text string | source string |
|---|---|
reliable: 0.363 suitable: 0.322 talents: 0.520 achieve: 0.494 eager: 0.343 prepared: 0.315 commanding: 0.516 precise: 0.471 fully: 0.342 accuracy: 0.253 bright: 0.479 preparedness: 0.442 highly: 0.322 talent: 0.253 thoroughly: 0.471 career: 0.439 automatic: 0.286 precise: 0.236 prowess: 0.455 powerful: 0.368 poised: 0.... | https://arxiv.org/abs/2505.20901v1 |
skilled: 0.324 expertise: 0.583 skills: 0.376 capability: 0.509 precise: 0.320 artistic: 0.476 strong: 0.376 sophisticated: 0.484 technological: 0.261 precise: 0.476 prowess: 0.376 expertise: 0.474 role: 0.250 intently: 0.402 accuracy: 0.351 enhanced: 0.402 professional: 0.197 cleaner: 0.349 precise: 0.351 intellectual... | https://arxiv.org/abs/2505.20901v1 |
0.616 accomplished: 0.866 expertise: 0.649 handworking: 0.631 sophisticated: 0.598 poised: 0.861 talent: 0.610 competence: 0.628 indiscernible: 0.598 skillfully: 0.822 sophisticated: 0.508 skillfully: 0.595 elegant: 0.582 skillful: 0.822 determined: 0.501 elegant: 0.568 bright: 0.563 capable: 0.819 enthusiasm: 0.491 co... | https://arxiv.org/abs/2505.20901v1 |
precise: 0.400 powerful: 0.434 characterized: 0.552 highly: 0.523 difficult: 0.329 intellectual: 0.396 quality: 0.467 consistent: 0.424 secured: 0.321 suitable: 0.332 preparedness: 0.377 words: 0.371 bright: 0.312 prepared: 0.306 enthusiastic: 0.349 enthusiastic: 0.363 advanced: 0.304 performance: 0.271 precise: 0.324 ... | https://arxiv.org/abs/2505.20901v1 |
intently: 0.656 strength: 0.333 accurate: 0.791 knowledge: 0.451 accurate: 0.589 prepared: 0.326 determined: 0.773 quiet: 0.444 diligently: 0.545 determination: 0.264 enhance: 0.642 learning: 0.436 beauty: 0.461 profession: 0.226 progress: 0.630 role: 0.416 knowledgeable: 0.267 job: 0.139 flair: 0.568 intellectual: 0.3... | https://arxiv.org/abs/2505.20901v1 |
0.650 fully: 0.547 confidence: 0.587 flair: 1.001 secured: 0.500 quality: 0.541 words: 0.587 strong: 0.967 prepared: 0.411 suitable: 0.527 knowledge: 0.528 competitive: 0.934 effort: 0.324 intellectual: 0.454 competence: 0.371 satisfaction: 0.660 profession: 0.307 strength: 0.382 intently: 0.340 skills: 0.660 thoughtfu... | https://arxiv.org/abs/2505.20901v1 |
insights: 1.433 abilities: 1.739 triumphant: 1.104 abilities: 1.083 highly: 1.308 talent: 0.680 rugged: 1.016 committed: 0.946 thoroughly: 1.130 equipped: 0.441 meticulously: 0.625 enthusiasm: 0.579 effort: 1.020 strong: 0.390 committed: 0.612 strong: 0.563 bright: 0.908 cleaner: 0.372 capability: 0.593 intellectual: 0... | https://arxiv.org/abs/2505.20901v1 |
Qwen 2.5 72B solid: 1.442 heated: 0.733 hot: 1.089 solid: 1.083 heated: 1.094 respect: 0.581 respect: 1.014 pleasant: 0.960 enjoying: 0.855 friendliness: 0.581 solid: 0.767 respect: 0.919 comfort: 0.480 enjoying: 0.329 active: 0.745 enjoying: 0.767 pleasant: 0.464 suitable: 0.287 gentle: 0.382 heated: 0.597 gentle: 0.4... | https://arxiv.org/abs/2505.20901v1 |
Table 20: Top 20 warmth-related words associated with ‘red’ Llama 3.2 11B Llama 3.2 90B InstructBLIP 7B InstructBLIP 13B pure: 1.603 fluffy: 0.874 harmonious: 0.997 friendly: 1.005 safely: 0.987 tasting: 0.645 taste: 0.582 bright: 0.980 skillfully: 0.925 fine: 0.289 skillful: 0.389 comfortable: 0.565 attractive: 0.796 ... | https://arxiv.org/abs/2505.20901v1 |
0.083 authenticity: 0.290 lively: 0.094 enthusiastic: 0.125 safe: 0.007 welcoming: 0.286 gentle: 0.085 enjoying: 0.112 neutral: -0.106 Pixtral 12B Pixtral Large Qwen 2.5 7B Qwen 2.5 72B freshness: 0.774 friendliness: 0.806 lively: 0.731 slicked: 0.990 good: 0.595 slicked: 0.624 prayerful: 0.579 enthusiastic: 0.797 good... | https://arxiv.org/abs/2505.20901v1 |
