arXiv subjects

Yibo Yang

Publications and source records attributed to Yibo Yang.

At least 19 recordsLinked to original sources

arXiv2026

HyperRepair: Recovering Forgotten General Capabilities of Fine-Tuned Language Models

Fine-tuning improves a language model's performance in a target domain but can weaken its general capabilities. Existing methods often face a trade-off between general capabilities and domain specialization, making it difficult to preserve both. We introduce HyperRepair, a framework for recovering general capabilities while retaining domain specialization at only inference cost. To make recovery generalizable across models sharing a pretrained backbone, HyperRepair learns a hypernetwork from repairs trained on source models. The hypernetwork uses activation fingerprints that summarize changes from the pretrained backbone to generate a tailored Repairing Side Net (RSN), whose prompt-conditioned gate applies corrections selectively. This enables recovery on models from unseen adaptation domains without target-domain examples or additional optimization. We evaluate HyperRpair across two model families, 12 LoRA adaptation datasets, and 15 general benchmarks. On the default split, HyperRepair achieves 93.4% of the base models' average general score while retaining 99.5% of the fine-tuned models' average domain score.

cs.CL↗

arXiv2026

MedConstBench: Benchmarking How Constitutions Steer LLMs in Medical Ethics

As large language models advance rapidly, they are increasingly used to answer medical questions, and many such models are now deployed locally. Once a model is deployed, its behavior is governed mainly by the ethical policy written into its constitution. However, most existing work focuses on factual correctness rather than the trade-off between competing ethical values. Even benchmarks that do evaluate ethics treat it as a fixed property of the model rather than as a variable, so how a constitution reshapes medical-ethics behavior remains largely unexplored. We introduce MedConstBench, a medical-ethics benchmark that treats the deployment constitution as a controlled variable, is organized around Beauchamp's four principles, and is scored by human-calibrated multi-model peer review. It includes MedDilemma, our own dataset of clinical dilemmas, and supports a systematic analysis of how constitutions steer LLMs. Evaluating ten open-source LLMs under four constitutions, we find that the effect of a constitution is dimension-selective and model-dependent, and that no constitution is universally optimal. Motivated by these findings, we propose a do-no-harm adaptive selector that chooses a constitution per model, matching the per-model oracle while guaranteeing a worst-case floor and generalizing to unseen models at low cost.

cs.AI↗

arXiv2026

Mitigating Model Merging Conflicts via Knowledge-Aware Task Vector Purification

Model merging aims to integrate task-specific abilities from individually fine-tuned models into a single model without extra training. In recent model merging methods, task vector has become a fundamental building block, as it can encapsulate the residual information from finetuning. However, the merged model often suffers from notable performance degradation due to the conflicts caused by task-irrelevant redundancy in task vectors. Existing efforts in overcoming redundancy by randomly dropping elements in the parameter space involves randomness and lacks knowledge awareness. To address these challenges, in this study, we propose Purifying TAsk Vectors (PAVE) in knowledge-aware subspace. Concretely, we sample some training examples from each task, and feed them into their corresponding fine-tuned models to acquire the covariance matrices before linear layers. We then perform a context-oriented singular value decomposition, which accentuates the weight components most relevant to the target knowledge. As a result, we can split fine-tuned model weights into task-relevant and redundant components in the knowledge-aware subspace, and purify the task vector by pruning the redundant components. To induce fair pruning efforts across models, we further introduce a spectral rank allocation strategy by optimizing a normalized activated pruning error. The task vector purification by our method as a plug-and-play scheme is applicable across various task vector-based merging methods to improve their performance. In experiments, we demonstrate the effectiveness of PAVE across a diverse set of merging methods, tasks, and model architectures.

cs.AI↗

arXiv2026

DeliMF: Gated Cross-Attention Fusion of Structured Time Series and Clinical Text for Early Intensive Care Delirium Prediction

