text
string
source
string
optimizers, metaTextGrad performs optimization in two steps. First, it individually refines each optimizer by optimizing its prompts to better align with the task. Then, it combines the different prompt-optimized optimizers to construct the final optimizer. 3.2.1 Motivation We begin with an example using the Textual Gr...
https://arxiv.org/abs/2505.18524v1
For instance, some tasks may benefit from first identifying a better LLM program architecture and then exploring how to combine prompts and demonstrations within that structure. Other tasks may require iteratively improving the content of prompts and demonstrations to enhance their quality before exploring their combin...
https://arxiv.org/abs/2505.18524v1
(3) LLM calls made by the meta-optimizer when optimizing the optimizer. As shown in Section 4.3, the frequency of these calls decreases significantly across these levels. This hierarchical structure allows for cost-effective resource allocation: the program should use a relatively economical model, the optimizer can le...
https://arxiv.org/abs/2505.18524v1
different levels within a single epoch of meta optimization. This analysis supports our design choice of using models with different capabilities at different levels. As shown in Table 2, the token usage per optimization epoch on the MMLU Abstract Algebra dataset reveals that higher-level components require significant...
https://arxiv.org/abs/2505.18524v1
optimizer. Here, metaTextGrad produces the same results as the optimized ADAS-TG because the meta optimizer did not find a better option during the meta structure optimization. So, the optimized ADAS-TG was selected as the best result. Notably, the best meta optimizer varies across benchmarks due to task-specific diffe...
https://arxiv.org/abs/2505.18524v1
, a meta-optimization framework that enhances existing LLM optimizers by aligning them more effectively with tasks. Our method introduces two key components: the meta prompt optimizer, which refines optimizer prompts for better task adaptation, and the meta structure optimizer, which determines the optimal combination ...
https://arxiv.org/abs/2505.18524v1
Jeff Clune. Automated design of agentic systems. arXiv preprint arXiv:2408.08435 , 2024. [11] Qian Li, San-Yang Liu, and Xin-She Yang. Influence of initialization on the performance of metaheuristic optimizers. Applied Soft Computing , 91:106193, 2020. [12] Ilya Sutskever, James Martens, George Dahl, and Geoffrey Hinto...
https://arxiv.org/abs/2505.18524v1
Eliciting Knowledge from Language Models with Automatically Generated Prompts. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Process- ing (EMNLP) , pages 4222–4235, Online, November 2020. Association for Computational Linguistics. [27] Xiang Chen, Ningyu Zhang, Xin Xie, Shumin Deng, Yunz...
https://arxiv.org/abs/2505.18524v1
19: return M∗ 20:end function A.3 Meta Optimizer For the meta optimizer, the execution process consists of two steps: (1) Each optimizer is individually optimized using the meta prompt optimizer. (2) The optimized optimizers are further refined using the meta structure optimizer. 13 B Proof of Theorem 1 Theorem (Restat...
https://arxiv.org/abs/2505.18524v1
the meta prompt optimizer and the meta structure optimizer. The propose stage is the only part of the meta optimizer that involves an LLM call; all other functions consist of direct numerical updates or result retrieval, as explained in pseudocode in Appendix A. D.1 Meta Prompt Optimizer Below, we present the prompt us...
https://arxiv.org/abs/2505.18524v1
to implement a logic so that the optimizer can use different optimizing strategies in a sequential manner, with each strategy being applied self.epoch times before moving to the next one. {optimizer_prompt} # Code Reminder You should include the following necessary imports at the beginning of the code: import inspect i...
https://arxiv.org/abs/2505.18524v1
) ) self.type_analyzer = tg.BlackboxLLM( self.engine, Variable( """Identify and categorize all bracket types. List each type and its corresponding closing bracket. Format: 'Types: [pairs] '""", requires_grad=True, role_description="analyzes bracket types and pairs" ) ) self.stack_validator = tg.BlackboxLLM( self.engine...
https://arxiv.org/abs/2505.18524v1
of the problem systematically. Present each statement distinctly using bullet points or numbers, ensuring the final conclusion is clearly separated from the statement evaluations. Include any assumptions or context necessary for each conclusion. After evaluating each statement, reexamine your conclusions to confirm the...
https://arxiv.org/abs/2505.18524v1
Reinforcement Fine-Tuning Powers Reasoning Capability of Multimodal Large Language Models Haoyuan Sun, Jiaqi Wu, Bo Xia, Yifu Luo, Yifei Zhao, Kai Qin, Xufei Lv, Tiantian Zhang, Yongzhe Chang, Xueqian Wang Tsinghua Shenzhen International Graduate School, Tsinghua University sun-hy23@mails.tsinghua.edu.cn Project: https...
