# post-training Published articles for post-training. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## Intel's BITCOS format compresses ternary LLM weights to 1.485 bits per weight DevFeed: [Intel's BITCOS format compresses ternary LLM weights to 1.485 bits per weight]() Original publisher: [Read original article]() Author: Amanda Caswell Published: 2026-09-17T20:51:57Z Content type: article Language: en Sources: [The New Stack]() Topics: [Large Language Model](), [quantization](), [Language models](), [intel](), [post-training](), [Decoding](), [qwen3]() Tags: [ai-infrastructure](), [ai-models](), [bits](), [decoding](), [hardware](), [intel](), [language](), [llm](), [model](), [post-training](), [quantization](), [qwen3]() ### AI overview Intel researchers developed BITCOS, a storage format for ternary language-model weights that uses the distribution of zero values to compress a checkpoint to 1.485 bits per weight without changing the model weights. The article reports decoding throughput improvements of up to 18% on CPUs and 27% on GPUs. ### Source excerpt The 1.58 in a 1.58-bit language model sounds like a hard limit, but Intel researchers pushed a ternary model below The post Intel squeezed a 1.58-bit LLM down to 1.485 bits without changing a single weight appeared first on The New Stack. ## PPO vs GRPO, Simply Explained DevFeed: [PPO vs GRPO, Simply Explained]() Original publisher: [Read original article]() Author: Dr. Ashish Bamania Published: 2026-09-17T11:47:38Z Content type: tutorial Language: en Sources: [Into AI]() Topics: [Large Language Model](), [post-training](), [Reinforcement learning](), [Algorithms]() Tags: [algorithm](), [alignment](), [human-feedback](), [llm](), [llm-training](), [post-training](), [reinforcement-learning](), [training]() ### AI overview A tutorial comparing PPO and GRPO as reinforcement learning algorithms used in LLM post-training. It explains PPO, including RLHF, policy-gradient updates, and clipped token-probability changes intended to keep model behavior close to its previous version. ### Source excerpt A simple lesson on two important LLM post-training algorithms. ## Salesforce Announces Koa Reasoning Model Built on NVIDIA Nemotron 3 Super DevFeed: [Salesforce Announces Koa Reasoning Model Built on NVIDIA Nemotron 3 Super]() Original publisher: [Read original article]() Author: Brian Caulfield Published: 2026-09-15T22:24:34Z Content type: news Language: en Sources: [NVIDIA Blog]() Topics: [Koa](), [Nemotron](), [Artificial Intelligence](), [Nvidia](), [post-training](), [ai safety]() Tags: [agentic](), [agentic-ai](), [ai](), [cloud](), [corporate](), [events](), [model](), [nemotron](), [nvidia](), [nvidia-nemo](), [post-training](), [safety](), [security](), [software]() ### AI overview At Salesforce Dreamforce, NVIDIA CEO Jensen Huang discussed AI infrastructure, safety, and enterprise adoption alongside Salesforce CEO Marc Benioff. The event coincided with the announcement of Koa, Salesforce's first CRM reasoning model, built by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic dataset derived from nearly three decades of enterprise CRM deployments. ### Source excerpt Know everything. Do anything. That was the message NVIDIA founder and CEO Jensen Huang brought to Salesforce Dreamforce Tuesday, joining CEO Marc Benioff onstage in an appearance that coincided with the announcement of Koa -- Salesforce's first CRM reasoning model, built on NVIDIA Nemotron 3 Super. Huang didn't just take the stage. He walked into [...] ## Autonomous LLM post-training with Tunix on TPUs DevFeed: [Autonomous LLM post-training with Tunix on TPUs]() Original publisher: [Read original article]() Author: Wei Wei Published: 2026-09-12T11:04:33.891311Z Content type: article Language: en Sources: [Google Developers Blog]() Topics: [post-training](), [AI Agent](), [Fine-tuning](), [Large Language Model](), [grpo](), [lora](), [gemma](), [Google](), [Reinforcement learning](), [dataset](), [Command-line interface]() Tags: [autonomous](), [cli](), [fine-tuning](), [gemma](), [git](), [google](), [grpo](), [llm](), [lora](), [post-training](), [reinforcement-learning](), [training]() ### AI overview This article presents autofinetune, an autonomous research loop for LLM post-training. Using AI agents and Google's AI stack, including Tunix, Gemma, Cloud TPUs, Antigravity CLI, and Gemini Flash 3.7, it automates supervised fine-tuning and reinforcement learning with GRPO, exploring hyperparameters such as LoRA