2025년 5월 7일
Practical Efficiency of Muon for Pretraining
(Essential AI: Ishaan Shah, Anthony M. Polloreno, Karl Stratos, Philip Monk, Adarsh Chaluvaraju, Andrew Hojel, Andrew Ma, Anil Thomas, Ashish Tanwer, Darsh J Shah, Khoi Nguyen, Kurt Smith, Michael Callahan, Michael Pust, Mohit Parmar, Peter Rushton, Platon Mazarakis, Ritvik Kapila, Saurabh Srivastava, Somanshu Singla, Tim Romanski, Yash Vanjani, Ashish Vaswani)
We demonstrate that Muon, the simplest instantiation of a second-order optimizer, explicitly expands the Pareto frontier over AdamW on the compute-time tradeoff. We find that Muon is more effective than AdamW in retaining data efficiency at large batch sizes, far beyond the so-called critical batch size, while remaining computationally efficient, thus enabling more economical training. We study the combination of Muon and the maximal update parameterization (muP) for efficient hyperparameter transfer and present a simple telescoping algorithm that accounts for all sources of error in muP while introducing only a modest overhead in resources. We validate our findings through extensive experiments with model sizes up to four billion parameters and ablations on the data distribution and architecture.
Muon과 Muon를 위한 muP에 대한 분석. 실험 규모가 크진 않은데 15% 정도의 데이터 효율성 증가를 이야기하고 있네요.
Essential AI가 어떤 회사인가 했는데 Ashish Vaswani가 창립한 회사군요.
Analysis of Muon and muP for Muon. The scale of experiments isn't very large, but they report about a 15% increase in data efficiency.
I was wondering what kind of company Essential AI is, and it turns out to be a company founded by Ashish Vaswani.
#optimizer #hyperparameter
Don't be lazy: CompleteP enables compute-efficient deep transformers
(Nolan Dey, Bin Claire Zhang, Lorenzo Noci, Mufan Li, Blake Bordelon, Shane Bergsma, Cengiz Pehlevan, Boris Hanin, Joel Hestness)
We study compute efficiency of LLM training when using different parameterizations, i.e., rules for adjusting model and optimizer hyperparameters (HPs) as model size changes. Some parameterizations fail to transfer optimal base HPs (such as learning rate) across changes in model depth, requiring practitioners to either re-tune these HPs as they scale up (expensive), or accept sub-optimal training when re-tuning is prohibitive. Even when they achieve HP transfer, we develop theory to show parameterizations may still exist in the lazy learning regime where layers learn only features close to their linearization, preventing effective use of depth and nonlinearity. Finally, we identify and adopt the unique parameterization we call CompleteP that achieves both depth-wise HP transfer and non-lazy learning in all layers. CompleteP enables a wider range of model width/depth ratios to remain compute-efficient, unlocking shapes better suited for different hardware settings and operational contexts. Moreover, CompleteP enables 12-34% compute efficiency improvements over the prior state-of-the-art.
깊이 변화에 대해서도 적용 가능한 Hyperparameter Transfer. α = 1이 최적 세팅이라 결과적으로는 Residual Branch에 대한 Reweighting과 비슷하네요. 너비-깊이 비율을 조정할 수 있는 여지도 커진다고 합니다. 이것도 깊은 모델에 대한 안정화 테크닉과 비슷한 점이라고 생각할 수 있을 것 같군요.
Hyperparameter transfer that can be applied to depth variations. As α = 1 is the optimal setting, it ultimately resembles reweighting for residual branches. It also allows for greater flexibility in adjusting the width-depth ratio. I think this can be considered as the result similar to stabilization techniques for deep models.
#hyperparameter
Llama-Nemotron: Efficient Reasoning Models
(NVIDIA)
We introduce the Llama-Nemotron series of models, an open family of heterogeneous reasoning models that deliver exceptional reasoning capabilities, inference efficiency, and an open license for enterprise use. The family comes in three sizes -- Nano (8B), Super (49B), and Ultra (253B) -- and performs competitively with state-of-the-art reasoning models such as DeepSeek-R1 while offering superior inference throughput and memory efficiency. In this report, we discuss the training procedure for these models, which entails using neural architecture search from Llama 3 models for accelerated inference, knowledge distillation, and continued pretraining, followed by a reasoning-focused post-training stage consisting of two main parts: supervised fine-tuning and large scale reinforcement learning. Llama-Nemotron models are the first open-source models to support a dynamic reasoning toggle, allowing users to switch between standard chat and reasoning modes during inference. To further support open research and facilitate model development, we provide the following resources: 1. We release the Llama-Nemotron reasoning models -- LN-Nano, LN-Super, and LN-Ultra -- under the commercially permissive NVIDIA Open Model License Agreement. 2. We release the complete post-training dataset: Llama-Nemotron-Post-Training-Dataset. 3. We also release our training codebases: NeMo, NeMo-Aligner, and Megatron-LM.
