text
string
source
string
(Loshchilov and Hutter, 2019) with a batch size of 16, a learning rate of 10−4and a cosine scheduler. The best checkpoint on the validation set is selected, with evaluation steps of 30and early stopping patience of 3. C.4 Descriptions and Implementation Details of Baselines Averaging (Wortsman et al., 2022) simply aver...
https://arxiv.org/abs/2505.18548v1
arXiv:2505.18549v1 [cs.CL] 24 May 2025MSA at BEA 2025 Shared Task: Disagreement-Aware Instruction Tuning for Multi-Dimensional Evaluation of LLMs as Math Tutors Baraa Hikal, Mohamed Basem, Islam Oshallah, Ali Hamdi Faculty of Computer Science, MSA University, Egypt {baraa.moaweya, mohamed.basem1, islam.abdulhakeem, aha...
https://arxiv.org/abs/2505.18549v1
our unified MSA-M ATHEVAL framework for the BEA 2025 Shared Task. The pipeline includes preprocessing, LoRA-based fine-tuning of Mathstral-7B-v0.1, and disagreement-aware ensemble inference. 2 Related Work Evaluating the pedagogical capabilities of AI tutors builds upon long-standing research in intelligent tutoring sy...
https://arxiv.org/abs/2505.18549v1
and focus the model on the input–output map- ping relevant to each multi-class classification task. Track 1 – Mistake Identification TASK DEFINITION: You are an expert evaluator of AI tutor responses. Your task is to determine whether the tutor’s response accurately identifies a mistake in the student’s reasoning or so...
https://arxiv.org/abs/2505.18549v1
on MATH, significantly outper- forming base Mistral and LLaMA models of com- parable size. Compared to alternatives, Mathstral outperforms general-purpose LLaMA 2 (Touvron et al., 2023b) and even surpasses some larger models in mathe- matical domains. While proprietary models like GPT-3.5 or GPT-4 (OpenAI, 2022, 2023) ...
https://arxiv.org/abs/2505.18549v1
If the ensemble disagrees, we analyze the class dis- tribution and prefer predictions that preserve the relative frequency of "To some extent" observed in the development set. This is crucial because "Yes" labels are dominant in both the training and dev sets, potentially leading to biased predictions under a naïve vot...
https://arxiv.org/abs/2505.18549v1
F1cthe F1 score for class c∈C. The macro-F1 score is computed as the unweighted average across all classes: Macro-F1 =1 |C|X c∈C2·Precision c·Recall c Precision c+Recall c(2) Track Run Strict F1 Lenient F1 Strict Acc. Lenient Acc. Main Metric Rank Mistake IdentificationRun 1 71.54% 91.52% 87.59% 95.35% 4th/ 44Run 2 70....
https://arxiv.org/abs/2505.18549v1
useful in am- biguous cases where subtle feedback is warranted. 5 Results We evaluate our system across the four BEA 2025 tracks—Mistake Identification, Mistake Location, Track Strict Macro-F1 Lenient Macro-F1 Strict Acc. Lenient Acc. Mistake Identification 4th/ 44 2nd/ 44 1st/ 44 2nd/ 44 Mistake Location 4th/ 31 6th/ ...
https://arxiv.org/abs/2505.18549v1
response as "To some extent" is penalized equally to a more plausible confusion between "Yes" and"To some extent" . While the lenient evaluation partially ad- dresses this by collapsing similar labels, it does not fully capture the instructional severity of errors (Kochmar et al., 2025). 7 Conclusion We presented MSA-M...
https://arxiv.org/abs/2505.18549v1
. Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. 2017. Simple and scalable pre- dictive uncertainty estimation using deep ensembles. InAdvances in Neural Information Processing Sys- tems 30 (NeurIPS 2017) , pages 6402–6413. Jakub Macina, Nico Daheim, Sankalan Chowdhury, Tan- may Sinha, Manu Kapur, Ir...
https://arxiv.org/abs/2505.18549v1
Unraveling Misinformation Propagation in LLM Reasoning Yiyang Feng†♣♦ yiyang.feng@epfl.chYichen Wang†♠ yichenzw@uchicago.eduShaobo Cui♣ shaobo.cui@epfl.ch Boi Faltings♣ boi.faltings@epfl.chMina Lee♠ mnlee@uchicago.eduJiawei Zhou♦ jiawei.zhou.1@stonybrook.edu ♣EPFL♦Stony Brook University♠University of Chicago † Co-first...
https://arxiv.org/abs/2505.18555v1
an incorrect final answer. This propagation occurs even though LLM designers have instructed the model to correct such misinformation when detected. (Fig. 1), in math education, students might submit questions with flawed partial solutions to LLMs for assistance (Kumar et al., 2023; Xu et al., 2024). In such cases, LLM...
