Unverified50% confidenceBenchmarkExact time
The effect size Cohen's d between task performance under the naive survival reward configuration and a random policy was -10.6.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/7/2026
First Seen
Valid until: 12/6/2026
Sources
Related Claims
Unverified研究者提出长度归一化奖励方法，即在计算奖励分数时除以回答长度，使模型无法通过单纯增加长度获取更高奖励64% similarUnverified由于LLM输出具有随机性且Agent会将上一轮失败结果并入下一轮上下文，基于哈希比对的传统去重机制在Agent场景下完全失效64% similarUnverifiedThe ε-greedy policy selects actions randomly with probability ε and selects the action with the highest current Q-value with probability 1-ε; ε is typically set to 1.0 at the beginning of training and then linearly decayed to 0.01 or 0.1 over millions of frames.63% similarUnverifiedAggregate metrics (such as multi-task average scores) have an inherent ability to dilute failures; when simpler tasks constitute the majority, the average score can mask significant degradation on a small number of high-difficulty tasks.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/870229API
curl https://kongchang.com/api/v1/knowledge/claims/870229MCP
get_claim(id=870229)