🚨Great Paper Alert🚨 GSPO (Group Sequence Policy Optimization)
tbh it's a bit tough on the math, but it's EXCELLENT at explaining the situation
it's an RL algorithm that fixes stability problems with GRPO (R1's algo) to enable training huge models easily
arxiv.org/abs/2507.180...
tbh it's a bit tough on the math, but it's EXCELLENT at explaining the situation
it's an RL algorithm that fixes stability problems with GRPO (R1's algo) to enable training huge models easily
arxiv.org/abs/2507.180...
July 27, 2025 at 12:54 AM
🚨Great Paper Alert🚨 GSPO (Group Sequence Policy Optimization)
tbh it's a bit tough on the math, but it's EXCELLENT at explaining the situation
it's an RL algorithm that fixes stability problems with GRPO (R1's algo) to enable training huge models easily
arxiv.org/abs/2507.180...
tbh it's a bit tough on the math, but it's EXCELLENT at explaining the situation
it's an RL algorithm that fixes stability problems with GRPO (R1's algo) to enable training huge models easily
arxiv.org/abs/2507.180...
Exciting addition to the RLHF Book on my DGX Spark arc is a bunch of single GPU, tinkering scripts for minimal examples of RL, reward models, etc. (and DPO like algorithms soon).
Right now has REINFORCE, RLOO, PPO, GRPO, Dr. GRPO, GSPO, CISPO, standard RM, PRM, ORM.
github.com/natolambert/...
Right now has REINFORCE, RLOO, PPO, GRPO, Dr. GRPO, GSPO, CISPO, standard RM, PRM, ORM.
github.com/natolambert/...
January 25, 2026 at 5:17 PM
Exciting addition to the RLHF Book on my DGX Spark arc is a bunch of single GPU, tinkering scripts for minimal examples of RL, reward models, etc. (and DPO like algorithms soon).
Right now has REINFORCE, RLOO, PPO, GRPO, Dr. GRPO, GSPO, CISPO, standard RM, PRM, ORM.
github.com/natolambert/...
Right now has REINFORCE, RLOO, PPO, GRPO, Dr. GRPO, GSPO, CISPO, standard RM, PRM, ORM.
github.com/natolambert/...
one thing they do very well with this GSPO paper is explaining (and proving!) the problems with GRPO as a means to show why GSPO works so well
it's such a phenomenal writing style. More papers should read like this. Juxtapose your thing against some other popular thing
it's such a phenomenal writing style. More papers should read like this. Juxtapose your thing against some other popular thing
July 27, 2025 at 12:54 AM
one thing they do very well with this GSPO paper is explaining (and proving!) the problems with GRPO as a means to show why GSPO works so well
it's such a phenomenal writing style. More papers should read like this. Juxtapose your thing against some other popular thing
it's such a phenomenal writing style. More papers should read like this. Juxtapose your thing against some other popular thing
z.ai's slime v0.10 - An open-source high-performance inference for large MoE models, powering models like GLM-4.5.
• More KV cache, higher concurrency
• CPU Adam for low-GPU training
• Supports full Megatron + DeepEP
• Supports GSPO for MoE training, TIS for FP8 rollout
• More KV cache, higher concurrency
• CPU Adam for low-GPU training
• Supports full Megatron + DeepEP
• Supports GSPO for MoE training, TIS for FP8 rollout
September 7, 2025 at 2:23 AM
z.ai's slime v0.10 - An open-source high-performance inference for large MoE models, powering models like GLM-4.5.
• More KV cache, higher concurrency
• CPU Adam for low-GPU training
• Supports full Megatron + DeepEP
• Supports GSPO for MoE training, TIS for FP8 rollout
• More KV cache, higher concurrency
• CPU Adam for low-GPU training
• Supports full Megatron + DeepEP
• Supports GSPO for MoE training, TIS for FP8 rollout
Alibaba Qwen team publishes a paper on Group Sequence Policy Optimization (GSPO)
Their stable, efficient, and performant RL algorithm that powers the large-scale RL training of the latest Qwen3 models (Instruct, Coder, Thinking).
Their stable, efficient, and performant RL algorithm that powers the large-scale RL training of the latest Qwen3 models (Instruct, Coder, Thinking).
July 26, 2025 at 12:36 AM
Alibaba Qwen team publishes a paper on Group Sequence Policy Optimization (GSPO)
Their stable, efficient, and performant RL algorithm that powers the large-scale RL training of the latest Qwen3 models (Instruct, Coder, Thinking).
Their stable, efficient, and performant RL algorithm that powers the large-scale RL training of the latest Qwen3 models (Instruct, Coder, Thinking).
✈️ Here’s the plane of the day!