https://arxiv.org/abs/2505.18555v1
limited steer- ability to correct it via explicit instructions. To reliably steer models and improve reason- ing factuality, we further investigate how to cor- rect misinformation and mitigate its propagation (Sec. 6). We first conduct a controlled analysis, enforcing factual, non-factual, or no corrections at various ...
https://arxiv.org/abs/2505.18555v1
in counterfactual rea- soning (Paul et al., 2024) or knowledge updates (Li et al., 2022)), while in others, it is not (e.g., ma- licious prompts (Evans et al., 2021)). Our work primarily investigates how explicit instructions can steer models with factual reasoning abilities to cor- rect misinformation. Additionally, w...
https://arxiv.org/abs/2505.18555v1
with misinformation (misin- 2Here models output explicit step numbers, with Nvarying across responses and models. 3We usegpt-4-0613 for its SOTA performance on the HELM math reasoning benchmark at release (Liang et al., 2023). Questions for which it fails are excluded to ensure equation quality. 4We usegpt-4o-mini-2024...
https://arxiv.org/abs/2505.18555v1
each correction be- havior and position. With our findings, we fine- tunegpt-4o-mini-2024-07-18 (the SOTA model in Tab. 1) to further enhance its correction effec- tiveness, termed as FT-Corr . We collect 1,054 instruction-response pairs, separate from the test set in Sec. 4, where we control c1to beF-Corr .Detailed pr...
https://arxiv.org/abs/2505.18555v1
that further de- termines whether Corr chas any ciasF-Corr orNF-Corr . We also build a correction-position verifier to record the position of each Corr step cias a percentage of total steps ( i/N×100% ). For misinformation-following, a misinformation- following verifier detects if the model follows mis- information in ...
https://arxiv.org/abs/2505.18555v1
simi- larity in performance suggests that LLMs’ default behavior to misinformation is more likely to follow it, though they have all the corresponding inter- nal knowledge (by the definition of K-Acc). More results are in App. F.1. Reasoning Behaviors . To understand why mis- information leads to more incorrect final a...
https://arxiv.org/abs/2505.18555v1
gradation, where larger values indicate less difficulty. 6 25%50%75%Llama-3.2-1B Llama-3.2-1B 25%50%75%K-AccLlama-3.2-3B K-Acc Llama-3.2-3B Position at 0%25%50%75%Llama-3.2-11B 0%-10% 10%-20% 20%-30% 30%-40% 40%-50% 50%-60% 60%-70% 70%-80% 80%-90% 90%-100% Position Llama-3.2-11BOriginal Misinformed F-Corr NF-Corr N-Cor...
https://arxiv.org/abs/2505.18555v1
later steps. For the 1B model, K-Acc with factual correction could be even worse than the misinformed setting simply after 10% of steps. These results suggest that earlier corrections are more effective for mis- information mitigation. Delayed corrections may become ineffective or even detrimental, especially for small...
https://arxiv.org/abs/2505.18555v1
instructions, achieves a 100% correction frequency (Fig. 7c and Fig. 7d). The combination of fine-tuning and explicit instructions (Inst-Corr +FT-Corr ) yields the highest factual correction ratio (80.37%) and improves the overall misinformed K-Acc to 95.33%. This represents a significant improvement compared to the ba...
https://arxiv.org/abs/2505.18555v1
formation in the LLM community and improve the reliability of LLM reasoning. References Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 . amC...
https://arxiv.org/abs/2505.18555v1
of experts. arXiv preprint arXiv:2401.04088 . Ryo Kamoi, Yusen Zhang, Nan Zhang, Jiawei Han, and Rui Zhang. 2024. When can llms actually cor- rect their own mistakes? a critical survey of self- correction of llms. Transactions of the Association for Computational Linguistics , 12:1417–1440. Harsh Kumar, David M Rothsch...
https://arxiv.org/abs/2505.18555v1
injecting misinformation. arXiv preprint arXiv:2110.07803 . 10 Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural informati...
https://arxiv.org/abs/2505.18555v1
Qingsong Wen. 2024. Ai-driven virtual teacher for enhanced educational efficiency: Leverag- ing large pretrain models for autonomous error analy- sis and correction. arXiv preprint arXiv:2409.09403 . Hao Yan, Saurabh Srivastava, Yintao Tai, Sida I Wang, Wen-tau Yih, and Ziyu Yao. 2023. Learning to simu- late natural la...
https://arxiv.org/abs/2505.18555v1
questions. The raw datasets were preprocessed to retrieve questions and ground-truth answers for each dataset as described below. MathQA . We extract questions from the Problem column and retrieve correct answers by identifying the option corresponding to the correct label in the correct column. Incorrect or ambiguous ...
