RM-R1: Reward Modeling as Reasoning
Xiusi Chen, Gaotang Li, Ziqi Wang +9 authors
Reasoning Reward Models (ReasRMs) enhance reward modeling for large language models by integrating reasoning tasks, improving interpretability and performance.
Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
50 篇论文 · 按点赞排序
Quartet: Native FP4 Training Can Be Optimal for Large Language Models
Roberto L. Castro, Andrei Panferov, Soroush Tabesh +5 authors
Quartet, a hardware-supported FP4 training approach for large language models, demonstrates state-of-the-art accuracy while significantly reducing computational costs compared to standard or FP8 precision.
Scaling Law for Quantization-Aware Training
Mengzhao Chen, Chaoyi Zhang, Jing Liu +8 authors
A unified scaling law for quantization-aware training (QAT) identifies key factors affecting quantization error, leading to improvements through mixed-precision quantization.
Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures
Chenggang Zhao, Chengqi Deng, Chong Ruan +12 authors
DeepSeek-V3 addresses hardware limitations through MLA, MoE, FP8 training, and Multi-Plane Network Topology, enabling efficient large-scale LLM training and inference.
SageAttention3: Microscaling FP4 Attention for Inference and An Exploration of 8-Bit Training
Jintao Zhang, Jia Wei, Pengle Zhang +6 authors
Efficiency enhancements for attention mechanisms, including leveraging FP4 Tensor Cores and developing an 8-bit attention method, improve inference and training performance.
Flow-GRPO: Training Flow Matching Models via Online RL
Jie Liu, Gongye Liu, Jiajun Liang +6 authors
Flow-GRPO combines online reinforcement learning with flow matching models through an ODE-to-SDE conversion and denoising reduction, improving sampling efficiency and performance across text-to-image tasks.
System Prompt Optimization with Meta-Learning
Yumin Choi, Jinheon Baek, Sung Ju Hwang
A meta-learning framework for optimizing system prompts in Large Language Models (LLMs) improves generalization across diverse tasks and datasets.
R2R: Efficiently Navigating Divergent Reasoning Paths with Small-Large Model Token Routing
Tianyu Fu, Yi Ge, Yichen You +6 authors
Roads to Rome (R2R) selectively utilizes large language models for critical reasoning tasks to enhance efficiency and performance in lightweight models.
Spatial-MLLM: Boosting MLLM Capabilities in Visual-based Spatial Intelligence
Diankun Wu, Fangfu Liu, Yi-Hsin Hung +1 authors
Spatial-MLLM improves spatial reasoning in multimodal large language models using a dual-encoder architecture with pretrained 2D and 3D structure encoders, achieving state-of-the-art performance on visual spatial tasks.
SynLogic: Synthesizing Verifiable Reasoning Data at Scale for Learning Logical Reasoning and Beyond
Junteng Liu, Yuanxiang Fan, Zhuo Jiang +12 authors
SynLogic, a data synthesis framework, enhances the logical reasoning capabilities of Large Language Models through RL, achieving state-of-the-art performance and improving generalization across various domains.
ZeroSearch: Incentivize the Search Capability of LLMs without Searching
Hao Sun, Zile Qiao, Jiayan Guo +6 authors
ZeroSearch is a reinforcement learning framework that improves LLM search capabilities without using real search engines, using supervised fine-tuning and curriculum-based rollouts to handle document quality and API cost challenges.
BizFinBench: A Business-Driven Real-World Financial Benchmark for Evaluating LLMs
Guilong Lu, Xuntao Guo, Rongjunchen Zhang +2 authors
BizFinBench is a benchmark for evaluating large language models in financial applications, revealing distinct performance patterns across various tasks.
Reasoning Model is Stubborn: Diagnosing Instruction Overriding in Reasoning Models
Doohyuk Jang, Yoonjeon Kim, Chanjae Park +2 authors
A diagnostic set examines and categorizes reasoning rigidity in large language models, identifying patterns where models ignore instructions and default to familiar reasoning.
One RL to See Them All: Visual Triple Unified Reinforcement Learning
Yan Ma, Linge Du, Xuyang Shen +7 authors
A unified reinforcement learning system, V-Triune, combines visual reasoning and perception tasks in vision-language models through a single training pipeline, achieving significant improvements across various tasks.
Scaling Reasoning, Losing Control: Evaluating Instruction Following in Large Reasoning Models
Tingchen Fu, Jiawei Gu, Yafu Li +2 authors
An empirical analysis of MathIF identifies a tension between enhancing reasoning capacity and maintaining instruction adherence in large language models.
Exploring the Latent Capacity of LLMs for One-Step Text Generation
Gleb Mezentsev, Ivan Oseledets
LLMs can generate long text segments in a single forward pass using learned embeddings, revealing a capability for multi-token generation without iterative decoding.
GuardReasoner-VL: Safeguarding VLMs via Reinforced Reasoning
Yue Liu, Shengfang Zhai, Mingzhe Du +9 authors
GuardReasoner-VL enhances VLM safety through a reasoning-based guard model trained with SFT and online RL, using diverse datasets and a length-aware safety reward.
Step1X-3D: Towards High-Fidelity and Controllable Generation of Textured 3D Assets
Weiyu Li, Xuanyang Zhang, Zheng Sun +15 authors
Step1X-3D addresses challenges in 3D generation through a hybrid VAE-DiT geometry generator and diffusion-based texture synthesis, achieving state-of-the-art results with an open-source framework.
Bielik v3 Small: Technical Report
Krzysztof Ociepa, Łukasz Flis, Remigiusz Kinas +2 authors
Bielik v3, a series of parameter-efficient generative text models, achieves high performance in Polish language processing with a custom tokenizer, Weighted Instruction Cross-Entropy Loss, and Adaptive Learning Rate.
WebThinker: Empowering Large Reasoning Models with Deep Research Capability
Xiaoxi Li, Jiajie Jin, Guanting Dong +5 authors
WebThinker, an autonomous deep research agent, improves LRM performance on complex tasks by integrating a Deep Web Explorer for dynamic web information gathering and using RL-based training through Direct Preference Optimization.