Author name cluster

Zefan Cai

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

10 papers

2 author rows

TMLR Journal 2026 Journal Article

From Preferences to Prejudice: The Role of Alignment Tuning in Shaping Social Bias in Video Diffusion Models

Zefan Cai
Haoyi Qiu
Haozhe Zhao
Ke Wan
Jiachen Li
Jiuxiang Gu
Wen Xiao
Nanyun Peng

Recent advances in video diffusion models have significantly enhanced text-to-video generation, particularly through alignment tuning using reward models trained on human preferences. While these methods improve visual quality, they can unintentionally encode and amplify social biases. To systematically trace how such biases evolve throughout the alignment pipeline, we introduce VideoBiasEval, a comprehensive diagnostic framework for evaluating social representation in video generation. Grounded in established social bias taxonomies, VideoBiasEval employs an event-based prompting strategy to disentangle semantic content (verbs and contexts) from actor attributes (gender and ethnicity). It further introduces multi-granular metrics to evaluate (1) overall ethnicity bias, (2) gender bias conditioned on ethnicity, (3) distributional shifts in social attributes across model variants, and (4) the temporal persistence of bias within videos. Using this framework, we conduct the first end-to-end analysis connecting biases in human preference datasets, their amplification in reward models, and their propagation through alignment-tuned video diffusion models. Our results reveal that alignment tuning not only strengthens representational biases but also makes them temporally stable, producing smoother yet more stereotyped portrayals. These findings highlight the need for bias-aware evaluation and mitigation throughout the alignment process to ensure fair and socially responsible video generation.

PDF Details

ICLR Conference 2025 Conference Paper

A Spark of Vision-Language Intelligence: 2-Dimensional Autoregressive Transformer for Efficient Finegrained Image Generation

Liang Chen 0024
Sinan Tan
Zefan Cai
Weichu Xie
Haozhe Zhao
Yichi Zhang 0010
Junyang Lin
Jinze Bai

This work tackles the information loss bottleneck of vector-quantization (VQ) autoregressive image generation by introducing a novel model architecture called the 2-Dimensional Autoregression (DnD) Transformer. The DnD-Transformer predicts more codes for an image by introducing a new direction, **model depth**, along with the sequence length. Compared to 1D autoregression and previous work using similar 2D image decomposition such as RQ-Transformer, the DnD-Transformer is an end-to-end model that can generate higher quality images with the same backbone model size and sequence length, opening a new optimization perspective for autoregressive image generation. Furthermore, our experiments reveal that the DnD-Transformer's potential extends beyond generating natural images. It can even generate images with rich text and graphical elements in a self-supervised manner, demonstrating an understanding of these combined modalities. This has not been previously demonstrated for popular vision generative models such as diffusion models, showing a spark of vision-language intelligence when trained solely on images. Code, datasets and models are open at https://github.com/chenllliang/DnD-Transformer.

Details

ICML Conference 2025 Conference Paper

AdaptiveStep: Automatically Dividing Reasoning Step through Model Confidence

Yuliang Liu
Junjie Lu
Chaofeng Qu
Zhaoling Chen
Zefan Cai
Jason Klein Liu
Chonghan Liu
Yunhui Xia

Current approaches for training Process Reward Models (PRMs) often involve deconposing responses into multiple reasoning steps using rule-based techniques, such as using predefined placeholder tokens or setting the reasoning step’s length to a fixed size. These approaches overlook the fact that certain words don’t usually indicate true decision points. To address this, we propose AdaptiveStep, a method that divides reasoning steps based on the model’s confidence in predicting the next word, offering more information on decision-making at each step, improving downstream tasks like reward model training. Moreover, our method requires no manual annotation. Experiments with AdaptiveStep-trained PRMs in mathematical reasoning and code generation show that the outcome PRM achieves state-of-the-art Best-of-N performance, surpassing greedy search strategy with token-level value-guided decoding, while also reducing construction costs by over 30% compared to existing open-source PRMs. We also provide a thorough analysis and case study on its performance, transferability, and generalization capabilities. We provide our code on https: //github. com/Lux0926/ASPRM.