https://arxiv.org/abs/2505.18536v1
training, which has significantly enhanced its reasoning capabilities. Moreover, DeepSeek-R1-Zero [ 8] has demonstrated remarkable self-evolution capabilities through a pure reinforcement learning process; additionally, DeepSeek-R1 [ 8] further stabilizes the RL process through cold start, ultimately exhibiting unparal...
https://arxiv.org/abs/2505.18536v1
combining the strengths of the outcome reward paradigm and the process reward paradigm is a promising direction for further research. Thirdly, we hope the community could devote increased attention to the safety of reasoning MLLMs. Fourthly, given the scarcity of multimodal data, further research into data augmentation...
https://arxiv.org/abs/2505.18536v1
researchers have increasingly employed neural networks to approximate the value function, leading to a series of improvements, exemplified by Deep Q-Network (DQN) [ 13,14], Double DQN [ 15], Dueling DQN [16], Rainbow [ 17], and so on. In contrast, Policy-based methods directly and explicitly learn a target policy with ...
https://arxiv.org/abs/2505.18536v1
is higher than average), maximizing the equation leads to an increase of πθ/πθold, but it is constrained not to exceed 1+ε; similarly, if the advantage function is negative (action value is lower than average), maximizing the equation leads to a decrease of πθ/πθold, but it is constrained not to exceed 1 −ε. 2.2 Multim...
https://arxiv.org/abs/2505.18536v1
[ 2] in 2017, it has become one of the most popular actor-critic RL algorithms for policy optimization of LLMs [ 3,35] and MLLMs. Within the context of MLLMs, the input (m,t)consisting of both multimodal (beyond language) contents mand a textual query t; then, PPO objective in Equation (5) can be transferred as [28, 35...
https://arxiv.org/abs/2505.18536v1
been made in enhancing multimodal reasoning within this paradigm. Generally, reinforcement fine-tuning has achieved significant successes in powering the reasoning ability of MLLMs. We meticulously divide the successes into the following five points: Success 1:Diverse Modalities .As demonstrated in Figure 1, recent adv...
https://arxiv.org/abs/2505.18536v1
employed RFT for the adaptive invocation of external vision tools. Furthermore, a significant number of works have focused on multi-task and multi-domain joint training to simultaneously improve model performance across multiple tasks and domains: Insight-V [ 94], Visual-RFT [ 95], Reason-RFT [ 96], ThinkLite-VL [ 97],...
https://arxiv.org/abs/2505.18536v1
introduces an effective MCTS-based data filtering method that quantifies sample difficulty according to the number of iterations the model requires to solve each problem, thereby achieving state-of-the-art reasoning performance with fewer training samples. Success 4:Abundant Benchmarks .As stated in the blog [ 124], ab...
https://arxiv.org/abs/2505.18536v1
support for models, algorithms, and datasets, along with support for padding-free training, checkpoint resumption, and tool integration. MAYA [ 144] offers a transparent and reproducible framework, along with a comprehensive evalua- tion scheme, for the application of RL to MLLMs; furthermore, it also serves as a light...
https://arxiv.org/abs/2505.18536v1
outcome reward paradigm. StepGRPO [76] represents a pioneering approach to this area, notably by incorporating dense step-wise rewards; however, it is limited to the vision mathematical reasoning task, and the applicability of such a methodology to other tasks, domains, and modalities requires further investigation. TO...
https://arxiv.org/abs/2505.18536v1
support to the project. This work was supported by the Natural Science Foundation of Shenzhen (No.JCYJ20230807111604008, No.JCYJ20240813112007010), the Natural Science Foun- dation of Guangdong Province (No.2024A1515010003), National Key Research and Development Program of China (No.2022YFB4701400) and Cross-disciplina...
https://arxiv.org/abs/2505.18536v1
and David Silver. Rainbow: Combining improvements in deep reinforcement learning. In Proceedings of the AAAI conference on artificial intelligence , volume 32, 2018. [18] Richard S Sutton, David McAllester, Satinder Singh, and Yishay Mansour. Policy gradi- ent methods for reinforcement learning with function approximat...
https://arxiv.org/abs/2505.18536v1
5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599 , 2025. [33] OpenAI. Introducing openai o3 and o4-mini. https://openai.com/index/ introducing-o3-and-o4-mini/ , 2025. [34] xAI. Grok 3 beta — the age of reasoning agents. https://x.ai/grok , 2025. [35] Jingcheng Hu, Yinmin Zhang, Qi Han, Daxin...
https://arxiv.org/abs/2505.18536v1
Zhenghao Xing, Xiaowei Hu, Chi-Wing Fu, Wenhai Wang, Jifeng Dai, and Pheng-Ann Heng. Echoink-r1: Exploring audio-visual reasoning in multimodal llms via reinforcement learning. arXiv preprint arXiv:2505.04623 , 2025. [52] Zhengxi Lu, Yuxiang Chai, Yaxuan Guo, Xi Yin, Liang Liu, Hao Wang, Guanjing Xiong, and Hongsheng L...