Delirium affects a large fraction of intensive care unit (ICU) patients and foreshadows longer stays, lasting cognitive decline, and higher mortality. Flagging the patients who will become delirious within the first day of admission would give clinicians a window to intervene. Yet state-of-the-art predictors rely only on the structured electronic health record (EHR) of vital signs, laboratory values, and medications, while overlooking the free-text clinical notes written in the same window, which carry information the numbers cannot encode. To this end, we present DeliMF, a multimodal framework that fuses the structured time series with clinical text. It encodes the structured signals with a temporal Transformer and the time-stamped radiology reports with a clinical text encoder, and fuses the two through a gated cross-attention module. The gate learns a per-patient weight on the text; when a patient has no note the gate closes and DeliMF reduces to its structured-only form, giving robustness to missing text by construction. On an incident-only MIMIC-IV cohort, DeliMF improves significantly over the baselines, particularly on patients who carry clinical text, while losing nothing on those without. Ablation studies confirm the contribution of multimodal fusion and gated cross-attention, and a case study further shows that fusing clinical text with the structured stream recovers delirium cases that the structured data alone misses. We hope DeliMF encourages early ICU prediction to make fuller use of the overlooked clinical text and contributes to the early identification and timely intervention of delirium.

cs.LG↗

arXiv2026

Self-Distillation for Supervised Fine-Tuning

Supervised fine-tuning treats the expert demonstration or gold response as a one-hot target, suppressing valid alternative continuations that the model would otherwise assign probability to. We propose Gold-Conditioned Distillation (GCD), a method that turns expert demonstrations into model-native supervision. GCD uses one model as both teacher and student: the teacher conditions on a prompt containing the gold response in context, the student conditions on the original prompt as at inference, and the student is trained to minimize the forward KL divergence to the teacher's next-token distribution at each gold response token. This lets the model use its prior knowledge to teach itself the new task: correctness is anchored by the gold response, while the soft target preserves plausible alternatives. Across various training tasks on Qwen and Llama models, GCD improves new-task accuracy over SFT and generally reduces degradation on general benchmarks.

cs.CL↗

arXiv2026

Optimizing Singular Spectrum for Large Language Model Compression

Large language models (LLMs) have demonstrated remarkable capabilities, yet prohibitive parameter complexity often hinders their deployment. Existing singular value decomposition (SVD) based compression methods simply deem singular values as importance scores of decomposed components. However, this importance ordered by singular values does not necessarily correlate with the performance of a downstream task. In this work, we introduce SoCo (Singular spectrum optimization for large language model Compression), a novel compression framework that learns to rescale the decomposed components of SVD in a data-driven manner. Concretely, we employ a learnable diagonal matrix to assign importance scores for singular spectrum and develop a three-stage training process that progressively refines these scores from initial coarse compression to fine-grained sparsification-thereby striking an effective balance between aggressive model compression and performance preservation. Thanks to the learnable singular spectrum, SoCo adaptively prunes components according to the sparsified importance scores, rather than relying on the fixed order of singular values. More importantly, the remaining components with amplified importance scores can compensate for the loss of the pruned ones. Experimental evaluations across multiple LLMs and benchmarks demonstrate that SoCo surpasses the state-of-the-art methods in model compression.

cs.CL↗

arXiv2026

COMPASS: Contextual Moment Prior Alignment for Generalizable AIGI Detection

Generalizable AI-generated image (AIGI) and deepfake detection remain challenging due to the asymmetry phenomenon, where adaptation to limited training distributions can degrade the transferable representations of Vision Foundation Models (VFMs). Recent subspace adaptation methods mitigate this issue by freezing high-magnitude principal components while adapting the residual subspace. However, such magnitude-based decomposition is data-agnostic and does not explicitly account for target-relevant semantics, potentially leading to semantic misalignment. We propose COMPASS (COntextual Moment Prior Alignment for Subspace Separation), a data-aware subspace adaptation framework that uses a small unlabeled Semantic Proxy Set to estimate target activation statistics. COMPASS incorporates activation covariance into the decomposition of pre-trained weights, preserving target-relevant representations as a stable semantic reference while the residual adapter learns forgery-related variations. The proxy calibration requires no gradient-based optimization. Extensive experiments demonstrate strong generalization across cross-dataset, cross-method, joint cross-dataset-method, and open-world cross-generator evaluations, achieving state-of-the-art performance on both Deepfake and AIGI detection benchmarks.

