arXiv subjects

Yuan Wu

Publications and source records attributed to Yuan Wu.

At least 37 records · Page 2Linked to original sources

arXiv2026

MMPCBench: Benchmarking Multimodal Large Language Models on Proactive Critique of Flawed Inputs

As Multimodal Large Language Models (MLLMs) evolve into sophisticated interactive assistants, their reliability depends not only on following instructions but also on validating them. We define Proactive Critique as the model's autonomous ability to identify, analyze and fix faulty user inputs without extra prompts. However, evaluations mainly test models under ideal circumstances or simple refusal behaviors, largely ignoring active error processing. To fill this gap, we propose MMPCBench, a comprehensive framework for evaluating MLLMs' proactive critique competence. It features a fine-grained taxonomy of 4 primary error types spanning 12 subcategories, ranging from cross-modal contradictions to missing visual premises. We adopt a hierarchical evaluation protocol to measure models' error detection, diagnosis and resolution performance, and apply alignment-aware metrics to assess the coherence between internal reasoning and final responses. Tests on 14 mainstream MLLMs show obvious weaknesses in proactive critique, especially in dealing with subtle visual anomalies. Notably, we identify a pervasive "consistency gap": reasoning models can often correctly identify and analyze errors during internal reasoning yet suppress these valid insights in final outputs to prioritize response compliance. The code and data is available at https://github.com/ALIENS32/MMPCBench.

cs.AI↗

arXiv2026

CaSKG: Counterfactual-Causal Skill Graphs for Scalable Agent Skill Retrieval

Reusable skill libraries allow large language model (LLM) agents to reuse procedural knowledge across tasks, but they also turn memory access into a challenging retrieval problem. Full-library prompting preserves coverage at high context cost, vector retrieval returns compact neighborhoods but treats skills as independent text, and graph-based retrieval can recover workflow context only when the edges that carry relevance are reliable. We propose CaSKG, a counterfactual-causal skill graph framework that calibrates procedural relations before retrieval. CaSKG first builds a high-recall directed candidate graph from semantic, lexical, input/output, and structural evidence, with repair evidence and an optional LLM judge further refining candidate scores. It then applies direction-conditioned textual counterfactual probes that remove, substitute, and reorder skill pairs, aggregates the evidence with Bayesian smoothing, and publishes a state-filtered weighted graph for task-conditioned expansion. The graph is constructed offline and used without changing the downstream agent policy or task interface. Across six LLM backbones on ALFWorld ID-140 and ScienceWorld U211, CaSKG achieves the highest task score in all twelve combinations of model and benchmark. Relative to Graph-of-Skills (GoS), it improves the six-model macro-average ScienceWorld score from 72.62 to 80.50 and ALFWorld success from 80.01\% to 86.79\%, while reducing mean environment steps on both benchmarks. Qualitative and ablation analyses further show that calibrated edges help retrieval preserve prerequisites, state-changing actions, verification routines, and final completion steps. These results position edge-confidence calibration as an effective route to compact and executable skill retrieval at scale\footnote{Code is available at: https://github.com/ZhiyuanLi218/Caskg }.

cs.AI↗

arXiv2026

Length-Controlled Margin-Based Preference Optimization without Reference Model

Direct Preference Optimization (DPO) is a widely adopted offline algorithm for preference-based reinforcement learning from human feedback (RLHF), designed to improve training simplicity and stability by redefining reward functions. However, DPO is hindered by several limitations, including length bias, memory inefficiency, and probability degradation. To address these challenges, we propose Length-Controlled Margin-Based Preference Optimization (LMPO), a more efficient and robust alternative. LMPO introduces a uniform reference model as an upper bound for the DPO loss, enabling a more accurate approximation of the original optimization objective. Additionally, an average log-probability optimization strategy is employed to minimize discrepancies between training and inference phases. A key innovation of LMPO lies in its Length-Controlled Margin-Based loss function, integrated within the Bradley-Terry framework. This loss function regulates response length while simultaneously widening the margin between preferred and rejected outputs. By doing so, it mitigates probability degradation for both accepted and discarded responses, addressing a significant limitation of existing methods. We evaluate LMPO against state-of-the-art preference optimization techniques on two open-ended large language models, Mistral and LLaMA3, across six conditional benchmarks. Our experimental results demonstrate that LMPO effectively controls response length, reduces probability degradation, and outperforms existing approaches. The code is available at https://github.com/gengxuli/LMPO.

