arXiv subjects

Yichi Zhang

Publications and source records attributed to Yichi Zhang.

At least 109 records · Page 6Linked to original sources

arXiv2026

Robust Training of Neural Networks at Arbitrary Precision and Sparsity

The discontinuous operations inherent in quantization and sparsification introduce a long-standing obstacle to backpropagation, particularly in ultra-low precision and sparse regimes. While the community has long viewed quantization as unfriendly to gradient descent due to its lack of smoothness, we pinpoint-for the first time-that the key issue is the absence of a proper gradient path that allows training to learn robustness to quantization noise. The standard Straight-Through Estimator (STE) exacerbates this with its well-understood mismatch: a quantization-aware forward pass but oblivious backward pass, leading to unmanaged error and instability. We solve this by explicitly modeling quantization as additive noise, making the full forward-backward path well-defined without heuristic gradient estimation. As one natural solution, we introduce a denoising dequantization transform derived from a principled ridge regression objective, creating an explicit, corrective gradient path that makes learning robust to the noise STE bypasses. We extend this to sparsification by treating it as a special form of quantization that zeros out small values. Our unified framework trains models at arbitrary precisions and sparsity levels with off-the-shelf recipes, enabling stable A1W1 and sub-1-bit networks where others falter. It yields state-of-the-art results, mapping efficiency frontiers for modern LLMs and providing a theoretically grounded path to hyper-efficient neural networks.

cs.LG↗

arXiv2026

UniHR: Hierarchical Representation Learning for Unified Knowledge Graph Link Prediction

Real-world knowledge graphs (KGs) contain not only standard triple-based facts, but also more complex, heterogeneous types of facts, such as hyper-relational facts with auxiliary key-value pairs, temporal facts with additional timestamps, and nested facts that imply relationships between facts. These richer forms of representation have attracted significant attention due to their enhanced expressiveness and capacity to model complex semantics in real-world scenarios. However, most existing studies suffer from two main limitations: (1) they typically focus on modeling only specific types of facts, thus making it difficult to generalize to real-world scenarios with multiple fact types; and (2) they struggle to achieve generalizable hierarchical (inter-fact and intra-fact) modeling due to the complexity of these representations. To overcome these limitations, we propose UniHR, a Unified Hierarchical Representation learning framework, which consists of a learning-optimized Hierarchical Data Representation (HiDR) module and a unified Hierarchical Structure Learning (HiSL) module. The HiDR module unifies hyper-relational KGs, temporal KGs, and nested factual KGs into triple-based representations. Then HiSL incorporates intra-fact and inter-fact message passing, focusing on enhancing both semantic information within individual facts and enriching the structural information between facts. To go beyond the unified method itself, we further explore the potential of unified representation in complex real-world scenarios. Extensive experiments on 9 datasets across 5 types of KGs demonstrate the effectiveness of UniHR and highlight the strong potential of unified representations. Code and data are available at https://github.com/zjukg/UniHR.

cs.CL↗

arXiv2026

Uni-LVC: A Unified Method for Intra- and Inter-Mode Learned Video Compression

Recent advances in learned video compression (LVC) have led to significant performance gains, with codecs such as DCVC-RT surpassing the H.266/VVC low-delay mode in compression efficiency. However, existing LVCs still exhibit key limitations: they often require separate models for intra and inter coding modes, and their performance degrades when temporal references are unreliable. To address this, we introduce Uni-LVC, a unified LVC method that supports both intra and inter coding with low-delay and random-access in a single model. Building on a strong intra-codec, Uni-LVC formulates inter-coding as intra-coding conditioned on temporal information extracted from reference frames. We design an efficient cross-attention adaptation module that integrates temporal cues, enabling seamless support for both unidirectional (low-delay) and bidirectional (random-access) prediction modes. A reliability-aware classifier is proposed to selectively scale the temporal cues, making Uni-LVC behave closer to intra coding when references are unreliable. We further propose a multistage training strategy to facilitate adaptive learning across various coding modes. Extensive experiments demonstrate that Uni-LVC achieves superior rate-distortion performance in intra and inter configurations while maintaining comparable computational efficiency.

