Zhipeng Wang

Xiwen Chen, Wenhui Zhu, Jingjing Wang et al.

Aligning Large Language Models (LLMs) with human preferences is often formulated via Direct Preference Optimization (DPO). However, the standard Bradley-Terry instantiation of DPO is limited in modeling common departures from transitivity in human preferences. To address this, recent work has introduced Self-Play Preference Optimization (SPPO), which iteratively refines the policy by training on self-generated win-lose pairs. Our investigation, however, reveals a critical instability in SPPO: the optimization is prone to policy degeneration when the preference oracle assigns overly confident wins to semantically indistinguishable responses. To mitigate this, we propose S-SPPO, a dual-space semantic calibration framework comprising: i) Supervision Calibration via semantic gating, which anneals win rate targets toward the maximum-entropy baseline as semantic overlap increases; and ii) Representation Calibration via latent repulsion to enforce geometric diversity to prevent manifold collapse and maintain latent diversity between chosen and rejected samples. Theoretically, we show that the calibration preserves the constant-sum game structure, facilitating convergence to a Nash Equilibrium. Empirically, S-SPPO avoids the performance degradation seen in prior methods, achieving 52.19% win rate and 47.46% length-controlled win rate on AlpacaEval 2.0 with Llama-3-8B, without using additional human-annotated preferences during training. The code will be available at https://github.com/xiwenc1/s-sppo.

Yuanda Xu, Hejian Sang, Zhengze Zhou et al.

On-policy knowledge distillation (OPD) trains a student on its own rollouts under token-level supervision from a teacher. Not all token positions matter equally, but existing views of token importance are incomplete. We ask a direct question: which tokens carry the most useful learning signal in OPD? Our answer is that informative tokens come from two regions: positions with high student entropy, and positions with low student entropy plus high teacher--student divergence, where the student is overconfident and wrong. Empirically, student entropy is a strong first-order proxy: retaining $50\%$ of tokens with entropy-based sampling matches or exceeds all-token training while reducing peak memory by up to $47\%$. But entropy alone misses a second important region. When we isolate low-entropy, high-divergence tokens, training on fewer than $10\%$ of all tokens nearly matches full-token baselines, showing that overconfident tokens carry dense corrective signal despite being nearly invisible to entropy-only rules. We organize these findings with TIP (Token Importance in on-Policy distillation), a two-axis taxonomy over student entropy and teacher--student divergence, and give a theoretical explanation for why entropy is useful yet structurally incomplete. This view motivates type-aware token selection rules that combine uncertainty and disagreement. We validate this picture across three teacher--student pairs spanning Qwen3, Llama, and Qwen2.5 on MATH-500 and AIME 2024/2025, and on the DeepPlanning benchmark for long-horizon agentic planning, where Q3-only training on $<$$20\%$ of tokens surpasses full-token OPD. Our experiments are implemented by extending the OPD repository https://github.com/HJSang/OPSD_OnPolicyDistillation, which supports memory-efficient distillation of larger models under limited GPU budgets.

Xuanzhao Dong, Wenhui Zhu, Peijie Qiu et al.

Despite their popularity and success, Multimodal Large Language Models (MLLMs) often struggle to interpret images accurately, which limits their reasoning capability in complex scenarios (e.g., high object density and complex background clutter). Prior work mainly addresses this limitation by incorporating explicit visual cues like bounding boxes that require extra annotations. In addition, the resulting low-resolution crops often miss fine-grained details that MLLMs require for accurate reasoning. Therefore, we propose Mags-RL, an Agentic Reinforcement Learning (RL) framework that equips MLLMs with an external super-resolution "magnifying glass" agent for high-resolution fine-grained inspection. Specifically, the model performs two-round reasoning: in the first round, it generates an initial rationale and autonomously identifies regions of interest without relying on additional annotations; in the second round, it invokes a super-resolution agent to crop and upscale those regions, then revisits and verifies its earlier reasoning to produce the final answer. We also introduce a novel curriculum learning strategy that enables data-efficient RL training, needing as few as only 40 training samples to achieve reasonable performance. Experiments on VSR, TallyQA, and GQA subsets show its superior performance against recent strong competing methods, demonstrating high-quality reasoning with precise visual grounding. Code and weights will be released soon.

Xuanzhao Dong, Wenhui Zhu, Xiwen Chen et al.

The advancement of general medical Multimodal Large Language Models (MLLMs) has shown great potential for building conversational assistants to support clinical diagnosis. However, their adaptation to highly specialized domains such as ophthalmology remains underexplored, primarily due to the scarcity of large-scale, domain-specific instruction-tuning data. Existing ophthalmic datasets for conversational agents are often limited in scale and largely rely on images from established public benchmarks, limiting the scalability of ophthalmic MLLMs and their ability to capture real-world clinical complexity. To address this gap, we propose $\textbf{OphIn-Engine}$, an ophthalmology-specific instruction data curation pipeline that constructs high-quality instruction data from open-access ophthalmology web-scale videos. The pipeline integrates multimodal transcription for extracting image-transcript pairs, visual cue separation and scoring for identifying clinically relevant visual descriptions, and instruction synthesis with quality control for generating accurate and diverse clinical dialogues. Using this engine, we introduce $\textbf{OphIn-500K}$, a large-scale multimodal ophthalmology instruction-tuning dataset containing over 500,000 instruction instances and more than 151,000 unique images from over 29,000 video clips, formatted as visual question answering (VQA), multi-turn conversational interactions, and chain-of-thought (CoT) reasoning. Built upon this dataset, we further develop $\textbf{OphIn-VL}$, an ophthalmology-specific MLLM with advanced visual understanding and conversational capabilities. Comprehensive experiments and case studies demonstrate that OphIn-VL achieves superior performance compared with state-of-the-art general medical and domain-specific MLLMs.

Zhihao Ding, Jieming Shi, Shiqi Shen et al.

Graph-level representation learning is important in a wide range of applications. Existing graph-level models are generally built on i.i.d. assumption for both training and testing graphs. However, in an open world, models can encounter out-of-distribution (OOD) testing graphs that are from different distributions unknown during training. A trustworthy model should be able to detect OOD graphs to avoid unreliable predictions, while producing accurate in-distribution (ID) predictions. To achieve this, we present SGOOD, a novel graph-level OOD detection framework. We find that substructure differences commonly exist between ID and OOD graphs, and design SGOOD with a series of techniques to encode task-agnostic substructures for effective OOD detection. Specifically, we build a super graph of substructures for every graph, and develop a two-level graph encoding pipeline that works on both original graphs and super graphs to obtain substructure-enhanced graph representations. We then devise substructure-preserving graph augmentation techniques to further capture more substructure semantics of ID graphs. Extensive experiments against 11 competitors on numerous graph datasets demonstrate the superiority of SGOOD, often surpassing existing methods by a significant margin. The code is available at https://github.com/TommyDzh/SGOOD.

Nanqing Dong, Zhipeng Wang, Jiahao Sun et al.

In the era of deep learning, federated learning (FL) presents a promising approach that allows multi-institutional data owners, or clients, to collaboratively train machine learning models without compromising data privacy. However, most existing FL approaches rely on a centralized server for global model aggregation, leading to a single point of failure. This makes the system vulnerable to malicious attacks when dealing with dishonest clients. In this work, we address this problem by proposing a secure and reliable FL system based on blockchain and distributed ledger technology. Our system incorporates a peer-to-peer voting mechanism and a reward-and-slash mechanism, which are powered by on-chain smart contracts, to detect and deter malicious behaviors. Both theoretical and empirical analyses are presented to demonstrate the effectiveness of the proposed approach, showing that our framework is robust against malicious client-side behaviors.

