arXiv subjects

Zhen Wu

Publications and source records attributed to Zhen Wu.

At least 19 recordsLinked to original sources

arXiv2026

Multidimensional Infinite-Horizon Markovian Backward Stochastic Differential Equations with Continuous Coefficients and Nonzero-Sum Differential Games

This paper addresses the solvability of multidimensional infinite-horizon Markovian backward stochastic differential equations with continuous coefficients. We do not require the Lipschitz property or other quantitative controls of differences, and each component of the generator is allowed to depend simultaneously on all components of the multidimensional backward variables. A local L^q-domination framework adapted to the infinite-horizon setting is developed and combined with our linear flow conjugacy analysis to obtain the required distributional control in the presence of unbounded linear drift. We also derive a common Markov representation valid for the entire family of initial states. The analytical obstacles arising from weak regularity and the multidimensional structure are thereby overcome. Furthermore, we consider a class of infinite-horizon nonzero-sum differential games with a general non linear structure, allowing player controls to be fully coupled and permitting payoff criteria that are not restricted to an exponential discount form, establish an equivalence between the game and the corresponding multidimensional infinite-horizon backward stochastic differential equations, and prove the existence of a Nash equilibrium.

math.PR↗

arXiv2026

Harness Policy Distillation: Aligning Agent Policies with Runtime Semantics

Deploying a general-purpose model as an agent means running it inside a product-specific runtime harness that decides which actions are executable, how state advances, and when the loop may terminate. Ordinary trajectory supervision shows which action was taken but leaves the runtime semantics behind that action implicit. We introduce Harness Policy Distillation (HPD), which makes these runtime control semantics an explicit training target alongside actions to align the harness policy. HPD replays reference trajectories through the harness to recover the runtime semantics that are implicit in ordinary trajectory data, and represents them as control documents at each model-call boundary. These documents provide privileged information for a teacher to generate actions, while harness verification ensures that the resulting targets are executable under the runtime contract. We jointly train a shared model on the runtime semantics recovered from control-state privileged information and native history, thereby internalizing these semantics into the model policy. At deployment, the trained model uses the original harness interface without requiring the control document or any additional supplement. On TRAJECT-Bench with the Pi harness, Qwen3-8B trained with HPD reaches 22.98% autonomous task success and 49.23% runtime compliance, against 3.09% and 19.90% for the unadapted model. Harness-verified teacher actions substantially improve the native deployment policy, while joint control-state supervision provides further gains in task completion and runtime compliance. HPD turns executable runtime semantics into dense supervision for learning harness-compatible actions at each decision boundary.

cs.LG↗

arXiv2026

WebNavigator: Global Web Navigation via Interaction Graph Retrieval

Despite significant advances in autonomous web navigation, current methods remain far from human-level performance in complex web environments. We argue that this limitation stems from Topological Blindness, where agents are forced to explore via trial-and-error without access to the global topological structure of the environment. To overcome this limitation, we introduce WebNavigator, which reframes web navigation from probabilistic exploration into deterministic retrieval and pathfinding. WebNavigator constructs Interaction Graphs via zero-token cost heuristic exploration offline and implements a Retrieve-Reason-Teleport workflow for global navigation online. WebNavigator achieves state-of-the-art performance on WebArena and OnlineMind2Web. On WebArena multi-site tasks, WebNavigator achieves a 72.9\% success rate, more than doubling the performance of enterprise-level agents. This work reveals that Topological Blindness, rather than model reasoning capabilities alone, is an underestimated bottleneck in autonomous web navigation.

cs.IR↗

arXiv2026

Mean-variance portfolio selection: connections between mean-field systems and forward-backward stochastic systems under initial-terminal cost coupling

This paper investigates the connection between mean-field systems and forward-backward stochastic systems arising from the mean-variance portfolio selection problem. By introducing the conditional expectation process of terminal wealth, the mean-field term is represented by a backward state, leading to a class of control problems for forward-backward stochastic systems with coupled initial-terminal values. The coupling between the terminal value of the forward state and the initial value of the backward state in the cost functional prevents classical methods for forward-backward stochastic optimal control from being directly applicable. To overcome this difficulty, we establish a stochastic maximum principle for control problems with running costs and coupled initial-terminal values, and provide sufficient conditions for optimality. Furthermore, we investigate a class of generalized stochastic linear-quadratic control problems with initial-terminal coupling, allowing the control weight in the running cost to be indefinite, and derive the optimal feedback representation via a Riccati equation. The classical mean-variance portfolio selection model is shown to be a special case, for which the corresponding efficient investment strategy and efficient frontier are obtained explicitly, and numerical simulations are presented to illustrate the results and examine parameter sensitivity and its financial implications.