natural: 0.146 enjoying: 0.216 bright: 0.091 gentle: -0.034 solid: 0.129 warmth: 0.182 neutral: 0.076 stylish: -0.035 stylish: 0.129 stylish: 0.178 friendly: 0.035 fresh: -0.045 Table 23: Top 20 warmth-related words associated with ‘woman’ Llama 3.2 11B Llama 3.2 90B InstructBLIP 7B InstructBLIP 13B admiration: 0.929 f... | https://arxiv.org/abs/2505.20901v1 |
trust: 0.346 fresh: -0.087 passionate: 0.360 strong: 0.342 warmth: -0.095 knowledgeable: 0.310 knowledgeable: 0.339 friendliness: -0.099 quality: 0.308 Pixtral 12B Pixtral Large Qwen 2.5 7B Qwen 2.5 72B solid: 1.720 friendliness: 0.925 stylish: 0.753 strong: 1.033 respect: 1.272 strong: 0.878 thoughtful: 0.718 warm: 0.... | https://arxiv.org/abs/2505.20901v1 |
-0.136 cheerful: -1.088 Table 26: Top 20 warmth-related words associated with ‘Indian’ Llama 3.2 11B Llama 3.2 90B InstructBLIP 7B InstructBLIP 13B slicked: 1.805 slicked: 1.277 pleasant: 1.104 suitable: 0.696 welcoming: 0.694 knowledgeable: 1.037 warm: 0.926 delicious: 0.651 friendly: 0.535 welcoming: 0.919 talented: ... | https://arxiv.org/abs/2505.20901v1 |
2.5 7B Qwen 2.5 72B respectful: 1.090 secured: 0.932 tasting: 1.218 secured: 1.231 gentle: 1.008 prayerful: 0.761 good: 0.902 pleasant: 1.121 secured: 0.679 fine: 0.699 thoughtful: 0.178 enjoying: 1.066 good: 0.553 secure: 0.375 kind: 0.174 welcoming: 0.869 enjoying: 0.515 solid: 0.317 fresh: 0.147 comfortable: 0.668 w... | https://arxiv.org/abs/2505.20901v1 |
Towards Objective Fine-tuning: How LLMs’ Prior Knowledge Causes Potential Poor Calibration? Ziming Wang1*, Zeyu Shi1*, Haoyi Zhou2,3†, Shiqi Gao1, Qingyun Sun1,Jianxin Li1,3 1SKLCCSE, School of Computer Science and Engineering, Beihang University 2School of Software, Beihang University 3Zhongguancun Laboratory, Beijing... | https://arxiv.org/abs/2505.20903v1 |
been shown as a critical factor af- fecting model adaptation (Gekhman et al., 2024; Kung et al., 2023; Huang et al., 2024; Seedat et al., 2023; Chen et al., 2024). Therefore, we try to ex- tend previous research by investigating the underly- ing mechanisms of poor calibration specifically in the context of fine-tuning,... | https://arxiv.org/abs/2505.20903v1 |
experiments on domain- specific multiple-choice and open-ended QA tasks with multiple models, using different fine-tuning methods, which demonstrate the effectiveness and generality of CogCalib in enhancing calibration. 2 Related Works 2.1 Confidence Calibration Confidence calibration methods can be categorized into th... | https://arxiv.org/abs/2505.20903v1 |
six ratios of un- known to known data in OBQA. While Figure 2 reveals irregular performance trends across knowl- edge bias levels, the calibration exhibits a clear directional pattern: lower knowledge bias consis- tently degrades calibration, whereas higher bias improves it , a phenomenon persistent across both in-doma... | https://arxiv.org/abs/2505.20903v1 |
bias adjustment fails to achieve con- sistent calibration improvements across all tasks. moving 25% of low-bias data improves calibration in ARC-C but degrades it in OBQA. This inconsis- tency may stem from the inherent characteristics of different datasets, making it challenging to find a universal optimal adjustment ... | https://arxiv.org/abs/2505.20903v1 |
the discrepancy in LLMs’ linguistic style and label formats prevents NLL from accu- rately assessing knowledge bias. Thus, the model requires a style adaptation process for calculating the initial threshold t0, based on findings (Zhang and Wu, 2024; Mai et al., 2024) that LLMs rapidly adapt to downstream task syntax du... | https://arxiv.org/abs/2505.20903v1 |
assess calibration. See Appendix E for more details of ECE. Baselines. We consider 4 baseline methods: (1) Vanilla SFT : We use standard LoRA or FFT as a lower performance bound. (2) MC-Dropout (MCD) (Gal and Ghahramani, 2016): We use a dropout rate of 0.02 during fine-tuning and per- form sampling 4 times. (3) Deep En... | https://arxiv.org/abs/2505.20903v1 |
