A new arXiv preprint explores whether midtraining, a stage between pretraining and post-training, can influence which behaviors a large language model carries forward. The authors propose a method called Inoculation Midtraining, which uses invented words to shape how desirable and undesirable properties generalize. The work is a research contribution and has not been peer reviewed.
A new arXiv preprint investigates how much of the benefit from on-policy distillation actually comes from the training data used. Testing the two teacher-student pairings most often seen in practice, the authors report findings that challenge assumptions about data's role in the method. The work is positioned as a closer look at a technique that has become standard in frontier post-training pipelines.
A new arXiv paper analyzes the actor-only variants of Proximal Policy Optimization that are commonly used to post-train large language models. The authors derive non-asymptotic global convergence guarantees for the clipped PPO objective, addressing how the clipping mechanism affects optimization. The work offers theoretical grounding for a method widely deployed in practice.
A new arXiv paper argues that training prompts in online reinforcement learning vary widely in how useful they are to the current policy, with some already solved and others too hard to give a dependable learning signal. The authors propose an exploration-guided prompt scaffolding method that selects or structures prompts for multimodal reinforcement post-training so rollouts are better spent. The work appears in both the cs.AI and cs.LG listings as arXiv:2609.15051v1.
A new arXiv preprint introduces a post-training method that uses solver feedback to guide self-distillation, aiming to improve how language models turn natural-language problem descriptions into operations research formulations. The approach is positioned as a way to go beyond training on verified answers alone when bootstrapping such models.
A new arXiv preprint describes a test-time adaptation technique for improving large language model reasoning on downstream tasks without expensive post-training. The method builds on predictive entropy as a model-derived signal but adds a stability-aware component to guide adaptation. The abstract presents the approach as a lightweight alternative to retraining or fine-tuning.
A new arXiv preprint addresses how to choose hyperparameters after a model has already been trained, such as inference-time settings and implementation-level options. The work aims to move this process beyond ad hoc tuning by providing statistical validity guarantees for the selected configuration. It targets deployment scenarios where pre-trained models still have multiple degrees of freedom that need to be fixed.
A new research paper tackles the shortage of training data suitable for reasoning-focused fine-tuning in the financial domain, noting that most available QA pairs lack explicit reasoning steps, sufficient context, or reliably checkable answers. The authors present a pipeline that mines financial text, distills it into reasoning-oriented training examples, and applies learning with verifiable answers to improve model performance on financial tasks.