arXiv subjects

Yuxuan Lu

Publications and source records attributed to Yuxuan Lu.

At least 19 recordsLinked to original sources

arXiv2026

Retry, Switch, or Abstain? Learning Strategy-Aware Tool-Use Policies via Controlled Error Injection

Tool-using LLM agents are commonly trained and evaluated in environments where tool calls succeed reliably, yet deployed tools can fail transiently, persistently, or silently. Robust recovery therefore requires more than repeated retries: an agent may need to retry the same path, switch to an alternative, or recognize that no viable path remains. We present BENCH2ROBUST, a framework that converts failure-free tool-use benchmarks into controlled stochastic environments with scenario-controlled solvability, where episodes explicitly require retrying, switching, or stopping after available paths are exhausted. We use BENCH2ROBUST to study two complementary interventions: structured runtime recovery context through Bayesian Tool Memory (BTM), and curriculum-controlled reinforcement learning. Across 7 models from 4 families and two multi-turn benchmark families, tool failures produce a near-universal robustness gap. On held-out Retail tasks, BTM improves robustness by up to 16.8 percentage points without retraining, while RL learns complementary recovery behavior that remains beneficial without inference-time BTM. Combining the two reaches 40.8-45.5% under injection while preserving failure-free performance. These results suggest that robust tool use benefits from combining environment-specific recovery knowledge with learned recovery behavior.

cs.AI↗

arXiv2026

FactorDrive: Adaptive Multi-Step Reasoning Driven by Planning-Critical Factors for End-to-End Autonomous Driving

Vision-language models (VLMs) have advanced scene understanding and enabled explicit reasoning in end-to-end autonomous driving. However, existing methods insufficiently integrate spatial-physical evidence into planning reasoning, while reasoning adaptation remains coarse-grained and falls short of scene-specific planning demands. Furthermore, reasoning-path optimization for higher planning quality remains largely unexplored in autonomous-driving post-training. To address these limitations, we propose FactorDrive, an end-to-end autonomous driving framework for adaptive multi-step reasoning driven by planning-critical factors (PCFs). We first perform large-scale driving-domain instruction tuning to establish foundational driving knowledge. Building on this foundation, we construct PCF-CoT, a chain-of-thought (CoT) dataset that grounds planning reasoning in trajectory-relevant spatial-physical evidence and organizes reasoning around scene-specific PCFs, enabling the composition and depth of reasoning paths to adapt to different planning demands. We further introduce Quality Search-Guided Group Relative Policy Optimization (QS-GRPO), which guides Monte Carlo Tree Search (MCTS) with trajectory-level planning rewards to discover reasoning paths with higher planning quality and uses the resulting responses to optimize the policy through GRPO, thereby improving trajectory planning performance. Extensive experiments on both open-loop (nuScenes) and closed-loop-oriented (NAVSIM) benchmarks demonstrate that FactorDrive achieves state-of-the-art planning performance.

cs.RO↗

arXiv2026

PADFormer: Pose-agnostic Anomaly Detection from Sparse View Images

Pose-agnostic Anomaly Detection (PAD) remains challenging as anomalies can appear under arbitrary viewpoints, requiring methods to handle significant pose variations. Existing approaches rely on complex 3D reconstruction, which are computationally expensive and require extensive multi-view data. We propose PADFormer, a novel image-space approach that leverages Vision Transformer (ViT) to directly reconstruct anomaly-free versions of query images while preserving pose information. Our key insight is to adapt cross-view masked reconstruction for anomaly detection through training exclusively on normal data, combined with dynamic patch selection and spatial alignment mechanisms that enable effective learning from sparse reference views under significant pose variations. During inference, we perform multiple forward passes with different masking patterns to generate an ensemble of anomaly-free reconstructions, ensuring comprehensive coverage of the query image. Anomalies are detected by comparing these reconstructions with the query image. PADFormer achieves state-of-the-art results on the PAD benchmark while maintaining comparable performance on classic few-shot anomaly detection (FSAD) tasks, demonstrating superior efficiency and generalization without requiring 3D reconstruction.