(-9.0) 8.20 (-7.9) 9.48 (-7.2) 9.83 (-7.6) 14.41 (-6.9) CoECP (∆TS) 7.30 (-2.6) 2.80 (-13.1) 4.90 (-5.5) 3.80 (-12.3) 3.46 (-13.2) 6.27 (-11.1) 9.51 (-11.8) ACC↑Vanilla SFT 84.80 79.10 84.10 79.20 79.52 77.63 73.47 MCD 83.60 78.92 84.22 80.78 79.22 76.24 73.38 Ensemble 88.00 79.35 87.37 80.32 79.52 78.39 75.11 TS 84.80... | https://arxiv.org/abs/2505.20903v1 |
which uniformly applies calibration loss to all data. (2) Random calibration, which randomly distinguishes between known and unknown data while maintain- ing a consistent number of known samples. As shown in Figure 7, in these tasks, our cogni- tive methods achieved optimal results in both fine- tuning performance and ... | https://arxiv.org/abs/2505.20903v1 |
on calibration while proposing a real-time calibration framework. All models and datasets utilized in this research are publicly available and have been widely adopted by the research community. The experimental results presented herein have been rigorously validated for accuracy and reproducibil- ity. Based on these c... | https://arxiv.org/abs/2505.20903v1 |
Downey, and Noah A Smith. 2020. Don’t stop pretraining: Adapt language models to domains and tasks. arXiv preprint arXiv:2004.10964 . Tianxing He, Jun Liu, Kyunghyun Cho, Myle Ott, Bing Liu, James Glass, and Fuchun Peng. 2021. Analyzing the forgetting problem in pretrain-finetuning of open- domain dialogue response mod... | https://arxiv.org/abs/2505.20903v1 |
Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavat- ula, and Yejin Choi. 2021. Winogrande: An adver- sarial winograd schema challenge at scale. Commu- nications of the ACM , 64(9):99–106. Sarah Sarabadani. 2019. Detection of adverse drug reac- tion mentions in tweets using elmo. In Proceedings of the Fourth Social Medi... | https://arxiv.org/abs/2505.20903v1 |
multi-stream network for retrosynthesis prediction. Frontiers of Computer Science , 18(2):182906. Xiao Zhang and Ji Wu. 2024. Dissecting learning and forgetting in language model finetuning. In The Twelfth International Conference on Learning Repre- sentations . Chiwei Zhu, Benfeng Xu, Quan Wang, Yongdong Zhang, and Zh... | https://arxiv.org/abs/2505.20903v1 |
us as explained in Appendix A. Both figures clearly demonstrate that calibration performance deteriorates significantly as bias decreases. These findings further corroborate our conclusion from Section 3.1, supporting the principle of "minimal bias, maximal overconfidence". Furthermore, addi- tional experiments are con... | https://arxiv.org/abs/2505.20903v1 |
fidence and improving calibration performance indeep neural networks, as follows: LI.S=−X k((1−ϵ)qk+ϵ K) logpk,(4) where ϵis label smoothing factor, Kdenotes the numbers of total classes, pis the softmax proba- bility predictions by model, which is computed as follows: p= (pk)1≤k≤K∈RK;pk=explk PK jexplj,(5) where l= (l... | https://arxiv.org/abs/2505.20903v1 |
of Style Adaptation As shown in Table 6, the AUROC scores for t0dis- crimination demonstrate significant improvement with style adaptation. This improvement can be attributed to better alignment between the model’s output style and fine-tuning data patterns after style adaptation, which enhances the model’s capability ... | https://arxiv.org/abs/2505.20903v1 |
It’s worth noting that for a fairer comparison of Metric Methods Physics Economics Health Law OBQA ARC-C ECE↓Vanilla SFT 22.87 18.44 20.00 28.82 12.71 15.87 MCD 18.15 15.44 15.94 24.88 6.93 10.72 Ensemble 20.43 15.52 17.68 25.45 7.49 10.11 TS 25.22 28.53 26.17 21.39 40.71 37.89 CoLS 11.51 7.88 8.89 15.14 5.39 4.24 CoMb... | https://arxiv.org/abs/2505.20903v1 |
