[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"glossary-reinforcement-learning::en":3,"gloss-cluster-reinforcement-learning::en":23,"gloss-next-reinforcement-learning::en":9},{"slug":4,"category":5,"name":6,"definition":7,"meta_desc":8,"faq":9,"schema_markup":9,"related":10},"reinforcement-learning","core-ai","Reinforcement Learning (RL)","Reinforcement learning is the setting where a model learns by acting and being scored, rather than by being shown correct answers. An agent observes a state, takes an action, receives a reward, and lands in a new state; over many episodes it learns a policy — a mapping from states to actions — that accumulates the most reward. Nobody labels the right move. The signal is only whether things went well, often long after the decision that mattered, which is what makes RL both powerful and awkward. That awkwardness has a name: credit assignment. If a game is lost on move sixty, which of the fifty-nine earlier moves was the mistake? RL spends its machinery answering that, and the answer is statistical, which is why RL is sample-hungry in a way supervised learning is not. This is also why its famous successes are in simulation — games, robotics rigs, traffic routing — where an agent can fail millions of times for free. In a live business system, exploration means deliberately making choices you expect to be worse in order to learn, and that has a real cost in real money. For most SaaS teams the relevance of RL is indirect but significant: it is how modern language models are aligned. Reinforcement learning from human feedback trains a reward model on human preferences between candidate responses, then uses RL to push the language model toward what that reward model likes. Direct preference optimization achieves a similar result without the separate RL loop. Both are why a raw pretrained model that merely continues text becomes an assistant that answers questions and declines harmful requests. The vocabulary is worth carrying for the failure modes it names. Reward hacking is a policy maximising the measured signal rather than the intended outcome, which shows up in aligned models as verbosity, hedging, and confident tone — behaviours a reward model rated highly. Exploration versus exploitation is the same tension behind A\u002FB testing and bandit algorithms, which are the RL-adjacent tools most product teams will actually reach for.","Reinforcement learning learns from rewards for its own actions rather than labelled answers — and it is how modern language models get aligned into assistants.",null,[11,14,17,20],{"slug":12,"name":13},"direct-preference-optimization","Direct Preference Optimization (DPO)",{"slug":15,"name":16},"machine-learning","Machine Learning (ML)",{"slug":18,"name":19},"reward-hacking","Reward Hacking",{"slug":21,"name":22},"rlhf","Reinforcement Learning from Human Feedback (RLHF)",[24,28,32,36,39,42,45,48,51,54,57,60],{"slug":25,"category":5,"name":26,"updated_at":27},"agentic","Agentic AI","2026-08-24T02:46:36+00:00",{"slug":29,"category":5,"name":30,"updated_at":31},"alignment-tax","Alignment Tax","2026-08-24T02:46:37+00:00",{"slug":33,"category":5,"name":34,"updated_at":35},"artificial-intelligence","Artificial Intelligence (AI)","2026-08-24T02:46:38+00:00",{"slug":37,"category":5,"name":38,"updated_at":27},"attention","Attention",{"slug":40,"category":5,"name":41,"updated_at":35},"beam-search","Beam Search",{"slug":43,"category":5,"name":44,"updated_at":31},"benchmark-contamination","Benchmark Contamination",{"slug":46,"category":5,"name":47,"updated_at":31},"catastrophic-forgetting","Catastrophic Forgetting",{"slug":49,"category":5,"name":50,"updated_at":35},"computer-vision","Computer Vision",{"slug":52,"category":5,"name":53,"updated_at":31},"constitutional-ai","Constitutional AI",{"slug":55,"category":5,"name":56,"updated_at":27},"context-window","Context Window",{"slug":58,"category":5,"name":59,"updated_at":35},"deep-learning","Deep Learning",{"slug":61,"category":5,"name":62,"updated_at":27},"diffusion-model","Diffusion Model"]