10 papers
GrowMTP: Can RL Grow Its Own Draft Head?
Minghua He, Lingzhe Zhang, Yuan Liu +2
Reinforcement learning (RL) post-training drives the frontier capabilities of large language models, with its wall-clock dominated by autoregressive rollout generation. Speculative…
Hidden Decoding at Scale: Latent Computation Scaling for Large Language Models
Aiwei Liu, Cheng Shi, Chuhan Wu +44
Scaling Large Language Models (LLMs) has been driven mainly by enlarging the Transformer backbone, but for an already-strong model this requires another round of costly pretraining…
RLCSD: Reinforcement Learning with Contrastive On-Policy Self-Distillation
Leyi Pan, Shuchang Tao, Yunpeng Zhai +5
On-policy self-distillation (OPSD) provides dense, token-level supervision for reasoning models by aligning a model's own distribution with that under privileged context, typically…
UniAudio-Token: Empowering Semantic Speech Tokenizers with General Audio Perception
Yuhan Song, Linhao Zhang, Aiwei Liu +6
Semantic speech tokenizers have become a widely used interface for Audio-LLMs, owing to their compact single-codebook design and strong linguistic alignment. However, their focus o…
How Much Parallelism Is "Free"? A Principle of Near-Free Parallelism for Parallel Decoding
Minghua He, Lingzhe Zhang, Yuan Liu +2
Parallel decoding improves generation efficiency by processing multiple decode positions within a single decode forward, but reported speedups conflate algorithmic token utilizatio…
DiffSpot: Can VLMs Spot Fine-Grained Visual Differences in Web Interfaces?
Linhao Zhang, Aiwei Liu, Yuan Liu +1
Vision-language models (VLMs) have made strong progress on high-level image-text alignment, yet their ability to perceive subtle visual differences remains limited. We study this p…
Beyond Transcription: Unified Audio Schema for Perception-Aware AudioLLMs
Linhao Zhang, Yuhan Song, Aiwei Liu +6
Recent Audio Large Language Models (AudioLLMs) exhibit a striking performance inversion: while excelling at complex reasoning tasks, they consistently underperform on fine-grained…
WeDLM: Reconciling Diffusion Language Models with Standard Causal Attention for Fast Inference
Aiwei Liu, Minghua He, Shaoxun Zeng +7
Autoregressive (AR) generation is the standard decoding paradigm for Large Language Models (LLMs), but its token-by-token nature limits parallelism at inference time. Diffusion Lan…
StableToken: A Noise-Robust Semantic Speech Tokenizer for Resilient SpeechLLMs
Yuhan Song, Linhao Zhang, Chuhan Wu +4
Prevalent semantic speech tokenizers, designed to capture linguistic content, are surprisingly fragile. We find they are not robust to meaning-irrelevant acoustic perturbations; ev…
WildSpeech-Bench: Benchmarking End-to-End SpeechLLMs in the Wild
Linhao Zhang, Jian Zhang, Bokai Lei +4
Recent multi-modal Large Language Models (LLMs) such as GPT-4o have demonstrated strong capabilities of direct speech interaction. However, the lack of specialized and comprehensiv…