eess.IV↗

arXiv2026

Guideline-Grounded Evidence Accumulation for High-Stakes Agent Verification

As LLM-powered agents have been used for high-stakes decision-making, such as clinical diagnosis, it becomes critical to develop reliable verification of their decisions to facilitate trustworthy deployment. Yet, existing verifiers usually underperform owing to a lack of domain knowledge and limited calibration. To address this, we establish GLEAN, an agent verification framework with Guideline-grounded Evidence Accumulation that compiles expert-curated protocols into trajectory-informed, well-calibrated correctness signals. GLEAN evaluates the step-wise alignment with domain guidelines and aggregates multi-guideline ratings into surrogate features, which are accumulated along the trajectory and calibrated into correctness probabilities using Bayesian logistic regression. Moreover, the estimated uncertainty triggers active verification, which selectively collects additional evidence for uncertain cases via expanding guideline coverage and performing differential checks. We empirically validate GLEAN with agentic clinical diagnosis across three diseases from the MIMIC-IV dataset, surpassing the best baseline by 12% in AUROC and 50% in Brier score reduction, which confirms the effectiveness in both discrimination and calibration. In addition, the expert study with clinicians recognizes GLEAN's utility in practice.

cs.AI↗

arXiv2026

Towards Safe Reasoning in Large Reasoning Models via Corrective Intervention

Although Large Reasoning Models (LRMs) have progressed in solving complex problems, their chain-of-thought (CoT) reasoning often contains harmful content that can persist even when the final responses appear safe. We show that this issue still remains in existing methods which overlook the unique significance of safe reasoning, undermining their trustworthiness and posing potential risks in applications if unsafe reasoning is accessible for and exploited by malicious users. We therefore shift our focus to aligning the safety of reasoning itself in this paper and explore process supervision as the solution. However, simply rewarding safe reasoning proves inadequate due to low rollout diversity and limited training signals. To tackle this challenge, we first delve into the characteristics of safe reasoning and uncover several critical insights that 1) safe reasoning is often consolidated by a few critical steps of safety triggers; 2) compliance cues strongly correlate with unsafe continuations; and 3) corrective interventions reliably steer unsafe trajectories towards safer traces. Motivated by these, we propose Intervened Preference Optimization (IPO), an alignment method that enforces safe reasoning by substituting compliance steps with safety triggers and constructing pairs for preference learning with strong signals. Experiments on jailbreak and adversarial safety benchmarks demonstrate that IPO remarkably improves overall safety regarding both reasoning and responses, outperforming SFT-based and RL-based baselines with a relative reduction of over 30% in harmfulness, while preserving excellent performance across diverse reasoning tasks. The results highlight the importance of explicit alignment for reasoning and provide a practical path to safer LRMs.

cs.AI↗

arXiv2026

Every Little Helps: Building Knowledge Graph Foundation Model with Fine-grained Transferable Multi-modal Tokens

Multi-modal knowledge graph reasoning (MMKGR) aims to predict the missing links by exploiting both graph structure information and multi-modal entity contents. Most existing works are designed for a transductive setting, which learns dataset-specific embeddings and struggles to generalize to new KGs. Recent knowledge graph foundation models (KGFMs) improve cross-KG transfer, but they mainly exploit structural patterns and ignore rich multi-modal signals. We address these gaps by proposing a token-based foundation model (TOFU) for MMKGR, which exhibits strong generalization across different MMKGs. TOFU discretizes structural, visual, and textual information into modality-specific tokens. TOFU then employs a hierarchical fusion architecture with mixture-of-message mechanisms, aiming to process these tokens and obtain transferable features for MMKGR. Experimental results on 17 transductive, inductive, and fully-inductive MMKGs show that TOFU consistently outperforms strong KGFM and MMKGR baselines, delivering strong performance on unseen MMKGs.

