Zhan Shi

Minhua Lin, Juncheng Wu, Zijun Wang et al.

LLM agents are increasingly deployed as systems built around editable external harnesses, including prompts, skills, memories and tools, that shape task execution without changing model parameters. Harness self-evolution adapts such agents by updating these harnesses from execution evidence. Yet it remains unclear whether a model's base capability in task-solving predicts its capabilities in harness self-evolution: which models produce useful harness updates, and which actually benefit from them? We analyze two harness self-evolution capabilities: (i) harness-updating, the capability to produce useful persistent harness updates from execution evidence; (ii) harness-benefit, the capability to benefit from updated harnesses during task solving. Our analysis reveals two findings. First, harness-updating is flat in base capability: models from different capability tiers produce harness updates that lead to surprisingly similar gains; even Qwen3.5-9B's updates yield gains comparable to those of Claude Opus~4.6. Second, harness-benefit is non-monotonic in base capability: weak-tier models benefit little from updated harnesses, mid-tier models benefit most, and strong-tier models benefit less than mid-tier. We trace low gains at the weak tier to two failure modes: weak-tier models may fail to activate relevant harness artifacts, or activate them but fail to follow them faithfully. These findings suggest investing capability budget in the task-solving agent rather than the evolver, and targeting harness invocation and long-horizon instruction following in agent training. Our source code is publicly available at https://github.com/A-EVO-Lab/a-evolve/tree/release/harness-evolution.

Zewen Liu, Zhan Shi, Yisi Sang et al.

Auto-harness systems such as A-Evolve, GEPA, and Meta-Harness improve LLM agents by optimizing prompts, skills, tools, memories, and supporting infrastructure from execution feedback, but they are typically evaluated on fixed offline benchmarks. Real deployments instead present open-ended task streams: histories grow without a fixed endpoint, heterogeneous tasks require different harnesses, and problem distributions shift over time. These challenges make a single repeatedly and densely updated harness brittle, causing performance degradation as accuracy peaks early and then declines. This motivates sustained harness construction with task-wise adaptation. We introduce Adaptive Auto-Harness, a framework and system for such streams. The framework decomposes the gap to an oracle harness into evolution loss and adaptation loss. The system addresses these losses with a stateful multi-agent evolver, a harness tree with solve-time routing, and human-steering hooks for cases where history lacks the needed signal. Across prediction-market, security-competition, and event-forecasting streams, Adaptive Auto-Harness outperforms five existing auto-harness baselines and ablations attribute gains to better construction, routing, or targeted human steering. Code is available in https://github.com/A-EVO-Lab/AdaptiveHarness .

Mohamed Ashraf Abdelsalam, Zhan Shi, Federico Fancellu et al.

The success of scene graphs for visual scene understanding has brought attention to the benefits of abstracting a visual input (e.g., image) into a structured representation, where entities (people and objects) are nodes connected by edges specifying their relations. Building these representations, however, requires expensive manual annotation in the form of images paired with their scene graphs or frames. These formalisms remain limited in the nature of entities and relations they can capture. In this paper, we propose to leverage a widely-used meaning representation in the field of natural language processing, the Abstract Meaning Representation (AMR), to address these shortcomings. Compared to scene graphs, which largely emphasize spatial relationships, our visual AMR graphs are more linguistically informed, with a focus on higher-level semantic concepts extrapolated from visual input. Moreover, they allow us to generate meta-AMR graphs to unify information contained in multiple image descriptions under one representation. Through extensive experimentation and analysis, we demonstrate that we can re-purpose an existing text-to-AMR parser to parse images into AMRs. Our findings point to important future research directions for improved scene understanding.

Tao Lv, Hao Ye, Quan Yuan et al.

We demonstrate a compact, cost-effective snapshot spectral imaging system named Aperture Diffraction Imaging Spectrometer (ADIS), which consists only of an imaging lens with an ultra-thin orthogonal aperture mask and a mosaic filter sensor, requiring no additional physical footprint compared to common RGB cameras. Then we introduce a new optical design that each point in the object space is multiplexed to discrete encoding locations on the mosaic filter sensor by diffraction-based spatial-spectral projection engineering generated from the orthogonal mask. The orthogonal projection is uniformly accepted to obtain a weakly calibration-dependent data form to enhance modulation robustness. Meanwhile, the Cascade Shift-Shuffle Spectral Transformer (CSST) with strong perception of the diffraction degeneration is designed to solve a sparsity-constrained inverse problem, realizing the volume reconstruction from 2D measurements with Large amount of aliasing. Our system is evaluated by elaborating the imaging optical theory and reconstruction algorithm with demonstrating the experimental imaging under a single exposure. Ultimately, we achieve the sub-super-pixel spatial resolution and high spectral resolution imaging. The code will be available at: https://github.com/Krito-ex/CSST.

Zhan Shi, Xin Ding, Peng Ding et al.

Ship orientation angle prediction (SOAP) with optical remote sensing images is an important image processing task, which often relies on deep convolutional neural networks (CNNs) to make accurate predictions. This paper proposes a novel framework to reduce the model sizes and computational costs of SOAP models without harming prediction accuracy. First, a new SOAP model called Mobile-SOAP is designed based on MobileNetV2, achieving state-of-the-art prediction accuracy. Four tiny SOAP models are also created by replacing the convolutional blocks in Mobile-SOAP with four small-scale networks, respectively. Then, to transfer knowledge from Mobile-SOAP to four lightweight models, we propose a novel knowledge distillation (KD) framework termed SOAP-KD consisting of a novel feature-based guidance loss and an optimized synthetic samples-based knowledge transfer mechanism. Lastly, extensive experiments on the FGSC-23 dataset confirm the superiority of Mobile-SOAP over existing models and also demonstrate the effectiveness of SOAP-KD in improving the prediction performance of four specially designed tiny models. Notably, by using SOAP-KD, the test mean absolute error of the ShuffleNetV2x1.0-based model is only 8% higher than that of Mobile-SOAP, but its number of parameters and multiply-accumulate operations (MACs) are respectively 61.6% and 60.8% less.

