arXiv subjects

Wenjie Wang

Publications and source records attributed to Wenjie Wang.

At least 55 records · Page 3Linked to original sources

arXiv2026

DR.GAP: Mitigating Bias in Large Language Models using Gender-Aware Prompting with Decoupled Reasoning

Large Language Models (LLMs) exhibit strong natural language understanding capabilities but also inherit and amplify societal biases, particularly gender bias, raising fairness concerns. Existing prompt-based debiasing strategies share a key limitation: they fail to disentangle gender information from task semantics. Bias steering compels models to overemphasize gender cues, while reasoning-based prompting induces gender-biased reasoning chains. To address these challenges, we propose DR.GAP (Decoupled Reasoning for Gender-Aware Prompting), an automated and model-agnostic pipeline that mitigates gender bias while preserving model performance. DR.GAP generates gender-neutral reasoning traces and applies them as in-context demonstrations during inference, effectively decoupling gender attributes from task semantics without modifying model parameters. Extensive experiments on coreference resolution and question-answering tasks across six LLMs demonstrate DR.GAP's effectiveness, generalizability, and robustness, supported by detailed mechanism analyses. Moreover, DR.GAP can be extended to vision-language models (VLMs), achieving substantial bias reduction.

cs.CL↗

arXiv2026

Hessian degeneracy of the torsion function on smooth simply connected nonconvex planar domains

We construct a family of bounded, smooth, simply connected, nonconvex planar domains $Ω_{a,ε}$. Let $u_{a,ε}$ be the corresponding torsion function satisfying \[ -Δu_{a,ε}=1\quad\text{in }Ω_{a,ε},\qquad u_{a,ε}=0\quad\text{on }\partialΩ_{a,ε}. \] There exists $a^*\in(0,1)$ such that, for every $a\in(a^*,1)$ and all sufficiently small $ε>0$, the origin is a unique global maximizer of $u_{a,ε}$. Moreover, \[ \lim_{a\downarrow a^*}\lim_{ε\to0} λ_{\max}\bigl(D^2u_{a,ε}(0)\bigr)=0. \] In addition, the ratios $\text{diam}(Ω_{a,ε}) / \text{inrad}(Ω_{a,ε})$ are uniformly bounded. Hence the Hessian estimate proved by Steinerberger (J. Funct. Anal. 274, 1611--1630, 2018) for convex planar domains cannot be extended to smooth, simply connected, nonconvex planar domains. \vskip0.2cm Our domains are star-shaped, symmetric with respect to both coordinate axes and convex in the horizontal direction. When the limiting slit is sufficiently long, we further show that certain superlevel sets of the torsion function are not star-shaped. This strengthens the counterexample to the star-shapedness question raised by Gladiali and Grossi (Amer. J. Math. 144, 1221--1240, 2022) under stronger geometric assumptions.

math.AP↗

arXiv2026

What to Remember, What to Reveal: Privacy-Aware Memory for Conversational Agents

Long-term memory enables personalized conversational agents to retain user information across sessions. However, existing memory architectures primarily optimize for utility while neglecting the risks of unnecessarily storing and reusing private attributes such as personally identifiable information (PII). Addressing privacy risks in personalized memory is challenging because simply removing sensitive values can undermine system utility. Therefore, privacy protection for memory agents should govern the full life cycle of sensitive values rather than only sanitizing individual records. To address this gap, we introduce Sanitized Privacy-Mapped Memory (SP-Mem), a privacy-aware memory architecture that decouples memory utility from exact private-value exposure. SP-Mem provides a full life-cycle privacy design that identifies and separates sensitive information from raw user inputs, stores sanitized content and exact private values in isolated structures, and selectively retrieves private values based on task requirements and user consent. We further introduce a privacy-aware memory benchmark that jointly evaluates response quality, privacy behavior, and inference cost. Extensive experiments across multiple LLM-based agents show that SP-Mem achieves stronger personalization while reducing unnecessary privacy exposure. Code and data are available at https://github.com/Jensassss/SP-Mem.

cs.CR↗

arXiv2026

Deep Koopman risk-preview supervised LTV-MPC for direct yaw moment control of distributed drive electric vehicles