math.OC↗

arXiv2026

Counterfactual Video Generation Enables Scalable Humanoid Loco-Manipulation

Teaching humanoids loco-manipulation skills, such as carrying diverse objects, via visual imitation is a promising path toward generalist robots. However, collecting diverse, high-quality interaction videos, such as clips that clearly show a person's full body and unoccluded interactions with objects, poses a practical barrier to scaling this approach. We propose PRISM, a real-to-sim-to-real framework that overcomes this limitation by amplifying a handful of real videos into a large, diverse training set. PRISM first generates hundreds of diverse "counterfactual" human-object interaction videos via video-to-video (V2V) generation from a few exemplar real videos. Our contact-anchored real-to-sim pipeline then reconstructs both human and object motions, retargeting this imperfect video data into physically plausible trajectories. The intra-class variability across these counterfactual videos lets us train a single policy that generalizes to unseen objects within each category. We demonstrate the full pipeline by deploying this policy on a real robot without any real-world fine-tuning. Using only onboard depth observations, our humanoid picks up, carries, and drops objects, including boxes, barrels, bins, and balls, across novel instances, sizes, and initial configurations.

cs.RO↗

arXiv2026

Master Equations for Mean Field Game of Controls: A Unification of Weak Solution Notions

In this manuscript, we study the master equation for mean field game of controls, assuming only that the data are Lipschitz continuous in the measure variable. Accordingly, we propose a weaker notion of solution called weak solutions for the master equation. We establish the global well-posedness of the master equation for such solution under the Lasry-Lions monotonicity and displacement $λ$-monotonicity conditions, respectively. The arguments rely on the analysis of the associated Pontryagin forward-backward stochastic differential equation system, especially the stability in the measure variable. Finally, we review several existing notions of non-smooth solution from the literature, namely good solutions, weak viscosity solutions, Lipschitz solutions, monotone solutions, and examine their relationship with our weak solution. We show that under appropriate assumptions all of them are equivalent by their definitions, and thus our manuscript provides a unification of weak solution notions for the master equations derived from mean field game of controls.

math.OC↗

arXiv2026

A minimal dynamical model of human-AI cognitive coevolution

The rapid deployment of generative AI raises the question of whether pervasive cognitive offloading drives a measurable, potentially irreversible decline in unassisted human reasoning. We examine this question with a phenomenological dynamical model based on the Cattell-Horn-Carroll distinction between fluid intelligence ($G_f$) and crystallized intelligence ($G_c$). Starting from the discrete-time Bivariate Latent Change Score formalism used in cognitive-aging research, we derive its continuous-time linear counterpart and introduce a fast third variable representing composite machine-intelligence capability. The machine variable enters the human subsystem through Hill-function modifications of the self-feedback and cross-coupling coefficients. We emphasize that our phenomenological model captures both the positive and negative effects of AI on human cognition. By formalizing the qualitative concepts of positive and negative impacts, we introduce definitions for AI-induced cognitive decline and AI-induced cognitive augmentation. A singular-perturbation reduction, justified via Tikhonov's theorem, collapses the resulting three-dimensional system onto a two-dimensional slow manifold. On this manifold we locate two bifurcations of the equilibrium. In one parameter regime, weakening the contribution of fluid reasoning to crystallized knowledge produces a saddle-node bifurcation that eliminates the stable positive equilibrium. In another, reversing the contribution of crystallized knowledge to fluid reasoning creates opposing feedback; increasing fluid self-feedback then reduces damping and produces a supercritical Hopf bifurcation with a stable periodic orbit. We discuss limitations and future directions, interpret our results as a phenomenological account of a coupled socio-technical system, and state the conditions under which each dynamical regime applies.

q-bio.NC↗

arXiv2026

From Refusal Tokens to Refusal Control: Discovering and Steering Category-Specific Refusal Directions

