Post-Training Modern LLMs
Pretraining produces a model that predicts text. Post-training is what turns it into something you can ship. This path walks the levers in the order you would actually reach for them: supervised fine-tuning and adapters, preference optimization without a reward model, the reinforcement-learning map from RLHF to verifiable rewards, RL against a verifier that cannot be talked out of its answer, and finally the inference-time compute that picks up where training leaves off. Every step names the ceiling it runs into.
Reasoning Models: How LLMs Learned to Think Before They Speak
Start here for why post-training became the interesting half of the pipeline. Reasoning models were not a new architecture, they were a new training signal, and this piece traces how chain-of-thought and reinforcement learning produced a capability jump that scaling alone had stopped delivering. Read it for the shape of the problem the rest of the path solves.
Fine-Tuning Transformer Models with Low-Rank Adaptation (LoRA)
The cheapest lever, and the one most teams should try first. LoRA and QLoRA freeze the base model and train under 1% of the weights, which is what makes fine-tuning a decision you can afford to get wrong twice. Focus on two claims that survive every method comparison: data quality dominates method choice, and fine-tuning teaches behaviour rather than facts.
Preference Optimization After DPO: IPO, KTO, and SimPO Compared
Supervised fine-tuning teaches the format, preference optimization teaches the preference. This step covers the family that removed the reward model from alignment, and it is deliberately placed before the RL chapters because DPO is where most production alignment actually happens. Focus on what each variant trades away, since IPO, KTO and SimPO are different failure boundaries rather than successive upgrades.
From RLHF to RLVR: How RL Post-Training Actually Evolved
Before the hands-on RL chapter, get the map. This piece lays out how post-training RL evolved from RLHF through GRPO to verifiable rewards, plus Constitutional AI and self-play, so the next step reads as one point on a lineage rather than a trick. Focus on why the field kept replacing the reward signal, which is the throughline that leads straight into RLVR.
RLVR in Practice: Build a Verifiable-Reward Loop with GRPO
Now swap the learned reward for a checker. RLVR is the technique behind the reasoning jump on math and code, and the reason it works is that a unit test or an answer comparison is far harder to game than a preference model. Focus on the central claim, that the verifier is the real product and the optimizer choice between GRPO, DAPO and GSPO is a stability question, not a capability one.
Test-Time Compute: Where More Thinking Stops Paying
Close the loop where training-time levers hand off to inference-time ones. Having spent the path making the model better, this step asks how much you can buy at serving time instead, and finds the same ceiling: the payoff from best-of-N and verifier-guided search is bounded by how well you can check an answer. Focus on where each curve saturates, because that is where the budget decision lives.
Latent Reasoning: The Open Problem of Thinking Without Words
Close with where the research goes once the token budget becomes the binding constraint. Every lever in this path, including the inference-time compute in the previous step, spends visible chain-of-thought tokens to buy accuracy, and latent reasoning asks whether the same computation can happen in the model's hidden state instead. Focus on the honest status of the work, a real capability with an unfinished research program behind it, and on what the field trades away in interpretability when the reasoning stops being text.