Petr Molodyk, Yuchen Zhu, Jaemoo Choi, Molei Tao, Yongxin Chen
Speculative decoding is used to reduce the inference cost of language models. One-step diffusion drafters propose a block of multiple tokens in parallel, reducing the sequential cost of speculative decoding. On-policy distillation trains these drafters using verifier feedback on their own proposals. However, parallel drafting makes this feedback difficult to use: the drafter predicts each token without observing the preceding tokens sampled within the same block, while the verifier conditions its prediction on those tokens. Different sampled prefixes therefore give different verifier targets for the same drafter prediction. We show that distilling against these targets separately can suppress tokens that are plausible after some prefixes but unlikely after others, harming draft acceptance. We explain the effects of this mismatch and introduce Multi-Draft On-Policy Distillation (MDraft), a new parallel drafter distillation method which samples several draft blocks from the same context and aggregates the verifier's full-vocabulary conditionals across their prefixes. We use all the drafts to collect information from the verifier, allowing support from alternative prefixes to influence the training target. The resulting algorithm improves the acceptance length of a parallel drafter beyond the level of pre-training, and alleviates the effects of the distributional mismatch. MDraft changes only training, retaining the drafter's architecture and inference procedure, allowing the resulting drafter to be used as-is in place of the pretrained checkpoint. Experiments on a suite of mathematical, coding, and conversational tasks show longer accepted drafts than prior distillation methods.
Liang-Bo Ning, Yuchen Zhu, Heqing Huang City University of Hong Kong, Xin Wang, Yi Chang Jilin University, China, Li Qing, Wenqi Fan The Hong Kong Polytechnic University
WebAgents have demonstrated strong capabilities in autonomously completing complex web tasks, yet their computational efficiency vulnerabilities have received limited attention. Adversaries can inject malicious prompts into web pages, causing WebAgents to generate unnecessarily long reasoning processes and incur excessive computational cost, termed Computational Cost Attacks (CCA). In this paper, to systematically study this vulnerability under realistic black-box settings, we propose CostBomb, a generation-then-selection attack framework that leverages large language models to generate diverse adversarial prompts and a reinforcement learning–enhanced selector to identify the most effective perturbations. Extensive experiments on multiple real-world web benchmarks reveal that existing WebAgents are highly vulnerable to CCA, suffering substantial increases in computational cost without compromising successful task completion. Our findings highlight an overlooked dimension of WebAgent robustness and underscore the urgent need for efficiency-aware defenses.
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
Xuan Wang Virginia Tech, Shuxiang Cao NVIDIA, Yuchen Zhuang Google DeepMind, Wenqi Shi University of Texas Southwestern Medical Center
Multi-agent systems powered by large language models (LLMs) offer a promising paradigm for tackling complex reasoning, decision-making, and problem-solving tasks. However, achieving both effectiveness and efficiency in such systems remains a critical challenge. This tutorial introduces recent advances in building effective and efficient multi-agent LLM systems, focusing on three core components. First, we discuss the design of individual LLM agents. We present state-of-the-art techniques for enabling capable agents using efficient and compact LLMs, including model distillation, dynamic routing, and memory- and compute efficient serving, providing a foundation for scalable and responsive agent design under resource constraints. Second, we cover coordination and communication among agents, crucial for collective performance, highlighting methods for improving multi-agent reasoning and decision-making through prompt and graph optimization, sycophancy mitigation, and structured LLM-based frameworks. Last, we explore real-world applications of LLM agents in areas such as industry, healthcare, quantum computing, and various scientific domains.
尚无 AI 解读，展示原文摘要
cs.CL · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 5: Tutorial Abstracts) · 2026 · 正式发表