Unverified50% confidenceBenchmarkExact time
In experiments on a tool-calling task using Qwen2.5-1.5B-Instruct, under the three-reward setting WAS-PO achieved a higher average training reward across all seven random seeds, by 3.8%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/3/2026
First Seen
Valid until: 2/1/2027
Sources
Related Entities
Related Claims
UnverifiedThe later stages of Qwen3 training made extensive use of outcome-reward-based reinforcement learning.73% similarUnverifiedThe Qwen3 training team adopted a progressive training strategy: first completing simple tasks, then scaling up, introducing multi-task training, and gradually increasing task difficulty.70% similarUnverifiedBoth Qwen models trained solely in WildAWM synthetic environments achieved significant improvements on all 8 OOD benchmarks, indicating that the acquired capabilities generalize to unseen computer-use tasks.68% similarUnverifiedThe Qwen2.5-Max-0902 update conducted targeted post-training optimization focused on two areas: programming and collaborative office work.67% similarUnverifiedMusk himself spoiled that Grok 4.8 has a total of 2.5 trillion parameters, is developed on a brand-new training architecture, and that pre-training has now concluded smoothly while the team is working full-speed on reinforcement learning tuning.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130536API
curl https://kongchang.com/api/v1/knowledge/claims/1130536MCP
get_claim(id=1130536)