Unverified50% confidenceFactExact time

The NGU algorithm superimposes external rewards with two types of intrinsic rewards: a short-term reward based on intra-episode novelty and a long-term reward based on lifetime novelty.

1

Sources

50%

Confidence

Long-term

Relevance

9/15/2026

First Seen

Sources

Reinforcement Learning for LLMs: A 'Never Give Up' Strategy for Hard Problems

Related Entities

Never Give Up (NGU)concept DeepMindcompany

Related Claims

Unverified贝尔曼方程将长期回报分解为即时奖励加上未来折扣回报的递归关系70% similar Unverified模型在PPO优化中通过增加输出长度获取更高奖励的现象被称为reward hacking（奖励攻击）70% similar UnverifiedReward engineering is essentially a highly iterative process: design the reward, observe behavior, discover loopholes, correct the reward, and repeat.70% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/924066

API

curl https://kongchang.com/api/v1/knowledge/claims/924066

MCP

get_claim(id=924066)