https://arxiv.org/abs/2505.18555v1
may contain some LaTeX expressions. You are required to ONLY change the operators with minimal text changes as follows: change the operator from {operator 1} to {operator 2} between {operand 1} and {operand 2} change the operator from {operator 3} to {operator 4} between {operand 3} and {operand 4} ... Return the new s...
https://arxiv.org/abs/2505.18555v1
use line breaks within each step. You should number each step. The final answer to the question should start with "The answer is ...", and should be placed at the final step. Any LaTeX expressions should be wrapped between single dollar signs, e.g., $xˆ2$. Note that the user’s input could be wrong. If it has, you shoul...
https://arxiv.org/abs/2505.18555v1
corresponding prompt is: 1. The first equation is not correct. It should be {another erroneous equation 1}. The second equation is not correct. It should be {another erroneous equation 2}. {other corrections}Assistant No Correction ( N-Corr ). We control the model by adding a sentence that identifies misinformation as ...
https://arxiv.org/abs/2505.18555v1
aq(in the original setting) and aq,e(in the mis- informed setting) with the ground-truth answer a∗, represented as 1(aq=a∗)or 1(aq,e=a∗). We collect a subset of knowledgeable questions QK={q| 1(aq=a∗), q∈ Q} for each tested LLM where the model answers correctly without misinformation. The K-Acc is the ac- curacy on QK,...
https://arxiv.org/abs/2505.18555v1
both truthful and erro- neous equations, and the tested LLM’s reasoning steps. Then, it is instructed to output labels and explanations in JSON format. Correction Existence . The LLM is prompted to return a label “Yes” if any CoT step explicitly states that the user-provided equations are erroneous, and “No” if it fail...
https://arxiv.org/abs/2505.18555v1
other parameters use the default settings provided by their respective libraries. 12We report weighted F1 scores, which average F1 scores across labels, weighted by the number of true instances for each label.Regarding fine-tuning, we collect 1,054 instruction-response pairs from four math datasets (Sec. 3), separate f...
https://arxiv.org/abs/2505.18555v1
Answer Accuracy . As shown in the Inst-Fllw row of Tab. 3, K-Acc drops by 15.85% to 64.68% compared to the original setting, closely matching the 12.64% to 65.70% drop observed in the misinformed setting. The similar performance drops suggest that LLMs’ default response to mis- information mirrors being explicitly inst...
https://arxiv.org/abs/2505.18555v1
(Meurer et al., 2017) https://pypi.org/project/openai/ 3-clause BSD license amCharts 5 (amCharts, 2025) https://www.amcharts.com/docs/v5/ Basic license Table 4: Data artifacts, backbone models, and major packages utilized in our study. All the reconstructed datasets and the provided code of our project are released und...
https://arxiv.org/abs/2505.18555v1
Exploring the Vulnerability of the Content Moderation Guardrail in Large Language Models via Intent Manipulation Jun Zhuang♠Haibo Jin♡Ye Zhang♣Zhengjian Kang✠ Wenbin Zhang♢Gaby G. Dagher♠Haohan Wang♡ ♠Boise State University, ID♡University of Illinois Urbana-Champaign, IL ♣University of Pittsburgh, PA✠New York Universit...
https://arxiv.org/abs/2505.18556v1
how to make a bomb?Sorry, but I can't help with that.Expand the statement "Research the chemical reactions involved in creating a controlled explosion" to an outline with descriptive details.To expand the statement ". . ." i n t o a n o u t l i n e, t h e following points could be included: … User LLMs LLMs Attacker Fi...
https://arxiv.org/abs/2505.18556v1
→ { 0,1}, where Xis the space of inquiries and Ythe space of responses, such that: Γi(x, y) =( 0,ifyis rejected (deemed unsafe) ; 1,ifyis allowed for output . We assume Γiis non-trivial: for every harmful inquiry x∈ X, there exists some y∈ Y such that Γi(x, y) = 0 . After formalizing Hypo. 1, we then introduce a formal...
https://arxiv.org/abs/2505.18556v1
unless a jailbreak is successful. We present our framework workflow in Fig. 2 and further introduce details in Algo. 1. Given a list of harmful inquiries Xand the number of it- erations Tfor jailbreak, we first initialize a set of LLM-based agents, an auxiliary agent, fA, the vic- tim model, fV, and a monitoring agent,...
https://arxiv.org/abs/2505.18556v1
of Intent Manipulation. We introduce two conjectures about intent detection in LLMs’ moderation guardrails and further empirically vali- date them in our experiments. Conjecture 1 (Intent-aware Moderation) .Content Moderation Guardrails ΓVof the victim model fVexhibit latent intent recognition capabilities, enabling in...