Language models are commonly fine-tuned for safety alignment to refuse harmful prompts. One approach fine-tunes them to generate categorical refusal tokens that distinguish different refusal types before responding. In this work, we leverage a version of Llama 3 8B fine-tuned with these categorical refusal tokens to enable inference-time control over fine-grained refusal behavior, improving both safety and reliability. We show that refusal token fine-tuning induces separable, category-aligned directions in the residual stream, which we extract and use to construct categorical steering vectors with a lightweight probe that determines whether to steer toward or away from refusal during inference. In addition, we introduce a learned low-rank combination that mixes these category directions in a whitened, orthonormal steering basis, resulting in a single controllable intervention under activation-space anisotropy, and show that this intervention is transferable across same-architecture model variants without additional training. Across benchmarks, both categorical steering vectors and the low-rank combination consistently reduce over-refusals on benign prompts while increasing refusal rates on harmful prompts, highlighting their utility for multi-category refusal control.

cs.AI↗

arXiv2026

Neutralizing Popularity Bias in LLM-based Recommendation via Counterfactual Reasoning Guidelines

In the era of generative AI, recommender systems are moving from precise prediction to trustworthy generation. Large language models (LLMs) support this shift by inferring user interests and producing natural-language explanations. However, LLM-based recommendation suffers from a fundamental obstacle: popularity bias. Through pre-training on massive corpora, LLMs tend to rely on global statistics and trend signals, yielding recommendations that follow popularity rather than genuine preference. As this bias is entangled in model parameters and is hard to remove directly, we propose Neutralizing Popularity Bias in LLM-based Recommendation via Counterfactual Reasoning Guidelines (NPRec), a model-agnostic framework that mitigates popularity bias through external semantic intervention. NPRec performs counterfactual refinement to causally separate intrinsic user interests from popularity-driven conformity, producing debiased textual guidelines that reflect actual user preferences. These guidelines are injected at inference time to shift the LLM from unconstrained generation to guided reasoning, without any parameter updates. Serving as explicit premises, they both ground faithful explanations and improve recommendation quality. Extensive experiments on three real-world datasets demonstrate that NPRec achieves promising performance in recommendation accuracy, explanation quality, and debiasing capability.

cs.AI↗

arXiv2026

CHIPSMORE: Compute-in-Interconnect and -Memory Chiplets for Multi-Mode Multi-Request LLM Inference Acceleration

Large language model (LLM) inference exhibits substantial variability across adaptation modes, context lengths, and request concurrency, creating challenges for maintaining high utilization, memory efficiency, and scalable performance on compute-in-memory (CIM) accelerators. This paper presents CHIPSMORE, a multi-mode and multi-request LLM inference accelerator that integrates compute-in-interconnect and CIM to support both base-mode and low-rank adaptation (LoRA) inference under diverse workloads. CHIPSMORE employs heterogeneous processing elements consisting of resistive RAM analog compute-in-memory (RRAM-ACIM) and static RAM digital compute-in-memory (SRAM-DCIM) interconnected through a programmable Inter-PE computational network (IPCN). A composable hierarchical key-value (KV) memory scheme dynamically allocates router scratchpad, SRAM-DCIM, and embedded DRAM (eDRAM) resources according to workload requirements, enabling scalable support for long-context and batched inference. Furthermore, a non-replicated multi-request execution pipeline exploits request-level parallelism without duplicating pretrained weights, while a state-aware resource reconfiguration mechanism selectively retains runtime states and power-gates inactive resources to improve energy efficiency. Evaluation using cycle-accurate hardware-software co-simulation demonstrates that CHIPSMORE effectively sustains high throughput across varying model sizes, context lengths, and batch sizes while maintaining favorable power scaling. Compared with Nvidia H100, CHIPSMORE achieves up to $2.38\times$ higher throughput and $27\times$ higher energy efficiency on Mistral-7B inference while eliminating weight replication for multi-request serving.

cs.AR↗

arXiv2026

Development of A Novel Compton Camera for MeV Gamma-Ray Measurement in Space

The astrophysical gamma rays in the MeV energy region have not yet been well-explored due to the limitation of detection technology in the past decades, and the famous gamma-ray "MeV gap" exists. Opening the window of MeV gamma-ray is not only critical for the gamma astronomy but also essential for rich frontier researches in astro-particle physics, such as detecting light dark matter, probing the primordial black hole and better understanding of nucleosynthesis. As a pilot experiment of the project for dark matter detection in space at Shanghai Jiao Tong University, a three-layer Compton camera with the energy resolution better than 4% and position resolution of ~2 mm is developed utilizing the novel scintillators. Here we show the design, detailed calibration and validation results of the novel Compton camera, and demonstrate its good ability of MeV gamma-ray source imaging for the upcoming in-orbit mission.

hep-ex↗

arXiv2026

OpenVisTool: An Open Recipe for Synthesizing Instructive Visual Tool-Use Trajectories