Zhan Shi, Hendra I. Nurdin

Kerr nonlinear cavities displaying optical thresholding have been proposed for the realization of ultra-low power photonic logic gates. In the ultra-low photon number regime, corresponding to energy levels in the attojoule scale, quantum input-output models become important to study the effect of unavoidable quantum fluctuations on the performance of such logic gates. However, being a quantum anharmonic oscillator, a Kerr-cavity has an infinite dimensional Hilbert space spanned by the Fock states of the oscillator. This poses a challenge to simulate and analyze photonic logic gates and circuits composed of multiple Kerr nonlinearities. For simulation, the Hilbert of the oscillator is typically truncated to the span of only a finite number of Fock states. This paper develops a quasi-principal components approach to identify important subspaces of a Kerr-cavity Hilbert space and exploits it to construct an approximate reduced model of the Kerr-cavity on a smaller Hilbert space. Using this approach, we find a reduced dimension model with a Hilbert space dimension of 15 that can closely match the magnitudes of the mean transmitted and reflected output fields of a conventional truncated Fock state model of dimension 75, when driven by an input coherent field that switches between two levels. For the same input, the reduced model also closely matches the magnitudes of the mean output fields of Kerr-cavity-based AND and NOT gates and a NAND latch obtained from simulation of the full 75 dimension model.

Zhan Shi, Hendra I. Nurdin

This paper presents formulae for Einstein-Podolsky-Rosen (EPR) entanglement generated from $N$ nondegenerate optical parametric amplifiers (NOPAs) interconnected in a linear coherent feedback (CFB) chain in the idealized lossless scenario and infinite bandwidth limit. The lossless scenario sets the ultimate EPR entanglement (two-mode squeezing) that can be achieved by this linear chain of NOPAs while the infinite bandwidth limit simplifies the analysis but gives an accurate approximation to the EPR entanglement at low frequencies of interest. Two adjustable phase shifts are placed at the outputs of the system to achieve the best EPR entanglement by selecting appropriate quadratures of the output fields.

Minhua Lin, Hanqing Lu, Zhan Shi et al.

As Large Language Models (LLMs) move from curated training sets into open-ended real-world environments, a fundamental limitation emerges: static training cannot keep pace with continual deployment environment change. Scaling training-time and inference-time compute improves static capability but does not close this train-deploy gap. We argue that addressing this limitation requires a new scaling axis-evolution. Existing deployment-time adaptation methods, whether parametric fine-tuning or heuristic memory accumulation, lack the strategic agency needed to diagnose failures and produce durable improvements. Our position is that agentic evolution represents the inevitable future of LLM adaptation, elevating evolution itself from a fixed pipeline to an autonomous evolver agent. We instantiate this vision in a general framework, A-Evolve, which treats deployment-time improvement as a deliberate, goal-directed optimization process over persistent system state. We further propose the evolution-scaling hypothesis: the capacity for adaptation scales with the compute allocated to evolution, positioning agentic evolution as a scalable path toward sustained, open-ended adaptation in the real world.

Yingqian Cui, Zhenwei Dai, Bing He et al.

Latent reasoning has been recently proposed as a reasoning paradigm and performs multi-step reasoning through generating steps in the latent space instead of the textual space. This paradigm enables reasoning beyond discrete language tokens by performing multi-step computation in continuous latent spaces. Although there have been numerous studies focusing on improving the performance of latent reasoning, its internal mechanisms remain not fully investigated. In this work, we conduct a comprehensive analysis of latent reasoning methods to better understand the role and behavior of latent representation in the process. We identify two key issues across latent reasoning methods with different levels of supervision. First, we observe pervasive shortcut behavior, where they achieve high accuracy without relying on latent reasoning. Second, we examine the hypothesis that latent reasoning supports BFS-like exploration in latent space, and find that while latent representations can encode multiple possibilities, the reasoning process does not faithfully implement structured search, but instead exhibits implicit pruning and compression. Finally, our findings reveal a trade-off associated with supervision strength: stronger supervision mitigates shortcut behavior but restricts the ability of latent representations to maintain diverse hypotheses, whereas weaker supervision allows richer latent representations at the cost of increased shortcut behavior.

Siyuan Xu, Shiyang Li, Xin Liu et al.

Existing synthetic tool-use corpora are primarily designed for offline supervised fine-tuning, yet reinforcement learning (RL) requires executable environments that support reward-checkable online rollouts. We propose COVERT, a two-stage pipeline that first generates reliable base tool-use trajectories through self-evolving synthesis with multi-level validation, and then applies oracle-preserving augmentations that systematically increase environmental complexity. These augmentations introduce distractor tools, indirect or ambiguous user queries, and noisy, multi-format, or erroneous tool outputs, while strictly preserving oracle tool calls and final answers as ground truth. This design enables automatic reward computation via reference matching for standard cases and lightweight judge-assisted verification for special behaviors such as error detection, supporting RL optimization of tool-calling policies. On Qwen2.5-Instruct-14B, COVERT-RL improves overall accuracy on BFCL v3 from 56.5 to 59.9 and on ACEBench from 53.0 to 59.3, with minimal regressions on general-ability benchmarks; when stacked on SFT, it further reaches 62.1 and 61.8, confirming additive gains. These results suggest that oracle-preserving synthetic environments offer a practical RL refinement stage, complementary to SFT, for improving tool-use robustness under ambiguity and unreliable tool feedback.

