Papers by Bohan Wang
Sparse-RL: Breaking the Memory Wall in LLM Reinforcement Learning via Stable Sparse Rollouts (2026.acl-long)
Copied to clipboard
Sijia Luo, Xiaokang Zhang, Yuxuan Hu, Bohan Zhang, Ke Wang, Jinbo Su, Mengshu Sun, Lei Liang, Jing Zhang
| Challenge: | Existing methods for storing key-value caches during long-horizon rollouts cause performance collapses. |
| Approach: | They propose a new training paradigm that empowers stable RL training under sparse rollouts. |
| Outcome: | The proposed model reduces rollout overhead while maintaining the performance. |
On the Sentence Embeddings from Pre-trained Language Models (2020.emnlp-main)
Copied to clipboard
| Challenge: | Pre-trained contextual representations like BERT have been widely used for NLP tasks. |
| Approach: | They propose to transform anisotropic sentence embedding distribution to smooth and isotropic Gaussian distribution by normalizing flows that are learned with an unsupervised objective. |
| Outcome: | The proposed method achieves significant performance gains over state-of-the-art embeddings on a variety of semantic textual similarity tasks. |
BSFA: Leveraging the Subspace Dichotomy to Accelerate Neural Network Training (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies highlight a fundamental dichotomy in deep learning optimization: parameter updates along the top eigendirections of the loss Hessian (Dom-space) capture most of the update magnitude, while updates in the orthogonal component (Bulk-space), have smaller magnitudes but drive most learning progress. |
| Approach: | They propose a plug-and-play framework that scales update components projected onto distinct subspaces and a block-wise strategy that applies this estimation on a per-parameter-block basis. |
| Outcome: | The proposed framework accelerates training by differentially scaling update components projected onto distinct subspaces, while enhancing stability by moderating updates in dominant subspace and boosting convergence speed by amplifying updates in bulk-space. |
A Survey on Efficient Large Language Model Training: From Data-centric Perspectives (2025.acl-long)
Copied to clipboard
Junyu Luo, Bohan Wu, Xiao Luo, Zhiping Xiao, Yiqiao Jin, Rong-Cheng Tu, Nan Yin, Yifan Wang, Jingyang Yuan, Wei Ju, Ming Zhang
| Challenge: | achieving data-efficient post-training of Large Language Models is a key research question. |
| Approach: | They propose a taxonomy of data-efficient LLM post-training methods from a data-centric perspective. |
| Outcome: | The proposed methods cover data selection, data quality enhancement, synthetic data generation, data distillation and compression, and self-evolving data ecosystems. |
MTP-RL: Acceleration of Reinforcement Learning Rollouts with Policy-Aligned Multi-Token Prediction (2026.findings-acl)
Copied to clipboard
| Challenge: | Reinforcement learning (RL) is widely applied to boost the performance of pretrained models, yet its training efficiency is severely constrained by rollout generation. |
| Approach: | They propose a framework that accelerates the rollout phase for diverse models by equipping a pipeline to equip the multi-layer parameter-sharing MTP for all models and an advantage-aware MTP optimization strategy. |
| Outcome: | The proposed framework achieves stable growth of acceptance length during RL training, and also accelerates RL rollouts, achieving an average 23.1%–55.3% reduction in rollout time compared to baselines. |
AgentAsk: Multi-Agent Systems Need to Ask (2026.acl-long)
Copied to clipboard
Bohan Lin, Kuo Yang, Zelin Tan, Yingchuan Lai, Chen Zhang, Guibin Zhang, Xinlei Yu, Miao Yu, Xu Wang, Yudong Zhang, Yang Wang
| Challenge: | Multi-agent systems fail to consistently outperform strong single-a agent baselines due to error propagation at inter-aggent message handoffs. |
| Approach: | They propose an edge-level error taxonomy that identifies four main errors in multi-agent interactions as data gaps, signal corruption, referential drift and capacity gaps as primary sources of failure. |
| Outcome: | The proposed module outperforms existing systems on five benchmarks and is architecture-agnostic. |
MetaPrompting: Learning to Learn Better Prompts (2022.coling-1)
Copied to clipboard
| Challenge: | Recent research on prompting moves from discrete tokens based "hard prompts" to continuous "soft prompts", which employ learnable vectors as pseudo prompt tokens and achieve better performance. |
| Approach: | They propose a generalized soft prompting method that uses model-agnostic meta-learning to find better initialization for soft prompts. |
| Outcome: | The proposed method improves on three datasets and brings new state-of-the-art performance. |
Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens (2026.findings-acl)
Copied to clipboard
Chengshuai Zhao, Zhen Tan, Pingchuan Ma, Dawei Li, Bohan Jiang, Yancheng Wang, Yingzhen Yang, Huan Liu
| Challenge: | Chain-of-Thought (CoT) prompting has been shown to be effective in eliciting structured reasoning from large language models (LLMs). |