cs.CL↗

arXiv2026

Learning to Evolve: Bayesian-Guided Continual Knowledge Graph Embedding

As social media and the World Wide Web become hubs for information dissemination, effectively organizing and understanding the vast amounts of dynamically evolving Web content is crucial. Knowledge graphs (KGs) provide a powerful framework for structuring this information. However, the rapid emergence of new hot topics, user relationships, and events in social media renders traditional static knowledge graph embedding (KGE) models rapidly outdated. Continual Knowledge Graph Embedding (CKGE) aims to address this issue, but existing methods commonly suffer from catastrophic forgetting, whereby older, but still valuable, information is lost when learning new knowledge (such as new memes or trending events). This means the model cannot effectively learn the evolution of the data. We propose a novel CKGE framework, BAKE. Unlike existing methods, BAKE formulates CKGE as a sequential Bayesian inference problem and utilizes the Bayesian posterior update principle as a natural continual learning strategy. This principle is insensitive to data order and provides theoretical guarantees to preserve prior knowledge as much as possible. Specifically, we treat each batch of new data as a Bayesian update to the model's prior. By maintaining the posterior distribution, the model effectively preserves earlier knowledge even as it evolves over multiple snapshots. Furthermore, to constrain the evolution of knowledge across snapshots, we introduce a continual clustering method that maintains the compact cluster structure of entity embeddings through a regularization term, ensuring semantic consistency while allowing controlled adaptation to new knowledge. We conduct extensive experiments on multiple CKGE benchmarks, which demonstrate that BAKE achieves the top performance in the vast majority of cases compared to existing approaches.

cs.CL↗

arXiv2026

Small-Margin Preferences Still Matter-If You Train Them Right

Preference optimization methods such as DPO align large language models (LLMs) using paired comparisons, but their effectiveness can be highly sensitive to the quality and difficulty of preference pairs. A common heuristic treats small-margin (ambiguous) pairs as noisy and filters them out. In this paper, we revisit this assumption and show that pair difficulty interacts strongly with the optimization objective: when trained with preference-based losses, difficult pairs can destabilize training and harm alignment, yet these same pairs still contain useful supervision signals when optimized with supervised fine-tuning (SFT). Motivated by this observation, we propose MixDPO, a simple yet effective difficulty-aware training strategy that (i) orders preference data from easy to hard (a curriculum over margin-defined difficulty), and (ii) routes difficult pairs to an SFT objective while applying a preference loss to easy pairs. This hybrid design provides a practical mechanism to leverage ambiguous pairs without incurring the optimization failures often associated with preference losses on low-margin data. Across three LLM-judge benchmarks, MixDPO consistently improves alignment over DPO and a range of widely-used variants, with particularly strong gains on AlpacaEval~2 length-controlled (LC) win rate.

cs.AI↗

arXiv2026

Reasoning as State Transition: A Representational Analysis of Reasoning Evolution in Large Language Models

Large Language Models have achieved remarkable performance on reasoning tasks, motivating research into how this ability evolves during training. Prior work has primarily analyzed this evolution via explicit generation outcomes, treating the reasoning process as a black box and obscuring internal changes. To address this opacity, we introduce a representational perspective to investigate the dynamics of the model's internal states. Through comprehensive experiments across models at various training stages, we discover that post-training yields only limited improvement in static initial representation quality. Furthermore, we reveal that, distinct from non-reasoning tasks, reasoning involves a significant continuous distributional shift in representations during generation. Comparative analysis indicates that post-training empowers models to drive this transition toward a better distribution for task solving. To clarify the relationship between internal states and external outputs, statistical analysis confirms a high correlation between generation correctness and the final representations; while counterfactual experiments identify the semantics of the generated tokens, rather than additional computation during inference or intrinsic parameter differences, as the dominant driver of the transition. Collectively, we offer a novel understanding of the reasoning process and the effect of training on reasoning enhancement, providing valuable insights for future model analysis and optimization.