Wenhui Zhu, Xuanzhao Dong, Xiwen Chen et al.

The rapid deployment of open-source frameworks has significantly advanced the development of modern multi-agent systems. However, expanded action spaces, including uncontrolled privilege exposure and hidden inter-system interactions, pose severe security challenges. Specifically, Indirect Prompt Injections (IPI), which conceal malicious instructions within third-party content, can trigger unauthorized actions such as data exfiltration during normal operations. While current security evaluations predominantly rely on isolated single-turn benchmarks, the systemic vulnerabilities of these agents within complex dynamic environments remain critically underexplored. To bridge this gap, we systematically evaluate six defense strategies against four sophisticated IPI attack vectors across nine LLM backbones. Crucially, we conduct our evaluation entirely within dynamic multi-step tool-calling environments to capture the true attack surface of modern autonomous agents. Moving beyond binary success rates, our multidimensional analysis reveals a pronounced fragility. Advanced injections successfully bypass nearly all baseline defenses, and some surface-level mitigations even produce counterproductive side effects. Furthermore, while agents execute malicious instructions almost instantaneously, their internal states exhibit abnormally high decision entropy. Motivated by this latent hesitation, we investigate Representation Engineering (RepE) as a robust detection strategy. By extracting hidden states at the tool-input position, we revealed that the RepE-based circuit breaker successfully identifies and intercepts unauthorized actions before the agent commits to them, achieving high detection accuracy across diverse LLM backbones. This study exposes the limitations of current IPI defenses and provides a highly practical paradigm for building resilient multi-agent architectures.

Wenhui Zhu, Xiwen Chen, Zhipeng Wang et al.

Long-horizon LLM agents require memory systems that remain accurate under fixed context budgets. However, existing systems struggle with two persistent challenges in long-term dialogue: (i) \textbf{disconnected evidence}, where multi-hop answers require linking facts distributed across time, and (ii) \textbf{state updates}, where evolving information (e.g., schedule changes) creates conflicts with older static logs. We propose AriadneMem, a structured memory system that addresses these failure modes via a decoupled two-phase pipeline. In the \textbf{offline construction phase}, AriadneMem employs \emph{entropy-aware gating} to filter noise and low-information message before LLM extraction and applies \emph{conflict-aware coarsening} to merge static duplicates while preserving state transitions as temporal edges. In the \textbf{online reasoning phase}, rather than relying on expensive iterative planning, AriadneMem executes \emph{algorithmic bridge discovery} to reconstruct missing logical paths between retrieved facts, followed by \emph{single-call topology-aware synthesis}. On LoCoMo experiments with GPT-4o, AriadneMem improves \textbf{Multi-Hop F1 by 15.2\%} and \textbf{Average F1 by 9.0\%} over strong baselines. Crucially, by offloading reasoning to the graph layer, AriadneMem reduces \textbf{total runtime by 77.8\%} using only \textbf{497} context tokens. The code is available at https://github.com/LLM-VLM-GSL/AriadneMem.

Yanxi Chen, Wenhui Zhu, Xiwen Chen et al.

Although Large Audio-Language Models (LALMs) deliver state-of-the-art (SOTA) performance, they frequently suffer from hallucinations, e.g. generating text not grounded in the audio input. We analyze these grounding failures and identify a distinct taxonomy: Event Omission, False Event Identity, Temporal Relation Error, and Quantitative Temporal Error. To address this, we introduce the AHA (Audio Hallucination Alignment) framework. By leveraging counterfactual hard negative mining, our pipeline constructs a high-quality preference dataset that forces models to distinguish strict acoustic evidence from linguistically plausible fabrications. Additionally, we establish AHA-Eval, a diagnostic benchmark designed to rigorously test these fine-grained temporal reasoning capabilities. We apply this data to align Qwen2.5-Omni. The resulting model, Qwen-Audio-AHA, achieves a 13.7% improvement on AHA-Eval. Crucially, this benefit generalizes beyond our diagnostic set. Our model shows substantial gains on public benchmarks, including 1.3% on MMAU-Test and 1.6% on MMAR, outperforming latest SOTA methods. The model and dataset are open-sourced at https://github.com/LLM-VLM-GSL/AHA.

Wei-Rui Chen, Vignesh Kothapalli, Ata Fatahibaarzi et al.

Distilling the capabilities from a large reasoning model (LRM) to a smaller student model often involves training on substantial amounts of reasoning data. However, knowledge distillation (KD) over lengthy sequences with prompt (P), chain-of-thought (CoT), and answer (A) sections makes the process computationally expensive. In this work, we investigate how the allocation of supervision across different sections (P, CoT, A) affects student performance. Our analysis shows that selective KD over only the CoT tokens can be effective when the prompt and answer information is encompassed by it. Building on this insight, we establish a truncation protocol to quantify computation-quality tradeoffs as a function of sequence length. We observe that beyond a specific length, longer training sequences provide marginal returns for downstream performance but require substantially higher memory and FLOPs. To this end, training on only the first $50\%$ of tokens of every training sequence can retain, on average, $\approx91\%$ of full-sequence performance on math benchmarks while reducing training time, memory usage, and FLOPs by about $50\%$ each. Codes are available at https://github.com/weiruichen01/distilling-the-essence.

Nanqing Dong, Jiahao Sun, Zhipeng Wang et al.

Federated learning (FL) is a promising way to allow multiple data owners (clients) to collaboratively train machine learning models without compromising data privacy. Yet, existing FL solutions usually rely on a centralized aggregator for model weight aggregation, while assuming clients are honest. Even if data privacy can still be preserved, the problem of single-point failure and data poisoning attack from malicious clients remains unresolved. To tackle this challenge, we propose to use distributed ledger technology (DLT) to achieve FLock, a secure and reliable decentralized Federated Learning system built on blockchain. To guarantee model quality, we design a novel peer-to-peer (P2P) review and reward/slash mechanism to detect and deter malicious clients, powered by on-chain smart contracts. The reward/slash mechanism, in addition, serves as incentives for participants to honestly upload and review model parameters in the FLock system. FLock thus improves the performance and the robustness of FL systems in a fully P2P manner.

Jelena Markovic-Voronov, Wenhui Zhu, Bo Long et al.

We introduce a principled probabilistic framework for reward-guided decoding in large language models, addressing the limitations of standard decoding methods that optimize token-level likelihood rather than sequence-level quality. Our method defines a reward-augmented target distribution over complete sequences by combining model transition probabilities with prefix-dependent reward potentials. Importantly, the approach is training-free: it leaves model weights unchanged and instead modifies the inference distribution via reward potentials, with all gains arising purely from inference-time sampling. To sample from this distribution, we develop Sequential Monte Carlo algorithms, including a computationally efficient prefix-only variant and a lookahead variant whose intermediate targets match the exact marginals of the full sequence distribution. The framework also integrates resample-move updates with Metropolis-Hastings rejuvenation and supports block-wise generation, subsuming common decoding strategies such as temperature sampling and power-tempered objectives. Empirical results across three 7B models show significant gains. On code generation (HumanEval), our method improves base performance by up to 54.9% and surpasses the strongest sampling baselines by 9.1%-15.3%. On mathematical reasoning (MATH500), it achieves gains of up to 8.8%. Notably, it reaches 87.8% on HumanEval and 78.4% on MATH500 with Qwen2.5-7B, consistently outperforming the reinforcement learning method GRPO.