| Approach: | They propose a data distribution lens to understand when and why CoT reasoning fails . they propose 'data-based' training that trains LLMs from scratch . |
| Outcome: | The proposed model enables models to generate reasoning trajectories that approximate those observed during training. |
InfiMM: Advancing Multimodal Understanding with an Open-Sourced Visual Language Model (2024.findings-acl)
Copied to clipboard
Haogeng Liu, Quanzeng You, Yiqi Wang, Xiaotian Han, Bohan Zhai, Yongfei Liu, Wentao Chen, Yiren Jian, Yunzhe Tao, Jianbo Yuan, Ran He, Hongxia Yang
| Challenge: | InfiMM is a multimodal large language model that adapts to complex vision-language tasks. |
| Approach: | They present a Multimodal Large Language Model that adapts to intricate vision-language tasks using large-scale training data and comprehensive training strategies. |
| Outcome: | Empirical evaluations across a variety of benchmarks underscore InfiMM’s remarkable capability in multimodal understanding. |
VideoScore: Building Automatic Metrics to Simulate Fine-grained Human Feedback for Video Generation (2024.emnlp-main)
Copied to clipboard
Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, Kai Wang, Quy Do, Yuansheng Ni, Bohan Lyu, Yaswanth Narsupalli, Rongqi Fan, Zhiheng Lyu, Bill Yuchen Lin, Wenhu Chen
| Challenge: | Existing video metrics are lagging behind in providing reliable scores over generated videos due to lack of large-scale human-annotated dataset. |
| Approach: | They propose to use VideoFeedback to train a human-annotated multi-aspect score over 37.6K synthesized videos from 11 existing video generative models. |
| Outcome: | The proposed model outperforms the prior best metrics by 50 points in the test. |
AdamS: Momentum Itself Can Be A Normalizer for LLM Pretraining and Post-training (2025.emnlp-main)
Copied to clipboard
| Challenge: | Empirically, AdamS demonstrates strong performance in various tasks . et al., 2023b): AdamS is efficient, efficient, and model-agnostic. |
| Approach: | They propose a model-agnostic alternative to Adam for large language model pretraining and post-training. |
| Outcome: | The proposed method matches memory footprint of SGD with momentum while delivering superior performance. |
BSCodec: A Band-Split Neural Codec for High-Quality Universal Audio Reconstruction (2026.findings-eacl)
Copied to clipboard
| Challenge: | Neural audio codecs have enabled high-fidelity reconstruction of speech, music and sound . however, speech-optimized codec systems suffer degradation on music or sound if they ignore spectral differences . |
| Approach: | They propose a neural audio codec that splits the spectral dimension into separate bands and compresses each band independently. |
| Outcome: | Experimental results show that BSCodec achieves better reconstruction quality on music and sound compared to existing codecs. |
MorphoBench: A Benchmark with Difficulty Adaptive to Model Reasoning (2026.findings-acl)
Copied to clipboard
Xukai Wang, Xuanbo Liu, Mingrui Chen, Haitian Zhong, Xuanlin Yang, Bohan Zeng, Jinbo Hu, Hao Liang, Junbo Niu, Xuchen Li, Ruitao Wu, Ruichuan An, Yang Shi, Liu Liu, Qiang Liu, Zhouchen Lin, Xu-Yao Zhang, Wentao Zhang, Bin Dong
| Challenge: | Existing benchmarks designed to evaluate the reasoning capabilities of large models are limited in scope and lack flexibility to adapt difficulty according to evolving reasoning capacities of models. |
| Approach: | They propose a benchmark that incorporates multidisciplinary questions to evaluate the reasoning capabilities of large models and can adjust and update question difficulty based on the reasoning abilities of advanced models. |
| Outcome: | The proposed benchmark incorporates multidisciplinary questions to evaluate the reasoning capabilities of large models and can adjust and update question difficulty based on the reasoning abilities of advanced models. |
Can Large Language Models Understand You Better? An MBTI Personality Detection Dataset Aligned with Population Traits (2025.coling-main)
Copied to clipboard
Bohan Li, Jiannan Guan, Longxu Dou, Yunlong Feng, Dingzirui Wang, Yang Xu, Enbo Wang, Qiguang Chen, Bichen Wang, Xiao Xu, Yimeng Zhang, Libo Qin, Yanyan Zhao, Qingfu Zhu, Wanxiang Che
| Challenge: | Existing data on MBTI personality detection are based on self-reported labels and fail to capture the full range of population personality traits. |
| Approach: | They construct a manually annotated MBTI personality detection dataset with soft labels under the guidance of psychologists and use them to identify the task. |
| Outcome: | The MBTIBench is the first manually annotated MBti personality detection dataset with soft labels under the guidance of psychologists. |
CoV: Chain-of-View Prompting for Spatial Reasoning (2026.findings-acl)
Copied to clipboard
Haoyu Zhao, Akide Liu, Zeyu Zhang, Weijie Wang, Feng Chen, Ruihan Zhu, Gholamreza Haffari, Bohan Zhuang
