Tongyi Lab Debuts FIPO Algorithm, 32B Model Inference Outperforms o1-mini

Tongyi Lab's Intelligent Computing team today unveiled a new post-training algorithm for large language models — FIPO (Future-KL Influenced Policy Optimization). This algorithm introduces a novel "Future-KL" mechanism that effectively resolves the common technical challenge of "inference length stagnation" in pure reinforcement learning (Pure RL) training.
During training for long-text reasoning and complex logical alignment, traditional reinforcement learning often fails to precisely identify key decision points in lengthy sequences. The FIPO algorithm, developed by the Tongyi team, applies differentiated reward allocation to key tokens, encouraging the model to adopt a more forward-looking approach during chain-of-thought (CoT) generation.
Experimental results indicate that, in a pure reinforcement learning setup with a 32B-scale model, the model using the FIPO algorithm has already outperformed similarly sized models like DeepSeek-Zero-MATH and OpenAI's o1-mini, representing significant advancement in the logical reasoning and mathematical computation abilities of domestic large models.
Currently, the competition among large models is shifting from pre-training scale to deep alignment at inference time. The release of FIPO not only offers a new way to assess the quality of "thinking processes" in logical reasoning models, but also signals that the open-source community and leading domestic labs are gradually forging an independent technological path in their pursuit of world-class reasoning models.
Related article
Stripe to Acquire OpenRouter as AI Model Routing Gains Momentum
Stripe has agreed to acquire OpenRouter, an AI model-routing platform that gives developers access to hundreds of models through a single interface. The deal adds model selection and routing to Stripe’s existing work around AI usage and token-based b
Former OpenAI Researcher Barrett Zoph Returns to Google to Support Gemini Development
Barret Zoph, who previously abruptly left and joined OpenAI due to a disagreement with Thinking Machines Lab CEO Miriam Mularati, has once again changed employers. Alphabet's Google announced that Zoph will return as Vice President of Research, bring
How Recursive Intelligence Raised $335M at $4B Valuation in 4 Months
The co-founders of startup Ricursive Intelligence seemed destined to be co-founders.Anna Goldie, CEO, and Azalia Mirhoseini, CTO, are so well-known in the AI community that they were among those AI engineers who “got those weird emails from Zuckerber
Related Special Topic Recommendations
Comments (0)
0/500

Tongyi Lab's Intelligent Computing team today unveiled a new post-training algorithm for large language models — FIPO (Future-KL Influenced Policy Optimization). This algorithm introduces a novel "Future-KL" mechanism that effectively resolves the common technical challenge of "inference length stagnation" in pure reinforcement learning (Pure RL) training.
During training for long-text reasoning and complex logical alignment, traditional reinforcement learning often fails to precisely identify key decision points in lengthy sequences. The FIPO algorithm, developed by the Tongyi team, applies differentiated reward allocation to key tokens, encouraging the model to adopt a more forward-looking approach during chain-of-thought (CoT) generation.
Experimental results indicate that, in a pure reinforcement learning setup with a 32B-scale model, the model using the FIPO algorithm has already outperformed similarly sized models like DeepSeek-Zero-MATH and OpenAI's o1-mini, representing significant advancement in the logical reasoning and mathematical computation abilities of domestic large models.
Currently, the competition among large models is shifting from pre-training scale to deep alignment at inference time. The release of FIPO not only offers a new way to assess the quality of "thinking processes" in logical reasoning models, but also signals that the open-source community and leading domestic labs are gradually forging an independent technological path in their pursuit of world-class reasoning models.
Stripe to Acquire OpenRouter as AI Model Routing Gains Momentum
Stripe has agreed to acquire OpenRouter, an AI model-routing platform that gives developers access to hundreds of models through a single interface. The deal adds model selection and routing to Stripe’s existing work around AI usage and token-based b
Former OpenAI Researcher Barrett Zoph Returns to Google to Support Gemini Development
Barret Zoph, who previously abruptly left and joined OpenAI due to a disagreement with Thinking Machines Lab CEO Miriam Mularati, has once again changed employers. Alphabet's Google announced that Zoph will return as Vice President of Research, bring
How Recursive Intelligence Raised $335M at $4B Valuation in 4 Months
The co-founders of startup Ricursive Intelligence seemed destined to be co-founders.Anna Goldie, CEO, and Azalia Mirhoseini, CTO, are so well-known in the AI community that they were among those AI engineers who “got those weird emails from Zuckerber





Home






