Haoli Bai
Research Subjects
Haoli Bai has a total of 53 co-authors. Here are the top co-authors by number of shared publications:
Lu Hou
Huawei Technologies
4 shared publicationsXianzhi Yu
Huawei Technologies
3 shared publications- Chun Yuan
Tsinghua University, Beijing, China
2 shared publications - Lu Hou
Huawei Noah's Ark Lab
2 shared publications Manyi Zhang
Huawei Technologies
2 shared publications- Carlo Vittorio Cannistraci
Tsinghua University, China
1 shared publication
Latest Publications
- Open Access
- 16 October 2026
- Published by MDPI AG in Preprints.org
Towards Next-Generation Architectures of Large Language Models: A Survey
Large Language Models (LLMs) have advanced rapidly through increases in model scale, data, and computation, while their underlying architectures have also undergone substantial evolution. Beyond the canonical Transformer architecture, recent designs seek greater efficiency, scalability, and flexibility by rethinking how sequence information is processed, represented, and transformed. We survey efficient attention and state management, modern recurrent and other linear-complexity sequence models, Mixture-of-Experts architectures, and advances in activation functions, normalization, residual connections, positional encoding, and tokenization. We also cover multi-token prediction and adaptive and recurrent depth. Across these areas, we highlight key design principles and trade-offs among representational power, optimization stability, and computational cost. Complementing this component-wise treatment, we take a system-level perspective. We trace representative open-weight LLM families to identify recurring patterns in how architectural choices are combined, and synthesize these developments along four dimensions—processing structure, transformation dynamics, state management, and output commitment—characterizing how recent designs preserve, refine, or relax the assumptions of the canonical decoder. Together, this survey aims to provide a systematic overview of LLM architectural evolution and to chart pathways toward next-generation designs.
- 16 October 2026
- Published by MDPI AG in Preprints.org
Open AccessTowards Next-Generation Architectures of Large Language Models: A Survey
Large Language Models (LLMs) have advanced rapidly through increases in model scale, data, and computation, while their underlying architectures have also undergone substantial evolution. Beyond the canonical Transformer architecture, recent designs seek greater efficiency, scalability, and flexibility by rethinking how sequence information is processed, represented, and transformed. We survey efficient attention and state management, modern recurrent and other linear-complexity sequence models, Mixture-of-Experts architectures, and advances in activation functions, normalization, residual connections, positional encoding, and tokenization. We also cover multi-token prediction and adaptive and recurrent depth. Across these areas, we highlight key design principles and trade-offs among representational power, optimization stability, and computational cost. Complementing this component-wise treatment, we take a system-level perspective. We trace representative open-weight LLM families to identify recurring patterns in how architectural choices are combined, and synthesize these developments along four dimensions—processing structure, transformation dynamics, state management, and output commitment—characterizing how recent designs preserve, refine, or relax the assumptions of the canonical decoder. Together, this survey aims to provide a systematic overview of LLM architectural evolution and to chart pathways toward next-generation designs.
- Open Access
What Makes Low-Bit Quantization-Aware Training Work for Reasoning LLMs? A Systematic Study
Reasoning models excel at complex tasks such as coding and mathematics, yet their inference is often slow and token-inefficient. To improve the inference efficiency, post-training quantization (PTQ) usually comes with the cost of large accuracy drops, especially for reasoning tasks under low-bit settings. In this study, we present a systematic empirical study of quantization-aware training (QAT) for reasoning models. Our key findings include: (1) Knowledge distillation is a robust objective for reasoning models trained via either supervised fine-tuning or reinforcement learning; (2) PTQ provides a strong initialization for QAT, improving accuracy while reducing training cost; (3) Reinforcement learning remains feasible for quantized models given a viable cold start and yields additional gains; and (4) Aligning the PTQ calibration domain with the QAT training domain accelerates convergence and often improves the final accuracy. Finally, we consolidate these findings into an optimized workflow (Reasoning-QAT), and show that it consistently outperforms state-of-the-art PTQ methods across multiple LLM backbones and reasoning datasets. For instance, on Qwen3-0.6B, it surpasses GPTQ by 44.53% on MATH-500 and consistently recovers performance in the 2-bit regime.
Open AccessWhat Makes Low-Bit Quantization-Aware Training Work for Reasoning LLMs? A Systematic Study
Reasoning models excel at complex tasks such as coding and mathematics, yet their inference is often slow and token-inefficient. To improve the inference efficiency, post-training quantization (PTQ) usually comes with the cost of large accuracy drops, especially for reasoning tasks under low-bit settings. In this study, we present a systematic empirical study of quantization-aware training (QAT) for reasoning models. Our key findings include: (1) Knowledge distillation is a robust objective for reasoning models trained via either supervised fine-tuning or reinforcement learning; (2) PTQ provides a strong initialization for QAT, improving accuracy while reducing training cost; (3) Reinforcement learning remains feasible for quantized models given a viable cold start and yields additional gains; and (4) Aligning the PTQ calibration domain with the QAT training domain accelerates convergence and often improves the final accuracy. Finally, we consolidate these findings into an optimized workflow (Reasoning-QAT), and show that it consistently outperforms state-of-the-art PTQ methods across multiple LLM backbones and reasoning datasets. For instance, on Qwen3-0.6B, it surpasses GPTQ by 44.53% on MATH-500 and consistently recovers performance in the 2-bit regime.