cs.CL↗

arXiv2026

VCIFBench: Evaluating Complex Instruction Following for Video Understanding

Multimodal large language models have made rapid progress in video understanding, yet existing benchmarks largely rely on simple prompts and provide limited evidence about whether models can satisfy explicit output constraints. We introduce VCIFBench, a benchmark for evaluating complex instruction following in video understanding. VCIFBench constructs constraint-rich instructions from both benchmark-adapted and directly video-grounded prompts, covering content, format, style, and structure requirements, and evaluates model outputs with a hybrid verification pipeline. The benchmark contains 306 satisfiable test instructions, 540 DPO training instances, and a 100-item diagnostic set for evaluating whether models can recognize instruction conflicts. Experiments on 10 MLLMs show that joint constraint satisfaction remains challenging. Preference optimization improves instruction following for two model families, while Conflict-100 reveals that models usually execute a satisfiable-looking subset instead of detecting global incompatibility.

cs.CL↗

arXiv2026

How Agents Represent Humans: Human-Directed Stereotypes in an Open Agent Social Network

LLM-based agents are increasingly deployed in persistent social environments, where generated claims can be posted, replied to, remembered, and reused. We study human-directed stereotypes on Moltbook, an open agent-native social platform, asking how agents construct humans as a social category. For this human-target analysis, we introduce an annotation framework with four evaluative dimensions---morality, friendliness, competence, and autonomy---and a second-stage subtype scheme for descriptive \textit{other} attributions. We find that competence dominates human-directed evaluations, while many \textit{other} attributions describe humans as epistemic, cultural, or embodied subjects. We further examine how these human representations appear in human--agent narrative contexts and platform-level circulation. As an auxiliary comparison, we analyze agent-internal community feedback through behavioral host affinity. Rather than reproducing the stable insider--outsider rejection often observed in human online communities, Moltbook feedback patterns are better explained by exposure, author visibility, and content selection. These findings suggest that bias in agent societies should be studied not only as isolated model output, but also as a discourse process.

cs.CL↗

arXiv2026

ToolRobustBench: Stage-Wise Perturbation Evaluation and Failure Diagnosis for Tool-Calling Agents

Large language models (LLMs) rely on tool calling as a fundamental agent capability, enabling them to invoke external systems and complete tasks beyond text generation. However, clean end-to-end (E2E) success cannot identify where a tool-use failure originates or how it propagates through a call. We introduce ToolRobustBench, a stage-wise diagnostic benchmark for tool-calling agents, where a tool-calling agent is an LLM system that selects a tool, supplies structured arguments, and interprets its returned feedback. ToolRobustBench aligns four perturbation families with the tool-use pipeline: tool-interface, user-intent, tool-output/observation, and runtime-environment perturbations. It attributes failures to tool selection, schema grounding, argument binding, tool-output/runtime-feedback handling, and E2E task success. Experiments on 15,456 single-family instances across 7 models, 16 sampled local tools, 4 perturbation families, and 14 subtypes show high but non-uniform clean performance and substantial robustness degradation, with tool-output/observation perturbation the dominant bottleneck. Mixed-family experiments reveal non-additive failure patterns that are not explained by isolated single-family results. Thus, ToolRobustBench provides a deterministic and cascade-aware benchmark for diagnosing robustness beyond clean tool-calling accuracy;