cs.CV↗

arXiv2026

Scoring Rules! Statistical and Strategic Alignment for Text Evaluation Metrics

Reference-based text evaluation metrics, which are widely used to assess natural language generation systems, score a candidate response by comparing it with a reference response. The reliability of an evaluation metric is usually judged by its statistical correlation with human ratings. However, as these metrics are increasingly used as optimization objectives, correlation alone is no longer sufficient: agents may strategically game the evaluation metric. We study this issue through two complementary notions of alignment. A metric is statistically aligned if it correlates with human ratings and strategically aligned if it resists perturbations that do not add task-relevant information. We make two contributions. First, we propose test principles for reference-based metrics consisting of human-rating correlation, degradation sensitivity, and manipulation robustness. These principles evaluate whether a metric agrees with human judgments, penalizes low-effort information loss, and resists strategic score inflation. Second, we develop a unified design framework for mutual-information-based metrics that decomposes existing and new metrics into four choices: information measure, estimation method, text representation, and prediction mechanism. Across peer review, summarization, and question answering, we find that strong human-rating correlation does not imply strategic alignment: LLM-as-a-Judge achieves high correlation but is susceptible to manipulation. In contrast, mutual-information-based metrics substantially improve manipulation robustness. Our framework also uncovers a new metric that achieves the strongest overall robustness in our experiments while remaining competitive on human-rating correlation.

cs.AI↗

arXiv2026

Scale Up Strategically: Learning Compositional Generalization via Bias-Aware Evaluation and Data Collection for Robotic Manipulation

Compositional generalization is essential for robot to follow diverse instructions. However, pretrained policies are known to take shortcuts, deferring to salient cues rather than grounding language. We introduce a diagnostic framework that localizes this failure to individual \textit{instruction factors}, \textit{e.g.,} reusable semantic components such as color, verb, object, size, and spatial attribute. Our framework formalizes instruction factor bias, the tendency of fine-tuned policies to over-rely on dominant factors as shortcuts, and quantifies it through two metrics: Factor Dominance Rate (FDR), capturing pairwise bias between factors, and Factor Dominance Hierarchy (FDH), aggregating these into a global ranking. Evaluation on six foundation policies reveals broadly consistent ordering, \textit{i.e.}, color $\geq$ object $\geq$ spatial $\geq$ verb $\geq$ size, with color dominant, and verb and size most under-grounded. We further show the diagnosis is actionable: a bias-aware data collection strategy that reallocates a fixed budget toward under-grounded factors outperforms baselines in simulation and on a real robot using half the demonstrations, thereby enabling more sample-efficient and generalizable policy learning.

cs.RO↗

arXiv2026

SpanUQ: Span-Level Uncertainty Quantification for Large Language Model Generation

Uncertainty estimation is essential not only for the trustworthy deployment of large language models (LLMs) but also as a foundation for self-refinement in LLM generation. However, existing approaches operate at suboptimal granularities: token-level scores lack semantic coherence, while sequence-level scores fail to localize errors. We formalize Span-Level Uncertainty Estimation (SLUE), a new task that targets the natural granularity for uncertainty: semantically coherent text spans, each conveying a single assessable unit of meaning. To address this task, we introduce SPANUQ, a lightweight probe that distills the uncertainty knowledge from expensive multi-sample inference into a single forward pass over LLM hidden states. SPANUQ employs a DETR-style span decoder to simultaneously detect spans and estimate their uncertainty via a Mixture of Beta distribution, trained with a principled combination of Beta NLL regression and contrastive ranking objectives. We construct SPANUQ-BENCH, the first span-level uncertainty benchmark comprising 20K prompts, 293K annotated spans, and continuous soft labels derived from multi-sample claim verification. Experiments on five LLM backbones show that SPANUQ consistently achieves the best span-level uncertainty quality, outperforming the strongest probe baseline and all sampling-based methods while being 10-20x faster. Its DETR-based span detector attains 0.910 F1, surpassing the best heuristic by 39.4%, enabling precise error localization that sequence-level methods cannot provide. The framework generalizes across five LLMs spanning two model families.

cs.CL↗

arXiv2026

SENTINEL: Failure-Driven Reinforcement Learning for Training Tool-Using Language Model Agents

Language model agents are increasingly effective in solving realistic tasks through multi-turn tool use. However, training reliable tool-using agents remains challenging in practice. While reinforcement learning provides an on-policy paradigm for improving agents from their own environment interactions, its effectiveness depends heavily on the training task distribution. When tasks are fixed before training, the task distribution can become increasingly mismatched with the policy's evolving capabilities, causing many rollouts to be spent on uninformative tasks. We propose SENTINEL, a failure-driven reinforcement learning framework that turns the Solver's rollout failures into targeted training tasks. SENTINEL follows a Controller--Proposer--Solver loop: the Controller analyzes failed trajectories and summarizes recurring error patterns, the Proposer generates executable tasks that stress these weaknesses, and the Solver is trained on the targeted tasks. On Tau2-Bench Retail with Qwen3-4B-Thinking-2507, SENTINEL improves Pass\^{}1 from 66.4 to 74.9 and outperforms RL on general synthetic tasks across Pass\^{}k metrics. These results demonstrate that model failures provide an effective and scalable source of targeted training signal for improving tool-using language model agents.