Plane: F-GSPO - Boeing 777-228(ER)
Flight: AF346
From: Paris (CDG) to Montreal (YUL)
#PlaneSpotting #PlaneOfTheDay #FlightToYUL #AviationPhotography #YULSpotting #AviationGeeks #Photography
(taken with a canon 90D with an EF-18-135mm f3.5-5.6)
Plane: F-GSPO - Boeing 777-228(ER)
Flight: AF346
From: Paris (CDG) to Montreal (YUL)
#PlaneSpotting #PlaneOfTheDay #FlightToYUL #AviationPhotography #YULSpotting #AviationGeeks #Photography
(taken with a canon 90D with an EF-18-135mm f3.5-5.6)
November 21, 2024 at 11:00 AM
✈️ Here’s the plane of the day!
Plane: F-GSPO - Boeing 777-228(ER)
Flight: AF346
From: Paris (CDG) to Montreal (YUL)
#PlaneSpotting #PlaneOfTheDay #FlightToYUL #AviationPhotography #YULSpotting #AviationGeeks #Photography
(taken with a canon 90D with an EF-18-135mm f3.5-5.6)
Plane: F-GSPO - Boeing 777-228(ER)
Flight: AF346
From: Paris (CDG) to Montreal (YUL)
#PlaneSpotting #PlaneOfTheDay #FlightToYUL #AviationPhotography #YULSpotting #AviationGeeks #Photography
(taken with a canon 90D with an EF-18-135mm f3.5-5.6)
Happy Pokemon Day! Explore all of the awesome Pokemon merch and deals at the links below!
EE ~ fnkpp.com/EE67
Amzn ~ fnkpp.com/582
GS ~ fnkpp.com/GSPo
FS ~ fnkpp.com/FS
#Ad #FPN #FunkoPOPNews #Funko #FunkoPOP #Pokemon #PokemonDay
EE ~ fnkpp.com/EE67
Amzn ~ fnkpp.com/582
GS ~ fnkpp.com/GSPo
FS ~ fnkpp.com/FS
#Ad #FPN #FunkoPOPNews #Funko #FunkoPOP #Pokemon #PokemonDay
February 27, 2025 at 4:52 PM
Happy Pokemon Day! Explore all of the awesome Pokemon merch and deals at the links below!
EE ~ fnkpp.com/EE67
Amzn ~ fnkpp.com/582
GS ~ fnkpp.com/GSPo
FS ~ fnkpp.com/FS
#Ad #FPN #FunkoPOPNews #Funko #FunkoPOP #Pokemon #PokemonDay
EE ~ fnkpp.com/EE67
Amzn ~ fnkpp.com/582
GS ~ fnkpp.com/GSPo
FS ~ fnkpp.com/FS
#Ad #FPN #FunkoPOPNews #Funko #FunkoPOP #Pokemon #PokemonDay
astonishing: using fp16 instead of bf16 results in more stable training runs as well as a smaller performance gap between training & inference
this is critical for RL, which is mostly inference and very sensitive to reproducible results
this is critical for RL, which is mostly inference and very sensitive to reproducible results
October 31, 2025 at 6:53 PM
astonishing: using fp16 instead of bf16 results in more stable training runs as well as a smaller performance gap between training & inference
this is critical for RL, which is mostly inference and very sensitive to reproducible results
this is critical for RL, which is mostly inference and very sensitive to reproducible results
Beyond PPO - The New Wave of Policy Optimization Techniques for LLM Post-Training by yadnyesh
He goes deep into the policy-optimization rabbit hole and dumped his “policy optmzn techniques beyond ppo” notes into one post. He has covered grpo, dr.grpo, gspo, dapo, cispo, gmpo, rspo, and sapo.
He goes deep into the policy-optimization rabbit hole and dumped his “policy optmzn techniques beyond ppo” notes into one post. He has covered grpo, dr.grpo, gspo, dapo, cispo, gmpo, rspo, and sapo.
Beyond PPO - The New Wave of Policy Optimization Techniques for LLM Post-Training
PPO used to be the default workhorse for RLHF because it’s reasonably stable and easy to reason about, but at LLM post-training scale its tradeoffs start to bite: the critic/value model is expensive t...
ydnyshhh.github.io
December 29, 2025 at 12:32 AM
Beyond PPO - The New Wave of Policy Optimization Techniques for LLM Post-Training by yadnyesh
He goes deep into the policy-optimization rabbit hole and dumped his “policy optmzn techniques beyond ppo” notes into one post. He has covered grpo, dr.grpo, gspo, dapo, cispo, gmpo, rspo, and sapo.
He goes deep into the policy-optimization rabbit hole and dumped his “policy optmzn techniques beyond ppo” notes into one post. He has covered grpo, dr.grpo, gspo, dapo, cispo, gmpo, rspo, and sapo.