NVIDIA의 추론 모델. Llama 3를 Distillation한 모델로 추론 RL과 RLHF까지 진행했군요.
NVIDIA's reasoning model. It's a distilled version of Llama 3 that has undergone reasoning RL and RLHF.
#rl #reasoning #distillation
Absolute Zero: Reinforced Self-play Reasoning with Zero Data
(Andrew Zhao, Yiran Wu, Yang Yue, Tong Wu, Quentin Xu, Yang Yue, Matthieu Lin, Shenzhi Wang, Qingyun Wu, Zilong Zheng, Gao Huang)
Reinforcement learning with verifiable rewards (RLVR) has shown promise in enhancing the reasoning capabilities of large language models by learning directly from outcome-based rewards. Recent RLVR works that operate under the zero setting avoid supervision in labeling the reasoning process, but still depend on manually curated collections of questions and answers for training. The scarcity of high-quality, human-produced examples raises concerns about the long-term scalability of relying on human supervision, a challenge already evident in the domain of language model pretraining. Furthermore, in a hypothetical future where AI surpasses human intelligence, tasks provided by humans may offer limited learning potential for a superintelligent system. To address these concerns, we propose a new RLVR paradigm called Absolute Zero, in which a single model learns to propose tasks that maximize its own learning progress and improves reasoning by solving them, without relying on any external data. Under this paradigm, we introduce the Absolute Zero Reasoner (AZR), a system that self-evolves its training curriculum and reasoning ability by using a code executor to both validate proposed code reasoning tasks and verify answers, serving as an unified source of verifiable reward to guide open-ended yet grounded learning. Despite being trained entirely without external data, AZR achieves overall SOTA performance on coding and mathematical reasoning tasks, outperforming existing zero-setting models that rely on tens of thousands of in-domain human-curated examples. Furthermore, we demonstrate that AZR can be effectively applied across different model scales and is compatible with various model classes.
데이터 없이 추론 RL을 진행하기. 프로그램 생성과 코드 실행을 기반으로 모델이 직접 문제를 생성하는 방식이군요.
Conducting reasoning RL without data. The main approach is to have the model generate problems based on program generation and code execution.
#rl #reasoning #synthetic-data
RM-R1: Reward Modeling as Reasoning
(Xiusi Chen, Gaotang Li, Ziqi Wang, Bowen Jin, Cheng Qian, Yu Wang, Hongru Wang, Yu Zhang, Denghui Zhang, Tong Zhang, Hanghang Tong, Heng Ji)
Reward modeling is essential for aligning large language models (LLMs) with human preferences, especially through reinforcement learning from human feedback (RLHF). To provide accurate reward signals, a reward model (RM) should stimulate deep thinking and conduct interpretable reasoning before assigning a score or a judgment. However, existing RMs either produce opaque scalar scores or directly generate the prediction of a preferred answer, making them struggle to integrate natural language critiques, thus lacking interpretability. Inspired by recent advances of long chain-of-thought (CoT) on reasoning-intensive tasks, we hypothesize and validate that integrating reasoning capabilities into reward modeling significantly enhances RM's interpretability and performance. In this work, we introduce a new class of generative reward models -- Reasoning Reward Models (ReasRMs) -- which formulate reward modeling as a reasoning task. We propose a reasoning-oriented training pipeline and train a family of ReasRMs, RM-R1. The training consists of two key stages: (1) distillation of high-quality reasoning chains and (2) reinforcement learning with verifiable rewards. RM-R1 improves LLM rollouts by self-generating reasoning traces or chat-specific rubrics and evaluating candidate responses against them. Empirically, our models achieve state-of-the-art or near state-of-the-art performance of generative RMs across multiple comprehensive reward model benchmarks, outperforming much larger open-weight models (e.g., Llama3.1-405B) and proprietary ones (e.g., GPT-4o) by up to 13.8%. Beyond final performance, we perform thorough empirical analysis to understand the key ingredients of successful ReasRM training. To facilitate future research, we release six ReasRM models along with code and data at https://github.com/RM-R1-UIUC/RM-R1.