Feiyang Li, Peng Fang, Zhan Shi et al.

Chain-of-thought (CoT) reasoning boosts large language models' (LLMs) performance on complex tasks but faces two key limitations: a lack of reliability when solely relying on LLM-generated reasoning chains and lower reasoning performance from natural language prompts compared with code prompts. To address these issues, we propose CoT-RAG, a novel reasoning framework with three key designs: (i) Knowledge Graph-driven CoT Generation, featuring knowledge graphs to modulate reasoning chain generation of LLMs, thereby enhancing reasoning credibility; (ii) Learnable Knowledge Case-aware RAG, which incorporates retrieval-augmented generation (RAG) into knowledge graphs to retrieve relevant sub-cases and sub-descriptions, providing LLMs with learnable information; (iii) Pseudo Program Prompting Execution, which promotes greater logical rigor by guiding LLMs to execute reasoning tasks as pseudo-programs. Evaluations on nine public datasets spanning three reasoning tasks reveal significant accuracy gains-ranging from 4.0% to 44.3%-over state-of-the-art methods. Furthermore, tests on four domain-specific datasets demonstrate exceptional accuracy and efficient execution, underscoring its practical applicability and scalability. Our code and data are available at https: //github.com/hustlfy123/CoT-RAG.

Jiuyun Zhang, Zhan Shi, Linsen Chen et al.

Existing computational spectral imaging systems typically rely on coded aperture and beam splitters that block a substantial fraction of incident light, degrading reconstruction quality under light-starved conditions. To address this limitation, we develop the Oscillating Dispersion Imaging Spectrometer (ODIS), which for the first time achieves near-full light throughput by axially translating a disperser between the conjugate image plane and a defocused position, sequentially capturing a panchromatic (PAN) image and a dispersed measurement along a single optical path. We further propose a PAN-guided Dispersion-Aware Deep Unfolding Network (PDAUN) that recovers high-fidelity spectral information from maskless dispersion under PAN structural guidance. Its data-fidelity step derives an FFT-Woodbury preconditioned solver by exploiting the cyclic-convolution property of the ODIS forward model, while a Dispersion-Aware Deformable Convolution module (DADC) corrects sub-pixel spectral misalignment using PAN features. Experiments show state-of-the-art performance on standard benchmarks, and cross-system comparisons confirm that ODIS yields decisive gains under low illumination. High-fidelity reconstruction is validated on a physical prototype.

Zhan Shi, Song Wang, Junbo Chen et al.

Visual grounding aims to identify objects or regions in a scene based on natural language descriptions, essential for spatially aware perception in autonomous driving. However, existing visual grounding tasks typically depend on bounding boxes that often fail to capture fine-grained details. Not all voxels within a bounding box are occupied, resulting in inaccurate object representations. To address this, we introduce a benchmark for 3D occupancy grounding in challenging outdoor scenes. Built on the nuScenes dataset, it integrates natural language with voxel-level occupancy annotations, offering more precise object perception compared to the traditional grounding task. Moreover, we propose GroundingOcc, an end-to-end model designed for 3D occupancy grounding through multi-modal learning. It combines visual, textual, and point cloud features to predict object location and occupancy information from coarse to fine. Specifically, GroundingOcc comprises a multimodal encoder for feature extraction, an occupancy head for voxel-wise predictions, and a grounding head to refine localization. Additionally, a 2D grounding module and a depth estimation module enhance geometric understanding, thereby boosting model performance. Extensive experiments on the benchmark demonstrate that our method outperforms existing baselines on 3D occupancy grounding. The dataset is available at https://github.com/RONINGOD/GroundingOcc.

Zhongshen Zeng, Yinhong Liu, Yingjia Wan et al.

Large language models (LLMs) have shown increasing capability in problem-solving and decision-making, largely based on the step-by-step chain-of-thought reasoning processes. However, evaluating these reasoning abilities has become increasingly challenging. Existing outcome-based benchmarks are beginning to saturate, becoming less effective in tracking meaningful progress. To address this, we present a process-based benchmark MR-Ben that demands a meta-reasoning skill, where LMs are asked to locate and analyse potential errors in automatically generated reasoning steps. Our meta-reasoning paradigm is especially suited for system-2 slow thinking, mirroring the human cognitive process of carefully examining assumptions, conditions, calculations, and logic to identify mistakes.MR-Ben comprises 5,975 questions curated by human experts across a wide range of subjects, including physics, chemistry, logic, coding, and more. Through our designed metrics for assessing meta-reasoning on this benchmark, we identify interesting limitations and weaknesses of current LLMs (open-source and closed-source models). For example, with models like the o1 series from OpenAI demonstrating strong performance by effectively scrutinizing the solution space, many other state-of-the-art models fall significantly behind on MR-Ben, exposing potential shortcomings in their training strategies and inference methodologies.

Chunlin Tian, Zhan Shi, Zhijiang Guo et al.

Adapting Large Language Models (LLMs) to new tasks through fine-tuning has been made more efficient by the introduction of Parameter-Efficient Fine-Tuning (PEFT) techniques, such as LoRA. However, these methods often underperform compared to full fine-tuning, particularly in scenarios involving complex datasets. This issue becomes even more pronounced in complex domains, highlighting the need for improved PEFT approaches that can achieve better performance. Through a series of experiments, we have uncovered two critical insights that shed light on the training and parameter inefficiency of LoRA. Building on these insights, we have developed HydraLoRA, a LoRA framework with an asymmetric structure that eliminates the need for domain expertise. Our experiments demonstrate that HydraLoRA outperforms other PEFT approaches, even those that rely on domain knowledge during the training and inference phases.

