15 papers
On-Policy Residual Learning for Diffusion Language Models
Yifan Yu, Yuqing Jian, Jisen Li +9
Residual context is key to unlocking the speed-quality Pareto frontier of diffusion language models (DLMs) by carrying forward predictive distributions across denoising steps. Yet…
Osprey: Target-agnostic Pre-training Makes Stronger Drafters in Speculative Decoding
Fengxiang Bie, Yuqing Jian, Yifan Yu +7
Speculative decoding is critical for accelerating LLM inference. However, the speedup is fragile: drafters are typically trained against a narrow distribution for a single target m…
ISO: An RLVR-Native Optimization Stack
Hanqing Zhu, Wenyan Cong, Zhizhou Sha +8
Reinforcement learning with verifiable rewards (RLVR) advances the reasoning capabilities of language models, yet the optimization layer that converts reward feedback into weight-s…
OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization
Zhongzhu Zhou, Donglin Zhuang, Jisen Li +4
INT2 KV-cache quantization is attractive for long-context LLM serving, but it remains difficult to make both accurate and deployable. Simple rotations such as Hadamard transforms r…
Search Your Block Floating Point Scales!
Tanmaey Gupta, Hayden Prairie, Xiaoxia Wu +10
Quantization has emerged as a standard technique for accelerating inference for generative models by enabling faster low-precision computations and reduced memory transfers. Recent…
SAW-INT4: System-Aware 4-Bit KV-Cache Quantization for Real-World LLM Serving
Jinda Jia, Jisen Li, Zhongzhu Zhou +8
KV-cache memory is a major bottleneck in real-world LLM serving, where systems must simultaneously support latency-sensitive small-batch requests and high-throughput concurrent wor…
Introspective Diffusion Language Models
Yifan Yu, Yuqing Jian, Junxiong Wang +12
Diffusion language models promise parallel generation, yet still lag behind autoregressive (AR) models in quality. We stem this gap to a failure of introspective consistency: AR mo…
Squeeze Evolve: Unified Multi-Model Orchestration for Verifier-Free Evolution
Monishwaran Maheswaran, Leon Lakhani, Zhongzhu Zhou +16
We show that verifier-free evolution is bottlenecked by both diversity and efficiency: without external correction, repeated evolution accelerates collapse toward narrow modes, whi…
CARE: Covariance-Aware and Rank-Enhanced Decomposition for Enabling Multi-Head Latent Attention
Zhongzhu Zhou, Fengxiang Bie, Ziyan Chen +6
Converting pretrained attention modules such as grouped-query attention (GQA) into multi-head latent attention (MLA) can improve expressivity without increasing KV-cache cost, maki…
: Unifying Generation and Self-Verification for Parallel Reasoners
Harman Singh, Xiuyu Li, Kusha Sareen +14
Test-time scaling for complex reasoning tasks shows that leveraging inference-time compute, by methods such as independently sampling and aggregating multiple solutions, results in…
When RL Meets Adaptive Speculative Training: A Unified Training-Serving System
Junxiong Wang, Fengxiang Bie, Jisen Li +14
Speculative decoding can significantly accelerate LLM serving, yet most deployments today disentangle speculator training from serving, treating speculator training as a standalone…
Understanding and Steering the Cognitive Behaviors of Reasoning Models at Test-Time
Zhenyu Zhang, Xiaoxia Wu, Zhongzhu Zhou +7
Large Language Models (LLMs) often rely on long chain-of-thought (CoT) reasoning to solve complex tasks. While effective, these trajectories are frequently inefficient, leading to…
Kitty: Accurate and Efficient 2-bit KV Cache Quantization with Dynamic Channel-wise Precision Boost
Haojun Xia, Xiaoxia Wu, Jisen Li +12
The KV cache is a dominant memory bottleneck for LLM inference. While 4-bit KV quantization preserves accuracy, 2-bit often degrades it, especially on long-context reasoning. We cl…
Beat the long tail: Distribution-Aware Speculative Decoding for RL Training
Zelei Shao, Vikranth Srivatsa, Sanjana Srivastava +12
Reinforcement learning(RL) post-training has become essential for aligning large language models (LLMs), yet its efficiency is increasingly constrained by the rollout phase, where…
Imitate Optimal Policy: Prevail and Induce Action Collapse in Policy Gradient
Zhongzhu Zhou, Yibo Yang, Ziyan Chen +7
Policy gradient (PG) methods in reinforcement learning frequently utilize deep neural networks (DNNs) to learn a shared backbone of feature representations used to compute likeliho…