cs.SE↗

arXiv2026

CD-LoRA: Consistency-Driven Low-Rank Adaptation for Multi-Task Fine-Tuning

While Multi-Task Learning (MTL) is essential for adapting Large Language Models (LLMs) to diverse domains, prevailing LoRA-based methods rely on complex routing mechanisms that partition task-specific knowledge. In this work, we reveal that such routing-based designs are prone to a training-inference discrepancy, where stochastic routing decisions under distribution shifts compromise inference stability. Driven by a second-order Taylor analysis that exposes the instability induced by routing variance, we challenge the training-inference discrepancy and propose Consistency-Driven Low-Rank Adaptation (CD-LoRA). By eliminating routers entirely, CD-LoRA employs a consistency-driven alignment mechanism to enforce representation congruence across tasks in a shared low-rank space. This paradigm fosters robust, task-agnostic features without explicit partitioning overhead. Extensive experiments show that CD-LoRA consistently outperforms state-of-the-art multi-adapter baselines, offering a simpler, router-free, and more stable solution for multi-task PEFT. The code is available at the anonymous link https://github.com/zhaqian21/CD-LoRA.

cs.LG↗

arXiv2026

The Asymmetric Harms of LLM Compression

Large language models (LLMs) compression reduces deployment costs, but standard aggregate metrics like perplexity and accuracy often mask underlying behavioral shifts. In this work, we systematically evaluate 3 LLMs across 11 compression methods to investigate the effects of compression on knowledge retention, model confidence, and social bias. We find that compression disproportionately reduces the relative retention of head knowledge compared to tail knowledge. Furthermore, compressed models often remain substantially confident in their incorrect answers on newly lost knowledge. Finally, we demonstrate that stable aggregate bias scores can conceal substantial, opposing shifts in stereotypical preferences across demographic subgroups. Together, these findings reveal asymmetric behavioral changes that aggregate performance measures fail to capture, highlighting the need for granular evaluation of compressed models before deployment.

cs.CL↗

arXiv2026

Cooperative Safety Intelligence over V2X Networks: A Survey

Vehicle-to-Everything (V2X) cooperation is reshaping traffic safety from an ego-centric sensing problem into a networked intelligence problem involving distributed sensing, cooperative perception, and coordinated decision-making. This survey reviews recent progress in V2X-enabled cooperative safety intelligence through a unified Sensor-Perception-Decision (SPD) framework, which characterizes how distributed observations are exchanged, fused, calibrated, and transformed into safety-ready evidence for risk-aware intervention. Within this framework, V2X networking conditions, multi-modal sensing, cooperative perception, and decision-making are analyzed as coupled components of a safety-intelligence pipeline. Cross-layer communication and networking constraints, including latency, synchronization, bandwidth, packet or feature loss, reliability, trust, and PQoS, are explicitly considered to assess whether shared evidence remains timely, robust, and actionable. Compared with prior V2X safety surveys, this work organizes the literature around a formal SPD safety loop and synthesizes representative methods, datasets, benchmarks, and platforms under communication-aware constraints from 2017 to 2026. It further examines semantic and task-oriented evidence exchange, bandwidth-efficient cooperative perception, ISAC-enabled V2X systems, safety-ready outputs, and evaluation practices under practical conditions such as delayed or lossy messages, pose misalignment, NLOS coverage, and heterogeneous deployment. The survey concludes with a roadmap toward scalable data infrastructure, PQoS-aware evaluation, embodied predictive intelligence, and trustworthy human-in-the-loop cooperation for next-generation V2X safety systems.

eess.SY↗

arXiv2026

Thinking vs. NoThinking: Towards Interpreting Reasoning Mechanisms of Large Language Models via Sparse Autoencoders