- Open Access
E$^3$-Pruner: Towards Efficient, Economical, and Effective Layer Pruning for Large Language Models
With the increasing size of large language models, layer pruning has gained increased attention as a hardware-friendly approach for model compression. However, existing layer pruning methods struggle to simultaneously address key practical deployment challenges, including performance degradation, high training costs, and limited acceleration. To overcome these limitations, we propose \name, a task-\underline{E}ffective, training-\underline{E}conomical and inference-\underline{E}fficient layer pruning framework. \namespace introduces two key innovations: (1) a differentiable mask optimization method using a Gumbel-TopK sampler, enabling efficient and precise pruning mask search; and (2) an entropy-aware adaptive knowledge distillation strategy that enhances task performance. Extensive experiments over diverse model architectures and benchmarks demonstrate the superiority of our method over state-of-the-art approaches. Notably, \namespace achieves 96\% accuracy, a mere 0.8\% drop from the original model (96.8\%) on MATH-500 when pruning 25\% layers of Qwen3-32B, outperforming existing SOTA (95\%), with a 1.33$\times$ inference speedup by consuming merely 0.5B tokens (0.5\% of the post-training data volume).
Open AccessE$^3$-Pruner: Towards Efficient, Economical, and Effective Layer Pruning for Large Language Models
With the increasing size of large language models, layer pruning has gained increased attention as a hardware-friendly approach for model compression. However, existing layer pruning methods struggle to simultaneously address key practical deployment challenges, including performance degradation, high training costs, and limited acceleration. To overcome these limitations, we propose \name, a task-\underline{E}ffective, training-\underline{E}conomical and inference-\underline{E}fficient layer pruning framework. \namespace introduces two key innovations: (1) a differentiable mask optimization method using a Gumbel-TopK sampler, enabling efficient and precise pruning mask search; and (2) an entropy-aware adaptive knowledge distillation strategy that enhances task performance. Extensive experiments over diverse model architectures and benchmarks demonstrate the superiority of our method over state-of-the-art approaches. Notably, \namespace achieves 96\% accuracy, a mere 0.8\% drop from the original model (96.8\%) on MATH-500 when pruning 25\% layers of Qwen3-32B, outperforming existing SOTA (95\%), with a 1.33$\times$ inference speedup by consuming merely 0.5B tokens (0.5\% of the post-training data volume).
- Open Access
The Synergy Dilemma of Long-CoT SFT and RL: Investigating Post-Training Techniques for Reasoning VLMs
Large vision-language models (VLMs) increasingly adopt post-training techniques such as long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL) to elicit sophisticated reasoning. While these methods exhibit synergy in language-only models, their joint effectiveness in VLMs remains uncertain. We present a systematic investigation into the distinct roles and interplay of long-CoT SFT and RL across multiple multimodal reasoning benchmarks. We find that SFT improves performance on difficult questions by in-depth, structured reasoning, but introduces verbosity and degrades performance on simpler ones. In contrast, RL promotes generalization and brevity, yielding consistent improvements across all difficulty levels, though the improvements on the hardest questions are less prominent compared to SFT. Surprisingly, combining them through two-staged, interleaved, or progressive training strategies, as well as data mixing and model merging, all fails to produce additive benefits, instead leading to trade-offs in accuracy, reasoning style, and response length. This ``synergy dilemma'' highlights the need for more seamless and adaptive approaches to unlock the full potential of combined post-training techniques for reasoning VLMs.
Open AccessThe Synergy Dilemma of Long-CoT SFT and RL: Investigating Post-Training Techniques for Reasoning VLMs
Large vision-language models (VLMs) increasingly adopt post-training techniques such as long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL) to elicit sophisticated reasoning. While these methods exhibit synergy in language-only models, their joint effectiveness in VLMs remains uncertain. We present a systematic investigation into the distinct roles and interplay of long-CoT SFT and RL across multiple multimodal reasoning benchmarks. We find that SFT improves performance on difficult questions by in-depth, structured reasoning, but introduces verbosity and degrades performance on simpler ones. In contrast, RL promotes generalization and brevity, yielding consistent improvements across all difficulty levels, though the improvements on the hardest questions are less prominent compared to SFT. Surprisingly, combining them through two-staged, interleaved, or progressive training strategies, as well as data mixing and model merging, all fails to produce additive benefits, instead leading to trade-offs in accuracy, reasoning style, and response length. This ``synergy dilemma'' highlights the need for more seamless and adaptive approaches to unlock the full potential of combined post-training techniques for reasoning VLMs.