cs.CV↗

arXiv2026

Beyond Outcome Feedback: Trajectory-Quality-Guided Experience-based Learning

Experience-based learning enables large language model (LLM) agents to improve over time by reusing knowledge from past executions. However, existing methods typically rely on final task outcomes as feedback for experience learning, overlooking the execution process and limiting their ability to evaluate and optimize experience based on trajectory quality. To address this limitation, we propose TQG-Exp, a trajectory-quality-guided experience-based learning framework that introduces process feedback. Specifically, we introduce Trajectory Quality Loss (TQL), which jointly considers task completion and process quality to provide a learning signal beyond final task outcomes. Based on this signal, we employ paired rollouts with and without retrieved experience to guide experience optimization. Representative trajectories are selected based on their outcomes and trajectory quality to extract task-level and task-type-level experiences. Meanwhile, differences in trajectory quality between the paired groups are used to estimate the contribution of retrieved experience and update its utility. The learned utility is then combined with semantic relevance to guide subsequent experience retrieval. Experiments on ALFWorld, WebShop, and ScienceWorld with two backbone LLMs show that TQG-Exp outperforms existing experience-based methods, improving Avg.@3 over the strongest baseline by 4.50 and 8.55 percentage points on DeepSeek-V4-Flash and Qwen3.5-Flash, respectively.

cs.AI↗

arXiv2026

The Thinking Leakage in Reinforcement Learning for LLMs

Hybrid reasoning models such as the Qwen3 series switch between a THINK mode and a NO-THINK mode through a few chat-template tokens. A popular recipe trains such models with reinforcement learning in NO-THINK mode and reports improved reasoning performance. We provide evidence that much of this gain can be explained by thinking leakage: the gain comes largely from partially reactivating the reasoning abilities of the original model's THINK mode instead of learning enhanced reasoning behavior. We train Qwen3-1.7B, 4B, and 8B with GRPO in NO-THINK mode. As accuracy improves, responses become longer and more reflective, are increasingly judged by the original model to resemble its own thinking, and the model's internal activations shift toward its THINK mode. We then investigate at the level of weights, activations, and attention heads: scaling the RL update further elicits THINK-mode behavior and approaches THINK-mode accuracy but does not exceed it; in Qwen3-8B, replacing a low-dimensional THINK-related activation subspace of the RL-trained model with its original-model value removes most of the RL gain, and the reverse exchange transfers a considerable fraction of the gain to the original model; and the RL-induced shift passes through the attention heads that read the empty think block. We further identify a possible mechanism in the GRPO training signal: responses favored by the advantage-weighted update are more THINK-like, and a gradient step computed from this signal moves the model toward its THINK state. Thus, gains in NO-THINK mode can largely reflect reactivation of reasoning abilities already present in the original model.

cs.CL↗

arXiv2026

Mental-R1: Aligning LLM Reasoning for Mental Health Assessment

Mental health problems such as anxiety, depression, and suicide remain urgent global challenges, where timely and accurate assessment is critical for effective intervention. Recently, large language models have been explored for mental health assessment. However, existing general-purpose post-training methods do not align with the cognitive processes of human assessment, which may lead to unreliable reasoning outcomes. To bridge this gap, we propose Cognitive Relative Policy Optimization (CRPO), a reinforcement learning framework tailored for the mental health domain. CRPO extends group relative policy optimization by integrating stage-dependent uncertainty modeling into the policy optimization process. Specifically, we introduce a stage-wise entropy regularization mechanism that encourages broad exploration in early reasoning phases and progressively enforces confident decision-making in later stages, mimicking the human cognitive shift from uncertainty to certainty. In addition, inspired by cognitive appraisal theory, we formalize cognitive reasoning stages, thereby guiding theory-grounded interpretable inference. Experiments on 8 mental health datasets show that CRPO achieves an average improvement of 10.4 percentage points in weighted F1-score over the best reinforcement learning baseline. Furthermore, the CRPO-trained model Mental-R1 demonstrates clear advantages compared with existing large language models on reasoning-intensive cases, suggesting that CRPO enhances reasoning capabilities for mental health assessment.