cs.CL↗

arXiv2026

Humans' ALMANAC: A Human Collaboration Dataset of Action-Level Mental Model Annotations for Agent Collaboration

Recent advances in LLM agents have enabled complex cognitive capabilities, such as multi-step reasoning, planning, and tool use, that increasingly position these agents as human collaborators. Effective collaboration, however, requires collaborators to continuously maintain and align mental models of their own reasoning,partners' intentions, and shared goals during the collaborative process. Today's agents rarely develop such capabilities since they are primarily optimized for task completion, and the community lacks authentic human collaboration data with action-level mental model annotations that could guide agents toward process-level collaborative competence. To bridge this gap, we present ALMANAC, a dataset of Action-Level Mental model ANnotations for Agent Collaboration built from the Map Task, a classic dyadic routing task from social science. ALMANAC contains 2,987 collaboration actions, each paired with theory-informed mental model annotations that record the participants' self-reasoning, perceived partner intent, and perceived team goal. We benchmark six LLMs on predicting humans' next-turn behavior and mental models. Our results demonstrate ALMANAC's utility in evaluating models' ability to simulate human collaborative behaviors and infer their underlying mental models.

cs.AI↗

arXiv2026

CollabSim: A CSCW-Grounded Methodology for Investigating Collaborative Competence of LLM Agents through Controlled Multi-Agent Experiments

Multi-agent systems (MAS) built on large language models have shown growing promise, with their effectiveness resting on agents' ability to coordinate through text-based channels much as human teams do. Yet recent study suggests that MAS often falter not because agents lack individual task-solving ability, but because they lack collaborative competence: the capacity to establish common ground, maintain shared task understanding, balance individual and collective incentives, and repair misalignment as interaction unfolds. Decades of research in Computer-Supported Cooperative Work have characterized these requirements for human teams coordinating under constrained communication, yet existing MAS evaluations focus mainly on task outcomes or single-agent proficiency in reasoning, planning, and tool use. To enable a systematic analysis of agents' collaborative competence in MAS, we introduce CollabSim, a configurable simulation framework that combines a theory-grounded definition of collaborative capabilities, controlled manipulation of interaction conditions, and action-level probing of agents' internal states. Experiments across four LLMs show that CollabSim can capture condition effects, separate model performance patterns, and reveal task-dependent effects of agent design.

cs.CL↗

arXiv2026

Firefly: Illuminating Large-Scale Verified Tool-Call Data Generation from Real APIs

Training tool-calling agents requires large-scale trajectory data with verifiable labels, yet existing approaches either synthesize environments that diverge from real API behavior or generate tasks without ground-truth outcomes for verification. We present FireFly, a pipeline for generating verified tool-call data from real-world MCP servers. Our key insight is to invert the standard synthesis pipeline: rather than generating tasks and hoping they are solvable, we first let a strong LLM explore real APIs along graph-guided DAG structures, then synthesize tasks backward from observed outcomes, guaranteeing label correctness by construction. To handle the scale of real-world tool spaces (${\sim}$1,000 tools), we build a pairwise tool graph and sample sub-DAGs to focus exploration on semantically coherent workflows. To address environment drift in live APIs, we construct a retrieval-augmented simulator that caches all exploration results and replays them during training and evaluation, enabling fully offline and reproducible RL. Applying this pipeline yields 5,144 verified tasks spanning 240 servers and 993 tools. A 4B-parameter model trained with GRPO on FireFly matches Claude Sonnet 4.6 on our held-out test set and shows improvements on multiple tool-calling benchmarks including Tau2-Bench, MCPMark, and MCP-Atlas.

cs.SE↗

arXiv2026

Jailbreaking LLMs via Calibration

Safety alignment in Large Language Models (LLMs) often creates a systematic discrepancy between a model's aligned output and the underlying pre-aligned data distribution. We propose a framework in which the effect of safety alignment on next-token prediction is modeled as a systematic distortion of a pre-alignment distribution. We cast Weak-to-Strong Jailbreaking as a forecast aggregation problem and derive an optimal aggregation strategy characterized by a Gradient Shift in the loss-induced dual space. We show that logit-arithmetic jailbreaking methods are a special case of this framework under cross-entropy loss, and derive a broader family of aggregation rules corresponding to other proper losses. We also propose a new hybrid aggregation rule. Evaluations across red-teaming benchmarks and math utility tasks using frontier models demonstrate that our approach achieves superior Attack Success Rates and lower "Jailbreak Tax" compared with existing methods, especially on the safety-hardened gpt-oss-120b.

