Buy Crypto
Pay with
USD
USD
Buy & Sell
Hot
Supports Visa, Mastercard, SEPA & more
P2P
0 Fees
Flexible trading, zero fees
Gate Card
Use your crypto for payments worldwide
Markets
Trade
Basic
Advanced
Futures
Futures
Access hundreds of perpetual contracts
CFD
Gold
One platform for global traditional assets
Options
Hot
Trade European-style vanilla options
Unified Account
Maximize your capital efficiency
Demo Trading
Introduction to Futures Trading
Learn the basics of futures trading
Futures Events
Join events to earn rewards
Demo Trading
Use virtual funds to practice risk-free trading
Stocks
CFD
U.S. stock CFD derivatives
US Stocks
Access real US stocks and ETFs
HK Stocks
Trade quality Hong Kong-listed stocks
Korean Stocks
SK Hynix
Real Korean stocks and top assets
Stock Futures
High leverage, 24/7 trading
Tokenized Stocks
Backed by real stock assets
IPO Access
Unlock full access to global stock IPOs
GUSD
Mint GUSD for Treasury RWA yields
Stocks Activities
Trade Popular Stocks and Unlock Generous Airdrops
Earn
Launch
CandyDrop
Collect candies to earn airdrops
Launchpool
Quick staking, earn potential new tokens
HODLer Airdrop
Hold GT and get massive airdrops for free
IPO Access
Unlock full access to global stock IPOs
Alpha Points
Trade on-chain assets and earn airdrops
Futures Points
Earn futures points and claim airdrop rewards
Investment
Square
Square
Post, share, and explore crypto trends
Live
moments live
Live crypto market analysis
Chat
Chat with crypto traders
News
What is happening in crypto
flower_head
More
Promotions
AI
Others
TradFi
2026 World Cup
Rewards
Square
Latest
Hot
News
Profile

V4 post-training upgrade: OPD replaces hybrid RL, distilling more than ten expert models into one

MeNews
robot
Abstract generation in progress
ME News reports that on April 24 (UTC+8), according to Dongcha Beating monitoring, DeepSeek V4's post-training methodology has undergone major changes: the mixed RL phase in V3.2 has been completely replaced by On-Policy Distillation (OPD). The new process consists of two steps. The first step involves training domain expert models separately for areas such as mathematics, code, Agent, and instruction following based on the V3.2 pipeline. Each expert is first fine-tuned and then undergoes reinforcement learning using GRPO. The second step uses multi-teacher OPD to distill the capabilities of more than ten experts into a unified model: on the trajectories generated by the student itself, reverse KL divergence-based full-vocabulary logit distillation is performed for each teacher. Through logit-level alignment, the weights of multiple experts are merged into a unified parameter space, avoiding the common capability conflicts seen in traditional weight merging and mixed RL. The report also introduces a Generative Reward Model (GRM): for tasks where rule-based verification is difficult, instead of training a traditional scalar reward model, the GRM is trained using rubric-guided RL data, allowing the actor network to simultaneously handle generation and evaluation capabilities. With a small amount of diverse human annotations, it can generalize to complex tasks. (Source: BlockBeats)
DEEPSEEK-2.53%
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • Comment
  • Repost
  • Share
Comment
Add a comment
Add a comment
No comments

  • Pinned

Sitemap