post-training

Deep Principle released Hydrogen-1.0, an open-weight model post-trained from GLM-5.3 for chemistry, materials science, and life sciences, showcased via a short film of it embedded in a daily research workflow alongside their Mira AI Scientist assistant.

Trajectory raised $45M in a Series A led by Sequoia at a $300M valuation to make continual learning and post-training of open source AI models accessible to nontechnical domain experts, with customers including Harvey, Rogo, and Nvidia.

The paper introduces Noura, a framework that turns executable personal clinical specifications into supervision for LLM-based diet planning, enabling compact models like Qwen3-14B to dramatically improve multi-day meal plan compliance and reduce rule violations compared to direct generation by larger models.

This paper explores using steering vectors as a training signal during LLM post-training, showing that injecting them into hidden states improves SFT and distillation (up to 12% gains) and that a correctness-contrast direction can repair advantages in GRPO to recover learning signals in zero-gradient groups.

A recent paper worth noting proposes a “unified training” algorithm that integrates pre-training, mid-training, and post-training into a single algorithmic workflow, simplifying infrastructure and providing a framework for better scientific research.

SWE-Trail is a history-guided trajectory collector for coding agents that uses earlier rollouts to steer where collection effort is spent without exposing that history to the agent, avoiding train-test mismatch. Trained on its trajectories, a 124B Ling-3.0-flash-base student gains 2.33 and 1.60 percentage points in resolution on SWE-bench Multilingual and SWE-bench Pro over independent sampling.

Rubricon is a pipeline from Handshake AI that automatically generates and refines discriminative rubrics for LLM evaluation and post-training without human labels or fine-tuning, achieving 9% better alignment with expert rubrics than prior synthetic methods and 4.5% better HealthBench performance when training Qwen3.5-9B.

IBM Research introduces LLMON (LLM Object Notation), a post-training methodology that separates executable instructions from passive data, reducing prompt-injection attack success rates from 53-100% to 0.02-5.6% when fine-tuning Granite-4.1 models while preserving general capabilities.

Modal announced first-class support for post-training Reflection's Beam model and released Modal Dojo, an open-source library and skill bundle for production-grade LLM training (including RL) on Modal, with an example training Qwen3.5-4B on GSM8K.

A primer on reinforcement learning environments using an escape room analogy, explaining how tools, tasks, and verifiers combine to generate fresh training signals for AI agents now that pre-training data is exhausted.

Anyscale is hosting a webinar on reinforcement learning for LLM post-training, covering RLVR with GRPO, on-policy distillation (OPD/MOPD), terminal agent post-training, and a deep dive into SkyRL's async RL capabilities.

UkisAI announces Swift GLM-5.3-Flash and Swift Ornith 9B, two reasoning-efficient models that cut thinking token usage significantly (-61.4% and -33.5% respectively) while preserving accuracy, using GSQ-RCO penalization of pathological overthinking tokens, RL-based adaptation, and on-policy distillation, with multiple quantization formats released.

Roboflow announced a partnership with NVIDIA and joined the Cosmos Coalition as a premiere provider of multi-modal data and post-training capabilities to advance frontier world model research for the physical economy.

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.

A Chinese-language explainer on reinforcement learning terminology for LLM readers, using the "Is OpenAI open?" example to walk through concepts such as policy, state, action, reward, return, value, advantage, and GAE, and explaining how scores are attributed to individual tokens when a judge only rates an entire response. Set against the backdrop of DeepSeek-R1-Zero's RLVR training, this is the fourth installment in the post-training series.

该论文提出 Recursive Knowledge Improvement (RKI)，将知识改进建模为序列化、状态感知的干预过程，通过区分知识状态别名问题来指导干预选择，在 HotpotEval 和 SafeInternalEval 上平均提升 18.41 分。

该论文研究预训练模型容量如何影响因果任务的适应效果，基于 Qwen 2.5 构建了 CausalLLM（标量势结果估计）和 CausalGen（生成式反事实建模）两个系统，探讨从标量预测到反事实后训练的容量-适应对齐问题，并被 NeurIPS 2026 Workshop 接收。

The paper proposes Teacher-Evolution Self-Distillation (TESD), which improves on-policy self-distillation by continuously optimizing the self-teacher with reward-guided objectives and adaptively down-weighting samples with unreliable teacher guidance, achieving consistent gains over token-level and sequence-level baselines across datasets and model scales.

RSI Arena introduces a two-stage study where ten frontier AI agents autonomously post-train language models from a shared base under fixed GPU and API budgets, guided by human preference feedback. The strongest agent (GPT-6 Astra) closed 76% of the base-to-Instruct gap, and most models improved in human preference from stage 1 to stage 2, demonstrating a step toward recursive self-improvement with human feedback (RSIHF).

该论文通过受控实验研究了不同意见混合比例的后训练语料如何在SFT、DPO和GRPO等五种后训练路径中放大LLM的价值偏好失衡，并发现这些偏好可通过蒸馏传递给学生模型，建议在开发和评估LLM时应同时考虑语料意见构成与后训练方法的选择。