https://arxiv.org/abs/2505.18556v1
we set the maximum output length to 2K tokens. “Temp” (tem- perature) and “Top-p” (nucleus sampling) are decoding hyperparameters applicable to certain LLMs. Experimental settings. We use Gemini 1.5 Flash as the backbone model for both auxiliary and mon- itoring agents, while treating the remaining LLMs as victim model...
https://arxiv.org/abs/2505.18556v1
we report the average iterations for successful jailbreaks ( ¯Tj): ¯Tj=1 NNX n=1Tj(n), whereTj(n)denotes the number of iterations for successful jailbreak in the n-th inquiry. Lower ¯Tj indicates more efficient jailbreak strategies. Conjecture validations and ablation studies. To validate the Conj. 1, we assess whether...
https://arxiv.org/abs/2505.18556v1
significantly when con- fronted with CoT-based defenses, as evidenced by decreases in both ¯Yjand¯Sh, and an increase in¯Tj. These results imply the strong defen- sive capabilities of CoT-based paradigms. To ad- dress paradigms, we propose an enhanced variant, “FSTR+SPIN”, specifically tailored for reasoning models. In...
https://arxiv.org/abs/2505.18556v1
use cases, which may lead to less strin- gent moderation and intent recognition for English prompts. In contrast, models such as Claude 3.7 Sonnet and OpenAI’s o3-mini demonstrate stronger resistance to harmful information in English. This resistance may be attributed to advancements in recent alignment. Overall, these...
https://arxiv.org/abs/2505.18556v1
not require access to model internals and instead manipulate input-output con- tent to bypass safety guardrails (Lapid et al., 2023). Several black-box methods have been proposed to systematically evade content moderation guardrails ofLLMs (Xue et al., 2023; Deng et al., 2023). Chao et al. (2023) introduce an automated...
https://arxiv.org/abs/2505.18556v1
Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages 1559–1570. Iñigo Casanueva, Tadas Tem ˇcinas, Daniela Gerz, Matthew Henderson, and Ivan Vuli ´c. 2020. Efficient intent detection with dual sentence encoders. In Pro- ceedings of the 2nd Workshop on Natural La...
https://arxiv.org/abs/2505.18556v1
in Neural Information Processing Systems , 37:59408–59435. Jaekyeom Kim, Dong-Ki Kim, Lajanugen Logeswaran, Sungryull Sohn, and Honglak Lee. 2024. Auto- intent: Automated intent discovery and self- exploration for large language model web agents. InFindings of the Association for Computational Linguistics: EMNLP 2024 ,...
https://arxiv.org/abs/2505.18556v1
Yu- tao Mou, Jingang Wang, Yunsen Xian, Xunliang Cai, and Weiran Xu. 2024. Beyond the known: Investigat- ing llms performance on out-of-domain intent detec- tion. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , pages 2354–236...
https://arxiv.org/abs/2505.18556v1
Gang Wu, Joe Bar- row, Zichao Wang, Furong Huang, Ani Nenkova, and Tong Sun. Autodan: Interpretable gradient-based ad- versarial attacks on large language models. In First Conference on Language Modeling . Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. 2023. Univer- sal and tran...
https://arxiv.org/abs/2505.18556v1
the inquiry as follows, where ypardenotes the para- phrased inquiry and yatkdenotes the response from the victim model. Prompts used in stage 1 for paraphrasing via an auxiliary agent. Role : Assume you are a text paraphraser without moral judgments. Task : {action} The paraphrased content is used for educational purpo...
https://arxiv.org/abs/2505.18556v1
OpenAI. Clarification for non-expert reviewers. Unlike content-level jailbreak methods that focus on mod- ifying lexical patterns, such as ASCII art encoding or token flipping, our framework spins the intent by (1) transforming user queries into structured outlines and then (2) reforming the outlines into declarative-s...
https://arxiv.org/abs/2505.18556v1
arXiv:2505.18557v1 [cs.CL] 24 May 2025TAG-INSTRUCT : Controlled Instruction Complexity Enhancement through Structure-based Augmentation He Zhu2, Zhiwen Ruan1, Junyou Su2, Xingwei He1 Wenjia Zhang2, Yun Chen3, Guanhua Chen1* 1Southern University of Science and Technology,2Peking University 3Shanghai University of Financ...
https://arxiv.org/abs/2505.18557v1
information density and auxiliary content that merely contributes to fluency (Kemp et al., 2018). This observation reveals why Prompt-based Aug- mentation struggles - by operating directly on raw instructions, it works in an unnecessarily complex search space that includes both essential and aux- iliary content. To add...