Après avoir terminé Kamen Rider Gaim, ma fille et moi venons de commencer à regarder un autre tokusatsu, Kaitou Sentai Lupinranger VS Keisatsu Sentai Patranger, chaque soir. Sauf le samedi, parce que c'est le soir de Kamen Rider Zeztz et d'Ultraman Omega.
Elle m'a rendue aussi addict qu'elle... 📺😅
Elle m'a rendue aussi addict qu'elle... 📺😅
a group of people are standing in front of a car that says gspo
ALT: a group of people are standing in front of a car that says gspo
media.tenor.com
October 4, 2025 at 2:26 AM
Après avoir terminé Kamen Rider Gaim, ma fille et moi venons de commencer à regarder un autre tokusatsu, Kaitou Sentai Lupinranger VS Keisatsu Sentai Patranger, chaque soir. Sauf le samedi, parce que c'est le soir de Kamen Rider Zeztz et d'Ultraman Omega.
Elle m'a rendue aussi addict qu'elle... 📺😅
Elle m'a rendue aussi addict qu'elle... 📺😅
I'm calling it now, GSPO will be the next big hype in LLM RL algos after GRPO.
It makes so much more sense intuitively to work on a sequence rather than on a token level when our rewards are on a sequence level.
It makes so much more sense intuitively to work on a sequence rather than on a token level when our rewards are on a sequence level.
July 26, 2025 at 7:41 PM
I'm calling it now, GSPO will be the next big hype in LLM RL algos after GRPO.
It makes so much more sense intuitively to work on a sequence rather than on a token level when our rewards are on a sequence level.
It makes so much more sense intuitively to work on a sequence rather than on a token level when our rewards are on a sequence level.
People of Bluesky! Let's get to know each other better. What is your all time favorite TV show?
Qoute share your answer only in GIF form.
Qoute share your answer only in GIF form.
March 24, 2025 at 10:53 PM
People of Bluesky! Let's get to know each other better. What is your all time favorite TV show?
Qoute share your answer only in GIF form.
Qoute share your answer only in GIF form.
July 30, 2026 at 8:52 PM
FIRSTだのNEXTだのには、コレが足りなかったんだよなあ。今更だけど。
www.youtube.com/watch?v=J45b...
www.youtube.com/watch?v=J45b...
【完全版】シン・仮面ライダー次回予告集
YouTube video by 帰ってきたGSPO
www.youtube.com
December 11, 2024 at 9:15 AM
FIRSTだのNEXTだのには、コレが足りなかったんだよなあ。今更だけど。
www.youtube.com/watch?v=J45b...
www.youtube.com/watch?v=J45b...
Anne Margot “Honey” Lantree (1943-2018) of the band ‘Honeycombs’ faced fierce sexual prejudice by those who wouldn’t believe that a woman could be a rock drummer. She blazed trails 1963-1967 for female drummers to follow. Ringo, for one, didn’t believe it.
www.independent.co.uk/news/obituar...
www.independent.co.uk/news/obituar...
Honey Lantree: Drummer of The Honeycombs who forced sexism to beat a retreat
In an era when men thought women should only sing in bands, there was incredulity that she could master the kit
www.independent.co.uk
May 19, 2026 at 9:16 AM
lrp: that game has such an awesome theme song
www.youtube.com/watch?v=GSpO...
www.youtube.com/watch?v=GSpO...
𝐑𝐎𝐀𝐃 𝐀𝐕𝐄𝐍𝐆𝐄𝐑 - Ost 𝐎𝐫𝐢𝐠𝐢𝐧𝐚𝐥 𝐒𝐨𝐮𝐧𝐝𝐭𝐫𝐚𝐜𝐤 𝐓𝐡𝐞𝐦𝐞 𝐒𝐨𝐧𝐠 𝐌𝐞𝐠𝐚-𝐂𝐃 - 𝐑𝐨𝐚𝐝 𝐁𝐥𝐚𝐬𝐭𝐞𝐫 𝐈𝐧𝐭𝐫𝐨 𝐀𝐫𝐜𝐚𝐝𝐞 𝐰𝐢𝐭𝐡 𝐋𝐲𝐫𝐢𝐜𝐬
YouTube video by 𝐃𝐌𝐓 | 𝔾𝕣𝕒𝕡𝕙𝕚𝕔 𝔸𝕕𝕧𝕖𝕟𝕥𝕦𝕣𝕖𝕤
www.youtube.com
July 18, 2025 at 8:43 PM
lrp: that game has such an awesome theme song
www.youtube.com/watch?v=GSpO...
www.youtube.com/watch?v=GSpO...
dni if you support patoranger over lupinranger
a group of people are standing in front of a car that says gspo
ALT: a group of people are standing in front of a car that says gspo
media.tenor.com
August 4, 2025 at 8:42 PM
dni if you support patoranger over lupinranger
This is the opening to the next Sentai series ain't it? LOL Enjoy Zack! =D
a group of people are standing in front of a car that says gspo
ALT: a group of people are standing in front of a car that says gspo
media.tenor.com
March 16, 2025 at 2:06 PM
This is the opening to the next Sentai series ain't it? LOL Enjoy Zack! =D
Artificial Intelligence #287
www.linkedin.com/pulse/artifi...
www.linkedin.com/pulse/artifi...