| Challenge: | Embodied question answering requires collecting context that is distributed across multiple viewpoints . most recent vision–language models (VLMs) are constrained to a fixed and finite set of input views . |
| Approach: | They propose a training-free, test-time reasoning framework that transforms a VLM into an active viewpoint reasoner through a coarse-to-fine exploration process. |
| Outcome: | The proposed framework improves LLM-Match performance by 11.98% on four mainstream VLMs. |
Act as you think: Reinforcing Consistent Reasoning in Medical Visual Question Answering (2026.acl-long)
Copied to clipboard
Songtao Jiang, Yuan Wang, Ruizhe Chen, Yan Zhang, Ruilin Luo, Bohan Lei, Yeying Jin, Sibo Song, ZhiBo Yang, Jimeng Sun, Jian Wu, Zuozhu Liu
| Challenge: | Recent advances have improved the accuracy of medical visual question answering (Med-VQA) however, the high stakes nature of the medical domain has precipitated a shift towards interpretability and transparency of reasoning processes. |
| Approach: | They propose a reinforcement learning from verifiable rewards framework that rewards internal consistency and logical coherence. |
| Outcome: | The proposed framework rewards internal consistency and logical coherence, and is highly versatile, the authors show. |
TableEval: A Real-World Benchmark for Complex, Multilingual, and Multi-Structured Table Question Answering (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing TableQA benchmarks focus on simple flat tables and suffer from data leakage . current benchmarks are monolingual and fail to capture cross-lingual variability . |
| Approach: | They propose a table-based TableQA benchmark to evaluate LLMs on real-world tasks. |
| Outcome: | The proposed benchmarks show that they achieve high agreement with human judgment . the proposed framework improves on the alignment between model responses and reference answers . |
Synthesizing Post-Training Data for LLMs through Multi-Agent Simulation (2025.acl-long)
Copied to clipboard
Shuo Tang, Xianghe Pang, Zexi Liu, Bohan Tang, Rui Ye, Tian Jin, Xiaowen Dong, Yanfeng Wang, Siheng Chen
| Challenge: | Extensive experiments demonstrate that our framework effectively generates both general and domain-specific data. |
| Approach: | They propose a multi-agent simulator that automatically generates diverse text-based scenarios, capturing a wide range of real-world human needs. |
| Outcome: | Experiments show that the proposed model outperforms Meta’s Llama-3-8B-Instruct model on AlpacaEval 2 and Arena-Hard benchmarks with just 20K instruction-response pairs. |
Causal Inference for Human-Language Model Collaboration (2024.naacl-long)
Copied to clipboard
| Challenge: | In this paper, we examine the collaborative dynamics between humans and language models where the interaction involves LMs proposing text segments and humans editing or responding to these segments. |
| Approach: | They propose a causal estimand to estimate the incremental stylistic effect (ISE) of various interaction strategies in dynamic human-LM collaborations. |
| Outcome: | The proposed estimand reduces confounding and significantly improves counterfactual estimation over a set of competitive baselines. |
Large Language Models for Data Annotation and Synthesis: A Survey (2024.emnlp-main)
Copied to clipboard
Zhen Tan, Dawei Li, Song Wang, Alimohammad Beigi, Bohan Jiang, Amrita Bhattacharjee, Mansooreh Karami, Jundong Li, Lu Cheng, Huan Liu
| Challenge: | Existing surveys focus on LLMs' specific utility for data annotation and synthesis. |
| Approach: | They propose to use large language models to generate annotations from raw data . they also propose to review learning strategies for models utilizing LLM-generated annotations . |
| Outcome: | The proposed models can be used to improve the efficacy of machine learning models by generating and labeling raw data with relevant information. |
MDIT-Bench: Evaluating the Dual-Implicit Toxicity in Large Multimodal Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Large Multimodal Models (LMMs) have raised concerns about model toxicity. |
| Approach: | They propose a model to measure the toxicity gap between models and their hard level to determine whether they can handle dual-implicit toxicity. |
| Outcome: | The proposed model can handle dual-implicit toxicity effectively on 13 prominent LMMs, but its performance drops significantly in hard level. |
Enhancing Open-Domain Task-Solving Capability of LLMs via Autonomous Tool Integration from GitHub (2025.acl-long)
Copied to clipboard
Bohan Lyu, Xin Cong, Heyang Yu, Pan Yang, Cheng Qian, Zihe Wang, Yujia Qin, Yining Ye, Yaxi Lu, Chen Qian, Zhong Zhang, Yukun Yan, Yankai Lin, Zhiyuan Liu, Maosong Sun
| Challenge: | Existing approaches lack flexibility to address diverse and ever-evolving user queries in open domains. |
| Approach: | They propose to evaluate LLMs on open-domain knowledge that requires tools to solve diverse and ever-evolving user queries. |
| Outcome: | The proposed system outperforms baselines in the open domain task-solving benchmark. |