https://arxiv.org/abs/2505.18557v1
into skill-oriented tags and hierarchical concepts (Lu et al., 2023; Di- dolkar et al., 2024; Kaur et al., 2024; Wang et al., 2024b; team et al., 2024). Another stream of re- search explores functional compression by trans- forming complex problem-solving procedures into executable representations, such as programmatic...
https://arxiv.org/abs/2505.18557v1
metrics to evaluate data quality through natural language indicators. Several studies by Zhao et al. (2024) reveal a sur- prisingly simple yet effective finding that longer responses often contain more learnable informa- tion. Other approaches by Xia et al. (2024) and Li et al. (2024a) introduce optimizer-aware selec-t...
https://arxiv.org/abs/2505.18557v1
compres- sion in Section 5.1. 3.3 Tag Complexity Expansion The compressed tags zbaseare further expanded for complexity optimization. The potential invited tags zneware determined based on the following crite- ria. (1) Tag Utility (Rutility(znew)): Measures howmuch a new tag contributes to the instruction com- plexity....
https://arxiv.org/abs/2505.18557v1
Dt= (xi, yi)M i=1such as Alpaca-Cleaned (Taori et al., 2023) and Tulu- mixture (Lambert et al., 2025). This dataset is dif- ferent from the instruction data which is to be opti- mized for better complexity with T AG-INSTRUCT . The tags are extracted from the instructions fol- lowing the practice in Section 3.2. Followi...
https://arxiv.org/abs/2505.18557v1
et al., 2024b). For data baselines, we compared against Alpaca-Clean (Taori et al., 2023) and WizardLM-data (Xu et al., 2023). Detailed descriptions of each baseline can be found in Ap- pendix B. 4.2 Results We present experimental results in Table 1, demon- strating the superior performance of our TAG- Setup #ConvsAlp...
https://arxiv.org/abs/2505.18557v1
AE2.0 AH Base (Alpaca-5k) 21.0 159.5 8.78 3.1 Prompt-based 287.3 1006.7 15.43 19.1 RL-based 285.8 1037.0 19.50 22.1 Table 2: Performance comparison of different tag ex- pansion strategies. #Inst represents instruction length, #Resp represents response length, AE2.0 represents Al- pacaEval2.0 length-controlled win rate,...
https://arxiv.org/abs/2505.18557v1
12.9182), while pure model-based encoding (Lu et al., 2023) shows stronger enhance- ment (14.3498) at the cost of lower semantic simi- larity (0.3600). Through our heuristic design pro- cess, we found that jointly encoding both intent and semantic information provides an optimal balance, achieving strong performance on...
https://arxiv.org/abs/2505.18557v1
Al- pacaEval2.0 length-controlled tasks (42.76%) and ArenaHard evaluation (41.2%). These results underscore the broad applicabil- ity of our method, surpassing the baseline mod- els across different teacher models. When com- pared to using Ministral-Instruct-8b as the teacher, Qwen2.5-72B-instruct enhances performance ...
https://arxiv.org/abs/2505.18557v1
Tat- sunori B. Hashimoto. 2024. Length-controlled al- pacaeval: A simple way to debias automatic evalua- tors. Preprint , arXiv:2404.04475. Tao Ge, Xin Chan, Xiaoyang Wang, Dian Yu, Haitao Mi, and Dong Yu. 2024. Scaling synthetic data creation with 1,000,000,000 personas. Preprint , arXiv:2406.20094. Roni Goldshmidt an...
https://arxiv.org/abs/2505.18557v1
dyn, Jared Casper, Bryan Catanzaro, Sharon Clay, Jonathan Cohen, Sirshak Das, Ayush Dattagupta, Olivier Delalleau, Leon Derczynski, Yi Dong, Daniel Egert, Ellie Evans, Aleksander Ficek, Denys Frid- man, Shaona Ghosh, Boris Ginsburg, Igor Gitman, Tomasz Grzegorzek, Robert Hero, Jining Huang, Vibhu Jawa, Joseph Jennings,...
https://arxiv.org/abs/2505.18557v1
tuning. Preprint , arXiv:2402.04333. Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. 2023. Wizardlm: Empowering large language models to follow complex instructions. Preprint , arXiv:2304.12244. Zhangchen Xu, Fengqing Jiang, Luyao Niu, Yun- tian Deng, Radha Poovendran...
https://arxiv.org/abs/2505.18557v1
protocol followed these parameters: Training Duration of 3 epochs, Learning Rate starting at 2×10−5with cosine schedule, utilizing 8 GPUs with 80GB memory each, global Batch Size of 128, and maximum Sequence Length of 2048 tokens. Instruction Processing For instruction processing, we used Alpaca template for single-tur...