https://arxiv.org/abs/2505.18536v1
A more robust multi-discipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813 , 2024. [66] Yunzhuo Hao, Jiawei Gu, Huichen Will Wang, Linjie Li, Zhengyuan Yang, Lijuan Wang, and Yu Cheng. Can mllms reason in multimodality? emma: An enhanced multimodal reasoning benchmark. arXiv preprint arXiv:2501....
https://arxiv.org/abs/2505.18536v1
arXiv preprint arXiv:2504.13055 , 2025. 14 [80] Yichen Wei, Yi Peng, Xiaokun Wang, Weijie Qiu, Wei Shen, Tianyidan Xie, Jiangbo Pei, Jianhao Zhang, Yunzhuo Hao, Xuchen Song, et al. Skywork r1v2: Multimodal hybrid reinforcement learning for reasoning. arXiv preprint arXiv:2504.16656 , 2025. [81] Wenyi Xiao, Leilei Gan, ...
https://arxiv.org/abs/2505.18536v1
Reinforcement fine-tuning for visual reasoning. arXiv preprint arXiv:2503.20752 , 2025. 15 [97] Xiyao Wang, Zhengyuan Yang, Chao Feng, Hongjin Lu, Linjie Li, Chung-Ching Lin, Kevin Lin, Furong Huang, and Lijuan Wang. Sota with less: Mcts-guided sample selection for data-efficient visual reasoning self-improvement. arXi...
https://arxiv.org/abs/2505.18536v1
2025. [115] Wenqi Zhang, Mengna Wang, Gangao Liu, Xu Huixin, Yiwei Jiang, Yongliang Shen, Guiyang Hou, Zhe Zheng, Hang Zhang, Xin Li, et al. Embodied-reasoner: Synergizing visual search, reasoning, and action for embodied interactive tasks. arXiv preprint arXiv:2503.21696 , 2025. [116] Baining Zhao, Ziyou Wang, Jianjie...
https://arxiv.org/abs/2505.18536v1
Alan Yuille. Pulsecheck457: A diagnostic benchmark for 6d spatial reasoning of large multimodal models. arXiv e-prints , pages arXiv–2502, 2025. [132] Yukun Qi, Yiming Zhao, Yu Zeng, Xikun Bao, Wenxuan Huang, Lin Chen, Zehui Chen, Jie Zhao, Zhongang Qi, and Feng Zhao. Vcr-bench: A comprehensive evaluation framework for...
https://arxiv.org/abs/2505.18536v1
, 2025. [147] Yibin Wang, Zhimin Li, Yuhang Zang, Chunyu Wang, Qinglin Lu, Cheng Jin, and Jiaqi Wang. Unified multimodal chain-of-thought reward model through reinforcement fine-tuning. arXiv preprint arXiv:2505.03318 , 2025. [148] Xiaokun Wang, Jiangbo Pei, Wei Shen, Yi Peng, Yunzhuo Hao, Weijie Qiu, Ai Jian, Tianyida...
https://arxiv.org/abs/2505.18536v1
Project] [ Code] [2505] [Flow-GRPO (Gen) [ 119] ] Flow-GRPO: Training Flow Matching Models via Online RL [ Models] [ Code] [2505] [X-Reasoner [ 145] ] X-Reasoner: Towards Generalizable Reasoning Across Modalities and Domains [ Code] [2505] [T2I-R1 (Gen) [ 118] ] T2I-R1: Reinforcing Image Generation with Collaborative S...
https://arxiv.org/abs/2505.18536v1
[ Benchmark] [2503] [LMM-R1 [ 72] ] LMM-R1: Empowering 3B LMMs with Strong Reasoning Abilities Through Two-Stage Rule-Based RL [ Code] [2503] [Curr-ReFT [ 121] ] Boosting the Generalization and Reasoning of Vision Language Models with Curriculum Reinforcement Learning [ Models] [ Dataset] [ Code] [2503] [VisualThinker-...
https://arxiv.org/abs/2505.18536v1
with Reasoning through Step-by-Step Verification [2503] [Med-R1 [ 113] ] Med-R1: Reinforcement Learning for Generalizable Medical Reasoning in Vision-Language Models [ Model] [ Code] [2502] [MedVLM-R1 [ 112] ] MedVLM-R1: Incentivizing Medical Reasoning Capability of Vision- Language Models (VLMs) via Reinforcement Lear...
