@seclink: Keywords, interested friends can query themselves: RLHF -> RLAIF -> RLTHF. Starting in 2025, RLTHF is adopted, reducing expert working hours by 93%.
Summary
The article briefly introduces the evolution of AI training methods from RLHF to RLAIF to RLTHF and predicts that by 2025, RLTHF will significantly reduce expert working hours.
View Cached Full Text
Cached at: 08/15/26, 11:45 AM
Key terms for those interested to look up:
RLHF → RLAIF → RLTHF.
Starting in 2025, RLTHF reduces expert work hours by 93%.
Similar Articles
@Xudong07452910: A classic challenge in RL training of LLM agents: after a long task fails, where should the model start learning? The final reward can usually only tell the agent 'success' or 'failure', but it's hard to pinpoint which intermediate judgments are worth keeping and which actions led the entire trajectory astray. This paper proposes SEED, using 'self-evolving online distillation...'
This paper proposes SEED, a method that internalizes post-hoc skills from trajectories into model parameters through self-evolving online distillation, solving the reward sparsity problem in long-horizon RL training, achieving significant improvements on benchmarks such as ALFWorld.
@0xcherry: https://x.com/0xcherry/status/2067610347633025281
This article analyzes the reasons behind the performance leap of Zhipu GLM-5.2, suggesting that its 40B activation parameters provide greater effective capacity after accounting for fixed overhead, making RL post-training more effective. It also reviews the history of Chinese AI model development and notes that the large model approach ultimately prevailed.
@freeman1266: You don't need math to understand most AI papers—just understand this chain: token → embedding → position encoding → attention → FFN → residual stream → next-token prediction. LLMs essentially stack Transf…
A Chinese science tweet that intuitively explains the core chain of LLMs (Large Language Models): from token, embedding, position encoding, attention, FFN to residual stream and next-token prediction, helping readers without a math background understand AI papers.
@Gorden_Sun: Achieving heuristic learning through coding agents. Continuously maintain and iterate a system of programmatic strategies using a coding agent to replace gradient updates in neural networks. In tests, this approach reached baseline levels of Deep RL. It may become the next paradigm following "pre-training → RLHF → large-scale RL." Heuristic learning has existed in the past, but...
The article proposes using coding agents to maintain and iterate a system of programmatic strategies to replace neural network gradient updates. This approach achieved baseline performance in Deep RL tests and is considered a potential new paradigm following pre-training and RLHF.
@VincentLogic: If Ilya Is Right, the Three Strongest Consensuses in AI Over the Past Few Years Might All Be Wrong: Scaling Is No Longer the Universal Answer. High Benchmark Scores Don't Equal True Intelligence. RL Might Even Be Making Models 'Dumber'. This Interview, Called 'the Last Interview Before Ilya Disappeared'...
Ilya Sutskever suggested in an in-depth interview that the three core consensuses of the AI industry over the past few years could all be mistaken: Scaling is no longer a silver bullet, high benchmark scores do not equate to real intelligence, and RL is instead making models 'dumber'. He believes the dividends from pre-training and RL are nearly exhausted, AI has re-entered the era of research, and true superintelligence should possess a strong learning capability like a gifted teenager, not a static repository of knowledge.