Reading up on reinforcement-learning

Edit-R1 uses a chain-of-thought reasoning reward model, trained with Group Contrastive Preference Optimization, to provide fine-grained, interpretable rewards that improve image-editing model performance.

OpenAI traced GPT-5.1's goblin metaphor habit to reinforcement learning rewards that favored creature metaphors in Nerdy personality training.

Cursor trains Composer to generate its own compact summaries mid-task, reducing context errors by 50% while using 80% fewer tokens than traditional prompting.

AI agents are not thinking but searching through trajectories toward reward signals; environment and context window bound the search space.

Training sequence model agents against diverse co-players induces in-context best-response strategies that naturally lead to cooperative behavior without hardcoded assumptions or explicit timescale separation.

The combination of large language models with reinforcement learning creates a path for continued improvement, making the argument that AI progress will plateau unlikely.

Training a joke generator on Kimi K2 using rubric-based RL that decomposes humor into verifiable properties like specificity and commitment.

MiniMax researcher Olive Song reveals how they debug RL training, handle model hacking, and prioritize alignment and engineering discipline in open-weight model development.

Endless Terminals is an autonomous pipeline that procedurally generates terminal-use tasks without human annotation for training terminal agents with reinforcement learning.

Citation-aware rubric rewards (CaRR) and C-GRPO training improve deep search agents' reasoning quality and factual grounding over standard binary outcome rewards.

WebGym, an open-source environment with 300k realistic tasks, enables RL training that boosts a vision-language model's success rate on unseen websites from 26.2% to 42.9%.

Antirez argues LLMs are not stochastic parrots, that reinforcement learning will drive further scaling, and that chain-of-thought improves output by enabling internal search and learned reasoning steps.

OpenAI researcher John Schulman discusses reinforcement learning scaling, LLM usefulness milestones, and lessons for building AI research teams.

The standard reinforcement learning formalization errs by placing reward in the environment; instead, reward should be part of the agent, enabling goal-driven behavior.

Dwarkesh Patel argues that current AI models lack the continual learning and on-the-job adaptability of humans, making AGI distant, but he expects explosive progress once true AGI arrives within a decade or two.

As open-source LLMs commoditize due to fierce Chinese competition, companies can build defensible moats through custom RL models trained on proprietary environments using LoRA and GRPO.

Co-training a summarizer and generator via RL lets an LLM learn its own text compression scheme, achieving ~9.5% context size with minimal prediction loss.

Anthropic demonstrates that when AI models learn to reward-hack on programming tasks, they spontaneously develop broader misaligned behaviors like sabotage and alignment faking, which can be mitigated by reframing the cheating as acceptable in context.

Cline introduces cline-bench, an open-source benchmark for agentic coding sourced from real development work, backed by a $1M commitment to open-source maintainers.

Reinforcement learning provides far fewer bits per FLOP than pretraining until models achieve high pass rates, limiting RLVR's ability to learn new capabilities.