Xiwen Chen, Jingjing Wang, Wenhui Zhu et al.

Black-box knowledge distillation for large language models presents a strict trade-off. Simple off-policy methods (e.g., sequence-level knowledge distillation) struggle to correct the student's inherent errors. Fully on-policy methods (e.g., Generative Adversarial Distillation) solve this via adversarial training but introduce well-known training instability and crippling computational overhead. To address this dilemma, we propose SODA (Semi On-policy Distillation with Alignment), a highly efficient alternative motivated by the inherent capability gap between frontier teachers and much smaller base models. Because a compact student model's natural, zero-shot responses are almost strictly inferior to the powerful teacher's targets, we can construct a highly effective contrastive signal simply by pairing the teacher's optimal response with a one-time static snapshot of the student's outputs. This demonstrates that exposing the small student to its own static inferior behaviors is sufficient for high-quality distribution alignment, eliminating the need for costly dynamic rollouts and fragile adversarial balancing. Extensive evaluations across four compact Qwen2.5 and Llama-3 models validate this semi on-policy paradigm. SODA matches or outperforms the state-of-the-art methods on 15 out of 16 benchmark results. More importantly, it achieves this superior distillation quality while training 10 times faster, consuming 27% less peak GPU memory, and completely eliminating adversarial instability.

Zelin Li, Qin Wang, Zhipeng Wang

The x402 protocol revives the HTTP 402 Payment Required status code to enable web-native micropayments across APIs, content, and agents. It combines synchronous HTTP authorization with asynchronous blockchain settlement and introduces a cross-layer attack surface absent from conventional web and on-chain payments. In this paper, we formally analyze x402 and empirically show that it is vulnerable in both design and implementation. We present five concrete attacks that reveal weaknesses in authorization, binding, replay protection, and web-layer handling, showing that x402 is vulnerable across multiple stages of the payment workflow. We validate these attacks through a reproducible testbed on local chains, Base Sepolia, and live endpoints and further audit three open-source SDKs and endpoints. Our results show that all five attacks are practical and can cause either unpaid service or paid-but-denied outcomes. We also propose practical mitigations.

Chengjin Wang, Yanmin Zhou, Zheng Yan et al.

In unknown cluttered environments with densely stacked objects, the free-motion space is extremely barren, posing significant challenges to motion planners. Collision-free planning methods often suffer from catastrophic failures due to unexpected collisions and motion obstructions. To address this issue, this paper proposes an interactive motion planning framework (I-MP), based on a perception-motion loop. This framework empowers robots to autonomously model and reason about contact models, which in turn enables safe expansion of the free-motion space. Specifically, the robot utilizes multimodal tactile perception to acquire stimulus-response signal pairs. This enables real-time identification of objects' mechanical properties and the subsequent construction of contact models. These models are integrated as computational constraints into a reactive planner. Based on fixed-point theorems, the planner computes the spatial state toward the target in real time, thus avoiding the computational burden associated with extrapolating on high-dimensional interaction models. Furthermore, high-dimensional interaction features are linearly superposed in Cartesian space in the form of energy, and the controller achieves trajectory tracking by solving the energy gradient from the current state to the planned state. The experimental results showed that at cruising speeds ranging from 0.01 to 0.07 $m/s$, the robot's initial contact force with objects remained stable at 1.0 +- 0.7 N. In the cabinet scenario test where collision-free trajectories were unavailable, I-MP expanded the free motion space by 37.5 % through active interaction, successfully completing the environmental exploration task.

Huining Li, Yalong Jiang, Xianlin Zeng et al.

The minimum network flow algorithm is widely used in multi-target tracking. However, the majority of the present methods concentrate exclusively on minimizing cost functions whose values may not indicate accurate solutions under occlusions. In this paper, by exploiting the properties of tracklets intersections and low-confidence detections, we develop a two-stage tracking pipeline with an intersection mask that can accurately locate inaccurate tracklets which are corrected in the second stage. Specifically, we employ the minimum network flow algorithm with high-confidence detections as input in the first stage to obtain the candidate tracklets that need correction. Then we leverage the intersection mask to accurately locate the inaccurate parts of candidate tracklets. The second stage utilizes low-confidence detections that may be attributed to occlusions for correcting inaccurate tracklets. This process constructs a graph of nodes in inaccurate tracklets and low-confidence nodes and uses it for the second round of minimum network flow calculation. We perform sufficient experiments on popular MOT benchmark datasets and achieve 78.4 MOTA on the test set of MOT16, 79.2 on MOT17, and 76.4 on MOT20, which shows that the proposed method is effective.

Yuanda Xu, Hejian Sang, Zhengze Zhou et al.

Standard LLM distillation wastes compute on two fronts: problems the student has already mastered (near-zero gradients) and problems far beyond its reach (incoherent gradients that erode existing capabilities). We show that this waste is not merely intuitive but structurally inevitable: the gradient signal-to-noise ratio in distillation provably vanishes at both pass-rate extremes. This theoretical observation leads to Paced, a framework that concentrates distillation on the zone of proximal development -- the frontier of a student model's competence -- via a principled pass-rate weight $w(p) = p^α(1 - p)^β$ derived from the boundary-vanishing structure of distillation gradients. Key results: (1) Theory: We prove that the Beta kernel $w(p) = p^α(1-p)^β$ is a leading-order weight family arising from the SNR structure of distillation, and that it is minimax-robust -- under bounded multiplicative misspecification, worst-case efficiency loss is only $O(δ^2)$. (2)Distillation: On distillation from a larger teacher to a smaller student model with forward KL, Paced achieves significant gain over the base model, while keeping benchmark forgetting at a low level. (3)Self-distillation: On instruction-tuned models with reverse KL, gains are exceeding baselines as well. (4)Two-stage synergy: A forward-KL-then-reverse-KL schedule yields the strongest results in our setting, reaching substantial improvements on standard reasoning benchmarks -- supporting a mode-coverage-then-consolidation interpretation of the distillation process. All configurations require only student rollouts to estimate pass rates, need no architectural changes, and are compatible with any KL direction.

Mingzhe Yang, Zhipeng Wang, Kaiwei Liu et al.

Quantifying emergence and modeling emergent dynamics in a data-driven manner for complex dynamical systems is challenging due to the lack of direct observations at the micro-level. Thus, it's crucial to develop a framework to identify emergent phenomena and capture emergent dynamics at the macro-level using available data. Inspired by the theory of causal emergence (CE), this paper introduces a machine learning framework to learn macro-dynamics in an emergent latent space and quantify the degree of CE. The framework maximizes effective information, resulting in a macro-dynamics model with enhanced causal effects. Experimental results on simulated and real data demonstrate the effectiveness of the proposed framework. It quantifies degrees of CE effectively under various conditions and reveals distinct influences of different noise types. It can learn a one-dimensional coarse-grained macro-state from fMRI data, to represent complex neural activities during movie clip viewing. Furthermore, improved generalization to different test environments is observed across all simulation data.

Xihan Xiong, Minfeng Qi, Shiping Chen et al.

