OPEN-1B: A Fully Auditable Training Run
为解决开源语言模型的可复现性问题，通过确定训练中的非确定性来源顺序，实现跨硬件的独立复现，引入完全可审计的透明度层级。
0 citationsRead paper
Selected work
Representative Papers
The Role of Feedback Alignment in Self-Distillation
The design mechanism of feedback context in self-distillation remains unclear. This work proposes a step-aligned critique feedback mechanism that provides structured guidance precisely at points of reasoning errors while avoiding interference with correct behaviors, thereby enhancing learning efficiency. Within a self-distillation framework, the study systematically compares three feedback forms: binary rewards (GRPO), reference solutions, and step-aligned critiques, complemented by per-token advantage analysis to evaluate their effectiveness. Experimental results demonstrate that step-aligned critique improves performance by 16.11 points over GRPO and by 5.27 points over reference solutions on the Avg@12 metric, confirming the critical importance of aligning feedback with the structural trajectory of reasoning.
0 citationsRead paper
IR3DE: A Linear Router for Large Language Models
Existing large language model (LLM) routing approaches suffer from limitations in cost efficiency or training overhead. This work proposes a lightweight linear routing mechanism based on ridge regression, which, to the best of our knowledge, is the first to apply linear models to dynamic selection among multi-domain expert LLMs. By leveraging input features derived from causal language modeling and reasoning tasks, the method enables low-overhead, highly generalizable routing decisions and supports dynamic addition or removal of expert models without retraining the router. Experimental results demonstrate that the approach matches baseline performance across two task categories and significantly outperforms existing methods in reasoning tasks, achieving 98.4% normalized performance.
0 citationsRead paper
DEI: Diversity in Evolutionary Inference for Quality-Diversity Search
This work addresses the limitations of traditional homogeneous parallel search, which is constrained by the inductive bias of a single language model and struggles to generate behavioral novelty. The authors propose the DEI framework, which for the first time employs heterogeneous large language models as mutation operators within distributed evolutionary nodes. By leveraging non-blocking collective communication to share local optima, DEI establishes a cross-model adversarial-cooperative mechanism that enhances both diversity and robustness. Empirical results demonstrate that model heterogeneity—not merely parallel scale—is the key driver of improved performance in language model–based quality-diversity (LLM-QD) optimization. On the Core War benchmark, a four-node heterogeneous system achieves a 124% increase in QD-Score (45.90 vs. 20.46) and a 28% improvement in coverage (80.6% vs. 63.0%) over the single-node baseline, consistently outperforming homogeneous parallel alternatives.
0 citationsRead paper
Training-Free Dynamic Upcycling of Expert Language Models
Training large language models is costly and struggles to balance expertise across diverse domains, while fine-tuning often leads to overfitting or catastrophic forgetting. To address these challenges, this work proposes Dynamic Upcycling Mixture of Experts (DUME), the first framework enabling dynamic fusion of expert models without any fine-tuning. DUME constructs a plug-and-play mixture-of-experts architecture via a closed-form ridge regression solution, eliminating the need for additional training or optimization. The method preserves 97.6% of expert model performance in causal language modeling and even achieves a 102.1% relative gain on inference tasks, significantly outperforming existing baselines. This approach enables efficient, scalable, and unified multi-task modeling with minimal computational overhead.
0 citationsRead paper
Recent publications
Latest Papers
OPEN-1B: A Fully Auditable Training Run
为解决开源语言模型的可复现性问题，通过确定训练中的非确定性来源顺序，实现跨硬件的独立复现，引入完全可审计的透明度层级。
0 citationsRead paper
The Role of Feedback Alignment in Self-Distillation
The design mechanism of feedback context in self-distillation remains unclear. This work proposes a step-aligned critique feedback mechanism that provides structured guidance precisely at points of reasoning errors while avoiding interference with correct behaviors, thereby enhancing learning efficiency. Within a self-distillation framework, the study systematically compares three feedback forms: binary rewards (GRPO), reference solutions, and step-aligned critiques, complemented by per-token advantage analysis to evaluate their effectiveness. Experimental results demonstrate that step-aligned critique improves performance by 16.11 points over GRPO and by 5.27 points over reference solutions on the Avg@12 metric, confirming the critical importance of aligning feedback with the structural trajectory of reasoning.
0 citationsRead paper
IR3DE: A Linear Router for Large Language Models
Existing large language model (LLM) routing approaches suffer from limitations in cost efficiency or training overhead. This work proposes a lightweight linear routing mechanism based on ridge regression, which, to the best of our knowledge, is the first to apply linear models to dynamic selection among multi-domain expert LLMs. By leveraging input features derived from causal language modeling and reasoning tasks, the method enables low-overhead, highly generalizable routing decisions and supports dynamic addition or removal of expert models without retraining the router. Experimental results demonstrate that the approach matches baseline performance across two task categories and significantly outperforms existing methods in reasoning tasks, achieving 98.4% normalized performance.
0 citationsRead paper
DEI: Diversity in Evolutionary Inference for Quality-Diversity Search
This work addresses the limitations of traditional homogeneous parallel search, which is constrained by the inductive bias of a single language model and struggles to generate behavioral novelty. The authors propose the DEI framework, which for the first time employs heterogeneous large language models as mutation operators within distributed evolutionary nodes. By leveraging non-blocking collective communication to share local optima, DEI establishes a cross-model adversarial-cooperative mechanism that enhances both diversity and robustness. Empirical results demonstrate that model heterogeneity—not merely parallel scale—is the key driver of improved performance in language model–based quality-diversity (LLM-QD) optimization. On the Core War benchmark, a four-node heterogeneous system achieves a 124% increase in QD-Score (45.90 vs. 20.46) and a 28% improvement in coverage (80.6% vs. 63.0%) over the single-node baseline, consistently outperforming homogeneous parallel alternatives.
0 citationsRead paper
Training-Free Dynamic Upcycling of Expert Language Models
Training large language models is costly and struggles to balance expertise across diverse domains, while fine-tuning often leads to overfitting or catastrophic forgetting. To address these challenges, this work proposes Dynamic Upcycling Mixture of Experts (DUME), the first framework enabling dynamic fusion of expert models without any fine-tuning. DUME constructs a plug-and-play mixture-of-experts architecture via a closed-form ridge regression solution, eliminating the need for additional training or optimization. The method preserves 97.6% of expert model performance in causal language modeling and even achieves a 102.1% relative gain on inference tasks, significantly outperforming existing baselines. This approach enables efficient, scalable, and unified multi-task modeling with minimal computational overhead.
0 citationsRead paper