While Large Language Models (LLMs) employing Chain-of-Thought (CoT) exhibit superior reasoning capabilities, the neural mechanisms distinguishing this explicit Thinking mode from direct answer generation (NoThinking mode) remain poorly understood. To deconstruct this cognitive process, we apply Top-K Sparse Autoencoders (SAEs) to the intermediate representations of DeepSeek-R1-Distill-Qwen-7B and examine the model's divergent behaviors across math-solving tasks of three distinct difficulty levels. Observationally, we identify a clear distinction in how the model functions under two reasoning modes: Thinking mode relies on sparse and high-intensity feature activations driving verbal deduction independent of problem complexity, whereas NoThinking mode exhibits an adaptive and diffuse pattern prioritizing symbolic manipulation. Causally, suppressing the three most active sparse features by Total Activation Volume reveals three principles: (i) reasoning and syntactic structure are tightly coupled, as interventions consistently degrade \LaTeX{} and boxed-solution formatting; (ii) Thinking responds to disruption with compensatory over-generation marked by increased metacognitive cues and repetitive, low-information continuations; and (iii) coherent CoT behavior depends on a fragile coordination among specialized features, yielding distinct failure modes under perturbation but a consistently impaired output structure.

cs.CL↗

arXiv2026

Structured Spectral Compression based Low-Bitrate Secure Speech Communications for Internet of Things assisted Non-Terrestrial Networks

This paper focuses on the Low-Bitrate Secure Speech Communications based on the Structured Spectral Compression (LB-S2C2). Specifically, the Mel spectral matrix of the speech signal is first encoded at the transmitter side through compressive sensing based on waveform segmentation and data quantization. Then, the Automatic Repeat Request (ARQ) is combined with forward error correction to achieve reliable transmission of speech signals over wireless channels. Thirdly, the received signals are recovered as the speech at the receiver side. Finally, we conduct a series of simulation experiments for the performance evaluation of LB-S2C2. Our simulations reveal that the dictionary matrix used for the speech reconstruction is different from the one used for the high-order matrix sparsification by even only approximately 0.1%, and then the accurate speech recovery fails. It implies that the speech data can be securely transmitted when the dictionary matrix is preserved. More importantly, the LB-S2C2 exhibits a very high privacy protection capability with the average voiceprint similarity to be only 0.3, which is much lower than the 0.8 of the semantic speech communication scheme DeepSC-S, and even lower than the 0.33 of the latest speech communication scheme OFI-OFCNB. In addition, our simulations reveal that the proposed structured speech coding boasts a time complexity of merely O(n), and the proposed speech recovery scheme requires the 12-bit memory storage only, which outperforms the traditional encryption algorithms proposed for speech communications. In comparison with the conventional compression techniques, our spectral compression method renders the coding rate of only 3.9kbps, which is lower than the current lowest speech coding rate of 6.3kbps achieved by G.723.

cs.NI↗

arXiv2026

Energy Minimization Oriented Resource Allocation for Integrated Sensing and Communication in Marine IoT Networks

Integrated sensing and communication (ISAC) has become a promising technical framework for Marine Internet of Things (MIoT) systems. Nevertheless, all devices rely on battery power, so energy efficiency becomes a core bottleneck limiting practical deployment. This paper investigates the energy consumption minimization problem of MIoT-oriented ISAC systems. In this system, an uncrewed aerial vehicle (UAV) uses non-orthogonal multiple access (NOMA) to simultaneously perform target sensing and collect data from uncrewed surface vehicles (USVs), then forwards processed sensing information and USV data to a shore-based base station (SBS). Subject to latency limits and sensing performance requirements, total system energy consumption can be minimized via joint optimization of multiple variables, UAV transmit beamforming, dedicated sensing signal, USV transmit power, UAV computation power, and time resource allocation for sensing and communication phases. To tackle this non-convex optimization problem, we build a layered solution architecture that divides the original problem into independent subproblems and optimizes each alternately according to its mathematical features. Specifically, we first derive closed-form USV transmit power solutions and conduct variable substitution. The successive convex approximation (SCA) method is adopted to convert remaining non-convex subproblems into convex forms, on which we design efficient iterative algorithms. Simulation results verify the validity and accuracy of our algorithm in reducing system energy consumption. Compared with orthogonal frequency division multiple access (OFDMA) and genetic algorithm benchmarks, our scheme lowers system energy consumption by 19.71% and 8%, respectively. In addition, our optimized energy value only has an 8.72% gap from the optimum solved by the LINGO solver.

