Expired50% confidenceSolutionExact time
The team proposed the Faithful GRPO method, which introduces consistency rewards and grounding rewards and balances constraints by learning Lagrangian multiplier coefficients, in order to maintain reasoning faithfulness while improving performance
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/17/2026
First Seen
This claim has expired. It is kept for the record and no longer describes the current state.
Valid until: 10/15/2026(expired)
Sources
Related Claims
UnverifiedAt the training level, reasoning-enhanced models typically leverage reinforcement learning algorithms such as GRPO (Group Relative Policy Optimization), using the correctness of the final answer as the reward signal75% similarUnverifiedJustGRPO的核心修复方案是在自回归顺序下训练GRPO，推理阶段保持并行解码65% similarUnverifiedThe Grok series adopts a Mixture of Experts (MoE) architecture, expanding parameter scale while maintaining lower inference costs by dynamically activating different expert sub-networks.64% similarUnverifiedGRPO (Group Relative Policy Optimization) is better suited than traditional RLHF for multi-step decision-making scenarios because it can simultaneously evaluate the relative merits of multiple execution paths.62% similarUnverified推理增强方法通常借助RLHF或GRPO等算法，在MATH、HumanEval、AIME等基准上取得突破62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/544830API
curl https://kongchang.com/api/v1/knowledge/claims/544830MCP
get_claim(id=544830)