cs.AI↗

arXiv2026

LeRF: Learning Reference Coordinate Frames for Perspective Taking Reasoning

Perspective taking is a fundamental component of spatial intelligence, requiring models interpret spatial relations from a specified viewpoint, such as that of another entity or an imagined observer. Despite the increasing spatial reasoning capabilities of Vision-Language Models (VLMs), they still struggle with perspective taking, often defaulting to the camera viewpoint when a query requires reasoning from a different perspective. We introduce Learning Reference Coordinate Frames for Perspective Taking (LeRF), a framework that trains VLMs to construct and use explicit reference frames for viewpoint-dependent reasoning. Given an image and a query, LeRF decides whether a coordinate frame is necessary. If so, it grounds the reference entity and predicts the frame's origin and entity-centered reference frame. A lightweight renderer overlays the frame onto the image, enabling subsequent reasoning over these visual cues without external perception models or explicit 3D reconstruction. To learn this process, we first perform supervised fine-tuning to teach selective tool invocation and reference coordinate frame prediction, followed by reinforcement learning on spatial VQA pairs to improve frame-guided reasoning. Across diverse perspective-taking benchmarks, LeRF consistently improves over its backbone and achieves strong performance against existing open-source methods. Further evaluations also show improved reference-frame grounding and orientation estimation, supporting the effectiveness of learned reference frames for viewpoint-dependent reasoning.

cs.CV↗

arXiv2026

Advances in Diffusion-Based Generative Compression

Popularized by their strong image generation performance, diffusion and related methods for generative modeling have found widespread success in visual media applications. In particular, diffusion methods have enabled new approaches to data compression, where realistic reconstructions can be generated at extremely low bit-rates. This article provides a unifying review of recent diffusion-based methods for generative lossy compression, with a focus on image compression. These methods generally encode the source into an embedding and use a diffusion model to iteratively refine it during decoding, so that the reconstruction approximately follows the true data distribution. The embedding can take various forms and is typically transmitted via an auxiliary entropy model, and recent methods also explore the use of diffusion models themselves for information transmission via channel simulation. We review representative approaches through the lens of rate-distortion-perception theory, highlighting the role of common randomness and connections to inverse problems, and identify open challenges.

eess.IV↗

arXiv2026

SpecGradFilter: A Spectral Gradient Filtering Framework for Taming Federated Heterogeneity

Federated Learning (FL) is fundamentally challenged by statistical heterogeneity, where non-identically distributed (non-IID) data induces client drift that severely hampers global convergence. While existing approaches attempt to mitigate this drift through spatial-domain gradient correction or regularization, they overlook the intrinsic spectral structure of optimization signals. In this work, we revisit client drift from a novel frequency-domain perspective and uncover a critical Spectral Bias of Drift: inter-client gradient divergence is predominantly concentrated in low-frequency components which encode client-specific distributional shifts, while high-frequency components representing fine-grained features remain relatively consistent. Motivated by this, we propose SpecGradFilter, a unified Spectral Gradient Filtering Framework that tames heterogeneity by suppressing discordant low-frequency signals. Crucially, we demonstrate that SpecGradFilter is a generalizable principle, effective not only via precise FFT-based truncation but also through spatial approximations like Gaussian detrending. Extensive experiments on benchmarks such as CIFAR-10/100 and Tiny-ImageNet demonstrate that SpecGradFilter significantly performs better performance in highly Non-IID settings with negligible communication overhead, establishing a new paradigm for robust federated optimization.

cs.LG↗

arXiv2026

Task-State Adaptation with Prototype Memory for Multi-Task Dense Prediction

Vision foundation backbones provide strong representations for dense prediction, yet a single shared feature still needs to support tasks with different, image-dependent adaptation requirements. We propose MemMTL, a multi-task dense prediction framework that estimates a compact task state from global visual context and refines it through a learnable task-state prototype memory. The refined state is converted into task-conditioned expert logits and combined with token-level logits before sparse top-$k$ selection over a local expert bank shared by all tasks. A separate task-agnostic residual bank provides a common adaptation path, and both paths are added once to the backbone feature before task-specific prediction. We specify a matched evaluation protocol on NYUD-v2 and PASCAL-Context with SAM 3 and ViT-L backbones to measure predictive quality, computational cost, and the contributions of task-state conditioning, prototype retrieval, and sparse routing. The numerical record in the present working draft predates this canonical implementation and must be regenerated before it can support empirical claims.

cs.CV↗

arXiv2026

Dual-Grained Agent Memory and Shapley Context Attribution for Multimodal Agentic Learner

