Searcharxiv⌕ Search

arXiv subjects

Peijie Qiu

Publications and source records attributed to Peijie Qiu.

At least 19 recordsLinked to original sources

Simplify, Agree, Distill: Label-Free Self-Distillation with Stepping Stones

Can a language model teach itself to reason without labels? On-policy self-distillation (OPSD) uses the model as its own teacher, but strengthens that teacher with a ground-truth solution. The natural label-free substitute, voting over the model's attempts, fails precisely on hard problems: attempts disagree, and their most common answer is often wrong. We propose STONE, a label-free alternative that moves agreement to easier related problems. The model writes these stepping stones, keeps those it solves consistently, and shows the teacher their worked solutions, with final answers removed; the student sees only the original problem. No step reads a label, and inference needs no extra context or model calls. Across five math benchmarks, the proposed method achieves the highest average accuracy for Qwen3-4B and Qwen3-8B in thinking mode. Without thinking mode, it gains 13.2 and 22.7 points, 2.5 and 2.7 times the gains of OPSD with ground-truth solutions.

cs.LG↗

FGCARL: Fisher-Guided Continuously Adaptive Representation Learning for Medical Image Segmentation

Deep learning (DL) has achieved remarkable success in medical image segmentation by learning data-driven representations from large-scale annotated datasets. However, medical images are often collected independently across clinical sites, resulting in substantial domain shifts due to differences in imaging modalities, acquisition protocols, and intensity distributions. Meanwhile, segmentation classes may vary across sites due to differences in annotation standards, leading to class shifts. These evolving domains and class definitions can substantially degrade segmentation performance on newly encountered data, while sequential adaptation may cause catastrophic forgetting of knowledge acquired from previously encountered domains and classes. Existing incremental learning methods for medical image segmentation primarily address either domain shifts or class shifts, with these two settings largely investigated independently. This separation leaves a critical gap in developing a unified approach that can simultaneously adapt to evolving domains and newly introduced classes while retaining previously acquired knowledge. To address this challenge, we propose Fisher-Guided Continuously Adaptive Representation Learning, a unified framework for domain- and class-incremental 3D volumetric medical image segmentation. FGCARL comprises two key components: Mutual Adaptive Instance Representation (MAIR), which adaptively aligns feature distributions across sequentially encountered domains through bidirectional normalization and denormalization, and Fisher-Guided Knowledge Retention (FGKR), which selectively preserves parameters important for previously learned domains and classes through Fisher-guided regularization. Experimental results demonstrate that FGCARL consistently outperformed state-of-the-art methods, effectively adapting to domain and class shifts while mitigating catastrophic forgetting.

eess.IV↗

ActiveMedAgent: Cost-Aware Trajectory Learning for Multimodal Medical Diagnosis

Clinical diagnosis is inherently sequential: clinicians escalate from cheap to costly tests only when additional evidence is expected to resolve diagnostic uncertainty. We present ActiveMedAgent, a framework that brings this cost-aware sequential logic to multimodal medical AI. Given a frozen, API-accessed vision-language model, ActiveMedAgent tracks probability distributions over candidate diagnoses and scores each acquisition by its per-step diagnostic utility minus cost. A lightweight MLP controller is then trained offline on these scored trajectories, learning when to request additional evidence and when to commit. Across three commonly used benchmarks, trajectory-based policy learning consistently outperforms both unguided acquisition and full-modality baselines. Notably, we identify an information overload effect. In 175 cases, the agent produces a correct diagnosis with fewer channels while the full-modality baseline fails, showing that learning what to omit can be as important as learning what to acquire.

cs.LG↗

LadderEdit: Edit-Level Residual Compression for Memory-Efficient Lifelong Editing of LLMs

Lifelong editing of LLMs requires storing thousands of edits after acquisition. A widely used family of approaches attaches one LoRA adapter per edit, which preserves behavior but grows linearly in storage. To address this challenge, we propose LadderEdit, a method that compresses each LoRA adapter after it is acquired. Each edit is first stored at low rank as a cheap sketch. We then check whether this sketch still satisfies the rewrite, generalization, and locality contract on probe prompts. Edits that pass keep the sketch; those that fail are promoted to a higher rank along a ladder until the contract is met. Because every edit retains some representation, coverage is maintained, and only hard edits consume more rank. Across ZsRE, CounterFact, and WikiBigEdit benchmarks on LLaMA-3-8B, Mistral-7B, and Qwen2.5-7B, LadderEdit tracks exact LoRA storage at 5.2x less memory and remains effective at 50,000 sequential edits.

