GSPO (Mixture-of-experts routing): superseded — cited as a baseline and beaten by newer methods. 1 paper(s) critique it, 3 beat it on benchmarks — #29 of 1370 most-superseded. Sub-problem: cluster led by SteerMoE. Newer alternatives in the same sub-problem include PADD, MESA, PR2, RouteHijack, MASCing.

Superseded baseline#29 of 1,370 most-superseded

GSPO

Mixture-of-experts routing

superseded — cited as a baseline and beaten by newer methods

1 papers critique it · 3 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites GSPO as a baseline.

“However, these methods do not explicitly control the impact of routing drift on off-policy updates.”
— Towards Stable and Effective Reinforcement Learning for Mixture-of-Experts

Beaten on benchmarks

Head-to-head results where a newer method reports beating GSPO. Values are copied from the source paper's tables — verify against the cited paper.

GMPO+RS (RSPO) beats GSPO · Pass@1 [Math]
77.1 vs 76.4
Towards Stable and Effective Reinforcement Learning for Mixture-of-Experts
GMPO+RS (RSPO) beats GSPO · Pass@1 [Code]
85.2 vs 80.4
Towards Stable and Effective Reinforcement Learning for Mixture-of-Experts
GRPO + \method beats GSPO · Average [Off-2]
46.80 vs 35.76
PR2: Predictive Routing Replay for MoE-Based LLM Reinforcement Learning
GRPO + \method beats GSPO · Average [Off-4]
46.24 vs 36.51
PR2: Predictive Routing Replay for MoE-Based LLM Reinforcement Learning
GRPO + \method beats GSPO · Average [Off-8]
41.90 vs 33.16
PR2: Predictive Routing Replay for MoE-Based LLM Reinforcement Learning
GRPO + \method beats GSPO · Average [Moonlight-16B-A3B Off-2]
68.88 vs 66.25
PR2: Predictive Routing Replay for MoE-Based LLM Reinforcement Learning
GRPO + \method beats GSPO · Average [Moonlight-16B-A3B Off-4]
62.64 vs 50.33
PR2: Predictive Routing Replay for MoE-Based LLM Reinforcement Learning
GRPO + \method beats GSPO · Average [Moonlight-16B-A3B Off-8]
53.48 vs 50.79
PR2: Predictive Routing Replay for MoE-Based LLM Reinforcement Learning
GRPO + \method beats GSPO · Average [OLMoE-1B-7B Off-2]
47.46 vs 47.06
PR2: Predictive Routing Replay for MoE-Based LLM Reinforcement Learning
GRPO + \method beats GSPO · Average [OLMoE-1B-7B Off-4]
47.89 vs 47.57
PR2: Predictive Routing Replay for MoE-Based LLM Reinforcement Learning
GRPO + \method beats GSPO · Average [OLMoE-1B-7B Off-8]
47.60 vs 47.07
PR2: Predictive Routing Replay for MoE-Based LLM Reinforcement Learning
PADD (Ours) beats GSPO · Average [Qwen family]
80.2 vs 76.3
PADD: Path-Aligned Decompression Distillation for Non-Router Teacher to Guide MoE Student Learning

Newer alternatives

Recent methods in the same sub-problem, not yet superseded in the knowledge base.