Artificial Intelligence #287
Hey, in this issue: Genie 3 from Google as a new frontier for world models; Qwen team claims DeepSeek's GRPO is ill-posed and proposes GSPO for Qwen3; AI took on the math olympiad—but mathematicians a...
www.linkedin.com
August 7, 2025 at 10:56 PM
Artificial Intelligence #287
www.linkedin.com/pulse/artifi...
www.linkedin.com/pulse/artifi...
THIS is part of why that excuse Simon Bennett gave for why Power Rangers can’t adapt Lupinranger Vs Patranger still pisses me off
March 19, 2025 at 8:54 PM
THIS is part of why that excuse Simon Bennett gave for why Power Rangers can’t adapt Lupinranger Vs Patranger still pisses me off
💡 Building on this insight, we adapt GRPO to online finetuning of DTs, introducing:
• Sub-trajectory optimization → better credit assignment
• Sequence-level likelihood objectives (concurrent w/ GSPO) → stability & efficiency
• Active sampling → improved exploration in uncertain regions
• Sub-trajectory optimization → better credit assignment
• Sequence-level likelihood objectives (concurrent w/ GSPO) → stability & efficiency
• Active sampling → improved exploration in uncertain regions
October 14, 2025 at 7:04 PM
💡 Building on this insight, we adapt GRPO to online finetuning of DTs, introducing:
• Sub-trajectory optimization → better credit assignment
• Sequence-level likelihood objectives (concurrent w/ GSPO) → stability & efficiency
• Active sampling → improved exploration in uncertain regions
• Sub-trajectory optimization → better credit assignment
• Sequence-level likelihood objectives (concurrent w/ GSPO) → stability & efficiency
• Active sampling → improved exploration in uncertain regions
Compared to GRPO, GSPO offers significant advantages in stability, efficiency, performance, and infra-friendliness. Furthermore, it fundamentally and naturally resolves the stability issues in the RL training of large MoE models
Paper: Group Sequence Policy Optimization
arxiv.org/abs/2507.18071
Paper: Group Sequence Policy Optimization
arxiv.org/abs/2507.18071
July 26, 2025 at 12:36 AM
Compared to GRPO, GSPO offers significant advantages in stability, efficiency, performance, and infra-friendliness. Furthermore, it fundamentally and naturally resolves the stability issues in the RL training of large MoE models
Paper: Group Sequence Policy Optimization
arxiv.org/abs/2507.18071
Paper: Group Sequence Policy Optimization
arxiv.org/abs/2507.18071
Как мы обеспечили +33% к точности на сложных SQL-запросах Генератор SQL на базе LLM — понятный продукт с понятной ...
#chase-sql #grpo #gspo #reasoning #sql #RL #skyrl-sql #sql-генератор #sqlfuse #генерация #sql
Origin | Interest | Match
#chase-sql #grpo #gspo #reasoning #sql #RL #skyrl-sql #sql-генератор #sqlfuse #генерация #sql
Origin | Interest | Match
Как мы обеспечили +33% к точности на сложных SQL-запросах
www.pvsm.ru
October 8, 2025 at 1:34 PM
Как мы обеспечили +33% к точности на сложных SQL-запросах Генератор SQL на базе LLM — понятный продукт с понятной ...
#chase-sql #grpo #gspo #reasoning #sql #RL #skyrl-sql #sql-генератор #sqlfuse #генерация #sql
Origin | Interest | Match
#chase-sql #grpo #gspo #reasoning #sql #RL #skyrl-sql #sql-генератор #sqlfuse #генерация #sql
Origin | Interest | Match
📬You have mail!📥It's the Media issue of the #gsportNewsletter feat. @espncricinfo.bsky.social scribe, @firdosem.bsky.social; Zimbabwe's Tatenda Gondo on winning at #gsport18 and advancing the African women in sport agenda; and former Basketball player and BSA Media Liaison, Mpho Mashego #GoGirls
November 27, 2024 at 4:51 PM
📬You have mail!📥It's the Media issue of the #gsportNewsletter feat. @espncricinfo.bsky.social scribe, @firdosem.bsky.social; Zimbabwe's Tatenda Gondo on winning at #gsport18 and advancing the African women in sport agenda; and former Basketball player and BSA Media Liaison, Mpho Mashego #GoGirls