Yixiao Li

Yixiao Li, Yifan Yu, Chen Liang et al. · gatech, microsoft-research

Quantization is an indispensable technique for serving Large Language Models (LLMs) and has recently found its way into LoRA fine-tuning. In this work we focus on the scenario where quantization and LoRA fine-tuning are applied together on a pre-trained model. In such cases it is common to observe a consistent gap in the performance on downstream tasks between full fine-tuning and quantization plus LoRA fine-tuning approach. In response, we propose LoftQ (LoRA-Fine-Tuning-aware Quantization), a novel quantization framework that simultaneously quantizes an LLM and finds a proper low-rank initialization for LoRA fine-tuning. Such an initialization alleviates the discrepancy between the quantized and full-precision model and significantly improves generalization in downstream tasks. We evaluate our method on natural language understanding, question answering, summarization, and natural language generation tasks. Experiments show that our method is highly effective and outperforms existing quantization methods, especially in the challenging 2-bit and 2/4-bit mixed precision regimes. The code is available on https://github.com/yxli2123/LoftQ.

Yixiao Li, Yifan Yu, Qingru Zhang et al. · gatech, microsoft-research

Transformer models have achieved remarkable results in various natural language tasks, but they are often prohibitively large, requiring massive memories and computational resources. To reduce the size and complexity of these models, we propose LoSparse (Low-Rank and Sparse approximation), a novel model compression technique that approximates a weight matrix by the sum of a low-rank matrix and a sparse matrix. Our method combines the advantages of both low-rank approximations and pruning, while avoiding their limitations. Low-rank approximation compresses the coherent and expressive parts in neurons, while pruning removes the incoherent and non-expressive parts in neurons. Pruning enhances the diversity of low-rank approximations, and low-rank approximation prevents pruning from losing too many expressive neurons. We evaluate our method on natural language understanding, question answering, and natural language generation tasks. We show that it significantly outperforms existing compression methods.

Alexander Bukharin, Yixiao Li, Pengcheng He et al.

Reward design is a fundamental, yet challenging aspect of reinforcement learning (RL). Researchers typically utilize feedback signals from the environment to handcraft a reward function, but this process is not always effective due to the varying scale and intricate dependencies of the feedback signals. This paper shows by exploiting certain structures, one can ease the reward design process. Specifically, we propose a hierarchical reward modeling framework -- HERON for scenarios: (I) The feedback signals naturally present hierarchy; (II) The reward is sparse, but with less important surrogate feedback to help policy learning. Both scenarios allow us to design a hierarchical decision tree induced by the importance ranking of the feedback signals to compare RL trajectories. With such preference data, we can then train a reward model for policy learning. We apply HERON to several RL applications, and we find that our framework can not only train high performing agents on a variety of difficult tasks, but also provide additional benefits such as improved sample efficiency and robustness. Our code is available at \url{https://github.com/abukharin3/HERON}.

Wei Zhou, Yixiao Li, Hadi Amirpour et al.

Single-image super-resolution (SR) has achieved remarkable progress with deep learning, yet most approaches rely on distortion-oriented losses or heuristic perceptual priors, which often lead to a trade-off between fidelity and visual quality. To address this issue, we propose an \textit{Efficient Perceptual Bi-directional Attention Network (Efficient-PBAN)} that explicitly optimizes SR towards human-preferred quality. Unlike patch-based quality models, Efficient-PBAN avoids extensive patch sampling and enables efficient image-level perception. The proposed framework is trained on our self-constructed SR quality dataset that covers a wide range of state-of-the-art SR methods with corresponding human opinion scores. Using this dataset, Efficient-PBAN learns to predict perceptual quality in a way that correlates strongly with subjective judgments. The learned metric is further integrated into SR training as a differentiable perceptual loss, enabling closed-loop alignment between reconstruction and perceptual assessment. Extensive experiments demonstrate that our approach delivers superior perceptual quality. Code is publicly available at https://github.com/Lighting-YXLI/Efficient-PBAN.

Yixiao Li, Xin Li, Chris Wei Zhou et al.

This paper presents the ISRGC-Q Challenge, built upon the Image Super-Resolution Generated Content Quality Assessment (ISRGen-QA) dataset, and organized as part of the Visual Quality Assessment (VQualA) Competition at the ICCV 2025 Workshops. Unlike existing Super-Resolution Image Quality Assessment (SR-IQA) datasets, ISRGen-QA places a greater emphasis on SR images generated by the latest generative approaches, including Generative Adversarial Networks (GANs) and diffusion models. The primary goal of this challenge is to analyze the unique artifacts introduced by modern super-resolution techniques and to evaluate their perceptual quality effectively. A total of 108 participants registered for the challenge, with 4 teams submitting valid solutions and fact sheets for the final testing phase. These submissions demonstrated state-of-the-art (SOTA) performance on the ISRGen-QA dataset. The project is publicly available at: https://github.com/Lighting-YXLI/ISRGen-QA.