Mingjie Ji, Zhan Shi, Kailai Zhou et al.

Self-supervised video denoising methods typically extend image-based frameworks into the temporal dimension, yet they often struggle to integrate inter-frame temporal consistency with intra-frame spatial specificity. Existing Video Blind-Spot Networks (BSNs) require noise independence by masking the center pixel, this constraint prevents the use of spatial evidence for texture recovery, thereby severing spatiotemporal correlations and causing texture loss. To address this, we propose Frames2Residual (F2R), a spatiotemporal decoupling framework that explicitly divides self-supervised training into two distinct stages: blind temporal consistency modeling and non-blind spatial texture recovery. In Stage 1, a blind temporal estimator learns inter-frame consistency using a frame-wise blind strategy, producing a temporally consistent anchor. In Stage 2, a non-blind spatial refiner leverages this anchor to safely reintroduce the center frame and recover intra-frame high-frequency spatial residuals while preserving temporal stability. Extensive experiments demonstrate that our decoupling strategy allows F2R to outperform existing self-supervised methods on both sRGB and raw video benchmarks.

Tao Ni, Xin Zhan, Tao Luo et al.

Road segmentation is a critical task for autonomous driving systems, requiring accurate and robust methods to classify road surfaces from various environmental data. Our work introduces an innovative approach that integrates LiDAR point cloud data, visual image, and relative depth maps derived from images. The integration of multiple data sources in road segmentation presents both opportunities and challenges. One of the primary challenges is the scarcity of large-scale, accurately labeled datasets that are necessary for training robust deep learning models. To address this, we have developed the [UdeerLID+] framework under a semi-supervised learning paradigm. Experiments results on KITTI datasets validate the superior performance.

Zichao Dong, Yilin Zhang, Xufeng Huang et al.

We introduce a novel MV-DETR pipeline which is effective while efficient transformer based detection method. Given input RGBD data, we notice that there are super strong pretraining weights for RGB data while less effective works for depth related data. First and foremost , we argue that geometry and texture cues are both of vital importance while could be encoded separately. Secondly, we find that visual texture feature is relatively hard to extract compared with geometry feature in 3d space. Unfortunately, single RGBD dataset with thousands of data is not enough for training an discriminating filter for visual texture feature extraction. Last but certainly not the least, we designed a lightweight VG module consists of a visual textual encoder, a geometry encoder and a VG connector. Compared with previous state of the art works like V-DETR, gains from pretrained visual encoder could be seen. Extensive experiments on ScanNetV2 dataset shows the effectiveness of our method. It is worth mentioned that our method achieve 78\% AP which create new state of the art on ScanNetv2 benchmark.

Shanglin Yang, Zhan Shi

Large language models provide rich semantic priors and strong reasoning capabilities, making them promising auxiliary signals for recommendation. However, prevailing approaches either deploy LLMs as standalone recommender or apply global knowledge distillation, both of which suffer from inherent drawbacks. Standalone LLM recommender are costly, biased, and unreliable across large regions of the user item space, while global distillation forces the downstream model to imitate LLM predictions even when such guidance is inaccurate. Meanwhile, recent studies show that LLMs excel particularly in re-ranking and challenging scenarios, rather than uniformly across all contexts.We introduce Selective LLM Guided Regularization, a model-agnostic and computation efficient framework that activates LLM based pairwise ranking supervision only when a trainable gating mechanism informing by user history length, item popularity, and model uncertainty predicts the LLM to be reliable. All LLM scoring is performed offline, transferring knowledge without increasing inference cost. Experiments across multiple datasets show that this selective strategy consistently improves overall accuracy and yields substantial gains in cold start and long tail regimes, outperforming global distillation baselines.

Zhan Shi, Jingwei Zhang, Jun Kong et al.

In digital pathology, the multiple instance learning (MIL) strategy is widely used in the weakly supervised histopathology whole slide image (WSI) classification task where giga-pixel WSIs are only labeled at the slide level. However, existing attention-based MIL approaches often overlook contextual information and intrinsic spatial relationships between neighboring tissue tiles, while graph-based MIL frameworks have limited power to recognize the long-range dependencies. In this paper, we introduce the integrative graph-transformer framework that simultaneously captures the context-aware relational features and global WSI representations through a novel Graph Transformer Integration (GTI) block. Specifically, each GTI block consists of a Graph Convolutional Network (GCN) layer modeling neighboring relations at the local instance level and an efficient global attention model capturing comprehensive global information from extensive feature embeddings. Extensive experiments on three publicly available WSI datasets: TCGA-NSCLC, TCGA-RCC and BRIGHT, demonstrate the superiority of our approach over current state-of-the-art MIL methods, achieving an improvement of 1.0% to 2.6% in accuracy and 0.7%-1.6% in AUROC.

Hang Ji, Tao Ni, Xufeng Huang et al.

This technical report introduces a targeted improvement to the StreamPETR framework, specifically aimed at enhancing velocity estimation, a critical factor influencing the overall NuScenes Detection Score. While StreamPETR exhibits strong 3D bounding box detection performance as reflected by its high mean Average Precision our analysis identified velocity estimation as a substantial bottleneck when evaluated on the NuScenes dataset. To overcome this limitation, we propose a customized positional embedding strategy tailored to enhance temporal modeling capabilities. Experimental evaluations conducted on the NuScenes test set demonstrate that our improved approach achieves a state-of-the-art NDS of 70.86% using the ViT-L backbone, setting a new benchmark for camera-only 3D object detection.

Weimin Lyu, Qingqiao Hu, Kehan Qi et al.

