Papers by Songyang Zhang
Capability Salience Vector: Fine-grained Alignment of Loss and Capabilities for Downstream Task Scaling Law (2025.acl-long)
Copied to clipboard
Qiming Ge, Shuhao Xing, Songyang Gao, Yunhua Zhou, Yicheng Zou, Songyang Zhang, Zhi Chen, Hang Yan, Qi Zhang, Qipeng Guo, Kai Chen
| Challenge: | Large language models have demonstrated impressive performance across a wide range of tasks, but this achievement comes with the trade-off of significant computational demands. |
| Approach: | They propose a scaling law that decomposes the overall validation loss and assigns different importance weights to tokens to assess a specific meta-capability. |
| Outcome: | The proposed model can predict the loss trending of models across different levels of computation without a gap between validation loss and model's downstream capabilities. |
LoRAMoE: Alleviating World Knowledge Forgetting in Large Language Models via MoE-Style Plugin (2024.acl-long)
Copied to clipboard
Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Wei Shen, Limao Xiong, Yuhao Zhou, Xiao Wang, Zhiheng Xi, Xiaoran Fan, Shiliang Pu, Jiang Zhu, Rui Zheng, Tao Gui, Qi Zhang, Xuanjing Huang
| Challenge: | Experimental results show that, as the instruction data increases, LoRAMoE can significantly improve the ability to process downstream tasks, while maintaining the world knowledge stored in the LLM. |
| Approach: | They propose a framework that introduces several low-rank adapters and integrates them by using a router network to freeze the backbone model and force a portion of LoRAs to focus on leveraging world knowledge to solve downstream tasks. |
| Outcome: | The proposed framework freezes the backbone model and forces a portion of LoRAs to focus on leveraging world knowledge to solve downstream tasks, to alleviate world knowledge forgetting. |
Ada-LEval: Evaluating long-context LLMs with length-adaptable benchmarks (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing long-text evaluation benchmarks, such as L-Eval and LongBench, focus on QA and summarization tasks. |
| Approach: | They propose a length-adaptable benchmark for evaluating the long-context understanding of large language models. |
| Outcome: | The proposed benchmarks do not cover ultralong settings (100k+ tokens) and are difficult to evaluate across different length ranges. |
OpenHuEval: Evaluating Large Language Model on Hungarian Specifics (2025.findings-acl)
Copied to clipboard
Haote Yang, Xingjian Wei, Jiang Wu, Noémi Ligeti-Nagy, Jiaxing Sun, Yinfan Wang, Győző Zijian Yang, Junyuan Gao, Jingchao Wang, Bowen Jiang, Shasha Wang, Nanjun Yu, Zihao Zhang, Shixin Hong, Hongwei Liu, Wei Li, Songyang Zhang, Dahua Lin, Lijun Wu, Gábor Prószéky, Conghui He
| Challenge: | Recent advances in Large Language Models (LLMs) represent significant strides toward artificial general intelligence (AGI). |
| Approach: | They introduce OpenHuEval, the first benchmark for LLMs focusing on the Hungarian language and specifics. |
| Outcome: | The framework reveals intrinsic patterns and mechanisms of LLMs in non-English languages, with Hungarian serving as an example. |
Self-Polish: Enhance Reasoning in Large Language Models via Problem Refinement (2023.findings-emnlp)
Copied to clipboard
Zhiheng Xi, Senjie Jin, Yuhao Zhou, Rui Zheng, Songyang Gao, Jia Liu, Tao Gui, Qi Zhang, Xuanjing Huang
| Challenge: | Existing prompting methods have been used to enhance multistep reasoning capabilities of large language models, but they have overlooked the potential of formulating higher-quality problems. |
| Approach: | They propose a method that starts from the problem side and refines problems to be more comprehensible and solvable for models. |
| Outcome: | The proposed method achieves notable and consistent effectiveness on five reasoning benchmarks across different models. |
AgentGym: Evaluating and Training Large Language Model-based Agents across Diverse Environments (2025.acl-long)
Copied to clipboard
Zhiheng Xi, Yiwen Ding, Wenxiang Chen, Boyang Hong, Honglin Guo, Junzhe Wang, Xin Guo, Dingwen Yang, Chenyang Liao, Wei He, Songyang Gao, Lu Chen, Rui Zheng, Yicheng Zou, Tao Gui, Qi Zhang, Xipeng Qiu, Xuanjing Huang, Zuxuan Wu, Yu-Gang Jiang
| Challenge: | Large language models (LLMs) are promising foundations to build generally-capable agents . however, the community lacks a unified interactive framework that covers diverse environments for comprehensive evaluation of agents. |
| Approach: | They propose a framework that features 7 real-world scenarios, 14 environments, and 89 tasks for unified, real-time, and concurrent agent interaction. |
| Outcome: | The proposed framework features 7 real-world scenarios, 14 environments, and 89 tasks for unified, real-time, and concurrent agent interaction. |