cs.CL↗

arXiv2026

Leveraging Second-Order Curvature for Efficient Learned Image Compression: Theory and Empirical Evidence

Training learned image compression (LIC) models entails navigating a challenging optimization landscape defined by the fundamental trade-off between rate and distortion. Standard first-order optimizers, such as SGD and Adam, struggle with \emph{gradient conflicts} arising from competing objectives, leading to slow convergence and suboptimal rate-distortion performance. In this work, we demonstrate that a simple utilization of a second-order quasi-Newton optimizer, \textbf{SOAP}, dramatically improves both training efficiency and final performance across diverse LICs. Our theoretical and empirical analyses reveal that Newton preconditioning inherently resolves the intra-step and inter-step update conflicts intrinsic to the R-D objective, facilitating faster, more stable convergence. Beyond acceleration, we uncover a critical deployability benefit: second-order trained models exhibit significantly fewer activation and latent outliers. This substantially enhances robustness to post-training quantization. Together, these results establish second-order optimization, achievable as a seamless drop-in replacement of the imported optimizer, as a powerful, practical tool for advancing the efficiency and real-world readiness of LICs.

eess.IV↗

arXiv2026

Beyond Correctness: Evaluating Subjective Writing Preferences Across Cultures

Current preference learning methods achieve high accuracy on standard benchmarks but exhibit significant performance degradation when objective quality signals are removed. We introduce WritingPreferenceBench, a dataset of 1,800 human-annotated preference pairs (1,200 English, 600 Chinese) across 8 creative writing genres, where responses are matched for objective correctness, factual accuracy, and length. On this benchmark, sequence-based reward models--the standard architecture for RLHF--achieve only 52.7% mean accuracy, while zero-shot language model judges perform at 53.9%. In contrast, generative reward models that produce explicit reasoning chains achieve 81.8% accuracy. We observe high within-model variance across genres: individual models range from 18.2% to 81.8% accuracy across different writing categories, with standard deviations averaging 10.1%. This variance persists regardless of model scale, with 27B parameter models showing no consistent improvement over 8B variants. Our results suggest that current RLHF methods primarily learn to detect objective errors rather than capture subjective quality preferences (e.g., creativity, stylistic flair, and emotional resonance), and that successful preference modeling may require intermediate reasoning representations rather than direct classification.

cs.CL↗

arXiv2026

Tensor Network Assisted Distributed Variational Quantum Algorithm for Large Scale Combinatorial Optimization Problem

Although quantum computing holds promise for solving Combinatorial Optimization Problems (COPs), the limited qubit capacity of NISQ hardware makes large-scale instances intractable. Conventional methods attempt to bridge this gap through decomposition or compression, yet they frequently fail to capture global correlations of subsystems, leading to solutions of limited quality. We propose the Distributed Variational Quantum Algorithm (DVQA) to overcome these limitations, enabling the solution of 1,000-variable instances on constrained hardware. A key innovation of DVQA is its use of the truncated higher-order singular value decomposition to preserve inter-variable dependencies without relying on complex long-range entanglement, leading to a natural form of noise localization where errors scale with subsystem size rather than total qubit count, thus reconciling scalability with accuracy. Theoretical bounds confirm the algorithm's robustness for p-local Hamiltonians. Empirically, DVQA achieves state-of-the-art performance in simulations and has been experimentally validated on the Wu Kong quantum computer for portfolio optimization. This work provides a scalable, noise-resilient framework that advances the timeline for practical quantum optimization algorithms.

quant-ph↗

arXiv2026

6D Movable Antenna Enhanced Cell-free MIMO: Two-timescale Decentralized Beamforming and Antenna Movement Optimization