Reasoning Reward Model. 채팅에 대해서는 Rubric을 생성하도록 했는데 이 점은 DeepSeek-GRM과 비슷하군요. (https://arxiv.org/abs/2504.02495) 추론 모델이 Reward Model로 사용되는 것은 자연스러운 흐름이겠죠.
Reasoning reward model. For chat, they let the model to generate rubrics, which is similar to DeepSeek-GRM (https://arxiv.org/abs/2504.02495). It seems natural that reasoning models are being used as reward models.
#reward-model #reasoning #rl
R1-Reward: Training Multimodal Reward Model Through Stable Reinforcement Learning
(Yi-Fan Zhang, Xingyu Lu, Xiao Hu, Chaoyou Fu, Bin Wen, Tianke Zhang, Changyi Liu, Kaiyu Jiang, Kaibing Chen, Kaiyu Tang, Haojie Ding, Jiankang Chen, Fan Yang, Zhang Zhang, Tingting Gao, Liang Wang)
Multimodal Reward Models (MRMs) play a crucial role in enhancing the performance of Multimodal Large Language Models (MLLMs). While recent advancements have primarily focused on improving the model structure and training data of MRMs, there has been limited exploration into the effectiveness of long-term reasoning capabilities for reward modeling and how to activate these capabilities in MRMs. In this paper, we explore how Reinforcement Learning (RL) can be used to improve reward modeling. Specifically, we reformulate the reward modeling problem as a rule-based RL task. However, we observe that directly applying existing RL algorithms, such as Reinforce++, to reward modeling often leads to training instability or even collapse due to the inherent limitations of these algorithms. To address this issue, we propose the StableReinforce algorithm, which refines the training loss, advantage estimation strategy, and reward design of existing RL methods. These refinements result in more stable training dynamics and superior performance. To facilitate MRM training, we collect 200K preference data from diverse datasets. Our reward model, R1-Reward, trained using the StableReinforce algorithm on this dataset, significantly improves performance on multimodal reward modeling benchmarks. Compared to previous SOTA models, R1-Reward achieves a 8.4% improvement on the VL Reward-Bench and a 14.3% improvement on the Multimodal Reward Bench. Moreover, with more inference compute, R1-Reward's performance is further enhanced, highlighting the potential of RL algorithms in optimizing MRMs.
Multimodal Reasoning Reward Model. 다들 이런 작업을 하고 있군요. 여기서는 학습이 불안정해서 RL 알고리즘에 대한 조정이 필요했다고 합니다.
Multimodal reasoning reward model. It seems everyone is working on similar projects. The authors mention that they experienced training instabilities, so they had to adjust RL algorithms.
#rl #reasoning #reward-model #multimodal
Rewriting Pre-Training Data Boosts LLM Performance in Math and Code
(Kazuki Fujii, Yukito Tajima, Sakae Mizuki, Hinari Shimada, Taihei Shiotani, Koshiro Saito, Masanari Ohi, Masaki Kawamura, Taishi Nakamura, Takumi Okamoto, Shigeki Ishida, Kakeru Hattori, Youmi Ma, Hiroya Takamura, Rio Yokota, Naoaki Okazaki)
The performance of large language models (LLMs) in program synthesis and mathematical reasoning is fundamentally limited by the quality of their pre-training corpora. We introduce two openly licensed datasets, released under the Llama 3.3 Community License, that significantly enhance LLM performance by systematically rewriting public data. SwallowCode (approximately 16.1 billion tokens) refines Python snippets from The-Stack-v2 through a novel four-stage pipeline: syntax validation, pylint-based style filtering, and a two-stage LLM rewriting process that enforces style conformity and transforms snippets into self-contained, algorithmically efficient examples. Unlike prior methods that rely on exclusionary filtering or limited transformations, our transform-and-retain approach upgrades low-quality code, maximizing data utility. SwallowMath (approximately 2.3 billion tokens) enhances Finemath-4+ by removing boilerplate, restoring context, and reformatting solutions into concise, step-by-step explanations. Within a fixed 50 billion token training budget, continual pre-training of Llama-3.1-8B with SwallowCode boosts pass@1 by +17.0 on HumanEval and +17.7 on HumanEval+ compared to Stack-Edu, surpassing the baseline model's code generation capabilities. Similarly, substituting SwallowMath yields +12.4 accuracy on GSM8K and +7.6 on MATH. Ablation studies confirm that each pipeline stage contributes incrementally, with rewriting delivering the largest gains. All datasets, prompts, and checkpoints are publicly available, enabling reproducible research and advancing LLM pre-training for specialized domains.
