Retrieval-Augmented Generation (RAG) im- proves the factuality of large language models (LLMs), yet existing RAG systems often struggle with complex, multi-step reasoning that requires adaptive retrieval and continuous revision of intermediate contexts. Recent reinforcement learning (RL)-based agentic RAG methods partially alleviate this issue, but typically rely on coarse-grained action spaces and trajectory-level rewards, resulting in weak reward assignment and a bias toward short-horizon, stereotyped reasoning template. To address, we propose AGENTICRAG-R1, a RL framework that deeply integrates reasoning, retrieval, and memory via a memory stack and fine-grained action space, supported by hierarchical action-aware rewards and an information-aware trajectory rejection strategy to enable effective long-horizon learning. Experiments across a diverse set of multi-hop, open-domain, and agentic reasoning benchmarks, spanning multiple backbone model sizes, demonstrate that AGENTICRAG- R1 consistently outperforms strong baselines. Moreover, AGENTICRAG-R1 learns more robust, interpretable, and memory-aware reasoning behaviors, highlighting the effect of fine-grained action modeling and information- aware optimization for long-horizon reasoning. Our code is anonymous available at https: //github.com/jiangxinke/Harness-RL/ tree/AgenticRAG-R1-Whitebox.

As training regimes centered on human-curated data encounter rising costs and an increasingly binding scaling ceiling, a core question in agent learning shifts from enlarging exogenous datasets to enabling agents to **self-evolve** under limited exogenous supervision. Prior surveys on self-evolution and self-improvement typically organize the literature by evolutionary stages or system architectures. While effective for summarizing implementation pipelines, these perspectives do not offer a unified explanation of what an agent actually observes during self-improvement, nor why such signals can serve as sufficient and reliable supervision. To address this gap, we introduce **endogenous observations**, defined as learning signals that arise from an agent’s own generation, evaluation, and interaction processes rather than exogenous labels. We systematize existing approaches into three categories—model-internal, peer-induced, and scenario-grounded observations—according to how these signals are produced, and we summarize the characteristic properties and typical application settings of each category. Finally, we discuss key challenges in reliability, scalability, and long-horizon stability, and outline open directions toward robust and sustainable self-evolving agents.

Open-ended deep research (OEDR) requires systems to acquire knowledge through multi-round retrieval and generate coherent long-form reports. The outline plays a central role as a structural scaffold that coordinates retrieval, evidence organization, and generation. However, existing methods either fix the outline before writing or refine it with local heuristics, leading to scaffold drift under continuous information accumulation and delayed feedback for evaluating outline modifications. We propose ScaffoldAgent, a utility-guided dynamic outline optimization framework for OEDR. ScaffoldAgent models outline evolution as a structured decision process with three operations: Expansion, Contraction, and Revision, enabling controlled updates to the report scaffold. It further introduces a utility-guided feedback mechanism that estimates the downstream value of each outline operation from retrieval gain, structural coherence, and trial-generation quality. The resulting utility signal guides node selection, operation scheduling, and termination during inference. Experiments on DeepResearch Bench and DeepResearch Gym show that ScaffoldAgent consistently improves long-form report generation and factual grounding over existing deep research agents.

Schema linking is a crucial step in Text-to-SQL systems, connecting natural language mentions to relevant database tables and columns. Despite recent advances in LLM-based Text-to-SQL models, current approaches face two fundamental limitations: (1) providing detailed column descriptions inflates the input context, often exceeding the model’s effective sequence length, and (2) predicting multi-token column names increases output complexity, making the task fragile to generation errors. To address these challenges, we propose SchemaToken, which represents each database column as a dedicated token in the LLM’s vocabulary. This design compresses column semantics into single embeddings, reducing context length and simplifying prediction. We introduce a two-stage training paradigm: (i) Semantic Alignment, which enables the LLM to understand the newly added schema tokens via structured embeddings, and (ii) Pseudo-task Alignment, which trains the LLM to generate these tokens as an unordered set using reinforcement learning with a Jaccard-based reward. Experiments on multiple benchmark datasets demonstrate that SchemaToken significantly improves schema linking performance, particularly for large and complex schemas, while maintaining computational efficiency. Our code is publicly available at https://github.com/MrBlankness/SchemaToken.

For program languages, testing plays a crucial role in the software development cycle, enabling the detection of bugs, vulnerabilities, and other undesirable behaviors. To perform software testing, testers need to write code snippets that execute the program under test. Recently, researchers have recognized the potential of large language models (LLMs) in software testing. However, there remains a lack of fair comparisons between different LLMs in terms of test case generation capabilities.In this paper, we propose TestEval, a novel benchmark for test case generation with LLMs. We collect 210 Python programs from an online programming platform, LeetCode, and design three different tasks: overall coverage, targeted line/branch coverage, and targeted path coverage. We further evaluate 17 popular LLMs, including both commercial and open-source ones, on TestEval. We find that generating test cases to cover specific program lines/branches/paths is still challenging for current LLMs, indicating a lack of ability to comprehend program logic and execution paths.

As safety remains a crucial concern throughout the development lifecycle of Large Language Models (LLMs), researchers and industrial practitioners have increasingly focused on safeguarding and aligning LLM behaviors with human preferences and ethical standards. LLMs, trained on extensive multilingual corpora, exhibit powerful generalization abilities across diverse languages and domains. However, current safety alignment practices predominantly focus on single-language scenarios, which leaves their effectiveness in complex multilingual contexts, especially for those complex mixed-language formats, largely unexplored. In this study, we introduce Multilingual Blending, a mixed-language query-response scheme designed to evaluate the safety alignment of various state-of-the-art LLMs (e.g., GPT-4o, GPT 3.5, Llama3) under sophisticated, multilingual conditions. We further investigate language patterns such as language availability, morphology, and language family that could impact the effectiveness of Multilingual Blending in compromising the safeguards of LLMs. Our experimental results show that, without meticulously crafted prompt templates, Multilingual Blending significantly amplifies the detriment of malicious queries, leading to dramatically increased bypass rates in LLM safety alignment (67.23% on GPT-3.5 and 40.34% on GPT-4o), far exceeding those of single-language baselines. Moreover, the performance of Multilingual Blending varies notably based on intrinsic linguistic properties, with languages of different morphology and from diverse families being more prone to evading safety alignments. These findings underscore the necessity of evaluating LLMs and developing corresponding safety alignment strategies in a complex, multilingual context to align with their superior cross-language generalization capabilities.