cs.CL↗

TEFM: Token-Efficient Faithful Modeling for Structured Data

In this paper, we solve two fundamental obstacles in applying LLMs to critical domains: token efficiency and faithfulness. To address both constraints jointly, we present TEFM (Token-Efficient Faithful Modeling), a framework designed for structured data analysis in critical domains. TEFM achieves token efficiency by compressing lengthy structured observations into compact Behavioral Code tokens, dramatically reducing token consumption with minimal information loss. Moreover, TEFM enables faithful rationalization through a dual-fidelity objective that jointly optimizes code-level reconstruction and prediction-level fidelity, identifying minimal sufficient feature subsets grounded in input data. Comprehensive experiments across various domain datasets and model backbones (Qwen3, Gemma-2, Phi-4) show that TEFM achieves competitive classification accuracy with dramatic token reduction (approximately 1\% token retention in clinical and 2\% in security domains) while producing faithful rationales.

cs.CL↗

SODA: Semi On-Policy Black-Box Distillation for Large Language Models

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.

cs.LG↗

DRA-GRPO: Your GRPO Needs to Know Diverse Reasoning Paths for Mathematical Reasoning

Post-training LLMs with Reinforcement Learning, specifically Group Relative Policy Optimization (GRPO), has emerged as a paradigm for enhancing mathematical reasoning. However, standard GRPO relies on scalar correctness rewards that are often non-injective with respect to semantic content: distinct reasoning paths receive identical rewards. This leads to a Diversity-Quality Inconsistency, where the policy collapses into a narrow set of dominant modes while ignoring equally valid but structurally novel strategies. To bridge this gap, we propose Diversity-aware Reward Adjustment (DRA), a theoretically grounded framework that calibrates the reward signal using the semantic density of sampled groups. By leveraging Submodular Mutual Information (SMI), DRA implements an Inverse Propensity Scoring (IPS) mechanism that effectively de-biases the gradient estimation. This creates a repulsive force against redundancy, driving the policy to achieve better coverage of the high-reward landscape. Our method is plug-and-play and integrates seamlessly with GRPO variants. Empirical evaluations on five math benchmarks demonstrate that DRA-GRPO consistently outperforms strong baselines, achieving an average accuracy of 58.2% on DeepSeek-R1-Distill-Qwen-1.5B with only 7,000 training samples and $55 cost, highlighting the critical role of diversity calibration in data-efficient alignment. The code is available at https://github.com/xiwenc1/DRA-GRPO.

cs.CL↗

S-SPPO: Semantic-Calibrated Self-Play Preference Optimization

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.

cs.AI↗

Active Source-free Domain Adaptation in Open-set Medical Image Segmentation via Decomposed Uncertainty and Prototype Discrepancy

Deep learning (DL) methods are challenged to demonstrate robust performance across different segmentation datasets due to domain shifts, but active domain adaptation techniques enhance their generalization performance by querying a few samples from target domains for adaptation training. However in clinical practice, target domains often include private classes of new anatomical structures or pathologies that are not presented in the source data, and existing methods implement closed-set segmentation where source and target domains have the same segmentation classes. Additionally, source data are often inaccessible during adaptation due to strict data privacy regulations. To address these limitations, we propose an Active Source-free Open-set Domain Adaptation (ASFOSDA) method which is the first work to implement active learning for adapting DL models in open-set medical image segmentation without the access to source data. This method employs an active open-set query strategy to select the most informative target samples for training models based on Class-aware Decomposed Uncertainty (CDU) and Class-agnostic Prototype Discrepancy (CPD). CDU measures sample aleatoric uncertainty and model epistemic uncertainty by employing test time augmentation in stochastic processes. CPD measures cross-domain and self-domain discrepancy for selecting diverse samples. Subsequently, to boost the adaptation performance by enhancing training samples, a Target-refined Self-training strategy is proposed to generate high-quality pseudo labels for unselected samples, thus combining them with labeled samples for a semi-supervised training. We evaluated our method on cross-domain open-set volumetric medical image segmentation tasks, and it outperformed state-of-the-art adaptation methods.

eess.IV↗

Mags-RL: Wearing Multimodal LLMs a Magnifying Glass via Agentic Reinforcement Learning For Complex Scene Reasoning

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.

cs.CV↗

FEFormer: Frequency-enhanced Vision Transformer for Generic Knowledge Extraction and Adaptive Feature Fusion in Volumetric Medical Image Segmentation

Accurate segmentation of organs and lesions in medical images is essential for clinical applications including diagnosis, prognosis, and treatment planning. While Vision Transformers (ViTs) have shown impressive segmentation performance, they face key challenges in module and architecture design. Specifically, self-attention struggles to capture fine-grained local features critical for understanding detailed anatomical structures, standard MLP modules lack explicit mechanisms to preserve spatial information, conventional encoder-decoder architectures rely on naive feature fusion strategies that cannot handle large semantic discrepancies, and existing designs lack explicit mechanisms to propagate low-level information from encoder to decoder. To address these limitations, we propose a Frequency-enhanced Vision Transformer (FEFormer) for robust and efficient volumetric medical image segmentation that explicitly models frequency information to jointly capture global context and fine structural details. FEFormer comprises four novel components: a Frequency-enhanced Dynamic Self-Attention (FDSA) module that jointly captures fine-grained local details and global long-range dependencies through locality-preserving convolution with frequency-domain attention; a Frequency-decomposed Gating MLP (FGMLP) that adaptively models low- and high-frequency components for enhanced semantic and structural representation; a Wavelet-guided Adaptive Feature Fusion (WAFF) module that enables semantically consistent encoder-decoder feature integration in the frequency domain; and a Frequency-enabled Cross-scale Stem Bridge (FCSB) that enhances low-level feature propagation across scales. Evaluated on four diverse volumetric medical image segmentation tasks, FEFormer achieved superior segmentation performance with high computational efficiency compared to state-of-the-art methods.

