Tag
该研究提出面向CEFR等级（A1/A2/B1）的俄语文本简化方法，通过拒绝采样和DPO微调YandexGPT，最佳模型达到87.15%的目标等级准确率和0.35 CEFR等级的MAE，模型已在Huggingface开源。
This paper proves lower bounds for smooth nonconvex zeroth-order optimization, showing that deterministic methods incur an extra factor of roughly d in query complexity compared to randomized methods under low curvature trace conditions, quantifying the price of determinism for tasks like LLM fine-tuning with function-value-only queries.
ToPA 提出了一种基于分块Toeplitz矩阵的参数高效微调方法，避免了低秩和稀疏结构的限制，在21个NLP和CV数据集、5种模型架构上验证了其表达力与训练稳定性。
该论文提出 FlowSpec 规范驱动的方法，微调一个 90 亿参数模型，使语音客服代理能够遵循从未见过的受监管催收流程，合规检查率从 67–71% 提升至 93–100%。
An explainer comparing SFT, DPO, and RLVR for customizing LLMs, with a hands-on experiment training Qwen3 0.6B on GSM8K using GRPO on Amazon SageMaker AI, achieving a ~2.9 percentage-point accuracy gain, plus guidance on choosing the right supervision method and an AWS serverless model-customization workshop.
ReMaRk 提出通过将生成列表与所有有效回复的标题联合评分来消除 GRPO 训练中的回复噪声，在保持或超越 Rank-GRPO 性能的同时显著降低 LLM 推荐模型的 RL 训练成本。
This paper proposes robust, adaptive Langevin Monte Carlo methods (R-LMC, A-LMC, RA-LMC) that handle generalized smoothness and state-dependent heavy-tailed noise, establishing Fisher information convergence guarantees without dissipativity assumptions, and demonstrates effective Bayesian fine-tuning of LLMs and vision-language models.
Vela提出以真实用户行为数据作为监督信号，将10万用户的匿名行为记录转化为可训练任务与轨迹，用于微调Qwen3.6-35B-A3B，在用户相关任务上最高提升11.5分，通用基准平均提升4分，表明真实用户行为是可扩展的监督来源。
该论文提出“人格层级模型”（Persona Hierarchy Model）来解释LLM微调中的上下文泛化现象：修改共享默认人格的微调会带来更广泛的迁移，而局部人格的改变则更具上下文特异性。作者还提出人格保持正则化（PPR），在RL训练中将奖励黑客行为从42–55%降至最多0.2%，同时保留准确率增益。
This paper shows that when selecting math training data for LLM fine-tuning, sharing a reasoning approach (e.g., invariant reasoning or double counting) yields significantly greater transfer to held-out targets than sharing a mathematical topic, with advantages of 8–16 percentage points across five base models in counterbalanced 2x2 designs.
SFTMill is an open-source tool that enables off-policy distillation of large language models using an OpenAI-compatible endpoint, allowing users to create training datasets from behavioral goals defined in a curriculum.
SlideDP presents an asynchronous data-parallel runtime for host-resident layer streaming that fine-tunes LLMs beyond GPU memory across multiple GPUs, achieving 1.46–2.64x throughput over SlideFormer, MegaTrain, and ZeRO-Offload and even exceeding FSDP2 peak throughput while fine-tuning Qwen2.5-72B on four RTX 4090s.
This paper studies the feedback loop where AI-generated reviews influence future training of AI reviewers, leading to reduced judgment diversity called 'scientific-judgment collapse,' and introduces TrustReviewer, an open-source system to mitigate this through curated training and activation steering.
The paper proposes FLoKD, an adaptive knowledge-distillation framework for federated LoRA fine-tuning of LLMs over wireless networks, reducing communication overhead by 50-65% while maintaining competitive performance.
FedSubMuon is a communication-efficient federated fine-tuning method for LLMs that optimizes compact coefficient matrices within structured subspaces to reduce upload costs while maintaining strong performance.
The paper identifies factual access failures in large language models after supervised fine-tuning and introduces Recall-Anchored Distillation (RAD) to preserve out-of-distribution factual recall without labeled data.
OraclePhys presents a systematic framework for fine-tuning large language models on structural mechanics, featuring a finite-element benchmark and dataset to investigate how training objectives causally determine model internalization.
Polaris trains LLMs to generate table descriptions directly from retrieval feedback using DPO, significantly outperforming previous methods like AutoDDG in table retrieval tasks.
This paper proposes a schema-constrained document-level event argument extraction method using fine-tuned mid-sized open LLMs with LoRA, role-set injection, and deterministic decoding, achieving state-of-the-art results on MAVEN-ARG.
PPO-HSC introduces a High-order Sampling Coverage reward to encourage exploration of diverse reasoning patterns in RL fine-tuning of LLMs, improving solution diversity and state-space coverage on math and code tasks.