Visual tool use has emerged as a fundamental capability for multimodal agents to actively acquire evidence beyond a fixed image encoding. The prevailing recipe learns this capability from teacher-generated trajectories filtered for answer correctness, implicitly assuming that every successful demonstration provides effective supervision. We argue this assumption is flawed: a strong teacher often reaches the correct answer without needing its tool calls, and imitating such trajectories teaches a student that tool calls accompany correct answers, not that tool observations ground them. We present OpenVisTool, an open framework for constructing instructive visual tool-use trajectories that provide effective supervision for tool learning. The key insight is that a trajectory should be retained only if its answer is correct (outcome validity) and its tool observations causally contribute to that answer (causal utility). The framework operates in three stages: difficulty screening to select queries that are not reliably answerable without tools, domain-specific trajectory synthesis to elicit coherent tool-use trajectories, and supervision verification to jointly test both conditions. Rather than encouraging models to imitate tool calls, the resulting supervision teaches when and how visual evidence should be acquired. Using this framework, we construct OpenVisTool-42K, a dataset spanning five visual reasoning domains, together with OpenVisTool-Bench, a benchmark covering the same domains. Across four backbones (4B-27B), fine-tuning on OpenVisTool-42K consistently improves visual tool-use performance and yields gains on two out-of-distribution benchmarks; the larger models approach leading closed-source systems. The evidence suggests that effective visual tool use is learned from causally grounded supervision rather than tool-calling patterns.

cs.CL↗

arXiv2026

MultiModal Code-Switching: Interleaving Visual Objects into Language for Explicit Object-Level Alignment

Existing Multimodal Large Language Models (MLLMs) predominantly rely on image-text pairs for modality alignment pretraining, mapping global image representations to long textual descriptions. However, this image-level alignment suffers from referential ambiguity: models struggle to infer the correspondences between multiple visual objects and textual entities from the global representation, leading to data inefficiency and suboptimal semantic grounding. To address this, we propose MultiModal Code-Switching (MMCS), a novel pretraining paradigm that provides explicit object-level supervision. Inspired by the linguistic phenomenon of code-switching, MMCS interleaves vision and language by replacing textual entities with their corresponding visual objects, enforcing local vision-language grounding. We further develop a scalable data synthesis pipeline to generate a pretraining dataset of 773K samples with accurate object-entity correspondences. Experiments show that MMCS is highly data-efficient: with only 50K samples, it matches or surpasses models trained on 600K image-text pairs. Furthermore, MMCS consistently improves visual grounding and perception capabilities across varying model scales.

cs.CV↗

arXiv2026

QuReC: All-in-One Image Restoration with Query-Specific Guidance and Local-Global Response Calibration

All-in-one image restoration aims to recover clean images degraded by multiple corruption types using a single unified model. Existing methods typically rely on image-level prompts or shared guidance to handle diverse degradations. However, such a paradigm becomes inadequate when degradations are spatially heterogeneous or even coexist in mixed forms within a single image. Yet spatially adaptive guidance alone is not sufficient, since accurate restoration also requires each spatial query to reliably aggregate complementary information from local neighborhoods and global contexts. To this end, we propose QuReC, a unified framework for all-in-one image restoration. QuReC consists of a Degradation-Guided Query Reconstruction Module (DQRM) and a Local-Global Response Calibration Module (LGRCM). Specifically, DQRM matches each spatial query against a degradation prototype space to reconstruct a query-specific degradation-aware representation, thereby providing fine-grained spatially adaptive restoration guidance. To further stabilize this query-wise matching process, we introduce a weakly supervised prototype matching learning strategy to improve optimization stability and degradation semantic consistency. Meanwhile, LGRCM performs local-global dual-branch aggregation and calibrates the aggregated responses with learnable priors, improving the reliability of feature aggregation and the coordination between local detail modeling and global context modeling. Extensive experiments demonstrate that QuReC achieves superior performance on multiple all-in-one image restoration benchmarks. The code is released at https://github.com/zhoushen1/QuReC.

cs.CV↗

arXiv2026

InCarEmo: A Multimodal Dataset for In-Cabin Emotion Recognition and Driver State Monitoring