configurations, learning rates, batch sizes, and rollout settings. ### Source excerpt Imagine going to sleep after writing a single Markdown specification and waking up to find that an A... ## Deploying AI You Control Doesn't Need to be So Hard DevFeed: [Deploying AI You Control Doesn't Need to be So Hard]() Original publisher: [Read original article]() Author: Jeetu Patel Published: 2026-09-10T09:00:50Z Content type: news Language: en Sources: [Cisco Blogs]() Topics: [Artificial Intelligence](), [AI Platforms/Deployment](), [AI Factory](), [Nvidia](), [Critical Infrastructure](), [post-training](), [data](), [Nemotron](), [on-prem](), [Fine-tuning](), [Cloud](), [Benchmark]() Tags: [ai](), [ai-factory](), [ai-infrastructure](), [ai-strategy](), [cisco-cloud-control-framework](), [cisco-secure-ai-factory](), [cloud](), [critical-infrastructure](), [data](), [executive-platform](), [fine-tuning](), [infrastructure](), [model](), [models](), [nemotron](), [nvidia](), [on-prem](), [post-training]() ### AI overview Cisco announces a collaboration with Palantir to deliver Palantir's Ontology for Cybersecurity through Cisco's Secure AI Factory, using NVIDIA as a preferred full-stack foundation for Palantir's Sovereign AI OS. The article argues that enterprise AI decisions should balance intelligence, cost, and control, including custom evaluations, post-training with proprietary data, and deployment in the cloud, at the edge, or on-premises. ### Source excerpt Announcing a collaboration with Palantir to deliver Cisco's Secure AI Factory with NVIDIA as a preferred full-stack foundation for Palantir's Sovereign AI OS. ## On-Policy Distillation, Simply Explained DevFeed: [On-Policy Distillation, Simply Explained]() Original publisher: [Read original article]() Author: Dr. Ashish Bamania Published: 2026-09-09T19:22:38Z Content type: tutorial Language: en Sources: [Into AI]() Topics: [LLM Techniques](), [Large Language Model](), [post-training](), [Reinforcement learning]() Tags: [fine-tuning](), [gpu](), [llms](), [post-training](), [training]() ### AI overview This tutorial explains on-policy distillation (OPD), a post-training method for large language models. A student model generates its own responses, which a stronger teacher scores token by token; the student is trained to reduce the reverse KL divergence between their next-token distributions. The article distinguishes OPD from reinforcement learning, conventional knowledge distillation, and supervised fine-tuning. ### Source excerpt On-Policy Distillation (OPD) has become a popular algorithm for post-training LLMs, and almost all recent open-weight LLMs (Qwen3, GLM-5.3, and Nemotron-Cascade 2) have used it to achieve amazing performance. ## This smart boxing band takes advantage of the new Arduino Nesso N1 DevFeed: [This smart boxing band takes advantage of the new Arduino Nesso N1]() Original publisher: [Read original article]() Author: Arduino Team Published: 2026-09-09T15:36:33Z Content type: article Language: en Sources: [Arduino Blog]() Topics: [Arduino](), [Hardware](), [data](), [real-time](), [Artificial Intelligence](), [post-training]() Tags: [ai](), [arduino](), [boxing](), [boxing-training](), [data](), [hardware](), [imu](), [nesso-n1](), [performance](), [post-training](), [real-time](), [smart-boxing-band](), [training]() ### AI overview An Arduino Nesso N1-based smart boxing band combines punch recognition, biometric tracking, and real-time feedback. It uses a pulse sensor and the board's IMU to relate punch types and movement to heart rate, while logging data for post-training analysis. ### Source excerpt Data is now a huge part of many sports, from F1 racing to football. Presenting that data to fans is secondary to the real purpose: helping teams and athletes maximize performance. No matter what sport you enjoy, you can benefit from that kind of quantified training. Manivannan proved that by building his Smart AI Boxing [...] The post This smart boxing band takes advantage of the new Arduino Nesso N1 appeared first on Arduino Blog. ## Understanding W8A8 INT8 LLM quantization: Half the size, better performance, same accuracy DevFeed: [Understanding W8A8 INT8 LLM quantization: Half the size, better performance, same accuracy]() Original publisher: [Read original article]() Author: Sana Fayyaz Published: 2026-09-07T07:01:34Z Content type: article Language: en Sources: [Red Hat Developer]() Topics: [quantization](), [Large Language Model](), [GPU](), [Inference](), [model-deployment](), [Post-training