Whole-slide images (WSIs) in pathology can reach up to 10,000 x 10,000 pixels, posing significant challenges for multimodal large language model (MLLM) due to long context length and high computational demands. Previous methods typically focus on patch-level analysis or slide-level classification using CLIP-based models with multi-instance learning, but they lack the generative capabilities needed for visual question answering (VQA). More recent MLLM-based approaches address VQA by feeding thousands of patch tokens directly into the language model, which leads to excessive resource consumption. To address these limitations, we propose Token Compression Pathology LLaVA (TCP-LLaVA), the first MLLM architecture to perform WSI VQA via token compression. TCP-LLaVA introduces a set of trainable compression tokens that aggregate visual and textual information through a modality compression module, inspired by the [CLS] token mechanism in BERT. Only the compressed tokens are forwarded to the LLM for answer generation, significantly reducing input length and computational cost. Experiments on ten TCGA tumor subtypes show that TCP-LLaVA outperforms existing MLLM baselines in VQA accuracy while reducing training resource consumption by a substantial margin.

Zhan Shi, Shanglin Yang

Fashion, deeply rooted in sociocultural dynamics, evolves as individuals emulate styles popularized by influencers and iconic figures. In the quest to replicate such refined tastes using artificial intelligence, traditional fashion ensemble methods have primarily used supervised learning to imitate the decisions of style icons, which falter when faced with distribution shifts, leading to style replication discrepancies triggered by slight variations in input. Meanwhile, large language models (LLMs) have become prominent across various sectors, recognized for their user-friendly interfaces, strong conversational skills, and advanced reasoning capabilities. To address these challenges, we introduce the Fashion Large Language Model (FLLM), which employs auto-prompt generation training strategies to enhance its capacity for delivering personalized fashion advice while retaining essential domain knowledge. Additionally, by integrating a retrieval augmentation technique during inference, the model can better adjust to individual preferences. Our results show that this approach surpasses existing models in accuracy, interpretability, and few-shot learning capabilities.

Chunlin Tian, Zhan Shi, Xinpeng Qin et al.

Federated Learning (FL) enables multiple devices to collaboratively train a shared model while ensuring data privacy. The selection of participating devices in each training round critically affects both the model performance and training efficiency, especially given the vast heterogeneity in training capabilities and data distribution across devices. To address these challenges, we introduce a novel device selection solution called FedRank, which is an end-to-end, ranking-based approach that is pre-trained by imitation learning against state-of-the-art analytical approaches. It not only considers data and system heterogeneity at runtime but also adaptively and efficiently chooses the most suitable clients for model training. Specifically, FedRank views client selection in FL as a ranking problem and employs a pairwise training strategy for the smart selection process. Additionally, an imitation learning-based approach is designed to counteract the cold-start issues often seen in state-of-the-art learning-based approaches. Experimental results reveal that \model~ boosts model accuracy by 5.2\% to 56.9\%, accelerates the training convergence up to $2.01 \times$ and saves the energy consumption up to $40.1\%$.

Yining Lu, Zilong Wang, Shiyang Li et al.

Prior works in multi-objective reinforcement learning typically use linear reward scalarization with fixed weights, which provably fail to capture non-convex Pareto fronts and thus yield suboptimal results. This limitation becomes especially critical in online preference alignment for large language models. Here, stochastic trajectories generated by parameterized policies create highly non-linear and non-convex mappings from parameters to objectives that no single static weighting scheme can find optimal trade-offs. We address this limitation by introducing dynamic reward weighting, which adaptively adjusts reward weights during the online reinforcement learning process. Unlike existing approaches that rely on fixed-weight interpolation, our dynamic weighting continuously balances and prioritizes objectives in training, facilitating effective exploration of Pareto fronts in objective space. We introduce two approaches of increasing sophistication and generalizability: (1) hypervolume-guided weight adaptation and (2) gradient-based weight optimization, offering a versatile toolkit for online multi-objective alignment. Our extensive experiments demonstrate their compatibility with commonly used online reinforcement learning algorithms (including GRPO, REINFORCE, and RLOO), effectiveness across multiple mathematical reasoning datasets, and applicability to different model families, consistently achieving Pareto dominant solutions with fewer training steps than fixed-weight linear scalarization baselines.

Zhining Liu, Ziyi Chen, Hui Liu et al.

Vision-Language Models (VLMs) achieve strong results on multimodal tasks such as visual question answering, yet they can still fail even when the correct visual evidence is present. In this work, we systematically investigate whether these failures arise from not perceiving the evidence or from not leveraging it effectively. By examining layer-wise attention dynamics, we find that shallow layers focus primarily on text, while deeper layers sparsely but reliably attend to localized evidence regions. Surprisingly, VLMs often perceive the visual evidence when outputting incorrect answers, a phenomenon we term ``seeing but not believing'' that widely exists in major VLM families. Building on this, we introduce an inference-time intervention that highlights deep-layer evidence regions through selective attention-based masking. It requires no training and consistently improves accuracy across multiple families, including LLaVA, Qwen, Gemma, and InternVL. These results show that VLMs encode reliable evidence internally but under-utilize it, making such signals explicit can bridge the gap between perception and reasoning, advancing the diagnostic understanding and reliability of VLMs.

Zhan Shi, Yefeng Yuan, Yuhong Liu et al.

