31 stories
I've been working on unsupervised tooling to discover what behaviors a model learns in post-training. The core idea is to use features from sparse autoencoders to identify differences in traces generated before and after training.
Large language models (LLMs) have achieved strong performance in mathematical and logical reasoning, yet they still suffer from the curse of complexity and struggle as problem difficulty increases with the number of constraint interactions and the size of the search space.
Alibaba's QwenGyre brings elastic GPU scheduling and trajectory deduplication to extreme-long-horizon RL, boosting NL2RepoBench scores by 6% and cutting training time up to 44%.
Vision-Language (VL) reasoning requires a model to both extract relevant and accurate information from an image (visual reasoning, VR), and to infer the answer from it (language reasoning, LR).
A verified reference solution provides a correct trajectory for training a reasoning model. Alternatively, a prefix of the reference can guide the model in generating a trajectory of its own. How much reference guidance should we provide?
Seahaven is an open source Python framework for building synthetic worlds for AI agents: stateful, resettable, parallel environments for evals and RL.
Training AI agents with reinforcement learning can be challenging because their tools, context, and decision-making are managed by complex frameworks. Agent Lightning connects existing agents to RL training, making it easier to improve them without rebuilding them.
Fine-tuning teaches a small search agent your tools and environment, giving it the reliability of a frontier model at lower latency and cost. In this post, we fine-tune an LLM-powered search agent with multi-turn reinforcement learning (MTRL) on Amazon SageMaker AI and share the gains we measured…
Voxel-based volumetric mapping is fundamental to 3D reconstruction, yet fixed-resolution grids remain inherently inefficient - wasting memory in uniform regions and losing detail in complex ones.
Article URL: https://blog.cloudflare.com/clef-decision-models/ Comments URL: https://news.ycombinator.com/item?id=49923692 Points: 218 # Comments: 78
Reinforcement learning (RL) is widely used in language-model post-training to maximize rewards assigned to individual model outputs, such as scores from binary verifiers or reward models trained on human feedback.
The pursuit of recursive self-improvement (RSI) toward general intelligence is divided between macro-level language model scaling and the interaction-driven principles of "Era of Experience".
Research Engineer / Research Scientist, RL Frontiers — San Francisco, CA · full-time · Hybrid · $500k-$850k
Research Engineer / Performance Engineer, RL Distributed Systems — San Francisco, CA · full-time · Hybrid · $500k-$850k
Yesterday I shared our open-source Clash Royale simulator and its recurrent PPO agent here. A training loop is easier to understand when you can watch it, so we put a small interactive version online:
On September 28, 2026 Huawei open-sourced openPangu-2.0 pretrain, SFT and post-training RL code for its Ascend-native MoE stack, complementing earlier Pro (505B) and Flash (92B) weight releases.
Huawei’s Ascend Tribe has released the pretraining, supervised fine-tuning and post-training reinforcement-learning code for its openPangu-2.0 model family. The code is designed for Huawei’s Ascend-based training ecosystem.
The opponent plans by simulation: every second it scores each candidate play by running the match 10 seconds ahead in the engine. Our PPO agent learned to park its Cannon behind its own King. Losing a building in a fight cost reward, and letting it decay cost nothing, so it found the loophole.
Showing 18 of 31 stories
Load more storiesGet the top AI news and tools straight to your inbox every week.
Weekly · AI news
Join 15,000+ AI enthusiasts
