Zifeng Wang
Zifeng Wang is a research scientist at Google, working on exciting machine learning algorithms and their applications. His research interests include efficient model adaptation, continual learning, and large language models. He received his PhD in machine learning from Northeastern University advised by Prof. Jennifer Dy.
Research Areas
Authored Publications
Sort By
Model Swarms: Collaborative Search of Adapted LLM Experts via Swarm Intelligence
Shangbin Feng
Yike Wang
Nathalie Rauschmayr
Yejin Choi
Yulia Tsvetkov
ICML 2025
Preview abstract
We propose Model Swarms, a collaborative search algorithm to adapt LLM experts via swarm intelligence. Specifically, Model Swarms starts with a pool of LLM experts and a utility function. Guided by the best-found checkpoints across models, diverse LLM experts collaboratively move in the weight space and search for adapted models that optimize the utility function. Compared to existing model composition approaches, Model Swarms offers modularity, works in low-data regimes, and doesn't need assumptions about existing experts and how they should be composed. Extensive experiments demonstrate that Model Swarms could flexibly adapt LLM experts to a single dataset, multi-dataset domains, reward models, as well as diverse human preferences. Further analysis reveals that LLM experts discover previously unseen capabilities in the search process and that Model Swarms enable the weak-to-strong transition of experts through the collaborative search process.
View details
Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting
Zilong Wang
Steven Zheng
Swaroop Mishra
Yuwei Zhang
Anush Mattapalli
Ankur Taly
Jingbo Shang
ICLR 2025
Preview abstract
Retrieval augmented generation (RAG) has attracted a lot of attention across both academia and industry due to its capability in inserting timely and accurate evidence to the generation by large language models. However, the introduction of retrieved evidence largely makes the input prompt longer, which would harm the understanding quality of large language models and make it slower in actual usage scenarios. To solve these issues, we propose SpeculativeRAG, which leverages a smaller LLM to conduct the retrieval augmented generation for a larger LLM. The smaller LLM can digest a few pieces of evidence and generate multiple pieces of drafts in parallel rapidly, and these drafts will be verified by a large LLM to guarantee the quality. We achieve a higher speed as well as a better quality in the RAG results.
View details
Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling
Wenda Xu
Dhruv Madeka
Lei Li
William Wang
Rishabh Agarwal
2025
Preview abstract
Recent advances in knowledge distillation (KD) have enabled smaller student models to approach the performance of larger teacher models. However, popular methods such as supervised KD and on-policy KD, are adversely impacted by the knowledge gaps between teacher-student in practical scenarios. Supervised KD suffers from a distribution mismatch between training with a static dataset and inference over final student-generated outputs. Conversely, on-policy KD, which uses student-generated samples for training, can suffer from low-quality training examples with which teacher models are not familiar, resulting in inaccurate teacher feedback. To address these limitations, we introduce Speculative Knowledge Distillation (SKD), a novel approach that leverages cooperation between student and teacher models to generate high-quality training data on-the-fly while aligning with the student’s inference-time distribution. In SKD, the student proposes tokens, and the teacher replaces poorly ranked ones based on its own distribution, transferring high-quality knowledge adaptively. We evaluate SKD on various text generation tasks, including translation, summarization, math, and instruction following, and show that SKD consistently outperforms existing KD methods across different domains, data sizes, and model initialization strategies
View details
Heterogeneous Swarms: Jointly Optimizing Model Roles and Weights for Multi-LLM Systems
Shangbin Feng
Yike Wang
Weijia Shi
Huang Xia
Luke Zettlemoyer
Yulia Tsvetkov
NeurIPS 2025
Preview abstract
We propose Heterogeneous Swarms, an algorithm to discover and adapt multi-LLM systems by jointly optimizing model roles and weights. Given a pool of LLM experts and a utility function, Heterogeneous Swarms employs two iterative steps: role-step and weight-step. For role-step, we interpret model roles as input-output relationships and optimize the directed acyclic graph (DAG) of LLMs representing a multi-LLM system. Starting from a swarm of randomly initialized continuous adjacency matrices, we decode them into discrete DAGs, call the LLMs in topological order with message passing, evaluate on the utility function, and optimize the adjacency matrices with swarm intelligence based on the utility score. For weight-step, we define JFK-score to evaluate the contribution of individual LLMs in the best-found DAG of the role-step, then optimize model weights with swarm intelligence based on the JFK-score. Extensive experiments demonstrate that Heterogeneous Swarms outperforms 15 baselines spanning role-based and weight-based approaches by 18.5% on average across 12 tasks and contexts. Further analysis reveals that Heterogeneous Swarms discovers multi-LLM systems with heterogeneous model roles and substantial collaborative gains, and benefits from the diversity of initial LLMs.