Frontier multimodal large language models (MLLMs) deliver impressive perception yet still falter on scientific and mathematical reasoning. Parameter-level adaptation is unavailable for closed-weight or on-device backbones, and stateless prompting forfeits any compounding benefit from problems already solved. We propose \textbf{DG-Mem}, a dual-grained agentic memory framework that augments a frozen MLLM with a non-parametric, externally stored memory built once from training-time rollouts and consulted read-only at test time. Motivated by the Complementary Learning Systems (CLS) account of human memory, DG-Mem factors its store into an instance-grounded exemplar memory and a category-level schema memory of IF-THEN rules, with a transient reflection store mediating their construction so that schemas are synthesized only from abstract reflections, never from exemplar text. Two design choices distinguish DG-Mem: an online concept categorizer that grows the category space incrementally during training rather than committing to a predefined taxonomy, and a Shapley context attribution procedure that decomposes correctness across the entire retrieved rule set and yields a per-rule utility that re-weights retrieval at test time. The pipeline introduces no gradient updates and is deployable on closed-weight or on-device backbones. Across MathVista, MMMU, and MMMU-Pro on four open-weight and proprietary backbones (Qwen3.5-27B, Qwen3.5-122B-A10B, GPT-5-Nano, Gemini-3-Flash), DG-Mem improves consistently over no-memory and competitive memory baselines.

cs.CV↗

arXiv2026

ShadowNet for Data-Centric Quantum System Learning

Understanding the dynamics of large quantum systems is hindered by the curse of dimensionality. Statistical learning offers new possibilities in this regime through neural network protocols and classical shadows, while both methods have limitations: the former suffers from incompatible dataset construction rules, resulting in substantial computational demands for data collection when addressing different tasks; the latter lacks the ability to distill knowledge from prior data to enhance subsequent learning endeavors. In this study, we propose a data-centric learning paradigm combining the strengths of these two approaches to advance quantum system learning (QSL). Central to our paradigm lies a unified dataset construction rule, achieved by classical shadows along with other easily obtainable information of quantum systems. To illustrate our approach, we present ShadowNet, implemented under both convolutional and attention mechanisms, to efficiently and faithfully tackle two pivotal QSL tasks: quantum state tomography (QST) and direct fidelity estimation (DFE). Numerical simulations on QST and DFE up to 60 qubits validate the efficacy of our proposal, showcasing how ShadowNet advances classical shadows with limited state copies, and highlighting how the varied neural networks impact the performance. Our work underscores the immense potential of a data-centric approach in comprehending novel and large quantum systems.

quant-ph↗

arXiv2026

Self-Guided Process Reward Optimization with Redefined Step-wise Advantage for Process Reinforcement Learning

Process Reinforcement Learning~(PRL) has demonstrated considerable potential in enhancing the reasoning capabilities of Large Language Models~(LLMs). However, introducing additional process reward models incurs substantial computational overhead, and there is no unified theoretical framework for process-level advantage estimation. To bridge this gap, we propose \textbf{S}elf-Guided \textbf{P}rocess \textbf{R}eward \textbf{O}ptimization~(\textbf{SPRO}), a novel framework that enables process-aware RL through two key innovations: (1) we show that process rewards can be derived intrinsically from the policy model itself, and (2) we redefine step-wise advantage by introducing well-defined Cumulative Process Rewards~(\textbf{CPR}) and \textbf{M}asked \textbf{S}tep \textbf{A}dvantage~(\textbf{MSA}), which facilitates rigorous step-wise action advantage estimation within shared-prompt sampling groups. Our experimental results show that SPRO outperforms vanilla GRPO with 3.4x higher training efficiency and a 12.9\% test accuracy improvement. Furthermore, SPRO maintains a stable and elevated policy entropy throughout training while achieving a considerable reduction in response length, evidencing sufficient exploration and prevention of reward hacking. Notably, SPRO incurs no additional computational overhead compared to outcome-supervised RL methods such as GRPO, which benefit industrial implementation.

cs.LG↗

arXiv2026

Self-Improvements in Modern Agentic Systems: A Survey

Self-improving autonomous agents are moving from research prototypes to deployed systems. The primary goal is controllable evolution, or adaptation, from experience with minimal or even no human input. This survey frames modern self-improving agents as adaptive systems that convert experience into accumulated capability gains. We offer a system-level framework that represents a modern agent as a configuration coupling a foundation model with an operational scaffold of prompts, memory, tools, and control logic. Within this framework, self-improvement is formalized as a self-induced update operator that obtains and commits updates to model parameters or scaffold components. We organize prior work by update target and by the signals that drive change, then review applications and discuss evaluation, before closing with open problems and future directions. For convenience, we track technical updates on https://github.com/selfimproving-agent/awesome-Self-Improving-Agents.

cs.AI↗