Intelligent agent interactions in real-world web environments are commonly constrained by request budgets,time delays, anti-crawling restrictions, and operational failure risks. Strategies solely optimizing task successrates often exhibit unusable phenomena such as "high success but high cost" or "low risk but conservativefailure."This paper proposes a constrained Agentic reinforcement learning model for Web Agents, unifyingpage access, search requests, and external API calls into a unified long-term decision-making framework withassociated costs. It simultaneously incorporates cost budget constraints and tail risk control into theoptimization objective: constructing a multidimensional cost vector comprising cumulative request count,total latency, and failure penalties to achieve budget compliance via Lagrange dual updates;while employing aCVaR risk term to suppress excessive exploration of high-failure-probability paths, thereby achievingadaptive balance among "completion rate, cost, and risk."Experiments were conducted across 30–70site/page templates and 800–1,500 end-to-end web tasks (including information extraction, pricecomparison, form submission, and cross-page navigation). Interaction sequences spanned 20–120 steps withtool scales of 30–200. Performance was benchmarked against unconstrained RL, budget-constrained RL, andrule-based/scripted web agents, quantifying task completion rates, cost-per-success, failure rates, and policystability.scripted web agents. We quantified task completion rates, cost-per-success, failure rates, and policystability.Results demonstrate that at equivalent completion rates, our method reduces C-PS by 22%–31% andlowers failure rates by 18%–26% under high failure penalties. Under fixed budgets, task completion ratesincrease by 10%–16%, highlighting the necessity and effectiveness of constraint modeling for practical WebAgent deployment.

To address the core shortcomings of existing memory reorganization algorithms in distributed change environments, such as trigger lag, performance fluctuations, and memory management imbalances, a Distribution-Aware Continuous Memory Reorganization (DAMCR) algorithm is proposed. This algorithm uses "distribution awareness - dynamic reorganization - performance feedback" as its core logic, employing a four-module architecture. It constructs a distribution change awareness factor by improving KL divergence, achieves adaptive memory unit management based on multi-dimensional value assessment, and combines a buffer mechanism and performance closed-loop feedback to ensure smooth reorganization and stable performance. Experimental scenarios are conducted using gradual and abrupt distribution changes, based on CIFAR-100 and a self-built navigation dataset, and compared with eight mainstream/classic algorithms (ST, AR, DIR, ADWIN, DDM, LRU, Reservoir, PER). Experimental results show that the DAMCR algorithm achieves an average decision accuracy of 92.3%, a maximum improvement of 6.6% compared to the comparison algorithms; the average reorganization latency is 12.7ms, a maximum reduction of 55.1%; and the performance fluctuation amplitude and memory utilization are reduced by a maximum of 73.1% and improved by 23.9% respectively compared to the comparison algorithms. Ablation experiments validated the necessity of each core module of DAMCR, and statistical significance analysis (independent samples t-tests, one-way ANOVA, 95% confidence interval/variance) with Cohen's d effect size confirmed that the performance advantage of DAMCR is statistically significant (p1.2). This algorithm can effectively optimize the memory structure of agents under distributed change environments while maintaining their performance, providing technical support for related engineering applications.

Existing Video-RAG systems often concatenate retrieved segments directly into input, leading toreasoning drift when hard negative samples are introduced. This paper proposes a Structured Post-Retrieval Reasoning (SPRR) module for Large Video Language Models (LVLMs), explicitly modelingthe post-retrieval process into three stages:(1) Evidence Validation: Generates "decidable" sub-problems (3–8) for Top-k=20 candidate clips, outputs binary/numeric scores, and filters to k′=4–6;(2) Conflict Resolution: Establishes consistency constraints (e.g., temporal order, entity attributeinvariance) for contradictory information across multiple clips, selecting the minimum conflictsubset to form a coherent evidence pool;(3) Temporal Aggregation: Indexed by event timestamps,evidence is serialized to generate interpretable reasoning chains (including referenced clip IDs andtemporal ranges).Evaluated on MLVU (3,102 QA) and LongVideoBench (6,678 MCQ) using open-ended and multiple-choice formats respectively, while measuring interpretability metrics (averageevidence count, conflict rate, reasoning chain length) and efficiency metrics (input tokens/reasoningsteps). This validates SPRR's benefits in "reducing noise, enhancing interpretability, and improvingstability.

Multi-agent systems often rely on long-term memory or shared knowledge bases to enhance collaborativeefficiency, yet this introduces risks of memory poisoning and cross-agent propagation. Addressing the covertdiffusion of poisoned information during collaboration, this study proposes a memory poisoning detection andrepair method tailored for multi-agent environments.This approach constructs an evidence graph based onmemory source credibility and content consistency to validate newly added memories. It combines contrastivelearning models to identify anomalous memories exhibiting command-induced characteristics. Upon detectingpoisoning, further propagation is suppressed through isolation, rewriting, and conflict resolution. Experimentsevaluated the method using 60 collaborative tasks, approximately 210,000 memory records, and 12,000injected poisoned samples.Results demonstrate an AUC of 0.94 in poisoning detection, reducing misbehaviorrates from 15.6% to 2.3% while decreasing cross-agent propagation by 78.1% on average, with minimal impacton overall task efficiency.

Hallucinations in large language models (LLMs) remain a critical challenge for deploying these systems in high-stakes domains. We introduce Calib-Abstain, a framework that combines token-level uncertainty estimation with multi-path consistency checking to generate calibrated confidence scores. When confidence falls below a learned threshold, the system triggers tiered mitigation actions: retrieval augmentation, query reformulation, or abstention with explanation. Our approach employs temperature scaling and isotonic regression to minimize Expected Calibration Error (ECE). Experiments on mathematical reasoning (GSM8K, MATH), code generation (HumanEval), and general knowledge tasks (MMLU) demonstrate that Calib-Abstain reduces hallucination rates among answered queries by 24--31% while maintaining 80--88% coverage, with ECE improvements of 0.08--0.12 across datasets.