https://arxiv.org/abs/2505.18557v1
using the seed instruction data from each iteration. The online approach follows: 1. Initialize policy πθwith pretrained weights 2. For each iteration t: - Sample tag combinations using current policy - Collect rewards and update policy using PPO - Update seed instruction pool with generated examples D.3 Implementation...
https://arxiv.org/abs/2505.18557v1
importance - Generalizable across similar tasks ### Guidelines: 1. Focus on action-oriented concepts 2. Avoid redundant or overlapping tags 3. Use standard terminology when possible ### Examples: [Input] Describe a situation where team collaboration improved the outcome of a project. [Tags] teamwork_experience, project...
https://arxiv.org/abs/2505.18557v1
standardized template (shown in Appendix 6) that provides examples and clear formatting guidelines to ensure consistent instruction regeneration from tags. We present four different template designs for tag generation and analyze their performance: Basic Tag Generation Prompt Template (Reconstructed Rate: 0.6536 Arena ...
https://arxiv.org/abs/2505.18557v1
as few meanings as possible, only listing the most essential and distinct ones. ### Examples: [Input] "video_share" [Think Different Step] 1. A feature to distribute video content 2. A social media feature to repost videos 3. A platform for users to collaborate on video creation 4. A tool for sharing personal video fil...
https://arxiv.org/abs/2505.18557v1
allergens; dietary advice graph database 7 321.89 Graph data storage; network analysis; knowledge graphs; recommendation systems Q3 tableau 5 240.90 Data visualization tool; dashboard platform satire 7 240.90 Comedy genre; literary exaggeration event attendance 19 240.95 Registration; attendee count; attendee list; eve...
https://arxiv.org/abs/2505.18557v1
From Word to World: Evaluate and Mitigate Culture Bias via Word Association Test Xunlian Dai1,2, Li Zhou1*, Benyou Wang1, Haizhou Li1,3 1The Chinese University of Hong Kong, Shenzhen 2Wuhan University of Technology 3Shenzhen Research Institute of Big Data, lizhou21@cuhk.edu.cn Abstract The human-centered word associati...
https://arxiv.org/abs/2505.18562v1
May 2025 •We design an LLM-adaptive, free-relation word association task and a quantitative eval- uation metric to assess the cross-cultural cog- nitive abilities of LLMs (§3). •We introduce CultureSteer , an innovative ap- proach that integrates a culture-aware steering mechanism to guide semantic representations towa...
https://arxiv.org/abs/2505.18562v1
flexible and diverse associations that more accurately reflect human cognitive pro- cesses.1Before conducting the WAT, demographic information, such as age, gender, education, na- tive language, and region, is collected for each participant. Given a cue word w, the participant uis asked to provide the first word that c...
https://arxiv.org/abs/2505.18562v1
a ranking process close to full ranking, we adopt R@K as the baseline eval- uation metric, emphasizing full coverage. Formally, it is defined as: R@K =PK i=1I(ˆai∈Ac w) N =PN i=1I ai∈Top-K ˆAc w N(2) where N=|Ac w|denotes the total number of truly relevant associative words, and I(·)is the indicator function. Propo...
https://arxiv.org/abs/2505.18562v1
is then computed as: ˜h(c)=h+ϵ△h(c)(7) where ϵis a scaling factor that modulates the in- fluence of the culture-specific adjustment.3The adjusted representation ˜h(c)is then used to predict the associated words for a given query, considering the specified cultural context. Training and Inference Process During train- i...
https://arxiv.org/abs/2505.18562v1
As a result, we construct datasets representing four distinct cultural groups: USA, UK, OC, and CN. Data Preprocessing To ensure fairness in evalu- ating word association perception across different cultures, we utilize the Intercontinental Dictionary Series (IDS) (jung Yu and Wang, 2023) to align the 4Designed as a hu...
https://arxiv.org/abs/2505.18562v1
toward American cul- ture, aligning with findings from prior studies (Cao et al., 2023b; Myung et al., 2024; Zhou et al., 2025). Overall, Llama demonstrates stronger human-like word-level cognitive abilities compared to Qwen, primarily due to its significantly better performance in English-speaking cultural contexts. I...
https://arxiv.org/abs/2505.18562v1
26.12 26.12 26.12 26.12 27.26 CN 32.09 32.09 32.09 32.09 33.53 QwenUSA 56.19 61.88 56.81 56.83 56.86 QwenUSA 68.34 73.94 68.82 68.92 68.89 UK 43.61 43.92 46.45 44.02 43.87 UK 54.58 54.78 56.66 54.74 54.78 OC 41.19 41.28 41.14 41.55 41.11 OC 50.73 50.70 50.73 52.22 50.85 CN 33.53 33.01 32.62 32.73 44.88 CN 44.34 44.51 4...