Details

TMLR Journal 2025 Journal Article

COMMA: A Communicative Multimodal Multi-Agent Benchmark

Timothy Ossowski
Danyal Maqbool
Jixuan Chen
Zefan Cai
Tyler J. Bradshaw
Junjie Hu

The rapid advances of multimodal agents built on large foundation models have largely overlooked their potential for language-based communication between agents in collaborative tasks. This oversight presents a critical gap in understanding their effectiveness in real-world deployments, particularly when communicating with humans. Existing agentic benchmarks fail to address key aspects of inter-agent communication and collaboration, particularly in scenarios where agents have unequal access to information and must work together to achieve tasks beyond the scope of individual capabilities. To fill this gap, we introduce COMMA: a novel puzzle benchmark designed to evaluate the collaborative performance of multimodal multi-agent systems through language communication. Our benchmark features a variety of multimodal puzzles, providing a comprehensive evaluation across four key categories of agentic capability in a communicative collaboration setting. Our findings reveal surprising weaknesses in state-of-the-art models, including strong proprietary models like GPT-4o and reasoning models like o4-mini. Many chain of thought reasoning models such as R1-Onevision and LLaVA-CoT struggle to outperform even a random baseline in agent-agent collaboration, indicating a potential growth area in their communication abilities.

PDF Details

ICLR Conference 2025 Conference Paper

No Preference Left Behind: Group Distributional Preference Optimization

Binwei Yao
Zefan Cai
Yun-Shiuan Chuang
Shanglin Yang
Ming Jiang 0018
Diyi Yang
Junjie Hu 0001

Preferences within a group of people are not uniform but follow a distribution. While existing alignment methods like Direct Preference Optimization (DPO) attempt to steer models to reflect human preferences, they struggle to capture the distributional pluralistic preferences within a group. These methods often skew toward dominant preferences, overlooking the diversity of opinions, especially when conflicting preferences arise. To address this issue, we propose Group Distributional Preference Optimization (GDPO), a novel framework that aligns language models with the distribution of preferences within a group by incorporating the concept of beliefs that shape individual preferences. GDPO calibrates a language model using statistical estimation of the group's belief distribution and aligns the model with belief-conditioned preferences, offering a more inclusive alignment framework than traditional methods. In experiments using both synthetic controllable opinion generation and real-world movie review datasets, we show that DPO fails to align with the targeted belief distributions, while GDPO consistently reduces this alignment gap during training. Additionally, our evaluation metrics demonstrate that GDPO outperforms existing approaches in aligning with group distributional preferences, marking a significant advance in pluralistic alignment.

Details

ICLR Conference 2025 Conference Paper

Not All Heads Matter: A Head-Level KV Cache Compression Method with Integrated Retrieval and Reasoning

Yu Fu 0009
Zefan Cai
Abedelkadir Asi
Wayne Xiong
Yue Dong 0002
Wen Xiao

Key-Value (KV) caching is a common technique to enhance the computational efficiency of Large Language Models (LLMs), but its memory overhead grows rapidly with input length. Prior work has shown that not all tokens are equally important for text generation, proposing layer-level KV cache compression to selectively retain key information. Recognizing the distinct roles of attention heads in generation, we propose HeadKV, a head-level KV cache compression method, and HeadKV-R2, which leverages a novel contextual reasoning ability estimation for compression. Our approach operates at the level of individual heads, estimating their importance for contextual QA tasks that require both retrieval and reasoning capabilities. Extensive experiments across diverse benchmarks (LongBench, LooGLE), model architectures (e.g., Llama-3-8B-Instruct, Mistral-7B-Instruct), and long-context abilities tests demonstrate that our head-level KV cache compression significantly outperforms strong baselines, particularly in low-resource settings (KV size = 64 & 128). Notably, our method retains just 1.5% of the KV cache while achieving 97% of the performance of the full KV cache on the contextual question answering benchmark.

Details

ICLR Conference 2025 Conference Paper

Omni-MATH: A Universal Olympiad Level Mathematic Benchmark for Large Language Models

Bofei Gao
Feifan Song 0001
Zhe Yang 0013
Zefan Cai
Yibo Miao
Qingxiu Dong
Lei Li 0039
Chenghao Ma

Recent advancements in large language models (LLMs) have led to significant breakthroughs in mathematical reasoning capabilities. However, existing benchmarks like GSM8K or MATH are now being solved with high accuracy (e.g., OpenAI o1 achieves 94.8% on MATH dataset), indicating their inadequacy for truly challenging these models. To bridge this gap, we propose a comprehensive and challenging benchmark specifically designed to assess LLMs' mathematical reasoning at the Olympiad level. Unlike existing Olympiad-related benchmarks, our dataset focuses exclusively on mathematics and comprises a vast collection of 4428 competition-level problems with rigorous human annotation. These problems are meticulously categorized into over 33 sub-domains and span more than 10 distinct difficulty levels, enabling a holistic assessment of model performance in Olympiad-mathematical reasoning. Furthermore, we conducted an in-depth analysis based on this benchmark. Our experimental results show that even the most advanced models, OpenAI o1-mini and OpenAI o1-preview, struggle with highly challenging Olympiad-level problems, with 60.54% and 52.55% accuracy, highlighting significant challenges in Olympiad-level mathematical reasoning.