Understanding driver emotion and state is critical for the next generation of intelligent in-cabin systems that ensure safety and enhance human-vehicle interaction. However, existing public datasets for in-cabin affective computing are largely limited to visual modalities and rarely include conversational information, making it difficult to capture the linguistic and interactive cues underlying driver emotion. To address these gaps, we introduce InCarEmo, a multimodal dataset for in-cabin emotion recognition and driver state monitoring. InCarEmo integrates RGB and infrared video, in-cabin audio, and dialogue text collected from scripted in-cabin scenarios designed to simulate realistic driver behaviors, covering diverse lighting conditions and driving contexts. The dataset supports three primary tasks: 1) multimodal emotion recognition, 2) fatigue detection, and 3) distraction monitoring. In addition to the original Chinese data, we construct an auxiliary English benchmark to support preliminary cross-lingual evaluation. We provide a unified benchmark with extensive baseline results across unimodal and multimodal methods, including analyses under modality-missing and noise conditions. Experimental results demonstrate the benefits of multimodal fusion and reveal remaining challenges under real-world noise and low-light conditions. By releasing InCarEmo, we aim to establish a comprehensive foundation for robust, interpretable, and human-centric in-cabin affective understanding, promoting safer and more empathetic driver-vehicle interaction.

cs.AI↗

arXiv2026

Optimizing RAG Rerankers with LLM Feedback via Reinforcement Learning

Rerankers play a pivotal role in refining retrieval results for Retrieval-Augmented Generation. However, current reranking models are typically optimized on static human annotated relevance labels in isolation, decoupled from the downstream generation process. This isolation leads to a fundamental misalignment: documents identified as topically relevant by information retrieval metrics often fail to provide the actual utility required by the LLM for precise answer generation. To bridge this gap, we introduce ReRanking Preference Optimization (RRPO), a reinforcement learning framework that directly aligns reranking with the LLM's generation quality. By formulating reranking as a sequential decision-making process, RRPO optimizes for context utility using LLM feedback, thereby eliminating the need for expensive human annotations. To ensure training stability, we further introduce a reference-anchored deterministic baseline. Extensive experiments on knowledge-intensive benchmarks demonstrate that RRPO significantly outperforms strong baselines, including the powerful list-wise reranker RankZephyr. Further analysis highlights the versatility of our framework: it generalizes seamlessly to diverse readers (e.g., GPT-4o), integrates orthogonally with query expansion modules like Query2Doc, and remains robust even when trained with noisy supervisors.

cs.CL↗

arXiv2026

SceneBot: Contact-Prompted General Humanoid Whole Body Tracking with Scene-Interaction

Current humanoid reinforcement-learning policies excel at free-space motions but struggle with contact-rich tasks, as pure kinematic tracking cannot resolve the physical ambiguities of interacting with objects and uneven terrain. To address this, we introduce SceneBot, a unified motion-tracking framework capable of handling freespace locomotion, terrain traversal, and whole-body manipulation. SceneBot conditions a single policy on both reference motions and per-link contact labels, explicitly defining expected environmental interactions. To overcome the lack of annotated interaction data, we propose a hindsight scene reconstruction approach that infers scene-interaction graphs from retargeted human motion. Trained on 7.5 hours of this reconstructed, contact-rich data, SceneBot successfully generalizes to unseen motions and environments. Our results demonstrate that SceneBot is the first general framework to seamlessly unify free-space and contact-rich behaviors executing complex, long-horizon tasks like carrying a box upstairs and establishing contact conditioning as a powerful interface for humanoid control. All code and data will be open-sourced. More demos and information are available at: https://ericcsr.github.io/scenebot/

cs.RO↗

arXiv2026

LatentCRS: A Variational EM Framework for Bridging Semantics and Behavior in LLM-based Conversational Recommendation

Conversational Recommender Systems (CRS) powered by Large Language Models (LLMs) enable users to articulate explicit and dynamic preferences, overcoming the limitations of fixed templates. However, despite their superior semantic proficiency, LLMs have not yet achieved corresponding improvements in recommendation accuracy. This discrepancy arises from a fundamental representation gap: while LLMs operate within a semantic space, they lack the behavioral grounding needed to encode user behavioral patterns, such as item co-occurrences, which are crucial for accurate recommendations. To address this, we propose a model-agnostic Variational EM Framework for Bridging Semantics and Behavior in LLM-based Conversational Recommendation (LatentCRS). Based on the observation that dialogue and interactions reflect the same latent intent, LatentCRS uses a variational expectation-maximization (EM) procedure, where user intent connects semantic representations with behavioral patterns. Extensive experiments on real-world datasets demonstrate that LatentCRS effectively bridges the representation gap and outperforms baselines.

cs.CL↗