Always-on direct yaw moment control (DYC) improves vehicle stability during critical maneuvers but can introduce unnecessary interventions under low-risk conditions. This paper proposes a Koopman risk-gated linear time-varying model predictive control (KRG-LTV-MPC) framework for low-intervention yaw stability assistance. Instead of replacing the physics-based execution model with a fully data-driven control predictor, this framework separates Koopman-based phase-risk preview from safety-critical execution. A Deep Koopman model predicts the nominal evolution of the sideslip-yaw rate phase risk to determine whether the constrained quadratic programming (QP) problem should be solved or skipped at each sampling time. When the gate is active, the LTV-MPC layer calculates the additional yaw moment; otherwise, the QP is skipped and the previously commanded moment is tapered to zero under a bounded-rate rule. Event-level shadow-mode evaluation shows that the Koopman predictor provides positive warning lead times of 0.19-0.28 s under low-friction and friction-transition conditions, whereas the LTV predictor gives delayed warnings. Under closed-loop low-friction conditions, KRG-LTV-MPC reduces the cumulative yaw-moment intervention by 43.9% relative to LTV-MPC and solves the QP for only 41.0% of the samples while maintaining vehicle stability within the phase plane. These results support the use of Koopman phase-risk information as an intelligent supervisory layer for low-intervention DYC.

eess.SY↗

arXiv2026

Electrostatic Superlattices beyond 1:1 Stoichiometry

Exotic nanoparticle superstructures can be accessed by harnessing nanoparticle softness and charge regulation, features often viewed as obstacles to structural control. Here, we show that regulated charge mismatch in polymer-grafted nanoparticles enables the assembly of high-stoichiometry cubic superlattices. By co-tuning grafting density, particle size, and bulk composition, we realize ionic-lattice analogues such as CaF2 and Th3P4, as well as single-component A3 and A7 superlattices without atomic counterparts. The A3 lattice has recently been identified theoretically as a photonic band-gap lattice. These phases emerge from a 1:1 "parent" lattice when local charge neutrality cannot be satisfied, driving either progressive interstitial filling or reorganization into a larger basis. For instance, the systematic occupation of ZnS tetrahedral sites yields CaF2, while ligand-swapping symmetry breaking converts CsCl into Th3P4. Upon heating, the assemblies exhibit reversible lattice contraction and pronounced negative thermal expansion. Furthermore, the energetic penalty for defects increases with nanoparticle size, facilitating the scalable production of high-quality, open superlattices for photonic applications.

cond-mat.mtrl-sci↗

arXiv2026

Can Large Language Models Derive New Knowledge? A Dynamic Benchmark for Biological Knowledge Discovery

Recent advancements in Large Language Model (LLM) agents have demonstrated remarkable potential in automatic knowledge discovery. However, rigorously evaluating an AI's capacity for knowledge discovery remains a critical challenge. Existing benchmarks predominantly rely on static datasets, leading to inevitable data contamination where models have likely seen the evaluation knowledge during training. Furthermore, the rapid release cycles of modern LLMs render static benchmarks quickly outdated, failing to assess the ability to discover truly new knowledge. To address these limitations, we propose DBench-Bio, a dynamic and fully automated benchmark designed to evaluate AI's biological knowledge discovery ability. DBench-Bio employs a three-stage pipeline: (1) data acquisition of rigorous, authoritative paper abstracts; (2) QA extraction utilizing LLMs to synthesize scientific hypothesis questions and corresponding discovery answers; and (3) QA filter to ensure quality based on relevance, clarity, and centrality. We instantiate this pipeline to construct a monthly-updated benchmark covering 12 biomedical sub-domains. Extensive evaluations of SOTA models reveal current limitations in discovering new knowledge. Our work provides the first dynamic, automatic framework for assessing the new knowledge discovery capabilities of AI systems, establishing a living, evolving resource for AI research community to catalyze the development of knowledge discovery.

cs.CL↗

arXiv2026

Physics from Video: Identifiability of Time-Invariant Second-Order ODEs under Minimal Trajectory Conditions

Bridging the gap between visual realism and physical understanding is a core challenge for video-based world models. We study the structural identifiability of continuous-time physical laws from raw pixels, focusing on whether an encoder-only pipeline can uniquely recover the parameters of second-order linear ODEs. We prove that a level-set slope-coverage condition ensures the learned latent space is locally affine to the true physical state, enabling exact parameter recovery. Our theory provides the first characterization of minimal data requirements across damping regimes, establishing that underdamped systems are identifiable from a single video clip, whereas other regimes require three diverse trajectories. We further introduce a variance-floor regularizer to stabilize the decoder-free objective and prevent latent collapse. Validated on synthetic and real-world data, our approach demonstrates that interpretable physical constants can be reliably estimated from video without the need for compute-intensive pixel reconstruction, ensuring both physical correctness and transparency. Code is available at https://github.com/wenjiewang3/PhysicsFromVideo.

cs.CV↗

arXiv2026

Reason4Rec: Deliberative User Preference Alignment of Large Language Models for Recommendation