This paper investigates a six-dimensional movable antenna (6DMA)-aided cell-free multi-user multiple-input multiple-output (MIMO) communication system. In this system, each distributed access point (AP) can flexibly adjust its array orientation and antenna positions to adapt to spatial channel variations and enhance communication performance. However, frequent antenna movements and centralized beamforming based on global instantaneous channel state information (CSI) sharing among APs entail extremely high signal processing delay and system overhead, which is difficult to be practically implemented in high-mobility scenarios with short channel coherence time. To address these practical implementation challenges and improve scalability, a two-timescale decentralized optimization framework is proposed in this paper to jointly design the beamformer, antenna positions, and array orientations. In the short timescale, each AP updates its receive beamformer based on local instantaneous CSI and global statistical CSI. In the long timescale, the central processing unit optimizes the antenna positions and array orientations at all APs based on global statistical CSI to maximize the ergodic sum rate of all users. The resulting optimization problem is non-convex and involves highly coupled variables, thus posing significant challenges for obtaining efficient solutions. To address this problem, a constrained stochastic successive convex approximation algorithm is developed. Numerical results demonstrate that the proposed 6DMA-aided cell-free system with decentralized beamforming significantly outperforms other antenna movement schemes with less flexibility and even achieves a performance comparable to that of the centralized beamforming benchmark.

eess.SP↗

arXiv2025

Meta-Router: Bridging Gold-standard and Preference-based Evaluations in Large Language Model Routing

In language tasks that require extensive human--model interaction, deploying a single "best" model for every query can be expensive. To reduce inference cost while preserving the quality of the responses, a large language model (LLM) router selects the most appropriate model from a pool of candidates for each query. A central challenge to training a high-quality router is the scarcity of reliable supervision. Gold-standard data (e.g., expert-verified labels or rubric-based scores) provide accurate quality evaluations of LLM responses but are costly and difficult to scale. In contrast, preference-based data, collected via crowdsourcing or LLM-as-a-judge systems, are cheaper and more scalable, yet often biased in reflecting the true quality of responses. We cast the problem of LLM router training with combined gold-standard and preference-based data into a causal inference framework by viewing the response evaluation mechanism as the treatment assignment. This perspective further reveals that the bias in preference-based data corresponds to the well-known causal estimand: the conditional average treatment effect. Based on this new perspective, we develop an integrative causal router training framework that corrects preference-data bias, address imbalances between two data sources, and improve routing robustness and efficiency. Numerical experiments demonstrate that our approach delivers more accurate routing and improves the trade-off between cost and quality.

cs.LG↗

arXiv2025

Standardized Evaluation of Automatic Methods for Perivascular Spaces Segmentation in MRI -- MICCAI 2024 Challenge Results

Perivascular spaces (PVS), when abnormally enlarged and visible in magnetic resonance imaging (MRI) structural sequences, are important imaging markers of cerebral small vessel disease and potential indicators of neurodegenerative conditions. Despite their clinical significance, automatic enlarged PVS (EPVS) segmentation remains challenging due to their small size, variable morphology, similarity with other pathological features, and limited annotated datasets. This paper presents the EPVS Challenge organized at MICCAI 2024, which aims to advance the development of automated algorithms for EPVS segmentation across multi-site data. We provided a diverse dataset comprising 100 training, 50 validation, and 50 testing scans collected from multiple international sites (UK, Singapore, and China) with varying MRI protocols and demographics. All annotations followed the STRIVE protocol to ensure standardized ground truth and covered the full brain parenchyma. Seven teams completed the full challenge, implementing various deep learning approaches primarily based on U-Net architectures with innovations in multi-modal processing, ensemble strategies, and transformer-based components. Performance was evaluated using dice similarity coefficient, absolute volume difference, recall, and precision metrics. The winning method employed MedNeXt architecture with a dual 2D/3D strategy for handling varying slice thicknesses. The top solutions showed relatively good performance on test data from seen datasets, but significant degradation of performance was observed on the previously unseen Shanghai cohort, highlighting cross-site generalization challenges due to domain shift. This challenge establishes an important benchmark for EPVS segmentation methods and underscores the need for the continued development of robust algorithms that can generalize in diverse clinical settings.