Siyuan Xu, Shiyang Li, Xin Liu et al.

Existing synthetic tool-use corpora are primarily designed for offline supervised fine-tuning, yet reinforcement learning (RL) requires executable environments that support reward-checkable online rollouts. We propose COVERT, a two-stage pipeline that first generates reliable base tool-use trajectories through self-evolving synthesis with multi-level validation, and then applies oracle-preserving augmentations that systematically increase environmental complexity. These augmentations introduce distractor tools, indirect or ambiguous user queries, and noisy, multi-format, or erroneous tool outputs, while strictly preserving oracle tool calls and final answers as ground truth. This design enables automatic reward computation via reference matching for standard cases and lightweight judge-assisted verification for special behaviors such as error detection, supporting RL optimization of tool-calling policies. On Qwen2.5-Instruct-14B, COVERT-RL improves overall accuracy on BFCL v3 from 56.5 to 59.9 and on ACEBench from 53.0 to 59.3, with minimal regressions on general-ability benchmarks; when stacked on SFT, it further reaches 62.1 and 61.8, confirming additive gains. These results suggest that oracle-preserving synthetic environments offer a practical RL refinement stage, complementary to SFT, for improving tool-use robustness under ambiguity and unreliable tool feedback.

Junxiao Liu, Zhijun Wang, Yixiao Li et al.

Long reasoning models often struggle in multilingual settings: they tend to reason in English for non-English questions; when constrained to reasoning in the question language, accuracies drop substantially. The struggle is caused by the limited abilities for both multilingual question understanding and multilingual reasoning. To address both problems, we propose TRIT (Translation-Reasoning Integrated Training), a self-improving framework that integrates the training of translation into multilingual reasoning. Without external feedback or additional multilingual data, our method jointly enhances multilingual question understanding and response generation. On MMATH, our method outperforms multiple baselines by an average of 7 percentage points, improving both answer correctness and language consistency. Further analysis reveals that integrating translation training improves cross-lingual question alignment by over 10 percentage points and enhances translation quality for both mathematical questions and general-domain text, with gains up to 8.4 COMET points on FLORES-200.

Yining Qian, Pengjie Wang, Yixiao Li et al.

Predicting drug-target affinity is fundamental to virtual screening and lead optimization. However, existing deep models often suffer from representation collapse in stringent cold-start regimes, where the scarcity of labels and domain shifts prevent the learning of transferable pharmacophores and binding motifs. In this paper, we propose Co-Diffusion, a novel affinity-aware framework that redefines DTA prediction as a constrained latent denoising process to enhance generalization. Co-Diffusion employs a two-stage paradigm: Stage I establishes an affinity-steered latent manifold by aligning drug and target embeddings under an explicit supervised objective, ensuring that the latent space reflects the intrinsic binding landscape. Stage II introduces modality-specific latent diffusion as a stochastic perturb-and-denoise regularizer, forcing the model to recover consistent affinity semantics from noisy structural representations. This approach effectively mitigates the reconstruction-regression conflict common in generative DTA models. Theoretically, we show that Co-Diffusion maximizes a variational lower bound on the joint likelihood of drug structures, protein sequences, and binding strength. Extensive experiments across multiple benchmarks demonstrate that Co-Diffusion significantly outperforms state-of-the-art baselines, particularly yielding superior zero-shot generalization on unseen molecular scaffolds and novel protein families-paving a robust path for in silico drug prioritization in unexplored chemical spaces.

Yixiao Li, Xiaoyuan Yang, Weide Liu et al.

As super-resolution (SR) techniques introduce unique distortions that fundamentally differ from those caused by traditional degradation processes (e.g., compression), there is an increasing demand for specialized video quality assessment (VQA) methods tailored to SR-generated content. One critical factor affecting perceived quality is temporal inconsistency, which refers to irregularities between consecutive frames. However, existing VQA approaches rarely quantify this phenomenon or explicitly investigate its relationship with human perception. Moreover, SR videos exhibit amplified inconsistency levels as a result of enhancement processes. In this paper, we propose \textit{Temporal Inconsistency Guidance for Super-resolution Video Quality Assessment (TIG-SVQA)} that underscores the critical role of temporal inconsistency in guiding the quality assessment of SR videos. We first design a perception-oriented approach to quantify frame-wise temporal inconsistency. Based on this, we introduce the Inconsistency Highlighted Spatial Module, which localizes inconsistent regions at both coarse and fine scales. Inspired by the human visual system, we further develop an Inconsistency Guided Temporal Module that performs progressive temporal feature aggregation: (1) a consistency-aware fusion stage in which a visual memory capacity block adaptively determines the information load of each temporal segment based on inconsistency levels, and (2) an informative filtering stage for emphasizing quality-related features. Extensive experiments on both single-frame and multi-frame SR video scenarios demonstrate that our method significantly outperforms state-of-the-art VQA approaches. The code is publicly available at https://github.com/Lighting-YXLI/TIG-SVQA-main.

