Shuaishuai Zu, Jihao Zhao, Pengfei Li, Guanghuan Fang, Ben Feng, Yuhan Liu, Biao Qin, Baochuan Yang
On-policy distillation (OPD) supplies dense token-level supervision by aligning a student to the teacher's next-token distribution on its own generated rollouts. Existing works mostly implement top-$k$ OPD, applying reverse Kullback-Leibler ($\RKL$) divergence between the teacher's and the student's renormalized distributions over teacher top-$k$ support. However, the teacher scores student-generated rollouts that are inherently off-policy for itself, leaving the reliability of its next-token guidance unclear. Restricting the divergence to the teacher top-$k$ support also discards the remaining mass outside that support. These gaps leave two questions implicit: 1) which tokens to align, and 2) how to align those tokens. To answer these questions, we propose Overlap-gated On-Policy Distillation (O$^{2}$PD). As for 1), we quantitatively analyze the reliability of teacher supervision and find that alignment on the student's trajectory prefix is more reliable than alignment on the full context. Therefore, our O$^{2}$PD is applied only to the 20\% prefix tokens of each student rollout. As for 2), we theoretically characterize two problems of top-$k$ OPD: the \emph{Teacher-support Uncovering} and the \emph{Dominant-Mode Mismatch}. Based on these analyses, we propose two support strategies: the \emph{Reset Support} strategy attaches a reset atom to every token, pooling mass outside the teacher top-$k$, and the \emph{Support Convergence} strategy converges the support to the already aligned intersection. Our O$^{2}$PD employs the designed overlap-gated $α$-divergence to interpolate between mode-covering and mode-seeking for per-token alignment, where the overlap is used as a gate to switch the support and adaptively modulate $α$.
YiJie Huang, Xiaocui Yang Northeastern University, Shi Feng Northeastern University, China, Wen Zhang Northeastern University, Kaisong Song, Yifei Zhang, Daling Wang
Efficiently aligning visual features with Large Language Models (LLMs) remains a critical bottleneck in Multimodal LLMs. Existing query-based alignment modules (e.g., Q-Former) rely on randomly initialized queries, resulting in an inefficient cold start exploration process. Furthermore, they enforce uniform cross-attention across all layers, leading to computational redundancy. Our empirical analysis reveals that query tokens initialized with language priors can rapidly capture global semantics, leading to early representation convergence after only a few layers. In this paper, we propose Cat-MoD, a Caption token Guided Asymmetric Mixture-of-Depths framework. It incorporates a Hybrid Query Construction module where Guide Tokens initialized from coarse-grained linguistic priors rapidly anchor global semantic context, and randomly initialized Explorer Tokens remain active to capture fine-grained visual details. Exploiting this early convergence, we introduce an Asymmetric Mixture-of-Depths mechanism, where a similarity-aware router dynamically prunes redundant tokens from expensive cross-attention layers while preserving their context in self-attention. Experiments on multiple benchmarks demonstrate that Cat-MoD matches or surpasses baseline performance, while substantially reducing alignment FLOPs by approximately 37% during both training and inference, offering a highly efficient solution for multimodal alignment. Code: https://github.com/JasonOrange0726/Cat-MoD.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Chaoyu Li Arizona State University, Yogesh Kulkarni Arizona State University, Pooyan Fazli Arizona State University
The computational cost of training multimodal large language models (MLLMs) grows rapidly with the number of processed tokens. Existing efficiency methods mainly target inference via token reduction or merging, offering limited benefits during training. We introduce ReGATE (Reference-Guided Adaptive Token Elision), an adaptive token pruning method for accelerating MLLM training. ReGATE adopts a teacher-student framework, in which a frozen teacher LLM provides per-token guidance losses that are fused with an exponential moving average of the student’s difficulty estimates. This adaptive scoring mechanism dynamically selects informative tokens while skipping redundant ones in the forward pass, substantially reducing computation without altering the model architecture. Across three representative MLLMs, ReGATE matches the peak accuracy of standard training on MVBench up to 2 \times faster, using only 38% of the tokens. With extended training, it even surpasses the baseline across multiple multimodal benchmarks, cutting total token usage by over 41%. Code and models will be released publicly.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yuhao Shen, Tianyu Liu, Junyi Shen, Jinyang Wu Tsinghua University, Quan Kong, Huan Li Zhejiang University, Cong Wang Zhejiang University
Parallel Speculative Decoding (PSD) accelerates traditional Speculative Decoding (SD) by overlapping draft generation with verification. However, it remains hampered by two fundamental challenges: (1) a theoretical speedup ceiling dictated by the speed ratio between the draft and target models, and (2) high computational waste and pipeline stall due to mid-sequence token rejections of early errors. To address these limitations, we introduce Double (Double Retrieval Speculative Parallelism). By bridging the gap between SD and PSD, our framework resolves the Retrieval Precision-Efficiency Dilemma through a novel synchronous mechanism. Specifically, we enable the draft model to execute iterative retrieval speculations to break the theoretical speedup limits; to alleviate rejections without rollback, the target model performs authoritative retrieval to generate multi-token guidance. Double is entirely training-free and lossless. Extensive experiments demonstrate state-of-the-art speedup of \textbf{5.3}\times on LLaMA3.3-70B and \textbf{2.8}\times on Qwen3-32B, significantly outperforming the advanced method EAGLE-3 that requires extensive model training. Our code is available at https://github.com/Sylvan820/Double1.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yubin Kim, Ken Gu, Chanwoo Park, Chunjong Park, Samuel Schmidgall, A. Ali Heydari, Yao Yan, Zhihan Zhang 等 20 位作者
Agents, language model-based systems capable of reasoning, planning, and acting are widely adopted in real-world tasks, yet how their performance changes as these systems scale across key dimensions remains underexplored. We introduce quantitative scaling principles for agent systems as a predictive model, capturing how performance varies with coordination, model capability, and measurable system and task factors. Across 260 configurations spanning six agentic benchmarks, five canonical architectures (Single-Agent and four Multi-Agent: Independent, Centralized, Decentralized, Hybrid), and three LLM families, we perform controlled evaluations, standardizing tools, prompts, and compute to isolate architectural effects. The resulting model achieves a cross-validated R^2=0.373 across all six benchmarks (R^2=0.413 with a task-grounded capability metric). We identify a robust capability-saturation effect and additional patterns: (1) a coordination yields diminishing returns once single-agent baselines exceed certain performance; (2) tool-heavy tasks appear to incur multi-agent overhead; and (3) architectures without centralized verification tend to propagate errors more than those with centralized coordination. Relative performance change compared to single-agent baseline ranges from +80.8% on decomposable financial reasoning to -70.0% on sequential planning, demonstrating that architecture-task alignment determines collaborative success. The framework identifies the best-performing architecture for 87% of held-out configurations and shows consistent relative architecture preferences on unseen frontier models. Agent effectiveness depends on alignment between coordination and task structure, and that mismatched coordination degrades the performance.