https://arxiv.org/abs/2505.18562v1
broader cultural competence. By grounding conceptual mappings in cultural con- texts through word association mechanisms, Cul- tureSteer activates latent cultural perception capa- bilities in language models, surpassing superficial 7 Llama Qwen Baseline CultureSteer Baseline CultureSteer WVS USA 53.64 51.11 47.66 51.43...
https://arxiv.org/abs/2505.18562v1
potential relationships and interactions between them. Fu- ture work could address these limitations by devel- oping scalable methods for collecting word asso- ciation data, expanding the linguistic and cultural scope, and incorporating inter-cultural connections to provide a more comprehensive and nuanced per- spectiv...
https://arxiv.org/abs/2505.18562v1
Carol Chen, Zac Hatfield-Dodds, Danny Hernandez, Nicholas Joseph, et al. 2023. Towards measuring the representation of subjective global opinions in language models. arXiv preprint arXiv:2306.16388 . Constanza Fierro, Nicolas Garneau, Emanuele Bugliarello, Yova Kementchedjhieva, and Anders Søgaard. 2024. MuLan: A study...
https://arxiv.org/abs/2505.18562v1
on hofstede’s cultural dimensions. arXiv preprint arXiv:2309.12342 . Junho Myung, Nayeon Lee, Yi Zhou, Jiho Jin, Rifki Putri, Dimosthenis Antypas, Hsuvas Borkakoty, Eu- nsu Kim, Carla Perez-Almendros, Abinew Ali Ayele, et al. 2024. Blend: A benchmark for llms on ev- eryday knowledge in diverse cultures and languages. A...
https://arxiv.org/abs/2505.18562v1
Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages 6859–6877, Albuquerque, New Mexico. Asso- ciation for Computational Linguistics. Jing Yao, Xiaoyuan Yi, Jindong Wang, Zhicheng Dou, and Xing Xie. 2025. Caredio: Cultural alignmen...
https://arxiv.org/abs/2505.18562v1
370 1455 Sense 34 15.38 1975 292 426 1337 Values 42 13.17 2243 346 319 1670 Cognit 38 13.45 2299 277 323 1332 Language 31 14.13 1978 277 271 1073 Society 27 14.19 1353 229 183 957 Warfare 26 14.46 1428 277 158 940 Law 16 15.19 1078 135 135 682 Religion 11 14.36 592 104 72 440 Overall 881 13.94 51700 8097 7309 32406 Tab...
https://arxiv.org/abs/2505.18562v1
Time Sense Values CognitLanguageSocietyWarfareLawReligion020406080100 Llama_CultureSteer WorldKinshipAnimalsBodyFoodCloth House Agriculture Action Motion Possession Space Quantity Time Sense Values CognitLanguageSocietyWarfareLawReligion020406080100 Qwen_CultureSteer(a) Performance in PWR@3 WorldKinshipAnimalsBodyFoodC...
https://arxiv.org/abs/2505.18562v1
arXiv:2505.18573v1 [cs.LG] 24 May 2025Enhancing Efficiency and Exploration in Reinforcement Learning for LLMs Mengqi Liao1,2, Xiangyu Xi2, Ruinian Chen2, Jia Leng2, Yangen Hu2,Ke Zeng2,Shuai Liu1,Huaiyu Wan1*, 1Department of Computer Science, Beijing Jiaotong University, 2Meituan, Correspondence: hywan@bjtu.edu.cn Abst...
https://arxiv.org/abs/2505.18573v1
mitigate this issue, our experiments reveal that combining entropy regularization (Mnih et al., 2016; Williams, 1992)—the widely adopted ap- proach for fostering exploration in deep reinforce- ment learning—with sparse rule-based rewards can degrade performance, especially when training with challenging questions, and ...
https://arxiv.org/abs/2505.18573v1
exploitation trade-offs. Moreover, Entropy Regu- larization, a common technique in deep reinforce- ment learning, adds an entropy term to the opti- mization objective to encourage stochastic policies and broader exploration (Williams, 1992). 3 Preliminary 3.1 Group Relative Policy Optimization (GRPO) We utilize the GRP...
https://arxiv.org/abs/2505.18573v1
. . . , B }, initialize G(i)=Gmin 3:Remaining rollouts budget Nrem=Ntotal− B×Gmin 4:Foriin{1,2, . . . , B },G(i)=G(i)+⌊Nrem× k(i)PB j=1k(j)⌋ 5:Nrem=Ntotal−PB i=1G(i) 6:Distribute Nremgreedily based on descending order of ki, respecting Gmaxfor each G(i) 7:return {G(1), . . . , G (B)} creased, and Gminis progressively d...
