AI Briefing#post-training
The latest AI and developer news about #post-training, with the original source and a short summary.
The latest AI and developer news about #post-training, with the original source and a short summary.
The method learns a binary mask over gate projections, resulting in checkpoints hundreds to thousands of times smaller than full fine-tuning.
A three-week engagement used near-policy SFT and SDPO to lift the model's quality by 20.7 points over the base and 8 points over the frontier competitor.
Figma shares how its latest unreleased editing model, post-trained on Qwen 3.8 27B, scores 77.8% on internal evals, surpassing the top third-party model, while the general availability agent sees 20% faster task completion.
RL environments generate infinite training signals by letting agents learn through experience, shifting post-training costs from per-task human labeling to one-time environment construction.
The method uses self-generated question-answer pairs and reasoning traces to boost performance across text and multimodal domains while retaining capabilities in untrained areas.
JOLT jointly trains a single model as both teacher and student to provide dense, aligned guidance, achieving higher accuracy and significantly better sample efficiency than GRPO on math, coding, and agentic benchmarks.
The new V2 model achieves 92.33% accuracy on a sampled MMLU-Pro subset while optimizing reasoning token usage for agent workloads.
AI21’s Forward-Deployed Researchers post-trained NVIDIA Nemotron 3.5 Super VL with three partners in three weeks, achieving frontier-level performance at a fraction of the cost.
Epoch AI has defined InnovationEval, a benchmark requiring AI agents to develop novel post-training algorithms that outperform a GRPO baseline within a constrained scope of algorithmic changes.
A study comparing base and post-trained models like Tulu 3 and Qwen2.5 found that while humor improved in 5 of 7 steps, joke variety significantly decreased.
Stanford researchers propose EXPO-FT, an algorithm for universal post-training in robotics that achieves 100% success rates on physical tasks by using a lightweight editing policy to stabilize reinforcement learning.
GLM 5.2 lowered the Tool-call failure rate by 21.2% through combined training using RFT on real user session data and Hint-guided Self-distillation.
Xiaomi has released a real-time post-training dashboard for Mimo 2.6, with the Pro model recording a score of 65.78 on DeepSWE v1.1.
Apple researchers confirmed that injecting specific values into LLMs improves safety but also increases sycophancy.
A practical guide focused on evaluation, harnesses, post-training, and online learning to reduce reliance on frontier labs and own your own intelligence.
Mercor and SkyRL have released the complete RL training process for a 397B-parameter model's knowledge work agent.
Apple improved performance by enhancing open-domain QA alignment using a search-distilled Rubric-Reward framework.
Harvey has post-trained the Kimi K3 model for long-running legal tasks.
RadixArk released Miles v0.1, a full-stack system for frontier-scale RL training.
AReno, a toolkit supporting RL, SFT/DPO training, and agent RL for LLMs on a single node, has been released.
Z.ai released GLM-5.3 with enhanced cybersecurity capabilities but delayed the open-weight release for two weeks for safety review.
Mirendil will conduct pre- and post-training of AI models using Google Cloud's TPU and NVIDIA GPU.
Mind Lab released the LoRA-based Macaron-V1, demonstrating the potential of continual learning.
The Ascend NPU-based SLAI T-Rex framework achieved 34.22% MFU in post-training the trillion-parameter DeepSeek-V4 MoE model.