The strangest thing about language model sampling is that "randomness" and "creativity" aren't the same thing — but the parameters that control them are often confused for being identical. Understanding why they're different is the key to understanding why they're both interesting. When a language model finishes processing your prompt, it doesn't "know" the next word. It produces a probability distribution across its entire vocabulary — maybe 50,000 tokens — where each token has some likelihood of being next. The word "Paris" might have a 40% chance, "London" 20%, "Rome" 15%, and then a long tail of thousands of other tokens with tiny probabilities. What happens next is the sampling decision, and this is where temperature and top-p live. Temperature reshapes the entire distribution before you sample from it. Mathematically, it divides every logit (the raw score before softmax) by a number. A temperature of 1.0 leaves the distribution untouched. A temperature of 0.5 sharpens it — the high-probability tokens get relatively higher, the low-probability tokens get crushed toward zero. A temperature of 2.0 flattens it — the distribution spreads out, making unlikely tokens more competitive. So temperature is literally a dial that compresses or expands the probability landscape. Low temperature makes the model more "decisive." High temperature makes it more "adventurous." Top-p (also called nucleus sampling) works differently. Rather than reshaping the distribution, it truncates it. You take all tokens sorted by probability, add them up until you hit some cumulative threshold — say, 90% — and then throw away everything below that cutoff. If the model is very confident and the top 5 tokens already account for 90% of the probability mass, you sample only from those 5. If the model is uncertain and it takes 500 tokens to reach 90%, you sample from 500. Top-p is adaptive: it responds to the model's own confidence level in a way temperature doesn't. The practical difference matters enormously. Temperature at 2.0 might make a confident prediction chaotic by amplifying the entire tail. Top-p at 0.9 would still restrict you to the high-confidence region in that same case. They're controlling different things: temperature adjusts how peaked the distribution is, top-p adjusts how wide the candidate pool is. Now here's where it gets genuinely interesting. In reinforcement learning, the explore-exploit tradeoff is one of the central problems: when should an agent try something new (explore) versus stick with what it knows works (exploit)? A classic solution is the epsilon-greedy strategy — with probability ε, take a random action; otherwise, take the best known action. Another is softmax action selection, where you sample actions proportional to their estimated value, with a temperature parameter controlling how much you favor high-value actions. Sound familiar? Temperature sampling in language models is softmax action selection. The RL temperature parameter and the LLM temperature parameter are literally the same mathematical operation applied to the same kind of problem: you have a scored set of options, and you need to decide how deterministically to pick the best one versus sample from the broader distribution. High temperature = explore. Low temperature = exploit. Top-p maps more loosely onto a different RL concept: confidence-based exploration, where you only consider actions within some threshold of the best known option. When an RL agent is uncertain (high entropy over action values), you expand the candidate set; when it's confident, you narrow it. This is exactly what top-p does adaptively. The deeper parallel is that both language models and RL agents face the same fundamental tension: the optimal next token (or action) is often not the most probable one. Pure exploitation — always taking the highest-probability token — produces text that's coherent but repetitive, the language model equivalent of an agent that found one good strategy and never discovered there were better ones. Pure exploration produces gibberish, the equivalent of an agent that never learns anything because it keeps trying random things. The practical upshot for inference engineering: these parameters aren't interchangeable dials. Temperature changes the shape of the distribution; top-p changes the size of the candidate pool. For creative tasks with high uncertainty, high top-p matters more than high temperature. For tasks where you want controlled variation around a confident prediction, temperature is the right lever. Most production systems combine both, treating them as orthogonal controls over different aspects of the same sampling problem. The thing about temperature and top-p is that they're not really "creativity dials" — they're uncertainty management strategies, and the fact that they mirror RL's explore-exploit solutions isn't a coincidence. Both problems are fundamentally about how to act under uncertainty when you have a probability distribution over outcomes and need to decide how much to trust the mode.