- Open Access
A Simple Linear Patch Revives Layer-Pruned Large Language Models
Layer pruning has become a popular technique for compressing large language models (LLMs) due to its simplicity. However, existing layer pruning methods often suffer from significant performance drops. We identify that this degradation stems from the mismatch of activation magnitudes across layers and tokens at the pruning interface. To address this, we propose LinearPatch, a simple plug-and-play technique to revive the layer-pruned LLMs. The proposed method adopts Hadamard transformation to suppress massive outliers in particular tokens, and channel-wise scaling to align the activation magnitudes. These operations can be fused into a single matrix, which functions as a patch to bridge the pruning interface with negligible inference overhead. LinearPatch retains up to 94.15% performance of the original model when pruning 5 layers of LLaMA-3-8B on the question answering benchmark, surpassing existing state-of-the-art methods by 4%. In addition, the patch matrix can be further optimized with memory efficient offline knowledge distillation. With only 5K samples, the retained performance of LinearPatch can be further boosted to 95.16% within 30 minutes on a single computing card.
Open AccessA Simple Linear Patch Revives Layer-Pruned Large Language Models
Layer pruning has become a popular technique for compressing large language models (LLMs) due to its simplicity. However, existing layer pruning methods often suffer from significant performance drops. We identify that this degradation stems from the mismatch of activation magnitudes across layers and tokens at the pruning interface. To address this, we propose LinearPatch, a simple plug-and-play technique to revive the layer-pruned LLMs. The proposed method adopts Hadamard transformation to suppress massive outliers in particular tokens, and channel-wise scaling to align the activation magnitudes. These operations can be fused into a single matrix, which functions as a patch to bridge the pruning interface with negligible inference overhead. LinearPatch retains up to 94.15% performance of the original model when pruning 5 layers of LLaMA-3-8B on the question answering benchmark, surpassing existing state-of-the-art methods by 4%. In addition, the patch matrix can be further optimized with memory efficient offline knowledge distillation. With only 5K samples, the retained performance of LinearPatch can be further boosted to 95.16% within 30 minutes on a single computing card.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 1-3
Efficient Inference for Large Language Models –Algorithm, Model, and System
Xuefei Ning, Guohao Dai, Haoli Bai, Lu Hou, Yu Wang. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Tutorial Abstracts. 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 1-3
Open AccessEfficient Inference for Large Language Models –Algorithm, Model, and System
Xuefei Ning, Guohao Dai, Haoli Bai, Lu Hou, Yu Wang. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Tutorial Abstracts. 2025.
- Open Access
- 27 May 2024
- Published by MDPI AG in Preprints.org
Plug-and-Play: An Efficient Post-training Pruning Method for Large Language Models
With the rapid growth of large language models (LLMs), there is increasing demand for memory and computation in LLMs. Recent efforts on post-training pruning of LLMs aim to reduce the model size and computation requirements, yet the performance is still sub-optimal. In this paper, we present a plug-and-play solution for post-training pruning of LLMs. The proposed solution has two innovative components: 1) Relative Importance and Activations (RIA), a new pruning metric that jointly considers the weight and activations efficiently on LLMs; and 2) Channel Permutation, a new approach to maximally preserve important weights under N:M sparsity. The two proposed components can be readily combined to further enhance the N:M semi-structured pruning of LLMs. Our empirical experiments show that RIA alone can already surpass all existing post-training pruning methods on prevalent LLMs, e.g., LLaMA ranging from 7B to 65B. Furthermore, N:M semi-structured pruning with channel permutation can even outperform the original LLaMA2-70B on zero-shot tasks, together with practical speed-up on specific hardware. Our code is available at: https://github.com/biomedical-cybernetics/Relative-importance-and-activation-pruning
- 27 May 2024
- Published by MDPI AG in Preprints.org
Open AccessPlug-and-Play: An Efficient Post-training Pruning Method for Large Language Models
With the rapid growth of large language models (LLMs), there is increasing demand for memory and computation in LLMs. Recent efforts on post-training pruning of LLMs aim to reduce the model size and computation requirements, yet the performance is still sub-optimal. In this paper, we present a plug-and-play solution for post-training pruning of LLMs. The proposed solution has two innovative components: 1) Relative Importance and Activations (RIA), a new pruning metric that jointly considers the weight and activations efficiently on LLMs; and 2) Channel Permutation, a new approach to maximally preserve important weights under N:M sparsity. The two proposed components can be readily combined to further enhance the N:M semi-structured pruning of LLMs. Our empirical experiments show that RIA alone can already surpass all existing post-training pruning methods on prevalent LLMs, e.g., LLaMA ranging from 7B to 65B. Furthermore, N:M semi-structured pruning with channel permutation can even outperform the original LLaMA2-70B on zero-shot tasks, together with practical speed-up on specific hardware. Our code is available at: https://github.com/biomedical-cybernetics/Relative-importance-and-activation-pruning