optimization](), [llama](), [LLM Techniques]() Tags: [ai-inference](), [artificial-intelligence](), [compression](), [deployment](), [gpu](), [inference](), [language-models](), [large-language-models](), [llama](), [llm](), [memory](), [performance](), [post-training](), [quantization]() ### AI overview This article explains INT8 W8A8 quantization for Llama 3.1 8B using SmoothQuant and GPTQ. It describes how quantization reduces memory requirements and can improve serving performance while preserving model accuracy, including a reported 46% reduction in model size and planned benchmark testing with vLLM and llm-compressor. ### Source excerpt Large language models are expensive to serve. A model like Llama 3.1 8B in Bfloat16 (BF16) precision occupies roughly 15 GB of GPU memory. In BF16, each of the 8 billion parameters takes 2 bytes to store, which adds up to roughly 15 GB for the weights--and that's not all. The GPU needs memory for the key-value (KV) cache to store context for active requests, alongside intermediate tensor outputs (activations, as we call them) generated during inference. The post Understanding W8A8 INT8 LLM quantization: Half the size, better performance, same accuracy appeared first on Red Hat Developer. ## From Preferences to Principles: Rubric-Based Alignment for Grounded Knowledge Answers DevFeed: [From Preferences to Principles: Rubric-Based Alignment for Grounded Knowledge Answers]() Original publisher: [Read original article]() Published: 2026-08-27T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [Large Language Model](), [LLM evaluation / benchmarking](), [Natural language processing](), [datasets]() Tags: [datasets](), [evaluation](), [models](), [post-training](), [research](), [responses](), [search]() ### AI overview The article introduces a rubric-based reward framework for grounded open-domain question answering. It generates query-specific rubrics from retrieved evidence and decomposes them into quality dimensions for fine-grained post-training supervision. The method improves composition, grounding, and instruction-following results over the stated baselines and evaluation datasets. ### Source excerpt Designing effective reward signals for open-domain question answering is challenging because high-quality responses must simultaneously satisfy multiple aspects of answer quality that are difficult to capture with a holistic scalar objective. We introduce a rubric-based reward framework that generates query-specific rubrics grounded in retrieved evidence and decomposed into multiple quality dimensions, providing fine-grained supervision during post-training. Averaged across three evaluation axes (composition, grounding, and instruction-following), our approach improves over the... ## PROOF-Gen: From Optimized Data to Better Distillation DevFeed: [PROOF-Gen: From Optimized Data to Better Distillation]() Original publisher: [Read original article]() Published: 2026-08-26T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [Fine-tuning](), [Prompt optimization](), [Artificial Intelligence](), [gemma4](), [Programming]() Tags: [agent](), [ai-agents](), [evaluation](), [fine-tuning](), [gemma](), [generate](), [models](), [on-device](), [post-training](), [prompt-optimization]() ### AI overview PROOF-Gen improves tool-calling model distillation by using per-scenario prompt optimization to recover successful trajectories from failed teacher attempts. The method strips corrective guidance before training, producing clean demonstrations and improving benchmark, deployed-pipeline, and on-device model performance. ### Source excerpt Supervised fine-tuning on teacher-generated trajectories is the standard first stage for distilling tool-calling capabilities into deployable models. Post-training pipelines that drive shipped tool-calling agents re-run this stage on a daily or weekly cadence, paying the frontier-teacher cost each cycle, yet the mechanism is generate-and-filter (keep the teacher's passing trajectories, discard the rest) and each cycle leaves behind the same hard scenarios because failures supply no signal. On τ 2-bench, 57% of teacher trials fail, two-thirds of them near-misses (most tool calls correct, undone... ## Granite 4.2 LLMs: How They're Built DevFeed: [Granite 4.2 LLMs: How They're Built]() Original publisher: [Read original article]() Author: Yousaf Shah; Swanand Kadhe; Riddhiman Moulick; Ashish Sunil Agrawal; Santosh Borse Published: 2026-08-25T15:14:14Z Content type: article Language: en Sources: [Hugging Face - Blog]() Topics: [Language models](), [AI Chat](), [AI-assisted coding](), [releases]() Tags: [agentic](), [agents](), [apache](), [chain-of-thought](), [code](), [fine-tuning](), [grouped-query-attention](), [ibm](), [llms](), [models](), [post-training](), [reasoning](), [reinforcement-learning](), [sglang](), [tool](), [training](), [transformer-architecture](), [vllm]() ### AI overview Granite 4.2 is a family of 3B, 8B, and 30B dense decoder-only reasoning language models. The article covers their training pipeline, thinking modes, native tool calling, and agentic reinforcement learning for the 8B and 30B models. ### Source excerpt Authors: Granite Team, IBM TL;DR: Granite 4.2 is our first family of dense, decoder-only reasoning LLMs, released in three sizes: 3B, 8B, and 30B. These models are post-trained from Granite-4.1 base models. Granite-4.1 base models were pre-trained from scratch on roughly 15T tokens with a five-phase strategy that extends the context window to 512K tokens, supervised fine-tuned on chain-of-thought, reasoning, and agentic-trajectory data, then post-trained with a multi-stage reinforcement... ## Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original DevFeed: [Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original]() Original publisher: [Read original article]() Author: Antonio Tiene; Iker García-Ferrero; Ali Hashemi; Bakbergen Ryskulov Published: 2026-08-25T11:39:24Z Content type: article Language: en Sources: [Hugging Face - Blog]() Topics: [AI Chat]() Tags: [architecture](), [benchmarks](), [blog](), [compression](), [fine-tuning](), [gpt](), [gpt-oss](), [hugging-face](), [llms](), [model](), [post-training](), [quantization](), [rlhf]() ### AI overview The article presents Quantization-Aware Healing (QAH), a method for recovering structurally compressed and 4-bit-quantized LLMs. It contrasts QAH with quantization-aware training and distillation, arguing that the latter can be limited when no independently trained full-precision version of the compressed architecture exists. ### Source excerpt A Blog post by Multiverse Computing on Hugging Face ## Post-Train NVIDIA Cosmos 3 Edge for On-Device Robot Control DevFeed: [Post-Train NVIDIA Cosmos 3 Edge for On-Device Robot Control]() Original publisher: [Read original article]() Author: Michelle Horton Published: 2026-08-19T16:00:00Z Content type: tutorial Language: en Sources: [NVIDIA Developer](), [NVIDIA Technical Blog]() Topics: [Cosmos](), [AI, ML & Data Engineering]() Tags: [agentic-ai-generative-ai](), [cosmos](), [deployment](), [edge](), [edge-computing](), [featured](), [foundation-models](), [inference](), [jetson](), [latency](), [nvidia](), [on-device](), [physical-ai](), [post-training](), [robotics](), [robotics-simulation](), [robots](), [simulation-modeling-design](), [thor](), [tutorial]() ### AI overview A tutorial on post-training NVIDIA Cosmos 3 Edge as an on-device robot manipulation policy, serving it on Jetson Thor, running receding-horizon inference, and evaluating it in closed-loop simulation. ### Source excerpt Robots need policies that can adapt to their sensors, environments, and tasks while running on onboard computing hardware. World models offer a foundation for... ## Developing Nemotron 3.5 Lightning NVFP4 with QAD Using NVIDIA Model Optimizer DevFeed: [Developing Nemotron 3.5 Lightning NVFP4 with QAD Using NVIDIA Model Optimizer]() Original publisher: [Read original article]() Author: Tanya Lenz Published: 2026-08-17T18:12:48Z Content type: tutorial Language: en Sources: [NVIDIA Developer](), [NVIDIA Technical Blog]() Topics: [Nemotron](), [NVFP4](), [quantization](), [Post-training optimization](), [Nvidia](), [post-training](), [Mamba]() Tags: [agentic-ai-generative-ai](), [benchmarks](), [compute](), [developer-tools-techniques](), [developers](), [edge-computing](), [evaluation](), [featured](), [latency](), [mamba](), [megatron](), [memory](), [model](), [model-optimizer](), [models](), [nemotron](), [nvfp4](), [nvidia](), [open](), [post-training](), [quantization](), [speed](), [training](), [training-ai-models]() ### AI overview This tutorial explains how quantization-aware distillation (QAD) creates the Nemotron 3.5 Lightning NVFP4 checkpoint using NVIDIA Model Optimizer. It covers post-training quantization, teacher-student distillation, and evaluation, showing how QAD can recover accuracy while reducing memory usage and increasing throughput. ### Source excerpt Teams customize their