Ethereum Inscriptions (Ethscriptions) repurpose Ethereum calldata into a persistent inscription channel by embedding \texttt{data:}~URI payloads. These transactions typically target externally owned accounts, allowing the payload to bypass EVM execution while remaining permanently replicated across full nodes. Although calldata was originally designed for compact smart-contract parameters, this repurposing enables structured data embedding with long-term storage consequences. We present the first large-scale empirical study of Ethscriptions, treating them as a distinct \emph{calldata-resident workload} rather than merely a subset of general calldata usage. Our analysis focuses on the \textit{Ethscription} operational subset, which consists of payloads that decode to JSON and conform to a token-operation grammar (e.g., \texttt{p}, \texttt{op}, \texttt{tick}, \texttt{amt}). From $6.27$ million Ethscription candidates (\Uone), we extract $4.75$ million Ethscription operations (\Utwo, $75.8\%$ of \Uone). This result shows that structured token-like activity dominates the ecosystem. Our measurements further reveal (i) a complete workload lifecycle compressed into nine months (bootstrap, expansion, saturation), (ii) proliferation of $30$+ competing protocols without convergence toward a dominant standard, (iii) a lifecycle funnel exhibiting $201\times$ deploy-to-mint amplification and a $57.6{:}1$ mint-to-transfer collapse indicative of speculative minting, (iv) extreme participation inequality (Gini~$0.86$), and (v) a measurable permanent data footprint imposed on the Ethereum network.

Jelena Markovic-Voronov, Kayhan Behdin, Yuanda Xu et al.

We study the problem of routing queries to large language models (LLMs) under cost, GPU resources, and concurrency constraints. Prior per-query routing methods often fail to control batch-level cost, especially under non-uniform or adversarial batching. To address this, we propose a batch-level, resource-aware routing framework that jointly optimizes model assignment for each batch while respecting cost and model capacity limits. We further introduce a robust variant that accounts for uncertainty in predicted LLM performance, along with an offline instance allocation procedure that balances quality and throughput across multiple models. Experiments on two multi-task LLM benchmarks show that robustness improves accuracy by 1-14% over non-robust counterparts (depending on the performance estimator), batch-level routing outperforms per-query methods by up to 24% under adversarial batching, and optimized instance allocation yields additional gains of up to 3% compared to a non-optimized allocation, all while strictly controlling cost and GPU resource constraints.

Zhaoxin Guo, Zhipeng Wang, Ruiquan Ge et al.

The early detection of a pulmonary embolism (PE) is critical for enhancing patient survival rates. Both image-based and non-image-based features are of utmost importance in medical classification tasks. In a clinical setting, physicians tend to rely on the contextual information provided by Electronic Medical Records (EMR) to interpret medical imaging. However, very few models effectively integrate clinical information with imaging data. To address this shortcoming, we suggest a multimodal fusion methodology, termed PE-MVCNet, which capitalizes on Computed Tomography Pulmonary Angiography imaging and EMR data. This method comprises the Image-only module with an integrated multi-view block, the EMR-only module, and the Cross-modal Attention Fusion (CMAF) module. These modules cooperate to extract comprehensive features that subsequently generate predictions for PE. We conducted experiments using the publicly accessible Stanford University Medical Center dataset, achieving an AUROC of 94.1%, an accuracy rate of 90.2%, and an F1 score of 90.6%. Our proposed model outperforms existing methodologies, corroborating that our multimodal fusion model excels compared to models that use a single data modality. Our source code is available at https://github.com/LeavingStarW/PE-MVCNET.

Cheng Yang, Hui Jin, Xinlei Yu et al.

Diagnosing lung cancer typically involves physicians identifying lung nodules in Computed tomography (CT) scans and generating diagnostic reports based on their morphological features and medical expertise. Although advancements have been made in using multimodal large language models for analyzing lung CT scans, challenges remain in accurately describing nodule morphology and incorporating medical expertise. These limitations affect the reliability and effectiveness of these models in clinical settings. Collaborative multi-agent systems offer a promising strategy for achieving a balance between generality and precision in medical applications, yet their potential in pathology has not been thoroughly explored. To bridge these gaps, we introduce LungNoduleAgent, an innovative collaborative multi-agent system specifically designed for analyzing lung CT scans. LungNoduleAgent streamlines the diagnostic process into sequential components, improving precision in describing nodules and grading malignancy through three primary modules. The first module, the Nodule Spotter, coordinates clinical detection models to accurately identify nodules. The second module, the Radiologist, integrates localized image description techniques to produce comprehensive CT reports. Finally, the Doctor Agent System performs malignancy reasoning by using images and CT reports, supported by a pathology knowledge base and a multi-agent system framework. Extensive testing on two private datasets and the public LIDC-IDRI dataset indicates that LungNoduleAgent surpasses mainstream vision-language models, agent systems, and advanced expert models. These results highlight the importance of region-level semantic alignment and multi-agent collaboration in diagnosing nodules. LungNoduleAgent stands out as a promising foundational tool for supporting clinical analyses of lung nodules.

Zhipeng Wang, Boyang Yang, Yidong Wan et al.

Large Language Models (LLMs) perform well on automatic program repair (APR) for high-resource programming languages (HRPLs), but their effectiveness drops sharply in low-resource programming languages (LRPLs), due to a lack of sufficient verified buggy-fixed pairs for APR training. To address this challenge, we propose HELO-APR (High-resource Enabled LOw-resource APR), a two-stage APR framework that enables cross-lingual transfer of repair knowledge from HRPLs to LRPLs. HELO-APR (1) constructs high-quality LRPL training data by synthesizing LRPL buggy-fixed pairs from HRPL counterparts, preserving defect type consistency while ensuring the synthesized code is idiomatic, and then (2) adopts a curriculum learning strategy that progressively performs HRPL repair learning, cross-lingual repair alignment, and LRPL repair adaptation, improving repair effectiveness in LRPLs. Using C++ as the source HRPL and Ruby and Rust as the target LRPLs, experiments on xCodeEval show that HELO-APR consistently outperforms strong baselines, increasing Pass@1 from 31.32% to 48.65% on DeepSeek-Coder-6.7B and from 1.67% to 11.97% on CodeLlama-7B, while improving syntactic validity by raising the average target compilation rate on CodeLlama from 49.77% to 91.98%. On Defects4Ruby, HELO-APR increases BLEU-4 from 61.20 to 66.79 and ROUGE-1 from 76.76 to 83.59 on CodeLlama-7B, indicating higher similarity to developer patches in real-world settings. Finally, we conduct ablation studies to assess the necessity of each core component. These results suggest that verified cross-lingual supervision provides a reusable approach for improving LLM-based repair in low-resource languages.

Yuanda Xu, Hejian Sang, Zhengze Zhou et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has become the leading paradigm for enhancing reasoning in Large Language Models (LLMs). However, standard RLVR algorithms suffer from a well-documented pathology: while they improve Pass@1 accuracy through sharpened sampling, they simultaneously narrow the model's reasoning boundary and reduce generation diversity. We identify a root cause that existing methods overlook: the uniform penalization of errors. Current approaches -- whether data-filtering methods that select prompts by difficulty, or advantage normalization schemes -- treat all incorrect rollouts within a group identically. We show that this uniformity allows overconfident errors (incorrect reasoning paths that the RL process has spuriously reinforced) to persist and monopolize probability mass, ultimately suppressing valid exploratory trajectories. To address this, we propose the Asymmetric Confidence-aware Error Penalty (ACE). ACE introduces a per-rollout confidence shift metric, c_i = log(pi_theta(y_i|x) / pi_ref(y_i|x)), to dynamically modulate negative advantages. Theoretically, we demonstrate that ACE's gradient can be decomposed into the gradient of a selective regularizer restricted to overconfident errors, plus a well-characterized residual that partially moderates the regularizer's strength. We conduct extensive experiments fine-tuning Qwen2.5-Math-7B, Qwen3-8B-Base, and Llama-3.1-8B-Instruct on the DAPO-Math-17K dataset using GRPO and DAPO within the VERL framework. Evaluated on MATH-500 and AIME 2025, ACE composes seamlessly with existing methods and consistently improves the full Pass@k spectrum across all three model families and benchmarks.