Fake Alignment: Are LLMs Really Aligned Well? (2024.naacl-long)
Copied to clipboard
Yixu Wang, Yan Teng, Kexin Huang, Chengqi Lyu, Songyang Zhang, Wenwei Zhang, Xingjun Ma, Yu-Gang Jiang, Yu Qiao, Yingchun Wang
| Challenge: | Existing studies on large language models have shown that they are poorly aligned in practice. |
| Approach: | They propose a framework to evaluate safety in large language models . they propose two new metrics to quantify fake alignment and obtain corrected performance estimation. |
| Outcome: | The proposed framework and two metrics show that some models with purported safety are poorly aligned in practice. |
BotChat: Evaluating LLMs’ Capabilities of Having Multi-Turn Dialogues (2024.findings-naacl)
Copied to clipboard
Haodong Duan, Jueqi Wei, Chonghua Wang, Hongwei Liu, Yixiao Fang, Songyang Zhang, Dahua Lin, Kai Chen
| Challenge: | Modern Large Language Models (LLMs) facilitate high-quality, multi-turn dialogues with humans, but human-based evaluation of such a capability requires substantial manual effort. |
| Approach: | They propose to evaluate LLMs' ability to emulate human-like, multi-turn conversations using an LLM-centric approach. |
| Outcome: | The proposed model emulates human-like, multi-turn conversations using an LLM-centric approach. |
RouteMoA: Dynamic Routing without Pre-Inference Boosts Efficient Mixture-of-Agents (2026.acl-long)
Copied to clipboard
Jize Wang, Han Wu, Zhiyuan You, Yiming Song, Yijun Wang, Zifei Shan, Yining Li, Songyang Zhang, Xinyi Le, Cailian Chen, Xinping Guan, Dacheng Tao
| Challenge: | Existing methods for mixing-of-agents (MoA) lack model selection criteria and struggle with large model pools. |
| Approach: | They propose a mixture-of-agents framework with dynamic routing that uses a lightweight scorer to perform initial screening and refines the model scores through self- and cross-assessment. |
| Outcome: | The proposed framework outperforms existing methods for large model pools and tasks . it reduces cost by 89.8% and latency by 63.6% in the large-scale model pool. |
CompassVerifier: A Unified and Robust Verifier for LLMs Evaluation and Outcome Reward (2025.emnlp-main)
Copied to clipboard
Shudong Liu, Hongwei Liu, Junnan Liu, Linchen Xiao, Songyang Gao, Chengqi Lyu, Yuzhe Gu, Wenwei Zhang, Derek F. Wong, Songyang Zhang, Kai Chen
| Challenge: | Existing approaches lack robustness to handle complex edge cases and generalizability across different domains. |
| Approach: | They develop an accurate and lightweight verifier model for evaluation and outcome reward that matches unstructured outputs against standard answers. |
| Outcome: | The proposed model can process multiple answer types including multi-subproblems, formulas, and sequence answers while identifying abnormal/invalid responses. |
LLaST: Improved End-to-end Speech Translation System Leveraged by Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | ***LLaST*** is a framework for building high-performance Large Language model based Speech-to-text Translation systems. |
| Approach: | They propose a framework for building high-performance Large Language model based Speech-to-text Translation systems. |
| Outcome: | The proposed model outperforms the CoVoST-2 benchmark and showcases exceptional scaling capabilities powered by LLMs. |
Are Your LLMs Capable of Stable Reasoning? (2025.findings-acl)
Copied to clipboard
Junnan Liu, Hongwei Liu, Linchen Xiao, Ziyi Wang, Kuikun Liu, Songyang Gao, Wenwei Zhang, Songyang Zhang, Kai Chen
| Challenge: | Existing evaluation protocols and metrics do not capture the full spectrum of LLM capabilities, especially in complex reasoning tasks. |
| Approach: | They propose a new evaluation metric that continuously assesses model performance across multiple sampling attempts, quantifying both the model’s potential capabilities and operational consistency. |
| Outcome: | The proposed evaluation metric measures model performance across multiple sampling attempts and provides comprehensive insights into their potential capabilities and operational consistency. |
MathBench: Evaluating the Theory and Application Proficiency of LLMs with a Hierarchical Mathematics Benchmark (2024.findings-acl)
Copied to clipboard
Hongwei Liu, Zilong Zheng, Yuxuan Qiao, Haodong Duan, Zhiwei Fei, Fengzhe Zhou, Wenwei Zhang, Songyang Zhang, Dahua Lin, Kai Chen
| Challenge: | Recent advances in large language models have showcased significant improvements in mathematics, but traditional benchmarks like GSM8k offer a unidimensional perspective. |
| Approach: | MathBench is a benchmark that rigorously assesses the mathematical capabilities of large language models. |