Aligning Large Language Models (LLMs) with recommendation tasks represents an emerging paradigm in recommendation domain, exhibiting promising performance overall. However, these aligned recommendation LLMs often struggle with complex scenarios due to limitations in the current alignment task formulation, which optimizes LLMs to directly generate user feedback without deliberation. To develop more reliable recommendation LLMs, we introduce a new \textit{Deliberative Recommendation} task, which enforces explicit reasoning about user preferences as an additional alignment objective. To address this task, we propose a \textit{Reasoning-powered Recommender} framework designed to enhance reasoning capabilities by leveraging verbalized user feedback in a step-wise manner. Specifically, this framework employs collaborative step-wise experts alongside specifically crafted expert-wise training strategies. Extensive experiments conducted on three real-world datasets demonstrate the rationality of the deliberative task formulation and the effectiveness of the proposed framework in improving both prediction accuracy and reasoning quality. Our implementation is publicly available on GitHub: https://github.com/Peter-Fy/Reason4Rec.

cs.IR↗

arXiv2026

Controllable LLM Reasoning via Sparse Autoencoder-Based Steering

Large Reasoning Models (LRMs) exhibit human-like cognitive reasoning strategies (\eg backtracking, cross-verification) during the reasoning process, which improves their performance on complex tasks. Currently, reasoning strategies are autonomously selected by LRMs themselves. However, such autonomous selection often produces inefficient or even erroneous reasoning paths. To make reasoning more reliable and flexible, it is important to develop methods for controlling reasoning strategies. Existing methods struggle to control fine-grained reasoning strategies due to conceptual entanglement in LRMs' hidden states. To address this, we leverage Sparse Autoencoders (SAEs) to decompose strategy-entangled hidden states into a disentangled feature space. To identify the few strategy-specific features from the vast pool of SAE features, we propose SAE-Steering, an efficient two-stage feature identification pipeline. SAE-Steering first recalls features that amplify the logits of strategy-specific keywords, filtering out over 99\% of features, and then ranks the remaining features by their control effectiveness. Using the identified strategy-specific features as control vectors, SAE-Steering outperforms existing methods by over 15\% in control effectiveness. Furthermore, controlling reasoning strategies can redirect LRMs from erroneous paths to correct ones, achieving a 7\% absolute accuracy improvement. Our code and data are available at https://github.com/Peter-Fy/SAE-Steering.

cs.AI↗

arXiv2026

Tracing cosmic star formation history through radio continuum spectral energy distribution and non-thermal emission

As a tracer of massive star formation unaffected by dust, the radio continuum emission provides a unique window into the formation of the first stars and galaxies in the Universe. Recent observations show that the integrated rest-frame mid-radio (~1-10 GHz) luminosity of galaxies serves as one of the most robust tracers of the star formation rate (SFR). These studies further demonstrate that the synchrotron spectral index and the shape of the radio spectral energy distribution (SED) evolves with redshift as a consequence of the cosmic evolution of star formation activity. These findings underscore the importance of deep multi-band radio continuum observations in calibrating the SFR of early galaxies and understanding the astrophysical processes governing their assembly and evolution over cosmic time. This chapter presents recent progress in radio SFR calibrations for star-forming galaxies (SFGs) and reviews radio-continuum studies of the cosmic star formation history (SFH). We highlight the transformative potential of SKA AA4, whose broad frequency coverage and high sensitivity will enable well-constrained radio SEDs for SFGs across a wide redshift range.

astro-ph.GA↗

arXiv2026

StanceFlip: A Comprehensive Multi-Dimensional Benchmark for Multimodal Conversational Stance Flipping Forecasting

Conversational stance detection has shifted from static text analysis to dynamic multimodal modeling. However, existing benchmarks exhibit three key limitations: failure to capture the dynamic evolution of beliefs, particularly during stance reversals; difficulty in disentangling affective states from logical reasoning; and neglect of the critical role of multimodal cues in resolving pragmatic ambiguities such as sarcasm. To address these limitations, we propose StanceFlip, a benchmark designed for multimodal conversational stance flipping forecasting over multi-turn dialogues across five modalities and multi-scenarios, which includes two novel subtasks: 1) Multimodal Stance Sextuple Extraction, extracting holder, target, emotion, sentiment, stance, and rationale as static state snapshots of dialogue to capture fine-grained cognitive structures. 2) Dynamic Stance Flip Attribution, tracking stance reversals across the conversation and identifying their underlying triggers. Alongside the dataset, we propose a dedicated framework, named ConStaFF, for Multimodal Conversational Stance Flipping Forecasting (MCSFF). Built upon a large language model, ConStaFF performs end-to-end stance reasoning, with a Thought-of-Stance (ToS) reasoning framework and a self-reflective verification mechanism integrated for structured stance modeling and faithful flip attribution. Specifically, ToS decomposes the reasoning process into specialized cognitive personas to formulate target propositions, resolve cross-modal conflicts, and infer historical stance trajectories. Extensive experiments show that our approach achieves state-of-the-art performance on both sextuple extraction and flip-trigger attribution, outperforming strong multimodal large language model baselines by substantial margins.