q-bio.QM↗

arXiv2025

Scaling Laws for Black box Adversarial Attacks

Adversarial examples exhibit cross-model transferability, enabling threatening black-box attacks on commercial models. Model ensembling, which attacks multiple surrogate models, is a known strategy to improve this transferability. However, prior studies typically use small, fixed ensembles, which leaves open an intriguing question of whether scaling the number of surrogate models can further improve black-box attacks. In this work, we conduct the first large-scale empirical study of this question. We show that by resolving gradient conflict with advanced optimizers, we discover a robust and universal log-linear scaling law through both theoretical analysis and empirical evaluations: the Attack Success Rate (ASR) scales linearly with the logarithm of the ensemble size $T$. We rigorously verify this law across standard classifiers, SOTA defenses, and MLLMs, and find that scaling distills robust, semantic features of the target class. Consequently, we apply this fundamental insight to benchmark SOTA MLLMs. This reveals both the attack's devastating power and a clear robustness hierarchy: we achieve 80\%+ transfer attack success rate on proprietary models like GPT-4o, while also highlighting the exceptional resilience of Claude-3.5-Sonnet. Our findings urge a shift in focus for robustness evaluation: from designing intricate algorithms on small ensembles to understanding the principled and powerful threat of scaling.

cs.LG↗

arXiv2025

Operation Veja: Fixing Fundamental Concepts Missing from Modern Roleplaying Training Paradigms

Modern roleplaying models are increasingly sophisticated, yet they consistently struggle to capture the essence of believable, engaging characters. We argue this failure stems from training paradigms that overlook the dynamic interplay of a character's internal world. Current approaches, including Retrieval-Augmented Generation (RAG), fact-based priming, literature-based learning, and synthetic data generation, exhibit recurring limitations in modeling the deliberative, value-conflicted reasoning that defines human interaction. In this paper, we identify four core concepts essential for character authenticity: Values, Experiences, Judgments, and Abilities (VEJA). We propose the VEJA framework as a new paradigm for data curation that addresses these systemic limitations. To illustrate the qualitative ceiling enabled by our framework, we present a pilot study comparing a manually curated, VEJA-grounded dataset against a state-of-the-art synthetic baseline. Using an LLM-as-judge evaluation, our findings demonstrate a significant quality gap, suggesting that a shift toward conceptually grounded data curation, as embodied by VEJA, is necessary for creating roleplaying agents with genuine depth and narrative continuity. The full dataset is available at https://github.com/HyouinKyoumaIRL/Operation-Veja

cs.CL↗

arXiv2025

Evaluating Long-Context Reasoning in LLM-Based WebAgents

As large language model (LLM)-based agents become increasingly integrated into daily digital interactions, their ability to reason across long interaction histories becomes crucial for providing personalized and contextually aware assistance. However, the performance of these agents in long context scenarios, particularly for action-taking WebAgents operating in realistic web environments, remains largely unexplored. This paper introduces a benchmark for evaluating long context reasoning capabilities of WebAgents through sequentially dependent subtasks that require retrieval and application of information from extended interaction histories. We develop a novel evaluation framework that simulates multi-session user interactions by injecting irrelevant task trajectories between dependent subtasks, creating contexts ranging from 25,000 to 150,000 tokens. Through extensive evaluation of four popular models, Claude-3.7, GPT-4.1, Llama 4, and o4-mini, we observe a dramatic performance degradation as context length increases, with success rates dropping from 40-50\% in baseline conditions to less than 10\% in long context scenarios. Our detailed error analysis reveals that agents primarily fail due to getting stuck in loops and losing track of original task objectives. We further propose an implicit RAG approach that provides modest improvements by generating task-relevant summaries, though fundamental limitations in long context reasoning persist. These findings highlight critical challenges for deploying WebAgents in realistic, long-term user interaction scenarios and provide insights for developing more robust agent architectures capable of maintaining coherent task execution across extended contexts.

cs.LG↗