Prefix Sliding for efficient test-time scaling
为解决长思考任务中模型内存占用过大的问题,提出Prefix Sliding方法,在推理过程中丢弃非关键的中间推理标记,从而有效支持长时间推理并提高效率。
为解决长思考任务中模型内存占用过大的问题,提出Prefix Sliding方法,在推理过程中丢弃非关键的中间推理标记,从而有效支持长时间推理并提高效率。
为解决长周期任务处理问题,Prime Agent采用递归语言模型和持续计算环境,通过子代理协调与人类管理实现高效执行。
Current AI programming evaluations are largely confined to short, isolated tasks and lack standardized benchmarks for assessing the ability to reproduce complete software systems end-to-end. This work proposes MirrorCode, a novel long-horizon programming benchmark that introduces a behavior-based reverse-engineering paradigm: under strict black-box conditions without access to source code, AI systems must precisely reconstruct the functionality of real-world software solely from its observable behavior. The benchmark encompasses 25 complex projects spanning Unix utilities, compilers, and bioinformatics tools. Leveraging large language model–driven autonomous agents and a high-fidelity test-validation framework, the strongest model achieves a 56% success rate, including the successful reproduction of large-scale tools such as gotree (16,000 lines), demonstrating a significant leap in AI’s capacity for complex software engineering tasks.
This work proposes the Trinity series of sparse Mixture-of-Experts (MoE) models—Trinity-Large, -Mini, and -Nano—to enhance parameter efficiency and training stability in large language models. The architecture integrates interleaved local-global attention, gated attention, depth-scaled Sandwich normalization, and a Sigmoid-based routing mechanism, and is trained using the Muon optimizer. A key innovation is the introduction of Soft-clamped Momentum Expert Bias Updates (SMEBU), a novel load-balancing strategy that significantly improves MoE training stability. All variants complete training without any loss spikes: Trinity-Nano and -Mini are pretrained on 10 trillion tokens, while Trinity-Large uses 17 trillion tokens. The codebase has been publicly released.
This work addresses the challenge of efficiently training large-scale Mixture-of-Experts (MoE) models while achieving strong generalization on mathematical reasoning, code generation, scientific problem-solving, and complex reasoning tasks. We propose a 106B-parameter MoE model with only 12B activated parameters per forward pass, trained via a custom end-to-end reinforcement learning (RL) paradigm. Our approach leverages Prime-RL—an open-source, asynchronous RL framework we introduce, supporting thousand-GPU scalability and multi-turn interactive training—integrated with a curated Verifier environment library and the GLM-4.5-Air-Base foundation model. The resulting model achieves state-of-the-art performance on multiple authoritative benchmarks at comparable parameter counts, outperforming larger frontier models despite its modest active parameter count. We fully open-source the model weights, RL training infrastructure, reproducible training recipes, and diverse verification environments—establishing a systematic, transparent, and scalable foundation for open MoE-RL research.
为解决长思考任务中模型内存占用过大的问题,提出Prefix Sliding方法,在推理过程中丢弃非关键的中间推理标记,从而有效支持长时间推理并提高效率。
为解决长周期任务处理问题,Prime Agent采用递归语言模型和持续计算环境,通过子代理协调与人类管理实现高效执行。
Current AI programming evaluations are largely confined to short, isolated tasks and lack standardized benchmarks for assessing the ability to reproduce complete software systems end-to-end. This work proposes MirrorCode, a novel long-horizon programming benchmark that introduces a behavior-based reverse-engineering paradigm: under strict black-box conditions without access to source code, AI systems must precisely reconstruct the functionality of real-world software solely from its observable behavior. The benchmark encompasses 25 complex projects spanning Unix utilities, compilers, and bioinformatics tools. Leveraging large language model–driven autonomous agents and a high-fidelity test-validation framework, the strongest model achieves a 56% success rate, including the successful reproduction of large-scale tools such as gotree (16,000 lines), demonstrating a significant leap in AI’s capacity for complex software engineering tasks.
This work proposes the Trinity series of sparse Mixture-of-Experts (MoE) models—Trinity-Large, -Mini, and -Nano—to enhance parameter efficiency and training stability in large language models. The architecture integrates interleaved local-global attention, gated attention, depth-scaled Sandwich normalization, and a Sigmoid-based routing mechanism, and is trained using the Muon optimizer. A key innovation is the introduction of Soft-clamped Momentum Expert Bias Updates (SMEBU), a novel load-balancing strategy that significantly improves MoE training stability. All variants complete training without any loss spikes: Trinity-Nano and -Mini are pretrained on 10 trillion tokens, while Trinity-Large uses 17 trillion tokens. The codebase has been publicly released.
This work addresses the challenge of efficiently training large-scale Mixture-of-Experts (MoE) models while achieving strong generalization on mathematical reasoning, code generation, scientific problem-solving, and complex reasoning tasks. We propose a 106B-parameter MoE model with only 12B activated parameters per forward pass, trained via a custom end-to-end reinforcement learning (RL) paradigm. Our approach leverages Prime-RL—an open-source, asynchronous RL framework we introduce, supporting thousand-GPU scalability and multi-turn interactive training—integrated with a curated Verifier environment library and the GLM-4.5-Air-Base foundation model. The resulting model achieves state-of-the-art performance on multiple authoritative benchmarks at comparable parameter counts, outperforming larger frontier models despite its modest active parameter count. We fully open-source the model weights, RL training infrastructure, reproducible training recipes, and diverse verification environments—establishing a systematic, transparent, and scalable foundation for open MoE-RL research.