Lawrence Liu, Inesh Chakrabarti, Yixiao Li et al.

Large language models (LLMs) exhibit remarkable performance across various natural language processing tasks but suffer from immense computational and memory demands, limiting their deployment in resource-constrained environments. To address this challenge, we propose NoWag (Normalized Weight and Activation Guided Compression), a unified framework for one-shot shape preserving compression algorithms. We apply NoWag to compress Llama-2 (7B, 13B, 70B) and Llama-3 (8B, 70B) models using two popular shape-preserving techniques: vector quantization (NoWag-VQ) and unstructured/semi-structured pruning (NoWag-P). Our results show that NoWag-VQ significantly outperforms state-of-the-art one-shot vector quantization methods, while NoWag-P performs competitively against leading pruning techniques. These findings highlight underlying commonalities between these compression paradigms and suggest promising directions for future research. Our code is available at https://github.com/LawrenceRLiu/NoWag

Yuqing Luo, Yixiao Li, Jiang Liu et al.

Image complexity assessment (ICA) is a challenging task in perceptual evaluation due to the subjective nature of human perception and the inherent semantic diversity in real-world images. Existing ICA methods predominantly rely on hand-crafted or shallow convolutional neural network-based features of a single visual modality, which are insufficient to fully capture the perceived representations closely related to image complexity. Recently, cross-modal scene semantic information has been shown to play a crucial role in various computer vision tasks, particularly those involving perceptual understanding. However, the exploration of cross-modal scene semantic information in the context of ICA remains unaddressed. Therefore, in this paper, we propose a novel ICA method called Cross-Modal Scene Semantic Alignment (CM-SSA), which leverages scene semantic alignment from a cross-modal perspective to enhance ICA performance, enabling complexity predictions to be more consistent with subjective human perception. Specifically, the proposed CM-SSA consists of a complexity regression branch and a scene semantic alignment branch. The complexity regression branch estimates image complexity levels under the guidance of the scene semantic alignment branch, while the scene semantic alignment branch is used to align images with corresponding text prompts that convey rich scene semantic information by pair-wise learning. Extensive experiments on several ICA datasets demonstrate that the proposed CM-SSA significantly outperforms state-of-the-art approaches. Codes are available at https://github.com/XQ2K/First-Cross-Model-ICA.

Ajay Jaiswal, Jianyu Wang, Yixiao Li et al.

Sparsely activated Mixture-of-Experts (SMoE) has shown promise in scaling up the learning capacity of neural networks. However, vanilla SMoEs have issues such as expert redundancy and heavy memory requirements, making them inefficient and non-scalable, especially for resource-constrained scenarios. Expert-level sparsification of SMoEs involves pruning the least important experts to address these limitations. In this work, we aim to address three questions: (1) What is the best recipe to identify the least knowledgeable subset of experts that can be dropped with minimal impact on performance? (2) How should we perform expert dropping (one-shot or iterative), and what correction measures can we undertake to minimize its drastic impact on SMoE subnetwork capabilities? (3) What capabilities of full-SMoEs are severely impacted by the removal of the least dominant experts, and how can we recover them? Firstly, we propose MoE Experts Compression Suite (MC-Suite), which is a collection of some previously explored and multiple novel recipes to provide a comprehensive benchmark for estimating expert importance from diverse perspectives, as well as unveil numerous valuable insights for SMoE experts. Secondly, unlike prior works with a one-shot expert pruning approach, we explore the benefits of iterative pruning with the re-estimation of the MC-Suite criterion. Moreover, we introduce the benefits of task-agnostic fine-tuning as a correction mechanism during iterative expert dropping, which we term MoE Lottery Subnetworks. Lastly, we present an experimentally validated conjecture that, during expert dropping, SMoEs' instruction-following capabilities are predominantly hurt, which can be restored to a robust level subject to external augmentation of instruction-following capabilities using k-shot examples and supervised fine-tuning.

Yixiao Li, Xianzhi Du, Ajay Jaiswal et al.