The performance of modern machine learning systems depends on access to large, high-quality datasets, often sourced from user-generated content or proprietary, domain-specific corpora. However, these rich datasets inherently contain sensitive personal information, raising significant concerns about privacy, data security, and compliance with regulatory frameworks. While conventional anonymization techniques can remove explicit identifiers, such removal may result in performance drop in downstream machine learning tasks. More importantly, simple anonymization may not be effective against inference attacks that exploit implicit signals such as writing style, topical focus, or demographic cues, highlighting the need for more robust privacy safeguards during model training. To address the challenging issue of balancing user privacy and data utility, we propose a reinforcement learning framework that fine-tunes a large language model (LLM) using a composite reward function that jointly optimizes for explicit and implicit privacy, semantic fidelity, and output diversity. To effectively capture population level regularities, the privacy reward combines semantic cues with structural patterns derived from a minimum spanning tree (MST) over latent representations. By modeling these privacy-sensitive signals in their distributional context, the proposed approach guides the model to generate synthetic rewrites that preserve utility while mitigating privacy risks. Empirical results show that the proposed method significantly enhances author obfuscation and privacy metrics without degrading semantic quality, providing a scalable and model-agnostic solution for privacy preserving data generation in the era of large language models.

Tevin Atwal, Chan Nam Tieu, Yefeng Yuan et al.

The increasing use of synthetic data generated by Large Language Models (LLMs) presents both opportunities and challenges in data-driven applications. While synthetic data provides a cost-effective, scalable alternative to real-world data to facilitate model training, its diversity and privacy risks remain underexplored. Focusing on text-based synthetic data, we propose a comprehensive set of metrics to quantitatively assess the diversity (i.e., linguistic expression, sentiment, and user perspective), and privacy (i.e., re-identification risk and stylistic outliers) of synthetic datasets generated by several state-of-the-art LLMs. Experiment results reveal significant limitations in LLMs' capabilities in generating diverse and privacy-preserving synthetic data. Guided by the evaluation results, a prompt-based approach is proposed to enhance the diversity of synthetic reviews while preserving reviewer privacy.

Maryam Aliakbarpour, Zhan Shi, Ria Stevens et al.

Estimating the density of a distribution from its samples is a fundamental problem in statistics. Hypothesis selection addresses the setting where, in addition to a sample set, we are given $n$ candidate distributions -- referred to as hypotheses -- and the goal is to determine which one best describes the underlying data distribution. This problem is known to be solvable very efficiently, requiring roughly $O(\log n)$ samples and running in $\tilde{O}(n)$ time. The quality of the output is measured via the total variation distance to the unknown distribution, and the approximation factor of the algorithm determines how large this distance is compared to the optimal distance achieved by the best candidate hypothesis. It is known that $α= 3$ is the optimal approximation factor for this problem. We study hypothesis selection under the constraint of differential privacy. We propose a differentially private algorithm in the central model that runs in nearly-linear time with respect to the number of hypotheses, achieves the optimal approximation factor, and incurs only a modest increase in sample complexity, which remains polylogarithmic in $n$. This resolves an open question posed by [Bun, Kamath, Steinke, Wu, NeurIPS 2019]. Prior to our work, existing upper bounds required quadratic time.

Zixin Ding, Junyuan Hong, Zhan Shi et al.

LLM-based prompt optimization, that uses LLM-provided "textual gradients" (feedback) to refine prompts, has emerged an effective method for automatic prompt engineering. However, its scalability and stability are unclear when using more data in training. We systematically investigate the potential and challenges of scaling training data in textual gradient descent. We show that naively scaling training examples is infeasible due to both explicit context-length limits and an implicit context wall, where long-context degradation yields diminishing returns. Inspired by prior wisdom in stochastic gradient descent, we propose Textual Stochastic Gradient Descent with Momentum (TSGD-M), which reweights updates through momentum sampling, using bootstrapped minibatch validation accuracy as importance weights over historical prompts. We introduce Gumbel-Top-$k$ sampling for prompt generation, balancing exploration--exploitation and improving sampling efficiency while maintaining a low-variance running mean estimator. TSGD-M integrates seamlessly into existing prompt optimization frameworks, including TextGrad, DSPy-COPRO, and AdalFlow, and achieves consistent gains across 5 benchmarks.

Subramaniam Vincent, Phoebe Wang, Zhan Shi et al.

Since the launch of ChatGPT in late 2022, the capacities of Large Language Models and their evaluation have been in constant discussion and evaluation both in academic research and in the industry. Scenarios and benchmarks have been developed in several areas such as law, medicine and math (Bommasani et al., 2023) and there is continuous evaluation of model variants. One area that has not received sufficient scenario development attention is journalism, and in particular journalistic sourcing and ethics. Journalism is a crucial truth-determination function in democracy (Vincent, 2023), and sourcing is a crucial pillar to all original journalistic output. Evaluating the capacities of LLMs to annotate stories for the different signals of sourcing and how reporters justify them is a crucial scenario that warrants a benchmark approach. It offers potential to build automated systems to contrast more transparent and ethically rigorous forms of journalism with everyday fare. In this paper we lay out a scenario to evaluate LLM performance on identifying and annotating sourcing in news stories on a five-category schema inspired from journalism studies (Gans, 2004). We offer the use case, our dataset and metrics and as the first step towards systematic benchmarking. Our accuracy findings indicate LLM-based approaches have more catching to do in identifying all the sourced statements in a story, and equally, in matching the type of sources. An even harder task is spotting source justifications.

Haochen Tan, Zhijiang Guo, Zhan Shi et al.