7.55 6.10 2.62 ARC-CACC↑ 70.90 70.73 72.10 70.90 71.16 71.50 70.56 ECE↓ 25.98 22.14 18.21 24.70 13.34 13.36 14.43 WG-SACC↑ 74.59 74.35 74.98 74.59 73.56 72.45 72.63 ECE↓ 16.96 15.73 15.95 14.90 7.99 8.34 16.00 WG-MACC↑ 82.08 81.61 82.40 82.08 80.82 81.22 81.06 ECE↓ 16.63 14.58 12.43 15.60 7.44 7.17 6.71 BoolQACC↑ 89.85... | https://arxiv.org/abs/2505.20903v1 |
is presented. Results are evaluated on Qwen2.5-7B model which is fine-tuned on the OBQA dataset. ble 11 and Table 12. Although Qwen demon- strated superior accuracy across all datasets and inherently low ECE compared to other LLMs, our CogCalib framework still achieved significant cal- ibration improvements over the ba... | https://arxiv.org/abs/2505.20903v1 |
65.25 Table 14: Comparison of our method’s performance against baselines on distribution shift datasets is presented. Results are evaluated on Mistral-7B model which is fine-tuned on the OBQA dataset. Dataset Metric Vanilla SFT MCD TS CoLS CoMbLS CoECP ARC-CACC↑ 66.81 65.96 66.81 70.82 72.18 70.22 ECE↓ 29.84 28.31 28.6... | https://arxiv.org/abs/2505.20903v1 |
to misclassify unknown samples, conse- quently applying calibration to these samples as well, which prevents LLM from effectively learn- ing critical knowledge. In detail, Table 17 presents the classification re- sults for both known and unknown data during the fine-tuning under different threshold calculation methods.... | https://arxiv.org/abs/2505.20903v1 |
Margin=5 on the ARC-C and WG-S datasets.G Implementation Details In this section, we present a detailed analysis of the SliCK method, the implementation of Temperature Scaling for both open-ended and multiple-choice data, along with our specific hyperparameter con- figurations. G.1 Details of SliCK In Section 3, we emp... | https://arxiv.org/abs/2505.20903v1 |
10 ECP β 0.15 Table 23: Calibration term’s hyperparameters for open- ended QA task. In our experimental setup, we fine-tune the LLM using both LoRA and FFT approaches. For LoRA implementation, we incorporate LoRA adapters into all linear layers of the LLM, maintaining the default PEFT configurations from Huggingface, a... | https://arxiv.org/abs/2505.20903v1 |
Automated Privacy Information Annotation in Large Language Model Interactions Hang Zeng1, Xiangyu Liu2, Yong Hu2, Chaoyue Niu1∗ Fan Wu1, Shaojie Tang3, Guihai Chen1 1Shanghai Jiao Tong University,2WeChat AI Tencent,3University at Buffalo {nidhogg, rvince}@sjtu.edu.cn Abstract Users interacting with large language model... | https://arxiv.org/abs/2505.20910v1 |
the query responsible for the privacy leakage need to be annotated. To further assist users who may still be uncertain, a concise explanation of what private information will be exposed by each annotated phrase should be provided. These notifications can help users quickly understand the nature of privacy leakage and t... | https://arxiv.org/abs/2505.20910v1 |
fine-tuning baselines. Benchmark results show that the fine-tuned 1B model with our dataset outperforms the directly prompted 72B model. We summarize the key contributions as follows: (1) We identify a new application requirement of privacy detection for real-name user interactions with LLMs; (2) we build the automated... | https://arxiv.org/abs/2505.20910v1 |
ALLTAG [ 9] targets privacy entity recognition in German emails, TAB [ 31] annotates privacy entities in court judgments, and JobStack [ 20] specializes in privacy entity annotation in job postings. Additionally, these datasets face challenges of data volume and they are manually labeled. In contrast, our dataset focus... | https://arxiv.org/abs/2505.20910v1 |