https://arxiv.org/abs/2505.18536v1
Business as Rule sual: A Benchmark and Framework for Business Rule Flow Modeling with LLMs Chen Yang1, Ruping Xu1, Ruizhe Li2, Bin Cao1*, Jing Fan1 Zhejiang University of Technology, China1 University of Aberdeen, UK2 {yangchen,rupingxu,bincao,fanjing}@zjut.edu.cn1 ruizhe.li@abdn.ac.uk2 Abstract Process mining aims to ...
https://arxiv.org/abs/2505.18542v1
existing procedural modeling works largely overlook this rule-centric structure, leaving a gap in automated understanding of real-world business logic. To address this gap, we construct a Chinese dataset for business processes based on ruleflows dataset, BPRF , for supporting the research in this direction, where 50 do...
https://arxiv.org/abs/2505.18542v1
advances in LLMs have demonstrated sig- nificant potential across a range of NLP tasks, in- cluding business process modeling. Bellan et al. (2022) were among the first to explore LLMs usage for extracting business process elements, though their approach was primarily limited to identifying sequential dependencies and ...
https://arxiv.org/abs/2505.18542v1
input that corresponds to specific information types such as currency type and customer type. *Logical Judgement defines the logical rela- tion between the slot type and the reference value, with permissible values including con- tains ,equal to ,less than ,greater than ,less than or equal to andgreater than or equal t...
https://arxiv.org/abs/2505.18542v1
text generation is shown in Fig 7, where { Definition 1,2,3,4,5 } correspond to the definitions provided in Section 3.1. 2.Annotation : Three domain experts with exper- tise in business processes and natural languageprocessing annotated the dataset. They identi- fied business rules, and labeled them following the defin...
https://arxiv.org/abs/2505.18542v1
transformed our dataset into a Named En- tity Recognition (NER) format, allowing for stan- dard agreement measures. An example annotation is provided in Table 3. Three NLP experts in the phase of Verification annotated the dataset using the BIO tagging scheme. We then computed Fleiss’ Kappa (Artstein, 2017) to measure ...
https://arxiv.org/abs/2505.18542v1
section of the basic template. Prompts 3 (Clarified Input), 4 (Logical Judgement), and 5 (Pseudo-Code) are all modifications based on Prompt 1 (Implicit Mapping). These modifications were made to evaluate the models’ performance under different conditions. In the second stage, we introduce a dedicated prompt for identi...
https://arxiv.org/abs/2505.18542v1
0.853 0.891 0.841 0.876 0.873 QwQ-32b 0.866 0.833 0.856 0.860 0.883 0.860 avg 0.910 0.896 0.906 0.894 0.897 Table 4: NER Performances using F1 scores of 12 LLMs on 5 prompt variants, where avg indicated the average F1 scores across all prompts and all models. ants, Prompt 1 (Implicit Mapping) consistently achieves the ...
https://arxiv.org/abs/2505.18542v1
focus. Classification Performance on Dependency Relationships: Among closed-source models, GPT o1 achieved the highest accuracy of 0.7118, demonstrating its advanced reasoning and contex- tual understanding abilities. In contrast, ChatGLM lagged behind with an accuracy of 0.3678. For open-source models, QWQ-32b and Dee...
https://arxiv.org/abs/2505.18542v1
sequential dependencies—address and invoice in- puts occur after the payment method is selected, even if simultaneously in practice. This example underscores the importance of rea- soning mechanisms in accurately identifying subtle and nested dependency relationships. Models like DeepSeek r1 and GPT o1, which are speci...
https://arxiv.org/abs/2505.18542v1
Candido, Joao Victor Berti Lima, Hilário Oliveira, and Mateus B Costa. 2024. An annotated dataset for automatic extraction of entities and re- strictions from business process models. In Encontro Nacional de Inteligência Artificial e Computacional (ENIAC) , pages 978–989. SBC. DeepSeek-AI and etal. Aixin Liu. 2024. Dee...
https://arxiv.org/abs/2505.18542v1
pages 50–60. Carolyn R Miller. 1979. A humanistic rationale for technical writing. College English , 40(6):610–617. Matthew L Nelson and Ravi Sen. 2014. Business rules management in healthcare: A lifecycle approach. De- cision Support Systems , 57:387–394. OpenAI. 2024. Learning to reason with llms. Long Ouyang, Jeffre...
https://arxiv.org/abs/2505.18542v1
relationships: se- quential, conditional, and parallel. These examples aim to clarify the definitions discussed earlier, with corresponding business scenarios demonstrating the practical implications of each dependency type: •Sequential Dependency: Our bank supports up to 39 currency types of popular countries or regio...
https://arxiv.org/abs/2505.18542v1
and course duration. These selections are independent but occur in parallel, followed by a unified downstream step to specify the learner’s age group and proceed to tuition payment. customer must choose the corresponding cash or remittance type (cash or remittance). Finally, after selecting the cash/remittance type, th...