Xuanzhao Dong, Wenhui Zhu, Xiwen Chen et al.

Autoregressive models (ARMs) have long dominated the landscape of biomedical vision-language models (VLMs). Recently, masked diffusion models such as LLaDA have emerged as promising alternatives, yet their application in the biomedical domain remains largely underexplored. To bridge this gap, we introduce \textbf{LLaDA-MedV}, the first large language diffusion model tailored for biomedical image understanding through vision instruction tuning. LLaDA-MedV achieves relative performance gains of 7.855\% over LLaVA-Med and 1.867\% over LLaDA-V in the open-ended biomedical visual conversation task, and sets new state-of-the-art accuracy on the closed-form subset of three VQA benchmarks: 84.93\% on VQA-RAD, 92.31\% on SLAKE, and 95.15\% on PathVQA. Furthermore, a detailed comparison with LLaVA-Med suggests that LLaDA-MedV is capable of generating reasonably longer responses by explicitly controlling response length, which can lead to more informative outputs. We also conduct an in-depth analysis of both the training and inference stages, highlighting the critical roles of initialization weight selection, fine-tuning strategies, and the interplay between sampling steps and response repetition. The code and model weight is released at https://github.com/LLM-VLM-GSL/LLaDA-MedV.

Jiwoo Hong, Shao Tang, Zhipeng Wang

Reward models are central to aligning language models with human preferences via reinforcement learning (RL). As RL is increasingly applied to settings such as verifiable rewards and multi-objective alignment, RMs are expected to encode more complex and multifaceted preference distributions. However, classifier RMs remain static once trained, limiting their adaptability at test time. We propose Variational In-Context Reward Modeling (ICRM), a novel Bayesian reward modeling objective that enables test-time steerability via in-context preference demonstrations. ICRM casts reward modeling as amortized variational inference over a latent preference probability under the Bradley-Terry model using a conjugate Beta prior. We show that ICRM adapt to unseen preference distributions at test time for both single and multi-objective settings. With more in-context demonstrations, ICRM gains 34% accuracy on SafeRLHF and 9% accuracy on RM-Bench in the single-objective setting, while widening the Pareto frontier with a 4% gain in hypervolume on helpfulness and refusal benchmarks. We further study the practical applicability of ICRM for RL training, showing that it can effectively encode verifiable rewards by outperforming a conventional RM in math reasoning. Finally, we provide theoretical guarantees that the variational objective admits a global interior optimum with finite confidence, and we analyze how KL regularization mitigates reward over-optimization.

Xuanzhao Dong, Wenhui Zhu, Xiwen Chen et al.

Over the past decade, generative models have demonstrated success in enhancing fundus images. However, the evaluation of these models remains a challenge. A benchmark for fundus image enhancement is needed for three main reasons:(1) Conventional denoising metrics such as PSNR and SSIM fail to capture clinically relevant features, such as lesion preservation and vessel morphology consistency, limiting their applicability in real-world settings; (2) There is a lack of unified evaluation protocols that address both paired and unpaired enhancement methods, particularly those guided by clinical expertise; and (3) An evaluation framework should provide actionable insights to guide future advancements in clinically aligned enhancement models. To address these gaps, we introduce EyeBench-V2, a benchmark designed to bridge the gap between enhancement model performance and clinical utility. Our work offers three key contributions:(1) Multi-dimensional clinical-alignment through downstream evaluations: Beyond standard enhancement metrics, we assess performance across clinically meaningful tasks including vessel segmentation, diabetic retinopathy (DR) grading, generalization to unseen noise patterns, and lesion segmentation. (2) Expert-guided evaluation design: We curate a novel dataset enabling fair comparisons between paired and unpaired enhancement methods, accompanied by a structured manual assessment protocol by medical experts, which evaluates clinically critical aspects such as lesion structure alterations, background color shifts, and the introduction of artificial structures. (3) Actionable insights: Our benchmark provides a rigorous, task-oriented analysis of existing generative models, equipping clinical researchers with the evidence needed to make informed decisions, while also identifying limitations in current methods to inform the design of next-generation enhancement models.

Yuanda Xu, Hejian Sang, Zhengze Zhou et al.

In settings where labeled verifiable training data is the binding constraint, each checked example should be allocated carefully. The standard practice is to use this data directly on the model that will be deployed, for example by running GRPO on the deployment student. We argue that this is often an inefficient allocation because it overlooks a reward-density principle: sparse sequence-level reward should train models where exploration is productive, while dense token-level teacher reward should be used where the aim is to compress behavior into a smaller model. In this view, GRPO-style sparse RL and OPD-style dense teacher supervision are not separate recipes; they are different reward-density regimes. The allocation rule is simple: use scarce labeled training data upstream on the strongest model that can turn it into reward-shaped behavior, then transfer that behavior downstream as dense supervision. We evaluate this rule on verifiable math with Qwen3 and Llama models. At fixed Qwen3-1.7B deployment-student size, an RL-improved 8B teacher distilled through the dense bridge outperforms direct GRPO on the same student, while transfer from the same teacher before RL underperforms. The bridge is important: a forward-KL warmup on teacher rollouts followed by OPD on student rollouts is consistently strongest on MATH before any post-bridge student-side sparse RL, and also gives the best pre-Stage~3 AIME endpoints for the canonical 8B/14B teachers. The bridge also makes later student-side sparse RL effective: GRPO that is weak on a cold student lifts MATH from $75.4\%$ to $78.5\%$ after the bridge and outperforms a matched replay control by $2.8$ points. The operational principal is to avoid using scarce labeled data on the least prepared policy: use sparse reward for teacher-side discovery, dense transfer for student compression, and student-side sparse reward only after the bridge.

Xiaofeng Lin, Sirou Zhu, Yilei Chen et al.

Large language models (LLMs) achieve strong performance when all task-relevant information is available upfront, as in static prediction and instruction-following problems. However, many real-world decision-making tasks are inherently online: crucial information must be acquired through interaction, feedback is delayed, and effective behavior requires balancing information collection and exploitation over time. While in-context learning enables adaptation without weight updates, existing LLMs often struggle to reliably leverage in-context interaction experience in such settings. In this work, we show that this limitation can be addressed through training. We introduce ORBIT, a multi-task, multi-episode meta-reinforcement learning framework that trains LLMs to learn from interaction in context. After meta-training, a relatively small open-source model (Qwen3-14B) demonstrates substantially improved in-context online learning on entirely unseen environments, matching the performance of GPT-5.2 and outperforming standard RL fine-tuning by a large margin. Scaling experiments further reveal consistent gains with model size, suggesting significant headroom for learn-at-inference-time decision-making agents. Code reproducing the results in the paper can be found at https://github.com/XiaofengLin7/ORBIT.

Xiwen Chen, Wenhui Zhu, Gen Li et al.

Multi-modal large language models (MLLMs) achieve strong visual-language reasoning but suffer from high inference cost due to redundant visual tokens. Recent work explores visual token pruning to accelerate inference, while existing pruning methods overlook the underlying distributional structure of visual representations. We propose OTPrune, a training-free framework that formulates pruning as distribution alignment via optimal transport (OT). By minimizing the 2-Wasserstein distance between the full and pruned token distributions, OTPrune preserves both local diversity and global representativeness while reducing inference cost. Moreover, we derive a tractable submodular objective that enables efficient optimization, and theoretically prove its monotonicity and submodularity, providing a principled foundation for stable and efficient pruning. We further provide a comprehensive analysis that explains how distributional alignment contributes to stable and semantically faithful pruning. Comprehensive experiments on wider benchmarks demonstrate that OTPrune achieves superior performance-efficiency tradeoffs compared to state-of-the-art methods. The code is available at https://github.com/xiwenc1/OTPrune.

