从RLHF到RLVR：大模型写作能力为何变差

一定程度解释了大模型写作能力越来越差的原因。

一定程度解释了大模型写作能力越来越差的原因。 Kun Chen @kunchenguid if you've been using latest frontier LLMs, it's almost certain that you would have noticed by now the newer models have become worse to talk to they're more robotic, they speak jargons, they spits out verbose text, and do stuff you didn't ask for how did that happen? well, i'm not the person who trained those models so i can't speak for certain, but i've known enough evidence that gives me a well-educated guess, and i thought it's interesting to share as a crash course of modern LLM training pipelines so here we go let's wind back to 2020. GPT-2 and GPT-3 already came out and were widely available, but they could only predict one token at a time - that's what LLMs are at their core token prediction was offered via API, but there was nothing you could "talk to". so while it generated a lot of excitement in the academic field due to the emergent intelligence, it didn't have any wide adoption in 2022, ChatGPT changed all that. the research work that led to ChatGPT was a model initially named "InstructGPT". it took GPT-3 as the intelligent base, and used reinforcement learning with human feedback (RLHF) to teach the models how to "chat" the core idea of RLHF is that you ask the model to generate a few responses, and then let real humans pick which one they like. do this over and over again, and you get a model that knows how to talk worth noting even as early as InstructGPT, research found that making the model more pleasant to talk to will reduce their pure academic capabilities. this was called "alignment tax", which is an interesting thing we'll come back to in a bit there were various techniques done to minimize the reliance on humans, but ultimately the reward is modeled after human preference, making these AI assistants easy to talk to so remember this: RLHF = training the model to be likable by humans in 2024, there was an inflection point introduced by claude sonnet 3.5 which was the first model that can kind of autonomously finish coding tasks. it led to the first wave of viable "coding agents" the way sonnet 3.5 achieved this was by training the model with a harness (now it's called an agent) that has bash and file editing tools, throw the agent into a virtual machine, give it a task, and let it try to complete it. these tasks all have a machine-verifiable outcome predefined, mostly via test cases, that can validate whether the model really finished the task or not then you let the model do billions and billions of attempts in such virtual environments, and some of them would succeed by chance. you keep the successful agent sessions, and use reinforcement learning to teach the model to do that more, and boom - you get a coding agent that is called reinforcement learning with verifiable rewards (RLVR). if you look closely, you'll see that in this RLVR process, the final text response from the model doesn't matter AT ALL, as long as the code written by the agent could pass the test. it could talk like a jerk and it would still be rewarded so remember this: RLVR = training the model to be accepted by machines late 2024 and early 2025, we saw o1 and deepseek R1 came out as the first wave of "reasoning models". this article is getting long so i'm not diving into reasoning models now, but just know that reasoning models also relied heavily on RLVR to scale the training process - let the model think before taking action, and if the thinking led to a machine verifiable outcome, reward the thinking trace and teach the model to think like that more often the biggest difference between RLVR and RLHF is that RLVR is more scalable. human feedback is expensive to get, especially in domains where only an expert can have a valid opinion on which result is good with RLHF, if we let the model generate 100 responses, then a human has to review all 100 responses to pick which is good with RLVR, the human (or sometimes an AI) would define a task and verifier only once, and the model can generate a million responses - the machine verifier will pick which responses are good in an automated way so as a result, RLVR is becoming more and more dominant in newer models' training pipeline if you put all these things together: - RLHF = training the model to be likable by humans - RLVR = training the model to be accepted by machines - RLVR is more scalable - "alignment tax" says "likable by humans" makes the model do worse on verifiable tasks now you see why the newer models are becoming less and less likable? this is not just a "frontier labs screwed up their model training" problem - this is a war between machines and humanity, and humanity is losing we chased after benchmarks, when none of the benchmarks measure whether humans actually enjoy working with the model we use machines to decide which AI response is better because that's easier and cheaper, when we have no way of making sure those machines actually represent what we humans want we let AI go dark in a virtual environment on its own and complete predefined tasks at all costs, when in reality we often cannot define a verifiable outcome upfront, and need AI to work with us along the way i don't have a good solution to this, but i want to call for awareness that we're starting to witness a failure in aligning super intelligence right in front of our eyes this war between machines vs humanity is one we really can't afford to lose 🔗 View Quoted Tweet 💬 2 🔄 0 ❤️ 2 👀 285 📊 2 ⚡