surveillance systems to detect moving targets by subtracting the static background components. Most roadside LiDAR object detection methods filter out foreground points by comparing new data points to pre-trained background references based on descriptive statistics over many frames (e.g., voxel density, number of neighbors, maximum distance). However, these solutions are inefficient under heavy traffic, and parameter values are hard to transfer from one scenario to another. In early studies, the probabilistic background modeling methods widely used for the video-based system were considered unsuitable for roadside LiDAR surveillance systems due to the sparse and unstructured point cloud data. In this paper, the raw LiDAR data were transformed into a structured representation based on the elevation and azimuth value of each LiDAR point. With this high-order tensor representation, we break the barrier to allow efficient high-dimensional multivariate analysis for roadside LiDAR background modeling. The Bayesian Nonparametric (BNP) approach integrates the intensity value and 3D measurements to exploit the measurement data using 3D and intensity info entirely. The proposed method was compared against two state-of-the-art roadside LiDAR background models, computer vision benchmark, and deep learning baselines, evaluated at point, object, and path levels under heavy traffic and challenging weather. This multimodal Weighted Bayesian Gaussian Mixture Model (GMM) can handle dynamic backgrounds with noisy measurements and substantially enhances the infrastructure-based LiDAR object detection, whereby various 3D modeling for smart city applications could be created.

cs.CV↗
arXiv2020

Importance-Aware Semantic Segmentation in Self-Driving with Discrete Wasserstein Training

Semantic segmentation (SS) is an important perception manner for self-driving cars and robotics, which classifies each pixel into a pre-determined class. The widely-used cross entropy (CE) loss-based deep networks has achieved significant progress w.r.t. the mean Intersection-over Union (mIoU). However, the cross entropy loss can not take the different importance of each class in an self-driving system into account. For example, pedestrians in the image should be much more important than the surrounding buildings when make a decisions in the driving, so their segmentation results are expected to be as accurate as possible. In this paper, we propose to incorporate the importance-aware inter-class correlation in a Wasserstein training framework by configuring its ground distance matrix. The ground distance matrix can be pre-defined following a priori in a specific task, and the previous importance-ignored methods can be the particular cases. From an optimization perspective, we also extend our ground metric to a linear, convex or concave increasing function $w.r.t.$ pre-defined ground distance. We evaluate our method on CamVid and Cityscapes datasets with different backbones (SegNet, ENet, FCN and Deeplab) in a plug and play fashion. In our extenssive experiments, Wasserstein loss demonstrates superior segmentation performance on the predefined critical classes for safe-driving.

cs.CV↗
arXiv2019

Unimodal-uniform Constrained Wasserstein Training for Medical Diagnosis

The labels in medical diagnosis task are usually discrete and successively distributed. For example, the Diabetic Retinopathy Diagnosis (DR) involves five health risk levels: no DR (0), mild DR (1), moderate DR (2), severe DR (3) and proliferative DR (4). This labeling system is common for medical disease. Previous methods usually construct a multi-binary-classification task or propose some re-parameter schemes in the output unit. In this paper, we target on this task from the perspective of loss function. More specifically, the Wasserstein distance is utilized as an alternative, explicitly incorporating the inter-class correlations by pre-defining its ground metric. Then, the ground metric which serves as a linear, convex or concave increasing function w.r.t. the Euclidean distance in a line is explored from an optimization perspective. Meanwhile, this paper also proposes of constructing the smoothed target labels that model the inlier and outlier noises by using a unimodal-uniform mixture distribution. Different from the one-hot setting, the smoothed label endues the computation of Wasserstein distance with more challenging features. With either one-hot or smoothed target label, this paper systematically concludes the practical closed-form solution. We evaluate our method on several medical diagnosis tasks (e.g., Diabetic Retinopathy and Ultrasound Breast dataset) and achieve state-of-the-art performance.

eess.IV↗
arXiv2019

A Linear LMP Model for Active and Reactive Power with Power Loss

Pricing the reactive power is more necessary than ever before because of the increasing challenge of renewable energy integration on reactive power balance and voltage control. However, reactive power price is hard to be efficiently calculated because of the non-linear nature of optimal AC power flow equation. This paper proposes a linear model to calculate active and reactive power LMP simultaneously considering power loss. Firstly, a linearized AC power flow equation is proposed based on an augmented Generation Shift Distribution Factors (GSDF) matrix. Secondly, a linearized LMP model is derived using GSDF and loss factors. The formulation of LMP is further decomposed into four components: energy, congestion, voltage limitation and power loss. Finally, an iterate algorithm is proposed for calculating LMP with the proposed model. The performance of the proposed model is validated by the IEEE-118 bus system.

eess.SY↗