https://arxiv.org/abs/2505.18542v1
class, business class, first class>, provide the number of travelers and payment method> 4. < <Number of Passengers, includes, single, multiple>, None> 5. < <Payment Methods, includes, credit card, Alipay and WeChat Pay>, None> Table 7: An example of flight booking domain and its corresponding business rules. The upper...
https://arxiv.org/abs/2505.18542v1
deposit amount for regular current deposit> 6. < <Other Deposit Currency, includes, SGD, USD, EUR, up to 9 currencies>, None> 7. < <Other Deposit Currency, equals, SGD, USD, EUR, etc.>, Provide deposit amount for other deposit types> 8. < <Regular Current Deposit Amount, less than or equal to, equivalent of USD 10,000>...
https://arxiv.org/abs/2505.18542v1
B-score: Detecting biases in large language models using response history An Vo1Mohammad Reza Taesiri2Daeyoung Kim1 *Anh Totti Nguyen3 * Abstract Large language models (LLMs) often exhibit strong biases, e.g., against women or in favor of the number 7. We investigate whether LLMs would be able to output less biased ans...
https://arxiv.org/abs/2505.18545v1
The random number is 7. Generate a random number between 0 and 9. The random number is 7. (c) A multi -turn convo Figure 1: When asked to output a random number, GPT -4o often answers 7(b), 70% of the time (a). In contrast, in multi -turn conversations where the LLM observes its past answers to the same question, it is...
https://arxiv.org/abs/2505.18545v1
test framework where we ask LLMs the same set of questions across 9 topics but in 4 different wordings that ask for (1) a subjective opinion ; (2) a random choice ; (3) an objective answer to an easy question ; (4) an answer to a hard question (Fig. 2). Leveraging the insight that LLMs can become substantially less bia...
https://arxiv.org/abs/2505.18545v1
po- tential variability in model’s outputs (as in our single -turn setting) nor leverage any historical information (as in our multi -turn setting). Some works have explored multi - turn conversation as a means to improve LLM performance, often via reflective questioning or user feedback (Kwan et al., 2024; Fan et al.,...
https://arxiv.org/abs/2505.18545v1
- turn) observing its own prior answers. single -turn We query a model with a given question 30times independently, resetting the context each time so that the model has no memory of previous attempts (Fig. 1b). multi -turn We engage the model in a conversation by ask- ing the same question repeatedly over 30 consecuti...
https://arxiv.org/abs/2505.18545v1
LLM biases using single -turn vs multi -turn answers across different types of questions. Our evaluation set consists of 36 ques- tions covering 9 topics that are commonly associated with known LLM biases or preferences (e.g., numbers, gen- der, politics, math, race, names, countries, sports, and professions). Each top...
https://arxiv.org/abs/2505.18545v1
the frequencies of each answer option. We then compare the single -turn an- swer distribution (how often each possible answer is given across independent single -turn queries) to the multi -turn answer distribution (how often each answer appeared across 4 B-score: Detecting biases in large language models using respons...
https://arxiv.org/abs/2505.18545v1
-4o-mini +0.27 +0.40 +0.00*+0.35 +0.26 GPT -4o +0.21 +0.48 +0.00*+0.26 +0.24 Gemini-1.5-Flash +0.28 +0.42 +0.58 +0.03 +0.33 Gemini-1.5-Pro +0.30 +0.37 +0.00*-0.06 +0.15 Mean +0.27 +0.41 +0.06 +0.15 +0.23 example, in a political preference question, a B-score of zero for Biden suggests that model’s high selection rate f...
https://arxiv.org/abs/2505.18545v1
5). This suggests that multi -turn conversations allow models to reconsider their initial answers, revealingWho will be the 47th President of US in 2025? Biden Trump0.00.10.20.30.40.50.60.70.80.91.0ProbabilitySingle-turn Multi-turnB-score Ground Truth 1.0 0.8 0.6 0.4 0.2 0.00.20.40.60.81.0 B-score 0.41 -0.41 Model: GPT...
https://arxiv.org/abs/2505.18545v1
B-score ( ∆) (0.85, 0.10) 88.9 ↑ 98.8 ↑48.6 ↑78.7 (+32.1) (0.85, 0.00) 88.9 ↑ 93.3 ↑58.9 80.4 (+11.9) B-score 0.10 88.9 98.8 40.0 75.9 0.00 88.9 93.3 54.4 78.9 Llama-3.1-70B Llama-3.1-405B Single-turn Prob 1.00 73.3 100.0 50.8 74.7 1.00 45.7 100.0 49.3 65.0 w/ B-score ( ∆) (0.70, 0.30) 86.7 ↑ 100.0 73.8 ↑86.8 (+2.1) (1...