Details

NeurIPS Conference 2025 Conference Paper

R-KV: Redundancy-aware KV Cache Compression for Reasoning Models

Zefan Cai
Wen Xiao
Hanshi Sun
Cheng Luo
Yikai Zhang
Ke Wan
Yucheng Li
Yeyang Zhou

Reasoning models have demonstrated impressive performance in self-reflection and chain-of-thought reasoning. However, they often produce excessively long outputs, leading to prohibitively large key-value (KV) caches during inference. While chain-of-thought inference significantly improves performance on complex reasoning tasks, it can also lead to reasoning failures when deployed with existing KV cache compression approaches. To address this, we propose Redundancy-aware KV Cache Compression for Reasoning models (R-KV), a novel method specifically targeting redundant tokens in reasoning models. Our method preserves nearly 100% of the full KV cache performance using only 10% of the KV cache, substantially outperforming existing KV cache baselines, which reach only 60% of the performance. Remarkably, R-KV even achieves 105% of full KV cache performance with 38% of the KV cache. This KV-cache reduction also leads to a 50% memory saving and a 2x speedup over standard chain-of-thought reasoning inference. Experimental results show that R-KV consistently outperforms existing KV cache compression baselines across two mathematical reasoning datasets.

PDF Details

ICLR Conference 2024 Conference Paper

MMICL: Empowering Vision-language Model with Multi-Modal In-Context Learning

Haozhe Zhao
Zefan Cai
Shuzheng Si
Xiaojian Ma 0001
Kaikai An
Liang Chen 0024
Zixuan Liu 0001
Sheng Wang 0012

Since the resurgence of deep learning, vision-language models (VLMs) enhanced by large language models (LLMs) have grown exponentially in popularity. However, while LLMs can utilize extensive background knowledge and task information with in-context learning, most VLMs still struggle with understanding complex multi-modal prompts with multiple images, making VLMs less effective in downstream vision-language tasks. In this paper, we address the limitation above by 1) introducing vision-language Model with **M**ulti-**M**odal **I**n-**C**ontext **L**earning(MMICL), a new approach to allow the VLM to deal with multi-modal inputs efficiently; 2) proposing a novel context scheme to augment the in-context learning ability of the VLM; 3) constructing the Multi-modal In-Context Learning (MIC) dataset, designed to enhance the VLM's ability to understand complex multi-modal prompts. Our experiments confirm that MMICL achieves new state-of-the-art zero-shot performance on a wide range of general vision-language tasks, especially for complex benchmarks, including MME and MMBench. Our analysis demonstrates that MMICL effectively tackles the challenge of complex multi-modal prompt understanding and emerges the impressive ICL ability. Furthermore, we observe that MMICL successfully alleviates language bias in VLMs, a common issue for VLMs that often leads to hallucination when faced with extensive textual context. Our code, dataset, dataset tool, and model are available at https://github.com/PKUnlp-icler/MIC.

Details

ICLR Conference 2023 Conference Paper

Compositional Task Representations for Large Language Models

Nan Shao
Zefan Cai
Hanwei Xu
Chonghua Liao
Yanan Zheng
Zhilin Yang 0001

Large language models have shown a remarkable cross-task generalization ability. Most prior work assumed that prompts effectively extract knowledge from language models to facilitate generalization to new tasks. This perspective led to numerous studies on improving prompts. In contrast, we introduce a new perspective, compositional generalization, that views each task as a composition of latent codes and generalizes to test tasks by a new composition of seen codes. To this end, we propose a novel prompt-free approach, Compositional Task Representations (CTR), that employs multi-task training to learn a discrete, compositional codebook. Empirically, our CTR substantially outperforms prompt-based methods in zero-label learning on average. According to our analysis, some of the learned CTR codes are interpretable to human and demonstrate a certain degree of controllability.

Details