cs.CL↗

arXiv2026

Trajectory2Task: Training Robust Tool-Calling Agents with Synthesized Yet Verifiable Data for Complex User Intents

Tool-calling agents are increasingly deployed in real-world customer-facing workflows. Yet most studies on tool-calling agents focus on idealized settings with general, fixed, and well-specified tasks. In real-world applications, user requests are often (1) ambiguous, (2) changing over time, or (3) infeasible due to policy constraints, and training and evaluation data that cover these diverse, complex interaction patterns remain under-represented. To bridge the gap, we present Trajectory2Task, a verifiable data generation pipeline for studying tool use at scale under three realistic user scenarios: ambiguous intent, changing intent, and infeasible intents. The pipeline first conducts multi-turn exploration to produce valid tool-call trajectories. It then converts these trajectories into user-facing tasks with controlled intent adaptations. This process yields verifiable task that support closed-loop evaluation and training. We benchmark seven state-of-the-art LLMs on the generated complex user scenario tasks and observe frequent failures. Finally, using successful trajectories obtained from task rollouts, we fine-tune lightweight LLMs and find consistent improvements across all three conditions, along with better generalization to unseen tool-use domains, indicating stronger tool-calling ability.

cs.CL↗

arXiv2025

DPRF: A Generalizable Dynamic Persona Refinement Framework for Optimizing Behavior Alignment Between Personalized LLM Role-Playing Agents and Humans

The emerging large language model role-playing agents (LLM RPAs) aim to simulate individual human behaviors, but the persona fidelity is often undermined by manually-created profiles (e.g., cherry-picked information and personality characteristics) without validating the alignment with the target individuals. To address this limitation, our work introduces the Dynamic Persona Refinement Framework (DPRF). DPRF aims to optimize the alignment of LLM RPAs' behaviors with those of target individuals by iteratively identifying the cognitive divergence, either through free-form or theory-grounded, structured analysis, between generated behaviors and human ground truth, and refining the persona profile to mitigate these divergences. We evaluate DPRF with five LLMs on four diverse behavior-prediction scenarios: formal debates, social media posts with mental health issues, public interviews, and movie reviews. DPRF can consistently improve behavioral alignment considerably over baseline personas and generalizes across models and scenarios. Our work provides a robust methodology for creating high-fidelity persona profiles and enhancing the validity of downstream applications, such as user simulation, social studies, and personalized AI.

cs.CL↗

arXiv2025

See, Think, Act: Online Shopper Behavior Simulation with VLM Agents

LLMs have recently demonstrated strong potential in simulating online shopper behavior. Prior work has improved action prediction by applying SFT on action traces with LLM-generated rationales, and by leveraging RL to further enhance reasoning capabilities. Despite these advances, current approaches rely on text-based inputs and overlook the essential role of visual perception in shaping human decision-making during web GUI interactions. In this paper, we investigate the integration of visual information, specifically webpage screenshots, into behavior simulation via VLMs, leveraging OPeRA dataset. By grounding agent decision-making in both textual and visual modalities, we aim to narrow the gap between synthetic agents and real-world users, thereby enabling more cognitively aligned simulations of online shopping behavior. Specifically, we employ SFT for joint action prediction and rationale generation, conditioning on the full interaction context, which comprises action history, past HTML observations, and the current webpage screenshot. To further enhance reasoning capabilities, we integrate RL with a hierarchical reward structure, scaled by a difficulty-aware factor that prioritizes challenging decision points. Empirically, our studies show that incorporating visual grounding yields substantial gains: the combination of text and image inputs improves exact match accuracy by more than 6% over text-only inputs. These results indicate that multi-modal grounding not only boosts predictive accuracy but also enhances simulation fidelity in visually complex environments, which captures nuances of human attention and decision-making that text-only agents often miss. Finally, we revisit the design space of behavior simulation frameworks, identify key methodological limitations, and propose future research directions toward building efficient and effective human behavior simulators.

cs.CY↗

arXiv2025

Customer-R1: Personalized Simulation of Human Behaviors via RL-based LLM Agent in Online Shopping

