GRPO

Evidence (6)

• 7Hacker News·9/16/2026research
  4B Model Beats Postgres Query Plans on Join Queries

  The article describes an experiment that post-trained a small 4B open-weights model to generate Postgres query plans that outperform the database's default optimizer on join-heavy workloads. Using supervised fine-tuning and reinforcement learning, the model reportedly achieved a 44.7% latency reduction across 113 queries and handled cases the initial model could not plan for.

• 7Hacker News·8/16/2026research
  LittleLearner studies LLMs trained only on K–5 material

  Researchers introduced LittleLearner, a set of language models trained from scratch on an 88B-token corpus filtered to U.S. elementary-school curriculum content up to grade 5. The study finds that scaling, post-training, and in-context learning improve in-scope performance but do not meaningfully extend capability beyond the curated knowledge boundary.

• 7Hacker News·7/28/2026research
  $500 RL fine-tune beats frontier models on catalog review

  A report describes a GRPO fine-tune of a 9B open-source model that outperformed multiple frontier model configurations on the same catalog-review workflow, using the same tools, images, and scoring setup. The result was achieved at about $0.50 per 1,000 listings, significantly cheaper than the frontier alternatives tested.

• 7Hugging Face Blog·5/6/2026framework_update
  vLLM V1 Migration: Fixing Rollout Logprobs Before RL Objective Changes

  Hugging Face describes a migration from vLLM V0 to V1 in an RL training pipeline, where mismatched rollout logprobs initially changed training behavior. The team restored parity by enabling processed logprobs, adjusting V1 runtime defaults, fixing the inflight weight-update path, and using an fp32 lm_head before considering any RL objective corrections.

• 7Hugging Face Blog·4/29/2026model_release
  Granite 4.1 LLMs: How They Are Built

  IBM’s Granite Team details how Granite 4.1 was trained, from data engineering through supervised fine-tuning and reinforcement learning. The family includes dense decoder-only models in 3B, 8B, and 30B sizes, trained on about 15 trillion tokens and extended to a 512K context window, with the 8B instruct model reportedly matching or exceeding the prior Granite 4.0-H-Small despite using fewer parameters.

• 7Hugging Face Blog·3/31/2026framework_update
  TRL v1.0 marks a stability shift for post-training tooling

  Hugging Face released TRL v1.0, presenting it as a more stable library for post-training workflows that now powers production use. The update emphasizes adapting to a fast-changing field, with support for more than 75 post-training methods including PPO, DPO-style approaches, and RLVR methods such as GRPO.