cs.NI↗

arXiv2026

Measurement of solar $pp$ neutrino flux with the new PandaX-4T data

We report a new measurement of the solar proton--proton ($pp$) neutrino flux via neutrino--electron elastic scattering using the PandaX-4T Run 2 data set collected between 2024 and 2026, corresponding to an exposure of 1.9 tonne$\cdot$yr. Before Run 2 data taking, the detector underwent a series of upgrades to improve its response and background conditions. Time variations of radioactive noble-gas impurities are constrained using the physics data themselves, complemented by measurements from the gas-assay system. The analysis introduced improvements in the data processing chain, detector response characterization, and background models. A blind spectral analysis was then performed on the electronic-recoil data across a wide energy range from 20 to 1000 keV. In combination with the Run 0 data published earlier, the fitted $pp$ flux is $(8.5 \pm 3.5)\times 10^{10}$ $\mathrm{cm^{-2}s^{-1}}$, consistent with the prediction of the Standard Solar Model. With a statistical significance of $2.2σ$ above background, this marks the first positive indication of solar $pp$ neutrino--electron scattering below an electronic-recoil energy of 165 keV.

hep-ex↗

arXiv2026

Overcoming the Speed-Fidelity Trade-off in Fast CZ Gates via Cyclic Control

High-fidelity quantum gates are essential for scalable quantum computation. However, at short durations, short-timescale waveform distortions break the time-reflection symmetry of control pulses, preventing the precise closure of cyclic evolution. This mechanism renders conventional symmetric protocols intrinsically over-constrained. Conventional strategies typically rely on smoothing the pulse envelopes or embedding the interaction pulse within a longer qubit pulse to bypass short-timescale distortions, which inevitably leads to a persistent speed-fidelity trade-off. To overcome this limitation, we introduce a cyclic control strategy based on parameter-space expansion, which restores controllability by incorporating an additional degree of freedom. We experimentally demonstrate this approach in a superconducting controlled-Z gate, achieving robust suppression of coherent errors without increasing gate duration, reducing the average coherent error from 0.27% to 0.12% across multiple two-qubit gates, as validated by cross-entropy benchmarking. Our results establish a general route to fast, high-fidelity cyclic quantum gates beyond the conventional speed-fidelity trade-off.

quant-ph↗

arXiv2026

LASER: Load-Aware Serving with Early-Exit for Reasoning LLMs at the Edge

Large reasoning models (LRMs) such as DeepSeek-R1 have achieved strong performance through extended chain-of-thought (CoT) generation. However, deploying them on edge devices raises a conflict between long CoT sequences and constrained resources. Recent confidence-based early exit methods reduce CoT length for individual requests, yet they apply fixed thresholds from a single-request perspective, ignoring multi-request concurrency and load fluctuation in edge serving. To bridge this gap, we propose \underline{L}oad-\underline{A}ware \underline{S}erving with \underline{E}arly-exit for \underline{R}easoning (LASER). LASER couples two complementary designs: (1) a load-aware adaptive exit threshold that adjusts the confidence bar based on real-time system load within an empirically validated robust range, and (2) a difficulty- and load-aware reasoning budget pre-allocation that assigns compute resources by request difficulty and system capacity. We formulate the problem as a joint optimization of reasoning quality and service latency. Experiments on two reasoning models, four benchmarks, and diverse load conditions show that LASER reduces average latency by 17--38\% and improves service-level objective (SLO) satisfaction by 3--6\% over fixed-threshold baselines, at an average accuracy cost of only 1\%.

cs.DC↗

arXiv2026

BA-LoRA: Bias-Alleviating Low-Rank Adaptation to Mitigate Catastrophic Inheritance in Large Language Models

