reinforcement learning
29 published articles
Document AI
The 0.8B model that beat every pipeline document parser at its own game
Alibaba's OvisOCR2, a compact 0.8B end-to-end model, achieves state-of-the-art scores on OmniDocBench (96.58) and PureDocBench (75.06), overtaking larger pipeline-based parsers. Its success comes from a data engine mixing filtered real documents with synthetic pages, reinforcement learning on a 4B teacher, and on-policy distillation into the small model.
2026-07-24
Closed-loop coding agents
Alibaba built a coding model that learns from its own users. The numbers are hard to ignore.
Alibaba's new coding model Qwen-Coder-Qoder beats Cursor Composer-1 on the Qoder Bench benchmark. Production metrics show a 3.85% code retention increase, a 61.5% drop in tool errors, and a 14.5% reduction in token use. The model trains on real agent traces through a rewarder-attacker framework against reward hacking.
2026-07-23
Biologically Plausible Learning
Sakana AI's MNIST-to-CIFAR flip exposes a blind spot in brain-like AI
Sakana AI scaled a backprop-free, Dale-compliant architecture past MNIST for the first time. The catch: which trick matters most reverses between datasets, a warning about how biologically plausible AI is measured.
2026-07-22
AI Research
Qwen just taught diffusion models a trick from the LLM playbook: RL beats supervised fine-tuning
The Qwen-Image-2.0-RL report details a post-training pipeline that combines RLHF, on-policy distillation, and composite reward models to improve text-to-image and image editing quality. The approach yields measurable gains across aesthetic quality, instruction following, and face identity preservation, borrowing techniques from LLM alignment research.
2026-07-20
Biologically plausible AI's hidden warning
The CIFAR-10 ablation that should make every AI researcher rethink their benchmarks
Sakana AI's Error Diffusion scales brain-like, backprop-free learning to CIFAR-10 and RL. The catch is in the ablation tables: which design choices matter reverses entirely between benchmarks, and the cost of Dale's principle grows with task difficulty.
2026-07-19
AI Research
RDPO: The two-word fix for reinforcement learning's self-sabotage problem
Multi-task reinforcement learning has a dirty secret: the reward signals that drive alignment often work against each other. An international team just published a method that prevents the system from fighting itself, and it costs almost nothing to add to existing pipelines.
2026-07-19
Embodied AI
Mistral's tiny 8B model just made lidar optional for office robots
Mistral AI's Robostral Navigate is an 8B model that uses only one RGB camera to achieve 76.6% on R2R-CE benchmarks, outperforming multi-sensor approaches by 4.5 points. Built with simulated data and a token-efficient prefix-caching method, it generalizes across robot types and adapts to unseen obstacles.
2026-07-17
AI Research
The AI agent bottleneck isn't exploration. It's knowing what good looks like.
Two new Hugging Face papers tackle the same core problem from opposite directions: how to make AI agents reliably evaluate their own actions. AJ-Bench builds a benchmark for environment-aware judge agents, while HeavySkill argues the best judge lives inside the model's parameters.
2026-07-17
Artificial Intelligence
Alibaba just open-sourced a world model that lets AI agents train inside a simulator
Alibaba's Qwen team released Qwen-AgentWorld, a language world model that simulates agent environments across seven domains. Its three-stage pipeline, CPT, SFT, RL, produces a simulator that beats GPT-5.4 on fidelity and enables agents to train via mental rehearsal rather than costly real-environment rollouts.
2026-07-16
Competitive programming
NousCoder-14B just opened the coding RL black box that OpenAI and DeepMind keep locked
Nous Research drops NousCoder-14B, a 14B competitive programming model with a fully open RL pipeline. The 68% Codeforces solve rate is notable, but the real story is that anyone can now replicate the stack.
2026-07-16
Artificial Intelligence
The reward-hacking collapse that nearly killed MiniMax's proof model
MiniMax details how M3's proof capabilities survived a reward-hacking crisis that nearly killed the project. The four-layer verifier and MaxProof test-time framework pushed scores above human gold-medal thresholds on IMO 2025 and USAMO 2026, offering a blueprint for any lab dealing with adversarial model behavior.
2026-07-16
Tooling for RL
The RL scaling headache just got a fix you can deploy with one docker-compose up
Nous Research releases tinker-atropos, an integration layer that decouples trainer, rollout, and environment services so scaling RL beyond a single machine becomes a config change, not a code rewrite.
2026-07-16