Large Language Models (LLMs) have succeeded remarkably in understanding long-form contents. However, exploring their capability for generating long-form contents, such as reports and articles, has been relatively unexplored and inadequately assessed by existing benchmarks. The prevalent evaluation methods, which predominantly rely on crowdsourcing, are recognized for their labor-intensive nature and lack of efficiency, whereas automated metrics, such as the ROUGE score, demonstrate discordance with human judgment criteria. In this paper, we propose ProxyQA, an innovative framework dedicated to assessing long-text generation. ProxyQA comprises in-depth human-curated meta-questions spanning various domains, each accompanied by specific proxy-questions with pre-annotated answers. LLMs are tasked to generate extensive content in response to these meta-questions, by engaging an evaluator and incorporating the generated texts as contextual background, ProxyQA assesses the generated content's quality through the evaluator's accuracy in addressing the proxy-questions. We examine multiple LLMs, emphasizing ProxyQA's demanding nature as a high-quality assessment tool. Human evaluation demonstrates that the proxy-question method is notably self-consistent and aligns closely with human evaluative standards. The dataset and leaderboard is available at \url{https://proxy-qa.com}.

Kevin Miller, John Mauro, Jason Setiadi et al.

We present a novel method for classification of Synthetic Aperture Radar (SAR) data by combining ideas from graph-based learning and neural network methods within an active learning framework. Graph-based methods in machine learning are based on a similarity graph constructed from the data. When the data consists of raw images composed of scenes, extraneous information can make the classification task more difficult. In recent years, neural network methods have been shown to provide a promising framework for extracting patterns from SAR images. These methods, however, require ample training data to avoid overfitting. At the same time, such training data are often unavailable for applications of interest, such as automatic target recognition (ATR) and SAR data. We use a Convolutional Neural Network Variational Autoencoder (CNNVAE) to embed SAR data into a feature space, and then construct a similarity graph from the embedded data and apply graph-based semi-supervised learning techniques. The CNNVAE feature embedding and graph construction requires no labeled data, which reduces overfitting and improves the generalization performance of graph learning at low label rates. Furthermore, the method easily incorporates a human-in-the-loop for active learning in the data-labeling process. We present promising results and compare them to other standard machine learning methods on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset for ATR with small amounts of labeled data.

Xiaoyu Yang, Xiaodan Zhu, Zhan Shi et al.

While recent research on natural language inference has considerably benefited from large annotated datasets, the amount of inference-related knowledge (including commonsense) provided in the annotated data is still rather limited. There have been two lines of approaches that can be used to further address the limitation: (1) unsupervised pretraining can leverage knowledge in much larger unstructured text data; (2) structured (often human-curated) knowledge has started to be considered in neural-network-based models for NLI. An immediate question is whether these two approaches complement each other, or how to develop models that can bring together their advantages. In this paper, we propose models that leverage structured knowledge in different components of pre-trained models. Our results show that the proposed models perform better than previous BERT-based state-of-the-art models. Although our models are proposed for NLI, they can be easily extended to other sentence or sentence-pair classification problems.

Hui Liu, Zhan Shi, Xiaodan Zhu

Conversation disentanglement aims to separate intermingled messages into detached sessions, which is a fundamental task in understanding multi-party conversations. Existing work on conversation disentanglement relies heavily upon human-annotated datasets, which are expensive to obtain in practice. In this work, we explore to train a conversation disentanglement model without referencing any human annotations. Our method is built upon a deep co-training algorithm, which consists of two neural networks: a message-pair classifier and a session classifier. The former is responsible for retrieving local relations between two messages while the latter categorizes a message to a session by capturing context-aware information. Both networks are initialized respectively with pseudo data built from an unannotated corpus. During the deep co-training process, we use the session classifier as a reinforcement learning component to learn a session assigning policy by maximizing the local rewards given by the message-pair classifier. For the message-pair classifier, we enrich its training data by retrieving message pairs with high confidence from the disentangled sessions predicted by the session classifier. Experimental results on the large Movie Dialogue Dataset demonstrate that our proposed approach achieves competitive performance compared to the previous supervised methods. Further experiments show that the predicted disentangled conversations can promote the performance on the downstream task of multi-party response selection.

Zhan Shi, Chirag Sakhuja, Milad Hashemi et al.

The use of deep learning has grown at an exponential rate, giving rise to numerous specialized hardware and software systems for deep learning. Because the design space of deep learning software stacks and hardware accelerators is diverse and vast, prior work considers software optimizations separately from hardware architectures, effectively reducing the search space. Unfortunately, this bifurcated approach means that many profitable design points are never explored. This paper instead casts the problem as hardware/software co-design, with the goal of automatically identifying desirable points in the joint design space. The key to our solution is a new constrained Bayesian optimization framework that avoids invalid solutions by exploiting the highly constrained features of this design space, which are semi-continuous/semi-discrete. We evaluate our optimization framework by applying it to a variety of neural models, improving the energy-delay product by 18% (ResNet) and 40% (DQN) over hand-tuned state-of-the-art systems, as well as demonstrating strong results on other neural network architectures, such as MLPs and Transformers.

Zhan Shi, Xu Zhou, Xipeng Qiu et al.

Image captioning is a multimodal problem that has drawn extensive attention in both the natural language processing and computer vision community. In this paper, we present a novel image captioning architecture to better explore semantics available in captions and leverage that to enhance both image representation and caption generation. Our models first construct caption-guided visual relationship graphs that introduce beneficial inductive bias using weakly supervised multi-instance learning. The representation is then enhanced with neighbouring and contextual nodes with their textual and visual features. During generation, the model further incorporates visual relationships using multi-task learning for jointly predicting word and object/predicate tag sequences. We perform extensive experiments on the MSCOCO dataset, showing that the proposed framework significantly outperforms the baselines, resulting in the state-of-the-art performance under a wide range of evaluation metrics.

Zac Cranko, Zhan Shi, Xinhua Zhang et al.