View details
Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling
Lei Li
Wenda Xu
Rishabh Agarwal
William Wang
Dhruv Madeka
ICLR 2025
Preview abstract
Recent knowledge distillation (KD) research made significant progress on improving smaller student models to match larger teachers' performances. Two noticeable methods, supervised KD and on-policy KD emerged as the state-of-the-art approaches. However, supervised KD for auto-regressive models suffers from distribution mismatch between training over fixed dataset and inference over student generated outputs. Conversely, on-policy KD, which uses student-generated samples for training, can suffer from low-quality training examples and the teacher's potential inaccuracies in assessing these samples. To address these limitations, we introduce Speculative Knowledge Distillation (SKD). Instead of solely training on teacher- or student-proposed samples, SKD leverages the student model to initially propose tokens following its own generation distribution. Subsequently, the teacher model is employed to replace tokens that are deemed out-of-distribution. Compared with supervised KD, the samples generated by SKD are more likely to align with the student's inference-time distribution, and 2) SKD can mitigate the generation of low-quality sequences by incorporating the teacher's feedback at each token. Furthermore, we demonstrate that SKD is a generic framework capable of implementing both supervised and on-policy knowledge distillation as specific instances. To validate SKD's effectiveness, we apply it to distill autoregressive large language models for various tasks, including translation, summarization, math, and instruction following. Our experiments consistently demonstrate SKD's superior performance compared to existing methods across different domains, tasks, data sizes, and model initialization strategies.
View details
In Prospect and Retrospect: Reflective Memory Management for Long-term Personalized Dialog Agents
Zhen Tan
George Lee
Anand Iyer
Tianlong Chen
Huan Liu
ACL 2025
Preview abstract
Large Language Models (LLMs) have made significant progress in open-ended dialogue, yet their inability to retain and retrieve relevant information from long-term interactions limits their effectiveness in applications requiring sustained personalization. External memory mechanisms have been proposed to address this limitation, enabling LLMs to maintain conversational continuity. However, existing approaches struggle with two key challenges. First, rigid memory granularity fails to capture the natural semantic structure of conversations, leading to fragmented and incomplete representations. Second, fixed retrieval mechanisms cannot adapt to diverse dialogue contexts and user interaction patterns. In this work, we propose Reflective Memory Management (RMM), a novel mechanism for long-term dialogue agents, integrating forward- and backward-looking reflections: (1) Prospective Reflection, which dynamically summarizes interactions across granularities—utterances, turns, and sessions—into a personalized memory bank for effective future retrieval, and (2) Retrospective Reflection, which iteratively refines the retrieval in an online reinforcement learning (RL) manner based on LLMs’ cited evidence. Experiments show that RMM demonstrates consistent improvement across various metrics and benchmarks. For example, RMM shows more than 10% accuracy improvement over the baseline without memory management on the LongMemEval dataset.
View details
Chain-of-Table: Evolves Tables in the LLM Reasoning Chain for Table Understanding
Zilong Wang
Hao Zhang
Chun-Liang Li
Jingbo Shang
ICLR (2024)
Preview abstract
Table-based reasoning with large language models (LLMs) is a promising direction to tackle many table understanding tasks, such as table-based question answering and fact verification. Compared with generic reasoning, table-based reasoning requires the extraction of underlying semantics from both free-form questions and semi-structured tabular data. Chain-of-Thought and its similar approaches incorporate the reasoning chain in the form of textual context, but it is still an open question how to effectively leverage tabular data in the reasoning chain. We propose the Chain-of-Table framework, where tabular data is explicitly used in the reasoning chain as a proxy for intermediate thoughts. Specifically, we guide LLMs using in-context learning to iteratively generate operations and update the table to represent a tabular reasoning chain. LLMs can therefore dynamically plan the next operation based on the results of the previous ones. This continuous evolution of the table forms a chain, showing the reasoning process for a given tabular problem. The chain carries structured information of the intermediate results, enabling more accurate and reliable predictions. Chain-of-Table achieves new state-of-the-art performance on WikiTQ, FeTaQA, and TabFact benchmarks across multiple LLM choices.