cs.CL↗

arXiv2026

Constrained Tensor Decomposition-Based Target Sensing for Sparse Non-Uniform Array-Enabled AFDM ISAC Systems

Sparse non-uniform array-enabled affine frequency division multiplexing (AFDM) is a promising candidate for integrated sensing and communication (ISAC), while its performance critically depends on accurate target parameter estimation. In this paper, we propose a constrained tensor decomposition-based sensing framework for delay, Doppler, and angle estimation. Specifically, a manifold-constrained alternating least squares (ALS) algorithm is developed by exploiting the sparse array geometry structure, enabling robust factor matrix extraction and direct angle estimation. From the decomposed factor matrices, we further apply an iterative one dimensional golden section search to refine delay and Doppler shift. Simulation results demonstrate that the proposed algorithm nearly attains Cramér-Rao bound (CRB) and significantly outperforms unconstrained ALS and conventional methods, validating its effectiveness for sparse non-uniform array-enabled AFDM ISAC systems.

eess.SP↗

arXiv2026

How Jailbreak Attacks Inform Safety Alignment: A Defender-Centric, Shapley-Based Evaluation of Jailbreak Contributions

Jailbreak attacks on large language models are usually evaluated by attacker-centric metrics such as attack success rate (ASR), yet an attack that breaks a model is not necessarily useful for improving its safety. We propose a defender-centric view of jailbreak evaluation, where attacks are evaluated by the downstream safety improvements they enable when used as red-teaming data for safety training. Building on this view, we introduce A-MESS (Minimal Effective Attack-Subset Selection), a setting-agnostic framework for attributing and selecting jailbreak attacks from black-box subset utility observations. A-MESS estimates AttackSHAP, a Shapley-based score that attributes marginal utility to individual attacks and selects compact attack subsets under user-specified budgets via greedy or surrogate-based optimization. Across controlled utility landscapes and real LLM safety settings, we find that ASR rankings are weakly aligned with defender-centric utility, that AttackSHAP can be estimated accurately with limited utility queries, and that directly optimizing subsets yields stronger safety utility than attacker-centric or attribution-only selection. These results suggest evaluating jailbreak attacks as resources for improving safety, not only as tools for breaking models.

cs.CR↗

arXiv2026

MemoHarness: Agent Harnesses That Learn from Experience

An agent harness is the external control layer that turns a base LLM into an executable agent by managing context, tools, orchestration, memory, decoding, and output handling. While harness design strongly affects agent behavior, most automatic improvement methods optimize narrower artifacts such as prompts, pipelines, or workflows, and deployed agents usually reuse a single global harness for all cases. We introduce MemoHarness, an adaptive harness optimization framework that learns from its own executions. MemoHarness decomposes the harness into six editable control dimensions, stores per-case diagnoses and distilled global patterns in a dual-layer experience bank, and adapts the learned harness to each test case using retrieved experience without test-time labels, feedback, or additional search. In our evaluation across shell-agent, code-generation, and analytical-reasoning benchmarks, MemoHarness improves over the fixed harnesses we compare against and shows selective transfer to unseen suites and base models. Its additional context can also remain cost-competitive when much of the retrieved experience is cacheable. These results provide evidence that execution experience is a practical substrate for building agent harnesses that are more adaptive than a single static configuration, while leaving broader claims about statistical robustness and component attribution to future work.

cs.AI↗

arXiv2026

From Passive Retrieval to Active Memory Navigation: Learning to Use Memory as a Structured Action Space

Long-term user memory is essential for personalized conversational agents, yet many memory systems still expose memory through passive retrieval interfaces, making the model a consumer of pre-selected evidence. We introduce NapMem, a framework for learning to use long-term user memory as a structured action space rather than passively retrieved context. NapMem organizes user history into a linked multi-granularity memory pyramid, where raw conversations, typed memory records, topic tracks, and user profiles are connected through provenance relations, and exposes these levels through memory tools. The agent is trained to select memory according to the query and intermediate evidence, allowing it to inspect different memory granularities before answering. Experiments on PersonaMem-v2, LongMemEval, and LoCoMo show that a NapMem agent trained with memory-tool reinforcement learning is competitive across diverse memory-intensive tasks, while evaluations on non-memory tasks suggest that the learned policy largely preserves general reasoning and tool-use abilities. Additional analyses examine storage, inference cost, tool-use behavior, and ablations over navigation, memory granularity, and RL training. Our results suggest that long-term user memory benefits from coupling structured storage with a learned policy for using memory at the appropriate granularity.