models to hit their targets for latency, speed, memory, and compute. With the open NVIDIA Nemotron family of models, developers can find... ## Empty shelves or lost keys? Recall is the bottleneck for parametric factuality DevFeed: [Empty shelves or lost keys? Recall is the bottleneck for parametric factuality]() Original publisher: [Read original article]() Published: 2026-08-12T09:51:00Z Content type: article Language: en Sources: [The latest research from Google]() Topics: [Google](), [LLM evaluation / benchmarking](), [Large Language Model](), [Benchmark](), [Hallucination detection](), [AI Inference]() Tags: [benchmark](), [classification](), [errors](), [generative-ai](), [google](), [inference](), [large-language-models-llms](), [natural-language-processing](), [post-training](), [research]() ### AI overview This Google Research article argues that many factual errors in frontier large language models arise from recall failures rather than missing encoded knowledge. It presents knowledge profiling, which separates encoding, recall, and recognition, and introduces WikiProfile, a benchmark of 2,150 Wikipedia-derived facts tested through questions targeting these abilities. ### Source excerpt Generative AI ## Training 100x Cheaper Retrieval models Neon and Castform DevFeed: [Training 100x Cheaper Retrieval models Neon and Castform]() Original publisher: [Read original article]() Author: Pranav Aurora Published: 2026-08-05T12:00:00Z Content type: article Language: en Sources: [Blog -- Neon Docs]() Topics: [Retrieval Augmented Generation (RAG)](), [Training AI Models](), [Machine learning](), [Open Source Models & Datasets](), [Large Language Model](), [Open Source](), [Latency](), [Database]() Tags: [agentic](), [agents](), [cost](), [embedding](), [infra](), [llms](), [machine-learning](), [model](), [models](), [open-source](), [performance](), [post-training](), [product](), [rag](), [retrieval](), [rl](), [scale](), [search](), [systems](), [tools](), [training]() ### AI overview The article explains how Castform uses reinforcement-learning post-training to improve open-weight models for agentic retrieval. It contrasts multi-step retrieval with one-shot embedding search, emphasizing the cost and latency of repeated frontier-model calls and the potential for smaller open models to perform specific search tasks more cheaply. ### Source excerpt "Most teams' best training data is just sitting in their databases. The problem is that turning raw data into something usable is hard, and letting agents read, search, and mutate data cheaply at scale requires advanced infra. Pointing Castform at Neon skips both." -- Ying Hang Seah, cofounder, Castform ## Beyond VLAs: How World Action Models Reshape Robot Manipulation DevFeed: [Beyond VLAs: How World Action Models Reshape Robot Manipulation]() Original publisher: [Read original article]() Author: Michelle Horton Published: 2026-08-04T16:00:00Z Content type: article Language: en Sources: [NVIDIA Developer](), [NVIDIA Technical Blog]() Topics: [Robotics](), [World models](), [vlm](), [post-training](), [Cosmos](), [Nvidia](), [NVIDIA Research]() Tags: [agentic-ai-generative-ai](), [cosmos](), [featured](), [nvidia](), [nvidia-research](), [physical-ai](), [post-training](), [robot-manipulation](), [robotics](), [simulation-modeling-design](), [thor](), [vlm](), [world-model](), [zero-shot]() ### AI overview The article explains how World Action Models (WAMs) use video world models as backbones for robot policies, addressing the physical-generalization limitations of vision-language-action models. It discusses post-training WAMs into specialized policies and presents NVIDIA Cosmos 3 as a foundation for building them. ### Source excerpt A central challenge in robotics is building policies that generalize beyond the demonstrations they're trained on. A policy that succeeds in a training scene... ## Data Engineering Weekly #281 DevFeed: [Data Engineering Weekly #281]() Original publisher: [Read original article]() Author: Ananth Packkildurai Published: 2026-08-03T12:34:40Z Content type: article Language: en Sources: [Data Engineering Weekly]() Topics: [data-engineering](), [AI, ML & Data Engineering](), [Loop Engineering](), [Artificial Intelligence](), [Large Language Model](), [post-training]() Tags: [agent-harness](), [ai](), [data-engineering](), [feature-engineering](), [genai](), [netflix](), [newsletter](), [post-training](), [weekly]() ### AI overview Data Engineering Weekly #281 covers building