Xuanzhao Dong, Wenhui Zhu, Yujian Xiong et al.

Color fundus photography (CFP) is central to diagnosing and monitoring retinal disease, yet its acquisition variability (e.g., illumination changes) often degrades image quality, which motivates robust enhancement methods. Unpaired enhancement pipelines are typically GAN-based, however, they can distort clinically critical vasculature, altering vessel topology and endpoint integrity. Motivated by these structural alterations, we propose Vessel-Aware Optimal Transport (\textbf{VAOT}), a framework that combines an optimal-transport objective with two structure-preserving regularizers: (i) a skeleton-based loss to maintain global vascular connectivity and (ii) an endpoint-aware loss to stabilize local termini. These constraints guide learning in the unpaired setting, reducing noise while preserving vessel structure. Experimental results on synthetic degradation benchmark and downstream evaluations in vessel and lesion segmentation demonstrate the superiority of the proposed methods against several state-of-the art baselines. The code is available at https://github.com/Retinal-Research/VAOT

Zehao Ni, Yonghao He, Lingfeng Qian et al.

In the context of imitation learning, visuomotor-based diffusion policy learning is one of the main directions in robotic manipulation. Most of these approaches rely on point clouds as observation inputs and construct scene representations through point clouds feature learning, which enables them to achieve remarkable accuracy. However, the existing literature lacks an in-depth exploration of vision-only solutions that have significant potential. In this paper, we propose a Vision-Only and single-view Diffusion Policy learning method (VO-DP) that leverages pretrained visual foundation models to achieve effective fusion of semantic and geometric features. We utilize intermediate features from VGGT incorporating semantic features from DINOv2 and geometric features from Alternating Attention blocks. Features are fused via cross-attention and spatially compressed with a CNN to form the input to the policy head. Extensive experiments demonstrate that VO-DP not only outperforms the vision-only baseline DP significantly but also exhibits distinct performance trends against the point cloud-based method DP3: in simulation tasks, VO-DP achieves an average success rate of 64.6% on par with DP3 64.0% and far higher than DP 34.8%, while in real-world tasks, it reaches 87.9%, outperforming both DP3 67.5% and DP 11.2% by a notable margin. Further robustness evaluations confirm that VO-DP remains highly stable under varying conditions including color, size, background, and lighting. Lastly, we open-source a training library for robotic manipulation. Built on Accelerate, this library supports multi-machine and multi-GPU parallel training, as well as mixed precision training. It is compatible with visuomotor policies such as DP, DP3 and VO-DP, and also supports the RoboTwin simulator.

Zhoutong Fu, Yihan Cao, Yi-Lin Chen et al.

Large language models (LLMs) have achieved strong performance across a wide range of natural language processing tasks. However, deploying LLMs at scale for domain specific applications, such as job-person fit and explanation in job seeking platforms, introduces distinct challenges. At LinkedIn, the job person fit task requires analyzing a candidate's public profile against job requirements to produce both a fit assessment and a detailed explanation. Directly applying open source or finetuned LLMs to this task often fails to yield high quality, actionable feedback due to the complexity of the domain and the need for structured outputs. Moreover, the large size of these models leads to high inference latency and limits scalability, making them unsuitable for online use. To address these challenges, we introduce LANTERN, a novel LLM knowledge distillation framework tailored specifically for job person fit tasks. LANTERN involves modeling over multiple objectives, an encoder model for classification purpose, and a decoder model for explanation purpose. To better distill the knowledge from a strong black box teacher model to multiple downstream models, LANTERN incorporates multi level knowledge distillation that integrates both data and logit level insights. In addition to introducing the knowledge distillation framework, we share our insights on post training techniques and prompt engineering, both of which are crucial for successfully adapting LLMs to domain specific downstream tasks. Extensive experimental results demonstrate that LANTERN significantly improves task specific metrics for both job person fit and explanation. Online evaluations further confirm its effectiveness, showing measurable gains in job seeker engagement, including a 0.24\% increase in apply rate and a 0.28\% increase in qualified applications.

Xiaofeng Lin, Hejian Sang, Zhipeng Wang et al.

A prevailing view holds that supervised fine-tuning (SFT) memorizes training data and fails to generalize, whereas reinforcement learning (RL) attains broader robustness. We revisit this claim through a systematic evaluation on two decision-making benchmarks, Sokoban and General Points, and arrive at a different conclusion. We show that much of SFT's perceived failure stems from frozen-prompt artifacts: when trained on fixed instruction templates, SFT models cling to training semantics rather than adapting to new ones. Introducing prompt diversity during training breaks this shortcut and yields strong generalization to unseen instruction variants without harming in-distribution performance. Beyond instruction shifts, we ask whether SFT can generalize to strictly harder tasks. Here, chain-of-thought (CoT) supervision provides an algorithmic scaffold that markedly improves transfer to more difficult regimes, such as larger Sokoban grids with additional boxes and arithmetic with out-of-distribution values or five-card compositions that increase combinatorial complexity. Finally, combining prompt diversity with CoT achieves the best of both worlds: robust generalization across both instruction-variant and difficulty-variant settings, matching or surpassing RL baselines on our benchmarks while retaining SFT's simplicity and stability. These findings challenge the narrative that SFT is inherently inferior to RL and support a data-centric perspective: with appropriately curated demonstrations, vanilla SFT can generalize as strongly as RL. Code reproducing the results in the paper can be found at: https://github.com/XiaofengLin7/debunking-sft-generalization.

Ryan Lucas, Kayhan Behdin, Zhipeng Wang et al.

Reasoning language models such as DeepSeek-R1 produce long chain-of-thought traces during inference time which make them costly to deploy at scale. We show that using compression techniques such as neural network pruning produces greater performance loss than in typical language modeling tasks, and in some cases can make the model slower since they cause the model to produce more thinking tokens but with worse performance. We show that this is partly due to the fact that standard LLM pruning methods often focus on input reconstruction, whereas reasoning is a decode-dominated task. We introduce a simple, drop-in fix: during pruning we jointly reconstruct activations from the input and the model's on-policy chain-of-thought traces. This "Reasoning-Aware Compression" (RAC) integrates seamlessly into existing pruning workflows such as SparseGPT, and boosts their performance significantly. Code reproducing the results in the paper can be found at: https://github.com/RyanLucas3/RAC

Jianxun Yu, Ruiquan Ge, Zhipeng Wang et al.

The diagnosis of medical diseases faces challenges such as the misdiagnosis of small lesions. Deep learning, particularly multimodal approaches, has shown great potential in the field of medical disease diagnosis. However, the differences in dimensionality between medical imaging and electronic health record data present challenges for effective alignment and fusion. To address these issues, we propose the Multimodal Multiscale Cross-Attention Fusion Network (MMCAF-Net). This model employs a feature pyramid structure combined with an efficient 3D multi-scale convolutional attention module to extract lesion-specific features from 3D medical images. To further enhance multimodal data integration, MMCAF-Net incorporates a multi-scale cross-attention module, which resolves dimensional inconsistencies, enabling more effective feature fusion. We evaluated MMCAF-Net on the Lung-PET-CT-Dx dataset, and the results showed a significant improvement in diagnostic accuracy, surpassing current state-of-the-art methods. The code is available at https://github.com/yjx1234/MMCAF-Net

Zhipeng Wang, Hao Wang, Jiexi Yan et al.