| Outcome: | MathBench spans a wide range of mathematical disciplines, offering a detailed evaluation of both theoretical understanding and practical problem-solving skills. |
ToolSword: Unveiling Safety Issues of Large Language Models in Tool Learning Across Three Stages (2024.acl-long)
Copied to clipboard
Junjie Ye, Sixian Li, Guanyu Li, Caishuang Huang, Songyang Gao, Yilong Wu, Qi Zhang, Tao Gui, Xuanjing Huang
| Challenge: | Existing research focuses on enhancing LLMs capabilities through tool utilization. |
| Approach: | They propose a framework to investigate safety issues in large language models in tool learning . they propose malicious queries and jailbreak attacks in the input stage . |
| Outcome: | The proposed framework investigates six safety scenarios for LLMs in tool learning . the data will be released upon acceptance of the proposed framework . |
Knowledge-to-Verification: Exploring RLVR for LLMs in Knowledge-Intensive Domains (2026.acl-long)
Copied to clipboard
Zhonghang Yuan, Zhefan Wang, Fang Hu, Zihong Chen, Jinzhe Li, Gang Li, Jie Ying, Huanjun Kong, Songyang Zhang, Nanqing Dong
| Challenge: | Recent large language models (LLMs) have demonstrated remarkable progress in reasoning, but their applications on knowledge-intensive domains have not been explored due to the scarcity of high-quality verifiable data. |
| Approach: | They propose a framework that extends reinforcement learning with verifiable rewards (RLVR) to knowledge-intensive domains through automated verififiability data synthesis while enabling verification of the LLM's reasoning process. |
| Outcome: | Extensive experiments show that the proposed framework enhances the reasoning of large language models in knowledge-intensive domains without significantly compromising the model’s general capabilities. |
LawBench: Benchmarking Legal Knowledge of Large Language Models (2024.emnlp-main)
Copied to clipboard
Zhiwei Fei, Xiaoyu Shen, Dawei Zhu, Fengzhe Zhou, Zhuo Han, Alan Huang, Songyang Zhang, Kai Chen, Zhixin Yin, Zongwen Shen, Jidong Ge, Vincent Ng
| Challenge: | LegalBench evaluated 20 LLMs in 162 legal tasks in 20 countries and jurisdictions. |
| Approach: | They present a comprehensive evaluation of 21 popular Large Language Models and the first comparative analysis of the empirical results. |
| Outcome: | The proposed benchmarks are based on the Bloom’s cognitive taxonomy and are compared to 21 popular LLMs. |
Benchmarking Chinese Commonsense Reasoning of LLMs: From Chinese-Specifics to Reasoning-Memorization Correlations (2024.acl-long)
Copied to clipboard
| Challenge: | Currently, many benchmarks evaluate the commonsense reasoning of large language models (LLMs), but most are English-based, limiting non-English evaluations. |
| Approach: | They propose to use Chinese commonsense reasoning to evaluate LLMs' commonsensing ability. |
| Outcome: | The proposed benchmark covers both globally known and Chinese-specific commonsense reasoning abilities and can be used as a reference for future research. |
Learning a Grammar Inducer from Massive Uncurated Instructional Videos (2022.emnlp-main)
Copied to clipboard
| Challenge: | aims to find more accurate syntactic grammars for accompanying text using video data. |
| Approach: | They build a video-aided grammar induction model that can learn video-span correlation without manual features. |
| Outcome: | The proposed model can learn video-span correlation without manual features adopted by previous systems. |
Inverse-Q*: Token Level Reinforcement Learning for Aligning Large Language Models Without Preference Data (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Reinforcement Learning from Human Feedback (RLHF) relies on complex methodologies like Proximal Policy Optimization (PPO) that require extensive hyper-parameter tuning and pose challenges in sample efficiency and stability. |
| Approach: | They propose an innovative framework that leverages direct preference optimization techniques but extends them by estimating the conditionally optimal policy directly from the model’s responses. |
| Outcome: | The proposed framework matches and exceeds the effectiveness of Proximal Policy Optimization (PPO) in terms of convergence speed and alignment of model responses with human preferences. |
RoTBench: A Multi-Level Benchmark for Evaluating the Robustness of Large Language Models in Tool Learning (2024.emnlp-main)
Copied to clipboard
Junjie Ye, Yilong Wu, Songyang Gao, Caishuang Huang, Sixian Li, Guanyu Li, Xiaoran Fan, Qi Zhang, Tao Gui, Xuanjing Huang
| Challenge: | Current research emphasizes LLMs’ capacity to utilize tools in well-structured environments while overlooking their stability when confronted with the inevitable noise of the real world. |