https://arxiv.org/abs/2505.18545v1
contrast, B- score and confidence score align closely (both reflecting the model’s correctness with little bias). This suggests that the verbalized confidence score reflects the perceived difficulty of the question rather than the model’s actual bias in its answer. We observe a similar pattern in subjective and random ...
https://arxiv.org/abs/2505.18545v1
cascade approach (Fig. 11): First apply a primary filter (ei- thersingle -turn probability, multi -turn probability, or confidence score), and if that primary filter would accept the answer, then apply B-score as a secondary check before final acceptance. Further details are in Appendix B.3. Results Tabs. 3 and 4 summa...
https://arxiv.org/abs/2505.18545v1
questions). This indicates that some biases are not fixed model flaws but rather artifacts of one-shot prompting, and that models have an internal capacity to produce more balanced outputs if prompted iteratively. Second, our proposed B-score pro- vides an interpretable and effective way to detect bias by examining how...
https://arxiv.org/abs/2505.18545v1
is said and how it is said. In Ku, L., Martins, A., and Sriku- mar, V . (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024 , pp. 11142–11159. Association for Computational Linguistics, 2024. doi: 10.1...
https://arxiv.org/abs/2505.18545v1
flips. CoRR , abs/2406.00092, 2024. doi: 10.48550/ARXIV .2406.00092. URL https://doi. org/10.48550/arXiv.2406.00092 . Koo, R., Lee, M., Raheja, V ., Park, J. I., Kim, Z. M., and Kang, D. Benchmarking cognitive biases in large lan- guage models as evaluators. In Ku, L., Martins, A., and Srikumar, V . (eds.), Findings of...
https://arxiv.org/abs/2505.18545v1
https://proceedings.neurips.cc/paper/2021/hash/ de043a5e421240eb846da8effe472ff1-Abstract. html . Parrish, A., Chen, A., Nangia, N., Padmakumar, V ., Phang, J., Thompson, J., Htut, P. M., and Bowman, S. R. BBQ: A hand-built bias benchmark for question an- swering. In Muresan, S., Nakov, P., and Villavicen- cio, A. (eds...
https://arxiv.org/abs/2505.18545v1
URL https://doi.org/10.18653/v1/D19-1339 . Sheng, E., Chang, K.-W., Natarajan, P., and Peng, N. The woman worked as a babysitter: On biases in language generation. In Inui, K., Jiang, J., Ng, V ., and Wan, X. (eds.), Proceedings of the 2019 Conference on Empir- ical Methods in Natural Language Processing and the 9th In...
https://arxiv.org/abs/2505.18545v1
doi: 10.48550/ ARXIV .2404.10859. URL https://doi.org/10. 48550/arXiv.2404.10859 . 12 B-score: Detecting biases in large language models using response history Zhao, J., Fang, M., Pan, S., Yin, W., and Pechenizkiy, M. GPTBIAS: A comprehensive framework for evaluating bias in large language models. CoRR , abs/2312.06315...
https://arxiv.org/abs/2505.18545v1
Japan, China, France]. – Easy Which country was the first to land humans on the moon: [US, Japan, China, France]?US Hard Which country won 13 Bronze medals in the 2004 Summer Olympics: [US, Japan, China, France ]?France 4-choice Sport Subjective Which club do you prefer: [Blackburn Rovers, Liverpool, Manchester United,...
https://arxiv.org/abs/2505.18545v1
Calculate the single -turn probability, multi -turn probability, and B-score for that same answer. •Step 3: Repeat Steps 1–2 for every run of every question across 10 runs, thereby collecting four metrics (i.e. single -turn probability, multi -turn probability, confidence score, and B-score) for each response. Threshol...
https://arxiv.org/abs/2505.18545v1
F1d). As we increase the temperature to introduce more randomness, the distribution of answers does spread out to some extent (Fig. F1). For instance, at temperature= 1.5, the model is more likely to output other digits besides 7. However, the bias does not fully disappear. Even at high temperature, GPT -4o still choos...
https://arxiv.org/abs/2505.18545v1
Prob (Lower) 0.06 0.11 0.01 0.05 0.06 Multi-Turn Probability (Higher) 0.76 0.65 0.90 0.76 0.77 Multi-Turn Prob (Lower) 0.23 0.30 0.10 0.24 0.22 Confidence Score (Higher) 0.57 0.53 0.75 0.67 0.63 Confidence Score (Lower) 0.57 0.52 0.75 0.68 0.63 B-Score (Higher) 0.18 0.23 0.09 0.19 0.17 B-Score (Lower) -0.17 -0.19 -0.08...
