Why does post-training matter so much now?
Because behavior is the current bottleneck, not knowledge. Two models with similar pre-training can feel generations apart after different post-training. Most of what distinguishes today’s reasoning models was built in post-training, largely with reinforcement learning against outcomes a checker can verify, guided by a reward model where no hard check exists.
How do the stages of post-training fit together?
As a progression from imitation to judgment. Supervised fine-tuning comes first: the model copies curated examples, which is quick but never exceeds what they show. Preference training, DPO or RLHF style, works on comparisons, this answer over that one, capturing qualities nobody can demonstrate perfectly, such as tone and refusal judgment, while inheriting whatever the raters systematically prefer. Last comes reinforcement learning on tasks with verifiable answers, math, code, tool use, with process supervision scoring intermediate steps when only the reasoning can be judged. Each stage raises the previous one’s ceiling and imports its own failure mode: imitation stays bland, preference training drifts sycophantic, and outcome RL exploits any gap the checker leaves open.
Is post-training something a company does?
Usually not. The lab post-trains the LLM; a company that adapts one performs fine-tuning, a small, targeted member of the same family, worth doing only after prompting and retrieval fall short. Buy behavior, do not train it is the sensible mid-market default, because post-training quality is what the model bill already pays for. What is worth understanding is the dependency: when a provider revises post-training, your system’s behavior can shift without any model number changing.
How does post-training show up when you are debugging?
As behavior you did not change. Refusal style, tool-call formatting and verbosity under pressure were all set in post-training, so two models given identical context can act like different colleagues. A regression suite over your own tasks is the only reliable early warning: prompt regression testing and AI agent evals, rerun on every model change. A failure about facts responds to retrieval and context; a model that stops following your format is showing a post-training difference, fixed in the prompt or the model choice.
How do our systems improve without post-training?
Our own agents learn without touching model weights: lessons captured from each working session are reviewed weekly into the written files every agent reads. A knowledge-capture system we reviewed learns the same way: every human correction is written back as a precedent that retrieval can surface on the next similar case. Before a build like that ships we check the loop twice: the correction must reappear for a near-identical input, and the golden set must still pass after the write, as described in Enterprise RAG on LangChain. No weights change, so a bad lesson is a record someone can read and remove.
When we evaluated two open-source agents that advertise self-learning, Hermes Agent and OpenClaw, in June 2026, both defined it as persistent memory plus self-written skills, not weight training.