cs.AI↗

arXiv2026

Autonomous Information Seeking: A Roadmap for Agentic Recommender Systems

The rapid integration of large language model-based agents into recommender systems has driven a shift from static, ranking-based pipelines toward autonomous and interactive systems that can reason, plan, and act. This survey provides a comprehensive overview of this emerging landscape by introducing a unified taxonomy grounded in the level of autonomy and three core paradigms of agentic recommender systems: agent-assisted recommendation, agent-as-recommender, and agent-as-user-simulator. The autonomy framework organizes existing methods along increasing capabilities in proactivity, context awareness, interaction flexibility, and adaptivity. Building on this framework, the survey analyzes how each paradigm adopts different agentic architectures and how agents enhance key components such as profiles, memory, tool use, workflows, and optimization mechanisms. We further examine evaluation methodologies for agentic recommendation, covering automated metrics, LLM-based judging, and simulation-based assessment, and discuss their limitations in capturing reasoning quality, user experience, and system behavior. Beyond existing evaluation protocols, we further discuss unresolved issues in evaluating agentic recommender systems, including trajectory-level assessment, agent contribution analysis, and calibration of user simulation. Lastly, the survey outlines open challenges in lifelong user modeling, contextual abstraction, multimodal alignment, controllability, trustworthiness, privacy, scalability, and efficiency. Together, these analyses establish a unified foundation for understanding the current progress of agentic recommender systems and highlight promising opportunities for developing more autonomous, reliable, and human-aligned recommendation agents.

cs.IR↗

arXiv2026

Optimizing Visual Generative Models via Distribution-wise Rewards

Conventional reinforcement learning strategies for visual generation typically employ sample-wise reward functions, yet this practice frequently results in reward hacking that degrades image diversity and introduces visual anomalies. To address these limitations, we present a novel framework that finetunes generative models using distribution-wise rewards, ensuring better alignment with real-world data distributions. Unlike rewards that evaluate samples individually, distribution-wise reward accounts for the data distribution of the samples, mitigating the mode collapse problem that occurs when all samples optimize towards the same direction independently. To overcome the prohibitive computational cost of estimating these rewards, we introduce a subset-replace strategy that efficiently provides reward signals by updating only a small subset of a generated reference set. Additionally, we apply RL to optimize post-hoc model merging coefficients, potentially mitigating the train-inference inconsistency caused by introducing stochastic differential equation (SDE) in regular RL practices. Extensive experiments show our approach significantly improves FID-50K across various base models, from 8.30 to 5.77 for SiT and from 3.74 to 3.52 for EDM2. Qualitative evaluation also confirms that our method enhances perceptual quality while preserving sample diversity.

cs.LG↗

arXiv2026

Synthetic Interaction Data for Scalable Personalization in Large Language Models

Personalized prompting offers large opportunities for deploying large language models (LLMs) to diverse users, yet existing prompt optimization methods primarily focus on task-level optimization while largely overlooking user-specific preferences and latent constraints of individual users. This gap is primarily due to (i) the absence of high-quality, privacy-sensitive data that capture personalized user-LLM interactions at scale, and (ii) the lack of robust reward signals for individual preferences. To overcome existing data limitations, we introduce a high-fidelity synthetic data generation framework called PersonaGym. Unlike prior work that treats personalization as static persona-preference pairs, PersonaGym models a dynamic preference process via an agentic LLM system to simulate realistic preference behaviors and semantic-aware noise in order to generate personalized multi-turn interaction trajectories. Using PersonaGym, we release PersonaAtlas, a large-scale, high-quality, and diverse synthetic dataset of high-fidelity multi-turn personalized interaction trajectories that closely mirror real-world preference expression and noise patterns. We further propose Personalized Prompt Optimization (PPOpt), a scalable and model-agnostic framework that optimizes user prompts based on interaction histories without modifying the deployed LLM. PPOpt adopts a reason-then-optimize paradigm that infers an explicit user profile and conditions prompt rewriting on the user profile to avoid reward hacking. Our training procedure for PPOpt integrates a cold-start supervised prior with outcome-driven multi-objective reinforcement learning. We present extensive experiments to demonstrate consistent improvements over state-of-the-art baselines in terms of task performance, personalization quality, and robustness to noisy as well as to sparse preference signals.

cs.LG↗