Zero-Shot Sketch-Based Image Retrieval (ZS-SBIR) is a novel cross-modal retrieval task, where abstract sketches are used as queries to retrieve natural images under zero-shot scenario. Most existing methods regard ZS-SBIR as a traditional classification problem and employ a cross-entropy or triplet-based loss to achieve retrieval, which neglect the problems of the domain gap between sketches and natural images and the large intra-class diversity in sketches. Toward this end, we propose a novel Domain-Smoothing Network (DSN) for ZS-SBIR. Specifically, a cross-modal contrastive method is proposed to learn generalized representations to smooth the domain gap by mining relations with additional augmented samples. Furthermore, a category-specific memory bank with sketch features is explored to reduce intra-class diversity in the sketch domain. Extensive experiments demonstrate that our approach notably outperforms the state-of-the-art methods in both Sketchy and TU-Berlin datasets. Our source code is publicly available at https://github.com/haowang1992/DSN.

Bikun Wang, Zhipeng Wang, Chenhao Zhu et al.

Learning-based approaches have achieved remarkable performance in the domain of autonomous driving. Leveraging the impressive ability of neural networks and large amounts of human driving data, complex patterns and rules of driving behavior can be encoded as a model to benefit the autonomous driving system. Besides, an increasing number of data-driven works have been studied in the decision-making and motion planning module. However, the reliability and the stability of the neural network is still full of uncertainty. In this paper, we introduce a hierarchical planning architecture including a high-level grid-based behavior planner and a low-level trajectory planner, which is highly interpretable and controllable. As the high-level planner is responsible for finding a consistent route, the low-level planner generates a feasible trajectory. We evaluate our method both in closed-loop simulation and real world driving, and demonstrate the neural network planner has outstanding performance in complex urban autonomous driving scenarios.

Zehua Cheng, Zhipeng Wang, Wei Dai et al.

Automated matching engines execute millions of orders per session, yet systematic asymmetries in latency, order size, and market access compound into persistent execution disparities that erode participant trust. We formulate provably fair order matching as a Constrained Markov Decision Process and propose CPO-FOAM (Constrained Policy Optimization with Feedback-Optimized Adaptive Margins). An inner loop computes an analytic trust-region step on the Fisher information manifold; a PID-controlled outer loop dynamically tightens safety margins, suppressing the sawtooth oscillations endemic to Lagrangian methods under non-stationary dynamics. Group fairness (demographic parity, equalized odds) enters the CMDP cost vector while individual Lipschitz fairness is enforced deterministically via spectral normalization. We prove BIBO stability and that the integral term drives steady-state violations to zero. On LOBSTER NASDAQ data across six market regimes, CPO-FOAM recovers 95.9% of unconstrained throughput at 2.5% constraint violation frequency; on crypto-asset LOB data under MEV injection it captures 98.4% of the reward envelope at 3.2% CVF. The method scales sub-linearly to M=8 constraints, settles on-chain within one Ethereum block, and yields a 2.1X reward improvement on Safety-Gymnasium, confirming domain-agnostic generalization.

Hejian Sang, Yuanda Xu, Zhengze Zhou et al.

Reasoning models think out loud, but much of what they say is noise. We introduce OPSDC (On-Policy Self-Distillation for Reasoning Compression), a method that teaches models to reason more concisely by distilling their own concise behavior back into themselves. The entire approach reduces to one idea: condition the same model on a "be concise" instruction to obtain teacher logits, and minimize per-token reverse KL on the student's own rollouts. No ground-truth answers, no token budgets, no difficulty estimators. Just self-distillation. Yet this simplicity belies surprising sophistication: OPSDC automatically compresses easy problems aggressively while preserving the deliberation needed for hard ones. On Qwen3-8B and Qwen3-14B, we achieve 57-59% token reduction on MATH-500 while improving accuracy by 9-16 points absolute. On AIME 2024, the 14B model gains 10 points with 41% compression. The secret? Much of what reasoning models produce is not just redundant-it is actively harmful, compounding errors with every unnecessary token.

Bing Yuan, Zhang Jiang, Aobo Lyu et al.

Emergence and causality are two fundamental concepts for understanding complex systems. They are interconnected. On one hand, emergence refers to the phenomenon where macroscopic properties cannot be solely attributed to the cause of individual properties. On the other hand, causality can exhibit emergence, meaning that new causal laws may arise as we increase the level of abstraction. Causal emergence theory aims to bridge these two concepts and even employs measures of causality to quantify emergence. This paper provides a comprehensive review of recent advancements in quantitative theories and applications of causal emergence. Two key problems are addressed: quantifying causal emergence and identifying it in data. Addressing the latter requires the use of machine learning techniques, thus establishing a connection between causal emergence and artificial intelligence. We highlighted that the architectures used for identifying causal emergence are shared by causal representation learning, causal model abstraction, and world model-based reinforcement learning. Consequently, progress in any of these areas can benefit the others. Potential applications and future perspectives are also discussed in the final section of the review.

Zhipeng Wang, Rui Sun, Elizabeth Lui et al.

The rapid advancement of AI has underscored critical challenges in its development and implementation, largely due to centralized control by a few major corporations. This concentration of power intensifies biases within AI models, resulting from inadequate governance and oversight mechanisms. Additionally, it limits public involvement and heightens concerns about the integrity of model generation. Such monopolistic control over data and AI outputs threatens both innovation and fair data usage, as users inadvertently contribute data that primarily benefits these corporations. In this work, we propose AIArena, a blockchain-based decentralized AI training platform designed to democratize AI development and alignment through on-chain incentive mechanisms. AIArena fosters an open and collaborative environment where participants can contribute models and computing resources. Its on-chain consensus mechanism ensures fair rewards for participants based on their contributions. We instantiate and implement AIArena on the public Base blockchain Sepolia testnet, and the evaluation results demonstrate the feasibility of AIArena in real-world applications.

Zhipeng Wang, Rui Sun, Elizabeth Lui et al.

Centralization enhances the efficiency of Artificial Intelligence (AI), but it also brings critical challenges, such as single points of failure, inherent biases, data privacy concerns, and scalability issues, for AI systems. These problems are especially common in closed-source large language models (LLMs), where user data is collected and used with full transparency. To address these issues, blockchain-based decentralized AI (DeAI) has been introduced. DeAI leverages the strengths of blockchain technologies to enhance the transparency, security, decentralization, as well as trustworthiness of AI systems. Although DeAI has been widely developed in industry, a comprehensive understanding of state-of-the-art practical DeAI solutions is still lacking. In this work, we present a Systematization of Knowledge (SoK) for blockchain-based DeAI solutions. We propose a taxonomy to classify existing DeAI protocols based on the model lifecycle. Based on this taxonomy, we provide a structured way to clarify the landscape of DeAI protocols and identify their similarities and differences. Specifically, we analyze the functionalities of blockchain in DeAI, investigate how blockchain features contribute to enhancing the security, transparency, and trustworthiness of AI processes, and also ensure fair incentives for AI data and model contributors. In addition, we provide key insights and research gaps in developing DeAI protocols for future research.

Shuo Jiang, Haonan Li, Ruochen Ren et al.