View details
LMDX: Language Model-based Document Information Extraction And Localization
Kai Kang
Florian Luisier
Xiaoyu Sun
Ramya Sree Boppana
Zilong Wang
Jiaqi Mu
Hao Zhang
Nan Hua
Findings of the Association for Computational Linguistics ACL 2024, Association for Computational Linguistics, Bangkok, Thailand and virtual meeting, pp. 15140-15168
Preview abstract
Large Language Models (LLM) have revolutionized Natural Language Processing (NLP), improving state-of-the-art and exhibiting emergent capabilities across various tasks. However, their application in extracting information from visually rich documents, which is at the core of many document processing workflows and involving the extraction of key entities from semi-structured documents, has not yet been successful. The main obstacles to adopting LLMs for this task include the absence of layout encoding within LLMs, which is critical for high quality extraction, and the lack of a grounding mechanism to localize the predicted entities within the document. In this paper, we introduce Language Model-based Document Information EXtraction and Localization (LMDX), a methodology to reframe the document information extraction task for a LLM. LMDX enables extraction of singular, repeated, and hierarchical entities, both with and without training data, while providing grounding guarantees and localizing the entities within the document. Finally, we apply LMDX to the PaLM 2-S and Gemini Pro LLMs and evaluate it on VRDU and CORD benchmarks, setting a new state-of-the-art and showing how LMDX enables the creation of high quality, data-efficient parsers.
View details
Preview abstract
Grounded generation aims to equip language models (LMs) with the ability to produce more credible and accountable responses by accurately citing verifiable sources. However, existing methods, by either feeding LMs with raw or preprocessed materials, remain prone to errors. To address this, we introduce CaLM, a novel verification framework. CaLM leverages the insight that a robust grounded response should be consistent with information derived solely from its cited sources. Our framework empowers smaller LMs, which rely less on parametric memory and excel at processing relevant information given a query, to validate the output of larger LMs. Larger LM responses that closely align with the smaller LMs' output, which relies exclusively on cited documents, are verified. Responses showing discrepancies are iteratively refined through a feedback loop. Experiments on three open-domain question-answering datasets demonstrate significant performance gains of 1.5% to 7% absolute average without any required model fine-tuning.
View details
CodecLM: Aligning Language Models with Tailored Synthetic Data
Chun-Liang Li
Jin Miao
NAACL 2024
Preview abstract
Instruction tuning has emerged as the key in aligning large language models (LLMs) with specific task instructions, thereby mitigating the discrepancy between the next-token prediction objective and users' actual goals. To reduce the labor and time cost to collect or annotate data by humans, researchers start to explore the use of LLMs to generate instruction-aligned synthetic data. Recent works focus on generating diverse instructions and applying LLM to increase instruction complexity, often neglecting downstream use cases. It remains unclear how to tailor high-quality data to elicit better instruction-following abilities in different target instruction distributions and LLMs. To this end, we introduce CodecLM, a general framework for adaptively generating high-quality synthetic data for LLM alignment with different downstream instruction distributions and LLMs. Drawing on the Encode-Decode principles, we use LLMs as codecs to guide the data generation process. We first encode seed instructions into metadata, which are concise keywords generated on-the-fly to capture the target instruction distribution, and then decode metadata to create tailored instructions. We also introduce Self-Rubrics and Contrastive Filtering during decoding to tailor data-efficient samples. Extensive experiments on four open-domain instruction following benchmarks validate the effectiveness of CodecLM over the current state-of-the-arts.
View details