The problem of adversarial examples has highlighted the need for a theory of regularisation that is general enough to apply to exotic function classes, such as universal approximators. In response, we give a very general equality result regarding the relationship between distributional robustness and regularisation, as defined with a transportation cost uncertainty set. The theory allows us to (tightly) certify the robustness properties of a Lipschitz-regularised model with very mild assumptions. As a theoretical application we show a new result explicating the connection between adversarial learning and distributional robustness. We then give new results for how to achieve Lipschitz regularisation of kernel classifiers, which are demonstrated experimentally.

Zhan Shi, Kevin Swersky, Daniel Tarlow et al.

As the performance of computer systems stagnates due to the end of Moore's Law, there is a need for new models that can understand and optimize the execution of general purpose code. While there is a growing body of work on using Graph Neural Networks (GNNs) to learn representations of source code, these representations do not understand how code dynamically executes. In this work, we propose a new approach to use GNNs to learn fused representations of general source code and its execution. Our approach defines a multi-task GNN over low-level representations of source code and program state (i.e., assembly code and dynamic memory states), converting complex source code constructs and complex data structures into a simpler, more uniform format. We show that this leads to improved performance over similar methods that do not use execution and it opens the door to applying GNN models to new tasks that would not be feasible from static code alone. As an illustration of this, we apply the new model to challenging dynamic tasks (branch prediction and prefetching) from the SPEC CPU benchmark suite, outperforming the state-of-the-art by 26% and 45% respectively. Moreover, we use the learned fused graph embeddings to demonstrate transfer learning with high performance on an indirectly related task (algorithm classification).

Yu-Ping Ruan, Xiaodan Zhu, Zhen-Hua Ling et al.

Winograd Schema Challenge (WSC) was proposed as an AI-hard problem in testing computers' intelligence on common sense representation and reasoning. This paper presents the new state-of-theart on WSC, achieving an accuracy of 71.1%. We demonstrate that the leading performance benefits from jointly modelling sentence structures, utilizing knowledge learned from cutting-edge pretraining models, and performing fine-tuning. We conduct detailed analyses, showing that fine-tuning is critical for achieving the performance, but it helps more on the simpler associative problems. Modelling sentence dependency structures, however, consistently helps on the harder non-associative subset of WSC. Analysis also shows that larger fine-tuning datasets yield better performances, suggesting the potential benefit of future work on annotating more Winograd schema sentences.

Zac Cranko, Simon Kornblith, Zhan Shi et al.

Robust risk minimisation has several advantages: it has been studied with regards to improving the generalisation properties of models and robustness to adversarial perturbation. We bound the distributionally robust risk for a model class rich enough to include deep neural networks by a regularised empirical risk involving the Lipschitz constant of the model. This allows us to interpretand quantify the robustness properties of a deep neural network. As an application we show the distributionally robust risk upperbounds the adversarial training risk.

Zac Cranko, Aditya Krishna Menon, Richard Nock et al.

The last few years have seen a staggering number of empirical studies of the robustness of neural networks in a model of adversarial perturbations of their inputs. Most rely on an adversary which carries out local modifications within prescribed balls. None however has so far questioned the broader picture: how to frame a resource-bounded adversary so that it can be severely detrimental to learning, a non-trivial problem which entails at a minimum the choice of loss and classifiers. We suggest a formal answer for losses that satisfy the minimal statistical requirement of being proper. We pin down a simple sufficient property for any given class of adversaries to be detrimental to learning, involving a central measure of "harmfulness" which generalizes the well-known class of integral probability metrics. A key feature of our result is that it holds for all proper losses, and for a popular subset of these, the optimisation of this central measure appears to be independent of the loss. When classifiers are Lipschitz -- a now popular approach in adversarial training --, this optimisation resorts to optimal transport to make a low-budget compression of class marginals. Toy experiments reveal a finding recently separately observed: training against a sufficiently budgeted adversary of this kind improves generalization.

Zhan Shi, Xinchi Chen, Xipeng Qiu et al.

Text generation is a crucial task in NLP. Recently, several adversarial generative models have been proposed to improve the exposure bias problem in text generation. Though these models gain great success, they still suffer from the problems of reward sparsity and mode collapse. In order to address these two problems, in this paper, we employ inverse reinforcement learning (IRL) for text generation. Specifically, the IRL framework learns a reward function on training data, and then an optimal policy to maximum the expected total reward. Similar to the adversarial models, the reward and policy function in IRL are optimized alternately. Our method has two advantages: (1) the reward function can produce more dense reward signals. (2) the generation policy, trained by "entropy regularized" policy gradient, encourages to generate more diversified texts. Experiment results demonstrate that our proposed method can generate higher quality texts than the previous methods.

Xinchi Chen, Zhan Shi, Xipeng Qiu et al.

Neural word segmentation has attracted more and more research interests for its ability to alleviate the effort of feature engineering and utilize the external resource by the pre-trained character or word embeddings. In this paper, we propose a new neural model to incorporate the word-level information for Chinese word segmentation. Unlike the previous word-based models, our model still adopts the framework of character-based sequence labeling, which has advantages on both effectiveness and efficiency at the inference stage. To utilize the word-level information, we also propose a new long short-term memory (LSTM) architecture over directed acyclic graph (DAG). Experimental results demonstrate that our model leads to better performances than the baseline models.

Xinchi Chen, Zhan Shi, Xipeng Qiu et al.

Different linguistic perspectives causes many diverse segmentation criteria for Chinese word segmentation (CWS). Most existing methods focus on improve the performance for each single criterion. However, it is interesting to exploit these different criteria and mining their common underlying knowledge. In this paper, we propose adversarial multi-criteria learning for CWS by integrating shared knowledge from multiple heterogeneous segmentation criteria. Experiments on eight corpora with heterogeneous segmentation criteria show that the performance of each corpus obtains a significant improvement, compared to single-criterion learning. Source codes of this paper are available on Github.