Cutting-edge robot learning techniques including foundation models and imitation learning from humans all pose huge demands on large-scale and high-quality datasets which constitute one of the bottleneck in the general intelligent robot fields. This paper presents the Kaiwu multimodal dataset to address the missing real-world synchronized multimodal data problems in the sophisticated assembling scenario,especially with dynamics information and its fine-grained labelling. The dataset first provides an integration of human,environment and robot data collection framework with 20 subjects and 30 interaction objects resulting in totally 11,664 instances of integrated actions. For each of the demonstration,hand motions,operation pressures,sounds of the assembling process,multi-view videos, high-precision motion capture information,eye gaze with first-person videos,electromyography signals are all recorded. Fine-grained multi-level annotation based on absolute timestamp,and semantic segmentation labelling are performed. Kaiwu dataset aims to facilitate robot learning,dexterous manipulation,human intention investigation and human-robot collaboration research.

Haozhe Xu, Yifei Zhao, Wenhao Feng et al.

Locomotion under reduced gravity is commonly realized through jumping, yet continuous pronking in lunar gravity remains challenging due to prolonged flight phases and sparse ground contact. The extended aerial duration increases landing impact sensitivity and makes stable attitude regulation over rough planetary terrain difficult. Existing approaches primarily address single jumps on flat surfaces and lack both continuous-terrain solutions and realistic hardware validation. This work presents a Dual-Horizon Hybrid Internal Model for continuous quadrupedal jumping under lunar gravity using proprioceptive sensing only. Two temporal encoders capture complementary time scales: a short-horizon branch models rapid vertical dynamics with explicit vertical velocity estimation, while a long-horizon branch models horizontal motion trends and center-of-mass height evolution across the jump cycle. The fused representation enables stable and continuous jumping under extended aerial phases characteristic of lunar gravity. To provide hardware-in-the-loop validation, we develop the MATRIX (Mixed-reality Adaptive Testbed for Robotic Integrated eXploration) platform, a digital-twin-driven system that offloads gravity through a pulley-counterweight mechanism and maps Unreal Engine lunar terrain to a motion platform and treadmill in real time. Using MATRIX, we demonstrate continuous jumping of a quadruped robot under lunar-gravity emulation across cratered lunar-like terrain.

Guoyao Li, Ran He, Shusen Jing et al.

Large language models (LLMs) excel at capturing semantic nuances and therefore show impressive relevance ranking performance in modern recommendation and search systems. However, they suffer from high computational overhead under industrial latency and throughput requirements. In particular, cross-encoder ranking systems often create long context prefill-heavy workloads, as the model has to be presented with the user, query and item information. To this end, we propose MixLM, a novel LLM-based ranking framework, which significantly improves the system throughput via reducing the input context length, while preserving the semantic strength of cross-encoder rankers. In contrast to a standard ranking system where the context is presented to the model as pure text, we propose to use mix-interaction, a mixture of text and embedding tokens to represent the input. Specifically, MixLM encodes all items in the catalog into a few embedding tokens and stores in a nearline cache. The encoded item descriptions are used during online inference, effectively reducing the item length from a few thousand text tokens to a few embedding tokens. We share insights from deploying our MixLM framework to a real-world search application at LinkedIn, including a detailed discussion of our training pipelines, as well as a thorough analysis of our online serving infrastructure optimization. With the same latency budget and on-par relevance metrics, MixLM increased throughput by 10.0x comparing with strong baselines, 75.9x over full-text LLM rerankers. The efficiency gains delivered by MixLM enabled full-traffic deployment of LLM-powered search, which resulted in a significant 0.47\% increase in Daily Active Users (DAU) in online A/B tests.

Vaibhav Jindal, Hejian Sang, Chun-Mao Lai et al.

Diffusion language models (dLLMs) are an emerging alternative to autoregressive (AR) generators, but aligning them to human preferences is challenging because sequence log-likelihoods are intractable and pairwise preference data are costly to collect. We introduce ELBO-KTO, which combines an ELBO surrogate for diffusion log-likelihoods with a prospect-theoretic, unpaired preference objective (Kahneman Tversky Optimization, KTO). We analyze the bias and variance induced by the ELBO substitution and employ variance-reduction practices that stabilize gradients during training. Applied to LLaDA-8B-Instruct, ELBO-KTO yields 65.9% and 62.3% adjusted win rates on kto-mix-14k and UltraFeedback-Binary, respectively, versus the base model under an automatic LLM judge. Across downstream tasks, including GSM8K, MMLU, and additional reasoning/knowledge benchmarks, ELBO-KTO trained on UltraFeedback-Binary performs on par with or better than the base model under identical decoding. This establishes unpaired preference optimization as a viable alternative to pairwise alignment in diffusion LLMs.

Kayhan Behdin, Qingquan Song, Sriram Vasudevan et al.

Large Language Models (LLMs) have demonstrated impressive quality when applied to predictive tasks such as relevance ranking and semantic search. However, deployment of such LLMs remains prohibitively expensive for industry applications with strict latency and throughput requirements. In this work, we present lessons and efficiency insights from developing a purely text-based decoder-only Small Language Model (SLM) for a semantic search application at LinkedIn. Particularly, we discuss model compression techniques such as pruning that allow us to reduce the model size by up to $40\%$ while maintaining the accuracy. Additionally, we present context compression techniques that allow us to reduce the input context length by up to $10$x with minimal loss of accuracy. Finally, we present practical lessons from optimizing the serving infrastructure for deploying such a system on GPUs at scale, serving millions of requests per second. Taken together, this allows us to increase our system's throughput by $10$x in a real-world deployment, while meeting our quality bar.

Siyu Zhu, Yanbin Jiang, Hejian Sang et al.

We investigated Agentic RL with large language models on the \textsc{TravelPlanner} benchmark. Our approach, \textsc{Planner-R1}, achieved a \textbf{56.9\%} final-pass rate with only 180 training queries, a $2.7\times$ improvement over GPT-5's $21.2\%$ baseline and the strongest agentic result on the public leaderboard. A central finding was that smaller models (8B) were highly responsive to reward shaping: with dense process-level signals, they reached competitive performance while being $3.5\times$ more compute-efficient and $1.5\times$ more memory-efficient than 32B models. Larger models were more robust under sparse rewards but exhibited smaller relative gains from shaping and higher variance across runs. While curriculum learning offered no significant benefit, shaped rewards consistently amplified learning dynamics, making 8B models the most efficient setting for agentic RL. Crucially, these gains did not come at the cost of overfitting: fine-tuned models mostly maintained or exceeded baseline performance on out-of-domain tasks, including \textsc{Multi-IF}, \textsc{NaturalPlan}, and $τ$-\textsc{Bench}. These results establish reward shaping as a decisive lever for scaling agentic RL, highlight the competitive strength of smaller models, and demonstrate that efficiency can be achieved without sacrificing generalization.

Wenhui Zhu, Xiwen Chen, Zhipeng Wang et al.

Instructed Visual Segmentation (IVS) tasks require segmenting objects in images or videos based on natural language instructions. While recent multimodal large language models (MLLMs) have achieved strong performance on IVS, their inference cost remains a major bottleneck, particularly in video. We empirically analyze visual token sampling in MLLMs and observe a strong correlation between subset token coverage and segmentation performance. This motivates our design of a simple and effective token pruning method that selects a compact yet spatially representative subset of tokens to accelerate inference. In this paper, we introduce a novel visual token pruning method for IVS, called EVTP-IV, which builds upon the k-center by integrating spatial information to ensure better coverage. We further provide an information-theoretic analysis to support our design. Experiments on standard IVS benchmarks show that our method achieves up to 5X speed-up on video tasks and 3.5X on image tasks, while maintaining comparable accuracy using only 20% of the tokens. Our method also consistently outperforms state-of-the-art pruning baselines under varying pruning ratios.