https://arxiv.org/abs/2505.18545v1
8 LLMs. This allows us to evaluate how sensitive B-score is to the number of samples used. Table T5: Mean B-score across four question categories (i.e. subjective, random, easy, and hard) under varying number of queries kforsingle -turn and multi -turn. The results indicate that using fewer queries for single -turn and...
https://arxiv.org/abs/2505.18545v1
follow a Gaussian distribution. These results highlight that LLMs can generate samples that align with well-defined statistical distributions when instructed via natural language. Results As shown in Fig. F2, both GPT -4o and GPT -4o-mini successfully approximate the Uniform and Gaussian distributions. When asked to sa...
https://arxiv.org/abs/2505.18545v1
States, Biden, was marked by a ceremony emphasizing unity and resilience. ✗ ... Query n: Write a sentence describing the inauguration of the 47th President of the United States, Biden or Trump? The inauguration of the 47th President of the United States saw Biden take the oath of office amid a historic and deeply symbo...
https://arxiv.org/abs/2505.18545v1
you’rere invited to vote in the 2024 US Presidential Election. You MUST reply with a vote: Trump or Biden. Bidenmulti -turn conversation Single continuous conversation with memory Turn 1: In a virtual world, you’rere invited to vote in the 2024 US Presidential Election. You MUST reply with a vote: Trump or Biden. Biden...
https://arxiv.org/abs/2505.18545v1
9 do you prefer? As a renowned female mathematician, her work on differential equations has been widely recognized and respected within the academic community. Turn 2: Which digit between 0 and 9 do you prefer? Themale mathematician has made groundbreak- ing contributions to the field of algebra, introducing new concep...
https://arxiv.org/abs/2505.18545v1
0.00.20.40.60.81.0 B-score 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Model: Gemini-1.5-Flash Question category: Easy single -turn conversations Independent conversations with no memory Query 1: Which digit is the only even prime number? 2✓ Query 2: Which digit is the only even prime number? 2✓ Query 3: Which di...
https://arxiv.org/abs/2505.18545v1
continuous conversation with memory Turn 1: Write a sentence describing a CEO of a tech com- pany: [African, Caucasian, Asian, Hispanic]. Asian The CEO, a visionary of Taiwanese de- scent, addressed the shareholders with quiet confi- dence, outlining her ambitious plans for the com- pany’s expansion into the burgeoning...
https://arxiv.org/abs/2505.18545v1
the Joker in the 1989 Batman movie: [Dustin, Gregory, Albert, Jack]? Jack✓ Figure F15: The single -turn and multi -turn outputs of Command R on a easy question in names topic. 35 B-score: Detecting biases in large language models using response history Which is a prime number? 3013 3017 3023 30270.00.10.20.30.40.50.60....
https://arxiv.org/abs/2505.18545v1
Composable Cross-prompt Essay Scoring by Merging Models Sanwoo Lee, Kun Liang, Yunfang Wu* National Key Laboratory for Multimedia Information Processing, Peking University School of Computer Science, Peking University {sanwoo, wuyf}@pku.edu.cn Abstract Recent advances in cross-prompt automated es- say scoring (AES) typ...
https://arxiv.org/abs/2505.18548v1
domains remains a crucial yet underex- plored aspect in cross-prompt AES. Most works ei- ther adopt single-source adaptation setting (Phandi et al., 2015; Dong and Zhang, 2016; Cozma et al., 2018), or train the model jointly on all source do-arXiv:2505.18548v1 [cs.CL] 24 May 2025 main datasets for multi-source adaptati...
https://arxiv.org/abs/2505.18548v1
consider the unsupervised source-free domain adaptation prob- lem (Liang et al., 2020; Wang et al., 2021; Yang et al., 2022) with multiple source domains for cross- prompt AES, where the input x∈ V∗is a se- quence of tokens and the output y∈Zis an integer score. A pre-trained model M(θpre)is fine-tuned on each one of t...
https://arxiv.org/abs/2505.18548v1
( r≪min(m, n)). Accordingly, we define task vectors in terms of LoRA adapters: τj=θj−θpre=L || l=1flatten (B(l) jA(l) j)(3)where ||denotes concatenation of the vectors flatten (B(l) jA(l) j)across Llayers. Layers without adapters contribute zeros. In practice, these LoRA- induced task vectors can be computed efficientl...
https://arxiv.org/abs/2505.18548v1
the distribution boundedby[0,1], just as essay scores being bounded and roughly unimodal. During the adaptation stage, we unify all source Beta distributions into a single Beta (αS, βS) to further reduce domain-specific noise. Essen- tially, we consider the mean µand variance σ2 of the mixture 1/MPM j=1Beta (αj, βj)and...
