Tag
该论文提出SHIP（Structured and Hierarchical Insight with Privacy），将差分隐私洞察形式化为可复用的查询回答表示，通过自动构建层级混合分类法并发布DP合成表，实现隐私保护下的文本数据分析，性能优于现有私有洞察基线。
The paper introduces TerminalHorizon, an automated data-synthesis engine that generates long-horizon terminal agent trajectories by reconstructing environments around real-world professional tasks, then fine-tunes Qwen3.5-27B to obtain TerminalHorizon-27B, achieving gains across 12 coding, agentic, and reasoning benchmarks.
本文提出 MagicBeanStalk，一种自顶向下分解与自底向上递归结合的 RLVR 数据合成算法，用以逃逸约88%数学语料落入的“零优势陷阱”，在多个数学与金融基准上显著提升 gpt-oss-120b、DeepSeek v4 Flash 和 Qwen3 的推理性能。
A reverse-engineering repo (REA) could let AI agents systematically decompile and reconstruct closed-source software, potentially feeding AI labs like DeepSeek and Z ai with vast datasets of real engineering problems and accelerating coding capabilities across the industry.
SynCo 提出了一种基于多智能体强化学习的数据合成协同训练框架，联合优化 Synthesizer 与 Reasoner 两个智能体，使训练数据分布随模型能力共同演化，并在八个数学推理基准上取得了最强的整体表现。
SAP Labs introduces Synthesis Through Simulation (STS), a schema-free data synthesis paradigm where an LLM agent generates enterprise data by executing operations against policy-enforcing APIs in simulated environments, achieving 0.88 average marginal fidelity and 100% constraint satisfaction across ten environments without DB schema access. The full framework and datasets are open-sourced.
提出了 TerminalHorizon，一个用于长程终端智能体的全自动数据合成引擎，构建 1.5K 任务环境并对 Qwen3.5-27B 进行监督微调，在 12 项基准上取得显著提升。
Eon's Era generates a complete simulated enterprise with exact ground-truth answers, writing records into Salesforce, Zendesk, Jira, Slack and Gong via MCP and REST interfaces, enabling deterministic evaluation of enterprise agents without human or LLM judges.
The paper proposes Recursive Self-Improvement (RSI), a training loop in which a unified multimodal model's text and visual abilities generate training data for each other, with program execution providing verified supervision. Evaluated on charts with a new BasicChartBench benchmark, four rounds of RSI improve performance from 45.7% to 60.2% while verified-program share rises from 48.9% to 95.2%.
PADMÉ reformulates meta-evaluation of LM agent evaluators as a preference alignment problem and introduces a data synthesis method using only small language models, boosting agreement with human judgment from 73% to 85% over a naive baseline while enabling evaluation of 25 common models across agentic domains.
AutoDataBench introduces a benchmark to evaluate if agents can write tasks for data pipelines that meet practical acceptance standards, aiming to enable scalable data synthesis for recursive self-improvement in AI.
The paper presents a systematic approach to tag-aware structured text translation by addressing data synthesis, capability building, and alignment, demonstrating significant improvements over existing methods.
提出了一种名为 Synthesis Through Simulation (STS) 的无模式数据合成范式，通过让 LLM 智能体在模拟企业环境中调用策略强制 API 来生成结构上天然有效的数据，其 Generalist Populator 智能体在十个环境中达到 0.88 的平均分布保真度和 100% 的约束满足率，且无需访问数据库 schema。
The paper proposes a method for proactively detecting implicit conflicts in user-side human-LLM dialogue, introducing a benchmark and synthesis approach to enhance lightweight LLMs' performance.
This paper introduces OmniVChat, a task for native audio-visual dialogue, and presents a data engine, benchmark, and reinforcement learning method to train and evaluate omni models, demonstrating improvements on synthesized and human-recorded data.
This paper introduces the Reflective Cognitive Alignment (RCA) framework to improve LLM-based agents for cognitive stimulation therapy in elderly care, focusing on protocol adherence and safety through structured reasoning and value alignment.
MIMIC is a framework that uses executable code to synthesize reasoning trajectories for LLMs, enhancing their deterministic reasoning through code-instrumented rewards and achieving improved performance on reasoning benchmarks.
This paper introduces MIFS, a pipeline for synthesizing RL-ready multimodal data to enhance instruction following in MLLMs, achieving an 8.13% average improvement on benchmarks and faster training convergence.
The paper proposes State-Guided Data Synthesis with Argument Provenance (SAP) to improve the synthesis of multi-turn tool-use data for training agentic models, and introduces SAP-4B, which performs competitively with larger models.
This paper demonstrates that causal fairness mechanisms, specifically edge cuts on causal graphs, are portable across various synthetic data generator families including GANs and diffusion models, with minimal impact on data fidelity and utility.