data platforms, emerging approaches to AI workflow architecture, data modernization, Netflix's GenRec recommendation system, AI infrastructure modernization, and evaluation practices for generative AI at scale. ### Source excerpt The Weekly Data Engineering Newsletter ## Post-Train NVIDIA Cosmos 3 in One Day Using Agent Skills DevFeed: [Post-Train NVIDIA Cosmos 3 in One Day Using Agent Skills]() Original publisher: [Read original article]() Author: Tanya Lenz Published: 2026-07-14T16:00:00Z Content type: tutorial Language: en Sources: [NVIDIA Developer](), [NVIDIA Technical Blog]() Topics: [Cosmos](), [Agent Skills](), [post-training](), [lora](), [Nvidia]() Tags: [agent-skills](), [agentic-ai-generative-ai](), [ai-agent](), [computer-vision-video-analytics](), [cosmos](), [featured](), [fine-tuning](), [lora](), [physical-ai](), [post-training](), [simulation-modeling-design](), [training-ai-models]() ### AI overview A tutorial on post-training NVIDIA Cosmos 3 Nano for video question answering with coding-agent skills, LoRA, and TAO AutoML configuration sweeps. ### Source excerpt What if autonomous coding AI agents could push your vision reasoning models above 90% accuracy with almost no manual effort? When adapting vision reasoning... ## Mastering Agentic Techniques: AI Agent Reinforcement Learning DevFeed: [Mastering Agentic Techniques: AI Agent Reinforcement Learning]() Original publisher: [Read original article]() Author: Elizabeth Goodman Published: 2026-07-01T17:04:02Z Content type: tutorial Language: en Sources: [NVIDIA Developer](), [NVIDIA Technical Blog]() Topics: [Language models](), [AI Bots]() Tags: [agentic](), [agentic-ai-generative-ai](), [ai-agent](), [ai-platforms-deployment](), [featured](), [language-models](), [nemo](), [nemotron](), [post-training](), [rag](), [reinforcement-learning](), [rlhf](), [rlvr](), [tool]() ### AI overview A guide to using reinforcement learning with verifiable rewards to post-train language models for specialized, long-running AI agents. It explains when prompting, RAG, tools, and agent harnesses are insufficient, and describes reward signals based on verifiers, execution, validation, models, and human feedback. ### Source excerpt Reinforcement learning (RL) is central to aligning language models, from reinforcement learning with human feedback (RLHF) within AI assistants to newer... ## GLM-5.2: Built for Long-Horizon Tasks DevFeed: [GLM-5.2: Built for Long-Horizon Tasks]() Original publisher: [Read original article]() Author: Z AI Published: 2026-06-17T09:01:25Z Content type: article Language: en Sources: [Hugging Face - Blog]() Topics: [long-context](), [coding](), [Benchmark](), [Optimization](), [Open Source](), [Software Engineering](), [post-training]() Tags: [benchmark](), [coding](), [long-context](), [open-source](), [optimization](), [post-training](), [research](), [software-engineering]() ### AI overview GLM-5.2 is presented as an open-source model designed for long-horizon engineering work, combining a stable 1M-token context, flexible coding effort levels, and architectural changes that reduce computation and improve speculative decoding. The article reports strong results across long-horizon and standard coding benchmarks, positioning it as the highest-ranked open-source model in the cited evaluations. ### Source excerpt - Solid 1M Context: A solid 1M-token context that stably sustains long-horizon work - Advanced Coding with Flexible Effort: Stronger coding capabilities with multiple thinking effort levels to balance performance and latency - Improved Architecture: We propose IndexShare, which reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9x at a 1M context length. ## Introducing OpenRL: A self-hosted post-training API for fine-tuning LLMs DevFeed: [Introducing OpenRL: A self-hosted post-training API for fine-tuning LLMs]() Original publisher: [Read original article]() Author: Google Open Source (noreply@blogger.com) Published: 2026-06-11T18:30:00Z Content type: release Language: en Sources: [Google Open Source Blog]() Topics: [API](), [Fine-tuning](), [Large Language Model](), [Self-hosted](), [post-training](), [Kubernetes](), [Orchestration](), [reliability]() Tags: [api](), [fine-tuning](), [gke](), [kubernetes](), [llms](), [open-source](), [orchestration](), [post-training](), [reinforcement-learning](), [reliability](), [rl](), [self-hosted](), [sft]() ### AI overview This article announces OpenRL, an open-source research preview from GKE Labs. OpenRL is a self-hosted