Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Belief-Calibrated Multi-Agent Consensus Seeking for Complex NLP Tasks
Authors: Wentao Deng, Jiahuan Pei, Zhiwei Xu, Zhaochun Ren, Zhumin Chen, Pengjie Ren
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on the MATH and MMLU benchmark datasets demonstrate that the proposed BCCS framework outperforms the best existing results by 2.23% and 3.95% of accuracy on challenging tasks, respectively. Our code and data are available at https://github.com/dengwentao99/BCCS. In the experimental implementation, we evaluate the effectiveness of belief-calibrated consensus seeking (BCCS) on two widely-used benchmarks: MATH [8] and MMLU [9]. Results demonstrate that BCCS improves accuracy by 2.23% on MATH and 3.95% on MMLU compared to existing best results on challenge tasks. Extensive experiments conducted on widely adopted benchmarks confirm the effectiveness of BCCS. Additionally, ablation studies are performed to quantify the impact of each core component. |
| Researcher Affiliation | Academia | Wentao Deng Shandong University EMAIL Jiahuan Pei Vrije Universiteit Amsterdam EMAIL Zhiwei Xu Shandong University EMAIL Zhaochun Ren Leiden University EMAIL Zhumin Chen Shandong University EMAIL Pengjie Ren Shandong University EMAIL |
| Pseudocode | Yes | A Algorithm Algorithm 1 Algorithm of belief-calibrated consensus seeking (BCCS) |
| Open Source Code | Yes | Our code and data are available at https://github.com/dengwentao99/BCCS. |
| Open Datasets | Yes | In the experimental implementation, we evaluate the effectiveness of belief-calibrated consensus seeking (BCCS) on two widely-used benchmarks: MATH [8] and MMLU [9]... MATH [8] is a mathematical reasoning benchmark that contains 5,000 cases covered 7 types of problems... MATH dataset is released under MIT License, which can be found in https://huggingface.co/datasets/HuggingFaceTB/MATH. MMLU [9] is an integrated reasoning benchmark that contains 57 subjects... MMLU dataset is released under MIT License, which can be found in https://people.eecs.berkeley.edu/~hendrycks/data.tar. |
| Dataset Splits | Yes | For each datasets, we randomly sample three groups of 500 examples with random seeds 100, 200, and 300 to conduct three independent experiments. |
| Hardware Specification | Yes | All experiments are conducted with Nvidia A800 GPUs with 80GB memory. Specifically, for experiments based on 7B and 14B models, the experiments need one A800 GPU, and the experiments based on 32B models, the experiments need two A800 GPUs. |
| Software Dependencies | No | No specific software dependencies with version numbers (like Python, PyTorch, TensorFlow, CUDA) are explicitly mentioned in the paper. The paper mentions backbone models like 'Qwen2.5-7B-Instruct' and methods like 'KMeans' and 'TF-IDF' but without specific software versions for these or other ancillary software components. |
| Experiment Setup | Yes | In the main experiments, the determination for optimal number of agents and iteration rounds is consistent with common practices in the existing multi-agent collaboration systems [18], where such hyperparameters are often set empirically. Specifically, the number of agents is n = 7 and the maximum iteration rounds is 3 across all methods. The number of leaders is set as nl = 2. A detailed ablation study of these hyperparameters is available in Appendix D.2. Unless stated otherwise, each agent employs Qwen2.5-7B-Instruct as the backbone model. To ensure the opinion diversity within MAS, we set the temperature as 0.7. The number of leaders nl is set as 2 and to ensure that at least one opinion group contains more than nl agents, and to allow for the potential coexistence of supportive and conflicting relationships among opinion groups, thus we set the number of opinion group clustering for KMeans [55] in BCCS as 3. |