eess.IV↗

Your Agent is More Brittle Than You Think: Uncovering Indirect Injection Vulnerabilities in Agentic LLMs

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.

cs.CL↗

OTPrune: Distribution-Aligned Visual Token Pruning via Optimal Transport

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.

cs.CV↗

LLaDA-MedV: Exploring Large Language Diffusion Models for Biomedical Image Understanding

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 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.

cs.CV↗

ImageRAGTurbo: Towards One-step Text-to-Image Generation with Retrieval-Augmented Diffusion Models

Diffusion models have emerged as the leading approach for text-to-image generation. However, their iterative sampling process, which gradually morphs random noise into coherent images, introduces significant latency that limits their applicability. While recent few-step diffusion models reduce the number of sampling steps to as few as one to four steps, they often compromise image quality and prompt alignment, especially in one-step generation. Additionally, these models require computationally expensive training procedures. To address these limitations, we propose ImageRAGTurbo, a novel approach to efficiently finetune few-step diffusion models via retrieval augmentation. Given a text prompt, we retrieve relevant text-image pairs from a database and use them to condition the generation process. We argue that such retrieved examples provide rich contextual information to the UNet denoiser that helps reduce the number of denoising steps without compromising image quality. Indeed, our initial investigations show that using the retrieved content to edit the denoiser's latent space ($\mathcal{H}$-space) without additional finetuning already improves prompt fidelity. To further improve the quality of the generated images, we augment the UNet denoiser with a trainable adapter in the $\mathcal{H}$-space, which efficiently blends the retrieved content with the target prompt using a cross-attention mechanism. Experimental results on fast text-to-image generation demonstrate that our approach produces high-fidelity images without compromising latency compared to existing methods.

cs.CV↗

Multimodal normative modeling in Alzheimers Disease with introspective variational autoencoders

Normative modeling learns a healthy reference distribution and quantifies subject-specific deviations to capture heterogeneous disease effects. In Alzheimers disease (AD), multimodal neuroimaging offers complementary signals but VAE-based normative models often (i) fit the healthy reference distribution imperfectly, inflating false positives, and (ii) use posterior aggregation (e.g., PoE/MoE) that can yield weak multimodal fusion in the shared latent space. We propose mmSIVAE, a multimodal soft-introspective variational autoencoder combined with Mixture-of-Product-of-Experts (MOPOE) aggregation to improve reference fidelity and multimodal integration. We compute deviation scores in latent space and feature space as distances from the learned healthy distributions, and map statistically significant latent deviations to regional abnormalities for interpretability. On ADNI MRI regional volumes and amyloid PET SUVR, mmSIVAE improves reconstruction on held-out controls and produces more discriminative deviation scores for outlier detection than VAE baselines, with higher likelihood ratios and clearer separation between control and AD-spectrum cohorts. Deviation maps highlight region-level patterns aligned with established AD-related changes. More broadly, our results highlight the importance of training objectives that prioritize reference-distribution fidelity and robust multimodal posterior aggregation for normative modeling, with implications for deviation-based analysis across multimodal clinical data.

cs.LG↗

AriadneMem: Threading the Maze of Lifelong Memory for LLM Agents

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.

cs.CL↗

U-Harmony: Enhancing Joint Training for Segmentation Models with Universal Harmonization

In clinical practice, medical segmentation datasets are often limited and heterogeneous, with variations in modalities, protocols, and anatomical targets across institutions. Existing deep learning models struggle to jointly learn from such diverse data, often sacrificing either generalization or domain-specific knowledge. To overcome these challenges, we propose a joint training method called Universal Harmonization (U-Harmony), which can be integrated into deep learning-based architectures with a domain-gated head, enabling a single segmentation model to learn from heterogeneous datasets simultaneously. By integrating U-Harmony, our approach sequentially normalizes and then denormalizes feature distributions to mitigate domain-specific variations while preserving original dataset-specific knowledge. More appealingly, our framework also supports universal modality adaptation, allowing the seamless learning of new imaging modalities and anatomical classes. Extensive experiments on cross-institutional brain lesion datasets demonstrate the effectiveness of our approach, establishing a new benchmark for robust and adaptable 3D medical image segmentation models in real-world clinical settings.

cs.CV↗