| Approach: | They propose a multi-level benchmark to evaluate the robustness of large language models in tool learning by establishing five external environments with varying levels of noise. |
| Outcome: | The proposed model outperforms the GPT-4 model in tool learning in three critical phases: tool selection, parameter identification, and content filling. |
Condor: Enhance LLM Alignment with Knowledge-Driven Data Synthesis and Refinement (2025.acl-long)
Copied to clipboard
Maosongcao Maosongcao, Taolin Zhang, Mo Li, Chuyu Zhang, Yunxin Liu, Conghui He, Haodong Duan, Songyang Zhang, Kai Chen
| Challenge: | Existing high-quality human-annotated SFT data is a bottleneck for Large Language Models (LLMs). |
| Approach: | They propose a two-stage synthetic data generation framework that incorporates World Knowledge Trees and Self-Reflection Refinement to produce high-quality SFT data at scale. |
| Outcome: | The proposed model fine-tuned on 20K condor-generated samples achieves superior performance compared to instruct model trained with RLHF. |
InternLM-Law: An Open-Sourced Chinese Legal Large Language Model (2025.coling-main)
Copied to clipboard
| Challenge: | InternLM-Law is a large language model (LLM) tailored for addressing diverse legal tasks related to Chinese laws. |
| Approach: | They introduce a large language model (LLM) tailored for addressing diverse legal tasks related to Chinese laws. |
| Outcome: | The proposed model performs better than existing models in a variety of legal tasks related to Chinese laws. |
RealBehavior: A Framework for Faithfully Characterizing Foundation Models’ Human-like Behavior Mechanisms (2023.findings-emnlp)
Copied to clipboard
Enyu Zhou, Rui Zheng, Zhiheng Xi, Songyang Gao, Xiaoran Fan, Zichu Fei, Jingting Ye, Tao Gui, Qi Zhang, Xuanjing Huang
| Challenge: | Existing studies on human-like behaviors in foundation models do not verify their faithfulness . a simple application of psychological tools cannot faithfully characterize all human-type behaviors . |
| Approach: | They propose a framework to characterize humanoid behaviors in foundation models . they argue that a simple application of psychological tools cannot faithfully characterize all human-like behaviors . |
| Outcome: | The proposed framework assesses the faithfulness of results based on reproducibility, internal consistency, and generalizability. |
Video-aided Unsupervised Grammar Induction (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing methods of multi-modal grammar induction focus on grammar inducing from text-image pairs, but videos provide even richer information, such as static objects and actions. |
| Approach: | They propose a video-aided grammar induction model which learns a constituency parser from unlabeled text and its corresponding video. |
| Outcome: | The proposed model outperforms existing systems on three benchmarks. |
ProSA: Assessing and Understanding the Prompt Sensitivity of LLMs (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Recent research has neglected instances-level prompt variations and their implications on subjective evaluations. |
| Approach: | They propose a framework to evaluate and comprehend prompt sensitivity in large language models. |
| Outcome: | The proposed framework evaluates and comprehends prompt sensitivity in large language models. |
Navigating the OverKill in Large Language Models (2024.acl-long)
Copied to clipboard
Chenyu Shi, Xiao Wang, Qiming Ge, Songyang Gao, Xianjun Yang, Tao Gui, Qi Zhang, Xuanjing Huang, Xun Zhao, Dahua Lin
| Challenge: | Recent studies have highlighted a tendency among large language models to refuse to answer benign queries. |
| Approach: | They propose a model-agnostic approach to reduce excessive attention to harmful words like ‘kill’ and a method to decode the next-token predictions by contrastive decoding. |
| Outcome: | The proposed approach reduces the refusal rate by 20% while having little impact on safety. |
T-Eval: Evaluating the Tool Utilization Capability of Large Language Models Step by Step (2024.acl-long)
Copied to clipboard
Zehui Chen, Weihua Du, Wenwei Zhang, Kuikun Liu, Jiangning Liu, Miao Zheng, Jingming Zhuo, Songyang Zhang, Dahua Lin, Kai Chen, Feng Zhao
| Challenge: | Existing studies evaluate the tool utilization ability of large language models based on the final output or only consider the single-step tool calling. |
| Approach: | They propose a new approach to evaluate the tool utilization capability of large language models (LLMs) they decompose the tool usage into multiple sub-processes, including instruction following, planning, reasoning, retrieval, understanding, and review. |
| Outcome: | The proposed model exhibits consistency with the outcome-oriented evaluation and provides a more fine-grained analysis of the capabilities of LLMs. |