training API for fine-tuning LLMs on a Kubernetes cluster, designed to separate post-training infrastructure from AI research workflows. The article describes potential benefits including concurrent reinforcement-learning jobs, improved GPU utilization, and simpler researcher workflows. ### Source excerpt by Sunil Arora, Shuby Mishra & Chuang Wang, GKE We are pleased to share a research preview of OpenRL, a new open-source project coming out of GKE Labs. OpenRL is a self-hosted training API for fine-tuning LLMs on your own Kubernetes cluster. Why we built it If you look at agentic RL on LLMs, it is incredibly easy to get bogged down in system complexity. To run a single RL loop, you have to coordinate a dozen different things: selecting and cleaning datasets, choosing RL environments, debugging training loops, managing reward signals, handling inference mismatches, allocating hardware, and managing infrastructure. Picture looks something like this: Figure shows an AI researcher and an infrastructure engineer staring at the hurdles in post training along the way to the summit. Each of these is a hard problem. But what makes it more complex is how tightly AI research and infrastructure concerns are mixed together in today's tooling and frameworks. We believe decoupling the infrastructure from AI research can make these problems more tractable so that infrastructure engineers and AI researchers can independently tackle them. We have seen this pattern with Kubernetes where Kubernetes abstracted out the infrastructure and made application developers and SREs life easier. So, can you abstract out post training infrastructure? We believe so and drew huge inspiration/validation from Tinker (from Thinking Machines). The Tinker APIs for post training hit that Goldilocks zone where it hides all the post training infrastructure behind four key APIs: Figure shows high level components and their interaction in a OpenRL based RL workflow So the end result of this abstraction is that AI Researchers get full flexibility on their RL loop and infrastructure engineers can focus on scaling, orchestration, and reliability. OpenRL allows you to run the same training APIs but on your own infrastructure. And this decoupling has other interesting benefits. Sharing GPUs Traditional RL loops ar ## Diverse reasoning traces teach LLMs to make better decisions DevFeed: [Diverse reasoning traces teach LLMs to make better decisions]() Original publisher: [Read original article]() Author: Sheng Jia; Xiao Wang; Shiva Kasiviswanathan Published: 2026-05-26T15:17:06Z Content type: article Language: en Sources: [Amazon Science homepage]() Topics: [Language models]() Tags: [benchmarks](), [chain-of-thought](), [code-generation](), [conversational-ai](), [deep-learning](), [evaluation](), [fine-tuning](), [iclr](), [iclr-2026](), [large-language-models-llms](), [llms](), [math-reasoning](), [parallel-reasoning](), [post-training](), [post-training-optimization](), [reasoning](), [reinforcement-learning](), [traces]() ### AI overview The article presents set-supervised fine tuning and global forking policy optimization to train LLMs on multiple distinct reasoning paths. It reports 5% to 7% single-shot accuracy gains on standard benchmarks. ### Source excerpt How to train language models to generate diverse, accurate reasoning paths using tokens that control distinct reasoning strategies. ## TRL v1.0: Post-Training Library Built to Move with the Field DevFeed: [TRL v1.0: Post-Training Library Built to Move with the Field]() Original publisher: [Read original article]() Author: Quentin Gallouédec; Steven Liu; Pedro Cuenca; Sergio Paniego Published: 2026-03-31T00:00:00Z Content type: article Language: en Sources: [Hugging Face - Blog]() Topics: [trl](), [post-training](), [dpo](), [grpo](), [rlvr](), [Artificial Intelligence](), [Code]() Tags: [announcement](), [artificial-intelligence](), [code](), [dpo](), [grpo](), [open-source](), [post-training](), [reinforcement-learning](), [rlvr](), [software](), [trl](), [v1]() ### AI overview TRL v1.0 presents a post-training library designed to remain useful as algorithms, models, and paradigms change. The article explains the evolution from PPO to DPO-style preference optimization and RLVR methods such as GRPO, emphasizing flexible abstractions and practical comparison over a supposedly permanent design. ### Source excerpt We're on a journey to advance and democratize artificial intelligence through open source and open science. [Next page]()