Parameter-efficient fine-tuning (PEFT) has become a de facto standard for adapting large language models (LLMs). However, we identify a critical vulnerability within popular low-rank adaptation methods such as LoRA: they can exacerbate "Catastrophic Inheritance" - the unchecked propagation of biases, noise, and data imbalances from pre-training. This phenomenon can degrade model robustness and fairness, undermining the benefits of efficient adaptation. To address this, we introduce Bias-Alleviating Low-Rank Adaptation (BA-LoRA). Our approach is founded on a principled decomposition of Catastrophic Inheritance into three core challenges: Knowledge Drift, Representation Collapse, and Overfitting to Noise. BA-LoRA systematically mitigates these issues by incorporating a trio of targeted regularizers: consistency, diversity, and an SVD-based term, designed to preserve core knowledge, promote representational richness, and encourage robust, low-rank output representations, respectively. We conduct comprehensive evaluations on a suite of Natural Language Generation (NLG) and Natural Language Understanding (NLU) tasks using diverse, prominent open-source language models (e.g., LLaMA-2-7B and DeBERTa-v3-base). Our results show that BA-LoRA not only outperforms state-of-the-art LoRA variants in terms of performance and stability, but also demonstrates superior robustness and bias mitigation on targeted evaluations. These results provide evidence that BA-LoRA can counteract the adverse effects of Catastrophic Inheritance.

cs.CL↗

arXiv2026

BV-Blend: Uncertainty-Weighted Historical Baselines for Stable Critic-Free RL with Verifiable Rewards

Critic-free reinforcement learning with verifiable rewards (RLVR), exemplified by Group Relative Policy Optimization (GRPO), avoids training a value function (critic) and reduces memory and compute overhead relative to critic-based PPO pipelines for aligning large language models. However, GRPO-style advantage estimation depends on prompt-local (within-prompt-group) reward statistics and can be unstable. In particular, when all rollouts in a prompt group receive identical rewards, the within-group reward variance becomes zero, and group normalization yields zero advantages for that group, impeding learning in cold-start regimes with binary verifiers. We introduce BV-Blend, a critic-free framework that stabilizes advantage estimation by combining prompt-local on-policy statistics with semantic-cluster-conditioned historical moments. BV-Blend maintains EMA-tracked reward moments for each cluster, derives a confidence weight from a standard error of the mean (SEM) proxy, and uses this weight to blend historical and prompt-local baseline and variance statistics into a standardized advantage for PPO-style clipped updates. Experiments on verifiable reasoning benchmarks show that BV-Blend improves training stability and performance, and remains robust in regimes where group-normalized methods may stall.

cs.AI↗

arXiv2026

SOS-LoRA: Static Orthogonal-Subspace Low-Rank Adaptation with Fixed Multi-Scale Scaling

Low-Rank Adaptation (LoRA) is a widely used parameter-efficient fine-tuning (PEFT) method for large language models. Under a fixed rank budget, LoRA parameterizes each adapted weight through a single low-dimensional input-side pathway, which may couple heterogeneous behaviors through shared input directions and induce interference during optimization. We propose Static Orthogonal Subspace LoRA (SOS-LoRA), a drop-in extension that reparameterizes a rank-rtot update as a sum of K static (always-on, non-routed) low-rank experts. SOS-LoRA (i) decomposes the total rank across experts, (ii) applies a fixed multi-scale scaling scheme to encourage scale-separated optimization dynamics, and (iii) promotes diverse input-side directions via cross-expert orthogonal initialization and a lightweight regularizer. SOS-LoRA remains fully mergeable, adding no inference-time parameters or latency after merging. Experiments on reasoning and knowledge-intensive benchmarks (Llama 2/3), encoder-based NLU (GLUE), and math reasoning (GSM8K/MATH) show consistent gains over matched-budget LoRA baselines and recent variants. Code is available at https://github.com/llm172/sos-lora.

cs.LG↗