supervised-fine-tuning

The paper proposes Stepwise-Querying, a method that elicits step-specific explanations from black-box teachers without accessing their internal chain-of-thought traces, assembling them into training data for supervised fine-tuning of smaller student models. It achieves the top score in 14 of 16 comparisons across two teachers, two students, and four math/code benchmarks, and introduces the SEIG diagnostic for measuring stepwise explanation information gain.

Hugging Face published a 5-step curriculum for post-training AI agents, covering SFT, distillation, GRPO-based RL, RL environments, and harness-scale training with livestreams, code, and docs for each step.

DynaFlow introduces a framework that lets a main agent synthesize executable workflows to coordinate sub-agents adaptively, achieving 88.08% accuracy and a 10.90-point gain over baselines with a Qwen3.5-27B backbone across six benchmarks.

该研究系统考察了在预训练与后训练阶段切换 AdamW 和 Muon 优化器的代价，发现最终性能主要由两个优化器各自独立的效果解释，切换优化器几乎没有额外的性能惩罚，后训练优化器的选择可基本独立于预训练优化器。

CoworkEnv 提出一个能力引导框架，为 AI 同事智能体合成可执行的训练环境，通过工作流引导的任务构建与程序化验证，支持 SFT 与 RL 训练紧凑的 PocketWorker 模型，在 Claw-Eval 多轮任务上取得显著提升。

Xiaohongshu (RedNote) 研究团队发布 Vela 框架，将真实用户行为记录转化为训练任务与轨迹，通过监督微调和强化学习提升 LLM 在用户相关任务（+11.5 分）及通用基准（平均 +4 分）上的表现。

GRADS 提出一种图引导的智能体数据合成框架，通过维护证据与推理依赖图实现可控难度的多跳任务生成；经 SFT 与 RL 后训练后，4B 模型在软件工程、科学与安全领域基准上匹配或超越 Sonnet 4。

该论文研究了工具使用型语言智能体在监督微调下的「选择性撤回」行为，提出一种能发现聚合准确率所掩盖的规则级回归的诊断方法，并表明微调本身并不足以保证智能体安全。

提出「在前沿LLM上迭代开发行为，再蒸馏到小模型（SLM）上线」的两阶段方法：用LoRA微调的Qwen3-8B在家庭护理排班智能体的生产决策上达到0.913一致率，优于教师模型GPT-5.4-nano自身重跑时72.7%的自我一致性，且延迟降低约2.5倍、成本大幅下降。论文已被NeurIPS 2026第二届企业任务智能体基准与应用研讨会（AABA4ET）接收。

AutoSFT 提出一个多智能体系统，将 SFT 数据研究中的执行、分析与合成分离，自动执行多样化的数据操作，相比单智能体在 IFEval、IFBench 和 MBPP+ 上取得显著提升。研究发现自动化 SFT 数据研究的关键在于覆盖多样化的数据操作，而非强化单一操作。

The paper proposes Agent-Reflex, a three-stage training framework (Reflection Skill Discovery, Reflective SFT, and On-Policy Reflective Policy Optimization) that teaches language agents to proactively reflect under noisy tool feedback and user interventions, achieving state-of-the-art results on long-horizon tool-use benchmarks with Qwen3-8B.

本文提出 CIG-SFT，一个信息论框架，通过估计每个示范 token 的任务条件信息增益来加权监督信号，从而在数学推理与代码生成基准上提升 SFT 的泛化能力，且无需额外数据或在线采样。

论文提出 TQS（Trajectory Quality Score），一种基于轨迹结构信号（信息聚焦、意图-行为对齐、策略动态）的线性时间评分框架，用于筛选高质量编码智能体轨迹进行监督微调。实验表明在 TQS 排名靠前的轨迹上训练的模型在 SWE-Bench Verified 上优于随机或低质量子集训练的结果。

该论文提出一种不确定性引导的渐进收缩框架，通过构建蒙特卡洛推理树评估过程级不确定性，并在SFT阶段进行在线路由，仅使用全量SFT约0.22%的优化预算（减少450倍以上的训练暴露）即可维持或提升LLM的推理能力。

论文提出 Adaptive Mutual Distillation (AMD) 框架，通过两个采用不同任务平衡策略的模型进行协同蒸馏，并利用短训练探针和任务级验证分数自适应调整各任务与传输方向的蒸馏权重，在六个基准和三个 LLM 骨干上取得优于 SFT 与现有任务平衡方法的效果，合并后平均提升 2.91 分。

AutoCompact trains coding agents to decide when to compact context, what working state to preserve, and how to resume, using judge-corrected trajectories for SFT followed by RL with task-success rewards, yielding +9.2% on SWE-bench Verified and +5.0% on SWE-PolyBench Verified.

Maxime Labonne shares Aayush Karan's research arguing SFT generalizes worse than RL because of off-policy data rather than the objective itself — rewriting expert trajectories to match the base model's own outputs can rival RL and on-policy distillation while forgetting less.

Authors argue supervised fine-tuning (SFT) is not obsolete, presenting a method that brings SFT on par with prevailing posttraining approaches like RL and OPSD, often generalizing better and forgetting less, by introducing sampling into the posttraining stack.

GraphForge introduces an evidence-graph framework that grounds both agent task generation and verification in real files, enabling synthesis of verifiable training trajectories for working agents. Fine-tuning Qwen3.6-27B on 2,169 GraphForge trajectories improves GDPVal, Workspace-Bench-Lite, and SpreadsheetBench II, with rejection fine-tuning yielding further gains.

The paper introduces GoldiMask, a method for supervised fine-tuning of diffusion language models that selects context tokens via submodular optimization and weights prediction targets based on context benefit, achieving improved accuracy on reasoning and code generation benchmarks across multiple backbones.