If you use an AI assistant at work, you have already felt the result of a step most owners never hear about. Before a model answers you, it goes through two stages. In the first, it reads an enormous amount of text and learns to predict what comes next. In the second, called post-training, it is shaped to behave: to follow instructions, to be useful, to say no to some things, to show its working on a hard problem.


Two methods do a lot of that second stage, and their names are easy to confuse: RLHF and RLVR. This post explains both in plain terms, as of November 2026. It is not a technical guide. It is meant to help you judge what an AI tool is good at, and where a person still needs to look.


The short version


  • RLHF is Reinforcement Learning from Human Feedback. People compare the model's answers, and the model is tuned toward the answers people prefer.
  • RLVR is Reinforcement Learning with Verifiable Rewards. A program checks the answer, and the model is rewarded only when the check passes.

One learns what people like. The other learns what is provably right, but only where "right" can be checked by a program.



In RLHF the judge is a model trained on human preferences. In RLVR the judge is a checker, such as a test or a calculation.


How RLHF works


The best-known description is the 2022 paper from OpenAI, "Training language models to follow instructions with human feedback." The authors started with GPT-3, showed it examples of good answers written by people, then collected rankings in which people compared several outputs for the same prompt. A second model learned to predict those rankings. The language model was then tuned to produce outputs that this "reward model" scored highly.


The headline result is worth knowing: in human evaluations, outputs from a 1.3 billion parameter model trained this way were preferred to outputs from the original 175 billion parameter GPT-3. The authors also said the model still made simple mistakes. Both halves matter. Feedback made the model far more useful to the people judging it. It did not make it reliable.


RLHF is why chat assistants feel helpful and polite instead of rambling on from your sentence. It is good at the things that have no single right answer: tone, clarity, how much detail to give, when to decline.


Where RLHF stops


  • People are not always right, and not always consistent. Raters can prefer a confident answer to a correct one, or a longer answer to a better one.
  • The reward model is a stand-in. It is a learned guess at what people like. A 2022 paper by Leo Gao, John Schulman and Jacob Hilton ("Scaling Laws for Reward Model Overoptimization") found that optimizing hard against an imperfect reward model can make the real quality worse, in line with Goodhart's law: when a measure becomes a target, it stops being a good measure.
  • It is expensive to collect. Human judgment takes time and money at scale.

How RLVR works


Some tasks have an answer a program can verify. A math problem has a result. Code either passes its tests or it does not. An instruction like "answer in exactly three bullet points" can be counted.


RLVR uses that. The name comes from Tülu 3, an open post-training recipe published in November 2024, which describes RLVR as a new method. The reward is simple: if the model's answer is verified correct, it gets a reward; otherwise it gets none. There is no reward model standing in for a person. The Tülu 3 authors report that this step improved results on math and instruction-following tests (GSM8K, MATH and IFEval), by modest amounts in their tables.


The idea reached a wider audience with DeepSeek-R1, published in January 2025 (and later in Nature). The authors report that reinforcement learning on tasks with verifiable answers, in mathematics, coding competitions and STEM fields, could develop reasoning behaviour such as checking its own work, without human-written examples of the reasoning steps. They report that the resulting model outperformed models trained on human demonstrations on those verifiable tasks. These are the authors' own results on their own tests.


Where RLVR stops


  • It only works where a checker exists. Whether an email sounds warm, or a quote is fair to the customer, has no program that returns pass or fail.
  • The checker can be wrong or gamed. A weak test can be satisfied without solving the real problem. The same logic as Goodhart's law applies: a model rewarded for passing a check will find the cheapest way to pass it.
  • A correct answer is not a good decision. Verifying that 2,340 plus 780 is 3,120 says nothing about whether you should send the invoice.

Why the difference matters for agents


An assistant that chats and an agent that does work have different needs. A lot of routine business work is checkable:


  • Does the invoice total equal the sum of its lines, and does it match the purchase order?
  • Does the bank deposit equal the payments you applied?
  • Does the record in the CRM match the signed form, field by field?
  • Did every required field get filled in?

This is where verifiable rewards are a good sign. Models shaped by RLVR are being trained on exactly the kind of work where the answer can be checked, which is part of why recent models are stronger at math, code and structured tasks. It is a reasonable bet, not a promise, that data entry, reconciliation and other rule-bound work benefits too. Your own tasks still need testing on your own records.


But look at what the examples have in common: the check sits outside the model. A good system does not trust the model to say it was right. It runs the check. Add up the lines. Compare the totals. Reject the draft if the numbers differ.


What this means for your business


1. Split your work into checkable and judgment work. Checkable work, such as totals, matching, required fields and dates, can be verified by code and is a good first place to let an agent run. Judgment work, such as tone, pricing, exceptions and what to say to an upset customer, needs a person.

2. Put a checker beside every agent. Ask any vendor: what verifies the output before it counts? If the answer is "the model is very accurate", keep asking.

3. Keep people in the approval seat for anything that leaves the building. RLHF makes models agreeable; RLVR makes them better at checkable tasks. Neither makes a model accountable. A person who approves a message, a price or a payment is.

4. Measure on your own examples. Take 30 to 50 real cases from your own work, with known right answers, and run any tool on them before you rely on it.

5. Do not read the acronym as a guarantee. "Trained with verifiable rewards" describes a method. It is not a promise about your invoices.


How this shows up at Humonix


Humonix does not train foundation models. We use models other labs have trained, so this is not a claim about our own training. We borrow the idea: wherever an output can be checked by code, it should be, and wherever it cannot, a person approves it.


In practice, that is why anything public, priced or paid waits in the approvals queue in Hue, and why each feature on a module page is labelled Live, Early access or Coming rather than all described as working. Know, the one record, is what a check compares against, so a draft is held to your facts and not to the model's memory. See Know, Hue and the white paper The Agentic Enterprise. If you want the plain-language terms, the glossary covers them.


None of this guarantees an outcome. It narrows what can go wrong and puts a person where the judgment is.


Sources



This post is general information, not technical or legal advice.