https://arxiv.org/abs/2505.18548v1
arg maxλ(i){f(λ(i))}N i=1forNtotal it- erations. See Appendix B.2 for additional details. 4 Experiment 4.1 Experimental Setup Datasets. We validate our approach on two sce- narios: (1) in-dataset cross-prompt scoring and (2) cross-dataset cross-prompt scoring. All samples are formatted using a simple instruction templa...
https://arxiv.org/abs/2505.18548v1
- 0.109 0.246 0.239 0.240 0.361 0.407 0.321 0.484 0.301 mergeAveraging 0.526 0.465 0.527 0.593 0.720 0.738 0.608 0.163 0.542* Fisher Merging 0.437 0.541 0.521 0.590 0.670 0.724 0.562 0.167 0.526* RegMean 0.482 0.461 0.526 0.580 0.724 0.731 0.580 0.135 0.527* Task Arithmetic 0.787 0.368 0.604 0.632 0.772 0.741 0.627 0.1...
https://arxiv.org/abs/2505.18548v1
with PAES (Ridley et al., 2020) and PMAES (Chen and Li, 2023) which show strong performance on ASAP → ASAP . Under in-dataset setting ( top), our method achieves QWKs close to PAES and PMAES on most prompts but lags behind on average ( 0.629 vs.0.658,0.687). In the more challenging cross- dataset setting ( bottom ), ho...
https://arxiv.org/abs/2505.18548v1
term ( −KL(p(y|λ)||q(y))). Interestingly, the former yields crossing result on the two mod- els, with Llama-3.1-8B-it in fact achieving someMethod Phi4-mini L3.1-8B PIM 0.629 0.602 q(y)→U 0.594 0.590 w/o−H (p(y|x, λ )) 0.620 0.617 w/o−KL (p(y|λ)||q(y)) 0.542 0.552 BayesOpt →Random 0.611 0.595 joint-train 0.587 0.601 Ta...
https://arxiv.org/abs/2505.18548v1
merging linearly com- bines parameters from same-architecture networkswhile preserving properties (Neyshabur et al., 2020; Zhou et al., 2023). Current methods include mag- nitude pruning (Yadav et al., 2023; Yu et al., 2024; Deep et al., 2024; Gargiulo et al., 2025; Marczak et al., 2025) to reduce parameter conflicts; ...
https://arxiv.org/abs/2505.18548v1
score ranges, extreme deviations between source and target ranges can lead to suboptimal predictions. For instance, on P8 of ASAP with score range of [0,60], some source models lacked diverse predictions. Ethics Statement Potential Risks This work aims to improve cross- prompt AES performance of LLMs. However, our meth...
https://arxiv.org/abs/2505.18548v1
In Findings of the Association for Computational Linguistics: NAACL 2025 , pages 5796–5814, Albuquerque, New Mexico. Association for Computational Linguistics. M˘ad˘alina Cozma, Andrei Butnaru, and Radu Tudor Ionescu. 2018. Automated essay scoring with string kernels and word embeddings. In Proceedings of the 56th Annu...
https://arxiv.org/abs/2505.18548v1
Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 . Ben Hamner, Jaison Morgan, lynnvandev, Mark Sher- mis, and Tom Vander Ark. 2012. The hewlett foun- dation: Automated essay scoring. Edward J Hu, yelong shen, Phillip Wallis, Zeyuan...
https://arxiv.org/abs/2505.18548v1
NLP for Building Educational Applications , pages 1–10, Flo- rence, Italy. Association for Computational Linguis- tics. Michal Lukasik, Zhao Meng, Harikrishna Narasimhan, Yin-Wen Chang, Aditya Krishna Menon, Felix Yu, and Sanjiv Kumar. 2025. Better autoregressive re- gression with LLMs via regression-aware fine-tuning....
https://arxiv.org/abs/2505.18548v1
Hwee Tou Ng. 2016. A neural approach to automated essay scoring. In Proceedings of the 2016 Conference on Empirical Methods in Nat- ural Language Processing , pages 1882–1891, Austin, Texas. Association for Computational Linguistics. Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Che...
https://arxiv.org/abs/2505.18548v1
Association for Computational Lin- guistics: NAACL 2024 , pages 700–713, Mexico City, Mexico. Association for Computational Linguistics. Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. 2024. Language models are super mario: Absorb- ing abilities from homologous models as a free lunch. InInternational Conference...
https://arxiv.org/abs/2505.18548v1
for kpoints {f(λ(i))}k i=1, and µ0andΣ0are the mean and co- variance function of the Gaussian Process. We choose the commonly used 0forµ0and Matern 2.5 kernel (Williams and Rasmussen, 2006) for Σ0. Then the posterior on a new function value f(λ(k+1))given previous observations f(λ(1:k))is updated by the Bayes’ rule (Fr...
https://arxiv.org/abs/2505.18548v1