Simulating step-wise human behavior with Large Language Models (LLMs) has become an emerging research direction, enabling applications in various practical domains. While prior methods, including prompting, supervised fine-tuning (SFT), and reinforcement learning (RL), have shown promise in modeling step-wise behavior, they primarily learn a population-level policy without conditioning on a user's persona, yielding generic rather than personalized simulations. In this work, we pose a critical question: how can LLM agents better simulate personalized user behavior? We introduce Customer-R1, an RL-based method for personalized, step-wise user behavior simulation in online shopping environments. Our policy is conditioned on an explicit persona, and we optimize next-step rationale and action generation via action correctness reward signals. Experiments on the OPeRA dataset emonstrate that Customer-R1 not only significantly outperforms prompting and SFT-based baselines in next-action prediction tasks, but also better matches users' action distribution, indicating higher fidelity in personalized behavior simulation.

cs.CL↗

arXiv2025

WEBSERV: A Full-Stack and RL-Ready Web Environment for Training Web Agents at Scale

Reinforcement learning (RL) for web agents demands environments that are both effective for evaluation and efficient enough for large-scale on-policy training. Current web environments fall short: server-side Docker setups are too resource-intensive for massive parallel rollouts, while browser-side interfaces produce noisy observations, execute actions unreliably under modern single-page applications, and omit visual interactivity cues. We introduce WebServ, a full-stack, RL-ready web environment that addresses these limitations end-to-end. On the server side, WebServ uses Incus containers with block-level copy-on-write, reducing launch latency by ~5x and persistent storage by ~240x, enabling 200+ concurrent isolated environments on a single host. On the browser side, WebServ provides a compact, site-agnostic observation and action interface derived automatically from the DOM with human-aligned interactivity cues, and a robust action execution backend using network-aware waiting for reliable SPA support. On WebArena-Lite, WebServ achieves state-of-the-art single-prompt results, with controlled comparisons confirming consistent gains across GPT-4o, OpenAI-o3, and Llama-3.1-8B over vanilla WebArena. We further train Qwen3-4B and Qwen3-30B-A3B with RL entirely within WebServ; the RL-trained 4B model achieves 55.5% mean accuracy, surpassing both Claude 4.5 Sonnet (50.0%) and the RL-trained 8B model from WebAgent-R1 (51.8%).

cs.LG↗

arXiv2025

Truthful Calibration Errors for Multi-Class Prediction

Calibrated predictions are useful because their numerical values can be interpreted as probabilities. Calibration errors are therefore widely used to evaluate, compare, and tune probabilistic predictors. Recently, Haghtalab et al. (2024) introduced an additional requirement for such measures: truthfulness. A calibration measure is truthful if a predictor minimizes its expected measured error by reporting the true conditional label distribution. Many standard empirical calibration errors are non-truthful: a predictor may appear better calibrated by distorting its probabilities rather than reporting them truthfully. We study the practical role of truthfulness for calibration measurement in multiclass prediction. First, we introduce perfectly truthful calibration errors for multidimensional linear properties of the label distribution, generalizing the truthful calibration error for binary predictions in Hartline et al. (2025). This framework includes full multiclass calibration and classwise calibration. We also identify a truthful correction for confidence calibration. Second, we characterize the decision-theoretic implications of these truthful errors. For calibrated predictors, truthful calibration errors preserve the Blackwell dominance: a more informative calibrated predictor receives no larger expected error. Third, we show that this decision-theoretic interpretation explains and mitigates the well-observed ranking robustness problem of binned calibration errors. Empirically, non-truthful confidence-based errors can reverse model rankings when the number of bins changes, while our truthful errors give more stable rankings across binning choices.

cs.LG↗

arXiv2025

LLaVAShield: Safeguarding Multimodal Multi-Turn Dialogues in Vision-Language Models

As Vision-Language Models (VLMs) move into interactive, multi-turn use, safety concerns intensify for multimodal multi-turn dialogue, which is characterized by concealment of malicious intent, contextual risk accumulation, and cross-modal joint risk. These characteristics limit the effectiveness of content moderation approaches designed for single-turn or single-modality settings. To address these limitations, we first construct the Multimodal Multi-turn Dialogue Safety (MMDS) dataset, comprising 4,484 annotated dialogues and a comprehensive risk taxonomy with 8 primary and 60 subdimensions. As part of MMDS construction, we introduce Multimodal Multi-turn Red Teaming (MMRT), an automated framework for generating unsafe multimodal multi-turn dialogues. We further propose LLaVAShield, which audits the safety of both user inputs and assistant responses under specified policy dimensions in multimodal multi-turn dialogues. Extensive experiments show that LLaVAShield significantly outperforms state-of-the-art VLMs and existing content moderation tools while demonstrating strong generalization and flexible policy adaptation. Additionally, we analyze vulnerabilities of mainstream VLMs to harmful inputs and evaluate the contribution of key components, advancing understanding of safety mechanisms in multimodal multi-turn dialogues.

cs.CV↗