training

This paper presents MyMentorLLM, a multimodal voice/text simulation environment using large language models for deliberate practice in psychotherapy training. It generated 2,100 CBT training sessions and evaluated emotional dynamics, therapeutic competence, and diagnostic accuracy across different LLMs.

Transformers is a model-definition framework providing unified APIs and centralized model definitions for running pretrained models across modalities, from checkpoint to inference or training.

Zing (知境) is a framework for measuring and improving social intelligence in large language models, introducing the SoMBench benchmark covering 3 primary dimensions and 71 task paradigms, a diagnosis-driven training recipe using supervised fine-tuning and rubric-based RL, and the Actio deployment-time architecture. Experiments show that the best model achieves only 72.08% accuracy, indicating substantial headroom, while Zing-series models consistently improve over base models across social-cognition benchmarks.

Humanoid robots are undergoing training in Wuxi, China to master household tasks, as reported by ChinaScience.

TRL now supports training on agent harnesses out of the box through OpenEnv integration, enabling training with harnesses like opencode.

An explainer on how AI models learn, using the analogy of human learning with practice, coaching, and reinforcement, covering pretraining, supervised fine-tuning, and reinforcement learning.

A personal project where a 0.5M parameter language model was trained on 1 billion tokens from the Fineweb-edu dataset.

OpenForgeRL is an open-source framework for training harness-based AI agents end-to-end in diverse environments, using a lightweight proxy and Kubernetes orchestrator to enable RL on any harness at scale. It achieves strong results on agentic benchmarks and shows that RL improves agent reliability.

Tests NVLink on dual RTX 3090s for AI inference and training, finding significant speedups for tensor parallel prompt processing (30%) and FSDP training (3x), but minimal effect on token generation or layer split inference.

Elon Musk announces that SpaceX's engineering data (excluding ITAR-restricted material) will be added to Grok's supplemental training, aiming to significantly improve Grok's engineering capabilities.

Microsoft's strategy for self-evolving agent skills, training them like neural networks with epochs, batch size, learning rates, and validation gates, fully open-source.

SkewAdam is a novel optimizer for mixture-of-experts models that tier allocates optimizer state across backbone, experts, and router, reducing memory footprint to 2.6% of AdamW while achieving better validation perplexity in controlled comparisons.

The article proposes that Transformers can generalize to new tasks through a well-designed harness that induces composition, without needing intrinsic model generalization. It shows RLMs can generalize from short tasks to 8-32x longer tasks and across domains.

An article explaining how LLMs can switch between low, medium, and high effort reasoning during inference and training.

The article introduces Tantalus, a free platform for learning how to secure agentic AI systems against indirect prompt injection and data exfiltration through realistic challenges.

A user released a BitNet trainer that Microsoft never published, along with custom kernels for training and inference, while also highlighting Microsoft's bitnet.cpp inference framework for fast 1-bit LLM inference on CPUs and GPUs.

Velo 3.0 is an AI video infrastructure platform designed to help explain, train, and sell faster.

Prime Intellect released verifiers v1 and prime-rl 0.7.0, an RL training tool with full support for verifiers, multiple algorithms like GRPO and OPD, and performance improvements.

A Substack article explains the 'doom loop' problem in LLMs where models repeat tokens endlessly, and introduces Final Token Preference Optimization (FTPO) from Liquid AI as a method to detect and fix such loops during fine-tuning.

A practical guide for building structured agent trajectory datasets for training tool-using agents, emphasizing the importance of designing trajectories with six key parts and treating them as data assets rather than logs.