프리트레이닝 데이터를 재작성하기. 이런 재작성은 다양성을 낮출 수 있다는 위험이 늘 존재하는데 원 데이터와 잘 결합하면 되지 않을까 싶기도 하네요.
Rewriting the pretraining data. There is always a risk that this kind of rewriting method could reduce diversity, but perhaps this could be mitigated by combining it with the original data.
#pretraining #corpus #synthetic-data
FormalMATH: Benchmarking Formal Mathematical Reasoning of Large Language Models
(Zhouliang Yu, Ruotian Peng, Keyi Ding, Yizhe Li, Zhongyuan Peng, Minghao Liu, Yifan Zhang, Zheng Yuan, Huajian Xin, Wenhao Huang, Yandong Wen, Ge Zhang, Weiyang Liu)
Formal mathematical reasoning remains a critical challenge for artificial intelligence, hindered by limitations of existing benchmarks in scope and scale. To address this, we present FormalMATH, a large-scale Lean4 benchmark comprising 5,560 formally verified problems spanning from high-school Olympiad challenges to undergraduate-level theorems across diverse domains (e.g., algebra, applied mathematics, calculus, number theory, and discrete mathematics). To mitigate the inefficiency of manual formalization, we introduce a novel human-in-the-loop autoformalization pipeline that integrates: (1) specialized large language models (LLMs) for statement autoformalization, (2) multi-LLM semantic verification, and (3) negation-based disproof filtering strategies using off-the-shelf LLM-based provers. This approach reduces expert annotation costs by retaining 72.09% of statements before manual verification while ensuring fidelity to the original natural-language problems. Our evaluation of state-of-the-art LLM-based theorem provers reveals significant limitations: even the strongest models achieve only 16.46% success rate under practical sampling budgets, exhibiting pronounced domain bias (e.g., excelling in algebra but failing in calculus) and over-reliance on simplified automation tactics. Notably, we identify a counterintuitive inverse relationship between natural-language solution guidance and proof success in chain-of-thought reasoning scenarios, suggesting that human-written informal reasoning introduces noise rather than clarity in the formal reasoning settings. We believe that FormalMATH provides a robust benchmark for benchmarking formal mathematical reasoning.
형식화된 수학에 대한 벤치마크.
A benchmark for formal math.
#benchmark #math
No Other Representation Component Is Needed: Diffusion Transformers Can Provide Representation Guidance by Themselves
(Dengyang Jiang, Mengmeng Wang, Liuzhuozheng Li, Lei Zhang, Haoyu Wang, Wei Wei, Guang Dai, Yanning Zhang, Jingdong Wang)
Recent studies have demonstrated that learning a meaningful internal representation can both accelerate generative training and enhance generation quality of the diffusion transformers. However, existing approaches necessitate to either introduce an additional and complex representation training framework or rely on a large-scale, pre-trained representation foundation model to provide representation guidance during the original generative training process. In this study, we posit that the unique discriminative process inherent to diffusion transformers enables them to offer such guidance without requiring external representation components. We therefore propose Self-Representation A}lignment (SRA), a simple yet straightforward method that obtain representation guidance through a self-distillation manner. Specifically, SRA aligns the output latent representation of the diffusion transformer in earlier layer with higher noise to that in later layer with lower noise to progressively enhance the overall representation learning during only generative training process. Experimental results indicate that applying SRA to DiTs and SiTs yields consistent performance improvements. Moreover, SRA not only significantly outperforms approaches relying on auxiliary, complex representation training frameworks but also achieves performance comparable to methods that heavily dependent on powerful external representation priors.
Diffusion 모델 학습에 DINOv2 같은 외부 모델 대신 Diffusion 모델의 EMA로 가이드를 줄 수 있다는 아이디어군요.
The idea is that instead of using external models like DINOv2, it's possible to use the EMA of diffusion models to guide the training of diffusion models.
#diffusion