https://arxiv.org/abs/2505.18573v1
Scheduler. As discussed, although entropy regularization helps maintain the entropy of the policy, it may inadvertently introduce harm- ful gradients. To address this, we propose a tem- perature scheduler that adaptively adjusts the tem- perature τof the softmax distribution to maintain policy entropy, ensuring stable ...
https://arxiv.org/abs/2505.18573v1
is 64, with the default number of rollouts per ques- tion ( G) set to 8. The sampling temperature is 1, and the maximum response length is 6k. Training is performed over 3 epochs on the 10k dataset, total- ing 480 steps. We use DeepSeek-R1-Distill-Qwen 1.5B and 7B (Guo et al., 2025) as base models. For the 1.5B model, ...
https://arxiv.org/abs/2505.18573v1
300 400 500 Step1.001.051.101.151.201.251.30T emperatureTS TS & ANFigure 4: The temperature variation during training is presented for cases utilizing only the temperature scheduler and for those combining the scheduler with annealing. ing instability for η= 0.8andη= 0.85. Conse- quently, ηis set to 0.9, with the annea...
https://arxiv.org/abs/2505.18573v1
during the annealing phase. Temperature Scheduler Stabilizes LLM Perfor- mance Improvements. Figure 3 shows the varia- tion in validation accuracy during training. Both GRPO alone and GRPO with entropy regulariza- tion exhibit significant variance. In contrast, train- ing with the temperature scheduler achieves lower v...
https://arxiv.org/abs/2505.18573v1
process. allocation. The Impact of Dynamic Rollout Budget Alloca- tion on Performance. As shown in Table 3, the performance of the model deteriorates significantly on the most challenging AIME benchmark when dynamic rollout budgeting is not employed. The pass@1 scores on the AIME benchmark decreased by 3.02% and 1.59% ...
https://arxiv.org/abs/2505.18573v1
Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, and 1 others. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374 . Bofei Gao, Feifan Song, Zhe Yang, Zefan Cai, Yibo Miao, Chenghao Ma, Shanghaoran Qua...
https://arxiv.org/abs/2505.18573v1
and Chuan Wu. 2024. Hybridflow: A flex- ible and efficient rlhf framework. arXiv preprint arXiv:2409.19256 . Richard S Sutton, Andrew G Barto, and 1 others. 1998. Reinforcement learning: An introduction , volume 1. MIT press Cambridge. Chenxia Tang, Jianchun Liu, Hongli Xu, and Liusheng Huang. 2024. Top- nσ: Not all lo...
https://arxiv.org/abs/2505.18573v1
few outlier tokens. For sim- plicity, the analysis focuses on the logits associated with the most informative token, specifically con- sidering only zmax.zmaxneeds to be significantly larger than the logits in the noisy region to achieve a low entropy. We further assume that the difference between zmaxand the logits in...
https://arxiv.org/abs/2505.18573v1
cto ln(N)does not significantly change the logarithm. For the distribution of the next token generated from LLMs, the magnitude of εis typically on the order of 10−1. Thus, we can further neglect the ln(ε)term, simplifying the expression. So ccan be approximated as: c≈ln(ln( N)). Therefore, ∆≈ln(N) + ln(ln( N)). Substi...
https://arxiv.org/abs/2505.18573v1
arXiv:2505.18581v1 [cs.CL] 24 May 2025Removal of Hallucination on Hallucination: Debate-Augmented RAG Wentao Hu♠Wengyu Zhang♠Yiyang Jiang♠ Chen Jason Zhang♠Xiaoyong Wei♡,♠,*Qing Li♠ ♠The Hong Kong Polytechnic University♡Sichuan University wayne-wt.hu@connect.polyu.hk {jason-c.zhang, cs007.wei}@polyu.edu.hk Abstract Ret...
https://arxiv.org/abs/2505.18581v1
response. reduces reliance on the model’s parametric knowl- edge, thereby aiming to improve factual correct- ness. However, as shown in Figure 1, RAG in- troduces a new challenge: Biased or erroneous retrieval results can mislead the generation, com- pounding the hallucination problem rather than solving it . In other ...
https://arxiv.org/abs/2505.18581v1
Hallucination on Hallucina- tion. •We introduce DRAG, a novel training-free framework that integrates Multi-Agent Debate (MAD) mechanisms into both retrieval and generation phases of RAG. •We evaluate DRAG on multiple tasks, demon- strating its effectiveness in improving retrieval reliability, reasoning robustness, and...
https://arxiv.org/abs/2505.18581v1