Recent advancements in large language models have intensified the need for efficient and deployable models within limited inference budgets. Structured pruning pipelines have shown promise in token efficiency compared to training target-size models from scratch. In this paper, we advocate incorporating enlarged model pretraining, which is often ignored in previous works, into pruning. We study the enlarge-and-prune pipeline as an integrated system to address two critical questions: whether it is worth pretraining an enlarged model even when the model is never deployed, and how to optimize the entire pipeline for better pruned models. We propose an integrated enlarge-and-prune pipeline, which combines enlarge model training, pruning, and recovery under a single cosine annealing learning rate schedule. This approach is further complemented by a novel iterative structured pruning method for gradual parameter removal. The proposed method helps to mitigate the knowledge loss caused by the rising learning rate in naive enlarge-and-prune pipelines and enable effective redistribution of model capacity among surviving neurons, facilitating smooth compression and enhanced performance. We conduct comprehensive experiments on compressing 2.8B models to 1.3B with up to 2T tokens in pretraining. It demonstrates the integrated approach not only provides insights into the token efficiency of enlarged model pretraining but also achieves superior performance of pruned models.

Yixiao Li, Julia Barth, Thomas Kiefer et al.

Multi-modal behavior cloning faces significant challenges due to mode averaging and mode collapse, where traditional models fail to capture diverse input-output mappings. This problem is critical in applications like robotics, where modeling multiple valid actions ensures both performance and safety. We propose EBGAN-MDN, a framework that integrates energy-based models, Mixture Density Networks (MDNs), and adversarial training. By leveraging a modified InfoNCE loss and an energy-enforced MDN loss, EBGAN-MDN effectively addresses these challenges. Experiments on synthetic and robotic benchmarks demonstrate superior performance, establishing EBGAN-MDN as a effective and efficient solution for multi-modal learning tasks.

Yixiao Li, Xiaoyuan Yang, Guanghui Yue et al.

Many super-resolution (SR) algorithms have been proposed to increase image resolution. However, full-reference (FR) image quality assessment (IQA) metrics for comparing and evaluating different SR algorithms are limited. In this work, we propose the Perception-oriented Bidirectional Attention Network (PBAN) for image SR FR-IQA, which is composed of three modules: an image encoder module, a perception-oriented bidirectional attention (PBA) module, and a quality prediction module. First, we encode the input images for feature representations. Inspired by the characteristics of the human visual system, we then construct the perception-oriented PBA module. Specifically, different from existing attention-based SR IQA methods, we conceive a Bidirectional Attention to bidirectionally construct visual attention to distortion, which is consistent with the generation and evaluation processes of SR images. To further guide the quality assessment towards the perception of distorted information, we propose Grouped Multi-scale Deformable Convolution, enabling the proposed method to adaptively perceive distortion. Moreover, we design Sub-information Excitation Convolution to direct visual perception to both sub-pixel and sub-channel attention. Finally, the quality prediction module is exploited to integrate quality-aware features and regress quality scores. Extensive experiments demonstrate that our proposed PBAN outperforms state-of-the-art quality assessment methods.

Ilgee Hong, Zichong Li, Alexander Bukharin et al.

Reinforcement learning from human feedback (RLHF) is a prevalent approach to align AI systems with human values by learning rewards from human preference data. Due to various reasons, however, such data typically takes the form of rankings over pairs of trajectory segments, which fails to capture the varying strengths of preferences across different pairs. In this paper, we propose a novel adaptive preference loss, underpinned by distributionally robust optimization (DRO), designed to address this uncertainty in preference strength. By incorporating an adaptive scaling parameter into the loss for each pair, our method increases the flexibility of the reward function. Specifically, it assigns small scaling parameters to pairs with ambiguous preferences, leading to more comparable rewards, and large scaling parameters to those with clear preferences for more distinct rewards. Computationally, our proposed loss function is strictly convex and univariate with respect to each scaling parameter, enabling its efficient optimization through a simple second-order algorithm. Our method is versatile and can be readily adapted to various preference optimization frameworks, including direct preference optimization (DPO). Our experiments with robotic control and natural language generation with large language models (LLMs) show that our method not only improves policy performance but also aligns reward function selection more closely with policy optimization, simplifying the hyperparameter tuning process.

Yixiao Li, Gloria Lin, Thomas Lau et al.

The objective of the change-point detection is to discover the abrupt property changes lying behind the time-series data. In this paper, we firstly summarize the definition and in-depth implication of the changepoint detection. The next stage is to elaborate traditional and some alternative model-based changepoint detection algorithms. Finally, we try to go a bit further in the theory and look into future research directions.