D. The goal is to extract the privacy phrase Pq={p1, ..., p m}and summarize specific privacy information Iq={I1, ..., I m}, where mis the number of the extracted phrases, pjis defined as a segment of qthat discloses private or sensitive information about the user, and Ijis a statement summarizing the specific details r... | https://arxiv.org/abs/2505.20910v1 |
GPT-4 to directly identify privacy phrases and the corresponding categories. We also provide an example to guide GPT-4. These results are then merged to create an extensive set of privacy categories. However, since extraction for each sample is independent, some categories may be expressed differently but hold the same... | https://arxiv.org/abs/2505.20910v1 |
rules: (1) the privacy phrase must be directly associated with the user or their close associates, and (2) the phrase must have an explicit reference, rather than be a general phrase such as “go a place” or “have ideas”. We use GPT-4 to evaluate these criteria and remove any phrases that do not meet the rules. The resu... | https://arxiv.org/abs/2505.20910v1 |
dataset as a case study, Figure 4 shows the relative frequency of top 20 different privacy categories. We can find that in open-domain interaction dialogues, information such as users’ status, opinions, preferences, and other potentially sensitive details is more likely to be inadvertently revealed, whereas categories ... | https://arxiv.org/abs/2505.20910v1 |
extraction task is to extract phrases based on the context of query that results in user’s privacy leakage. For each data sample, we focus on both missed and incorrect extractions. For missed extractions, we define the recall score of sample qas the ratio of the intersection between the predicted phrases and the ground... | https://arxiv.org/abs/2505.20910v1 |
detailed training and test set statistics are shown in Appendix E. During the training process, we balance the number of samples with and without privacy leakage in the training data, ensuring a ratio of approximately 1:1. Models. We take the instruction-tuned version of different series of language models with various... | https://arxiv.org/abs/2505.20910v1 |
further insert five examples from training data based on the template for inference with local privacy detection LLM. After the examples, the current user query is added for privacy detection. All corresponding prompts are provided in Appendix G. Tuning-based Baseline. We use Supervised Fine-Tuning (SFT) with instructi... | https://arxiv.org/abs/2505.20910v1 |
Case Study Table 4: Case study with Qwen2.5-7B-Instruct. Query :( I am worried I will screw up the computer because i am bad on coding. ZGPhrase 1: I am bad on coding Info 1: The user is worried about their coding abilities. ICLPhrase 1: bad on coding Info 1: The user expresses concern about their coding abilities and ... | https://arxiv.org/abs/2505.20910v1 |
Dario Amodei. Language models are few-shot learners. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtua... | https://arxiv.org/abs/2505.20910v1 |
IEEE, 2021. [15] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net, 2022. 10 [1... | https://arxiv.org/abs/2505.20910v1 |
Muresan, Preslav Nakov, and Aline Villavicencio, editors, Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022 , pages 8424–8445. Association for Computational Linguistics, 2022. [26] Marvin Li, Jason Wang, Jeffrey Wa... | https://arxiv.org/abs/2505.20910v1 |
Smaranda Muresan, Preslav Nakov, and Aline Villavicencio, editors, Findings of the Association for Computational Linguistics: ACL 2022, Dublin, Ireland, May 22-27, 2022 , pages 1035–1047. Association for Computational Linguistics, 2022. [34] Nafis Sadeq, Zhouhang Xie, Byungkyu Kang, Prarit Lamba, Xiang Gao, and Julian ... | https://arxiv.org/abs/2505.20910v1 |
Kai Rannenberg. Privacybot: Detecting privacy sensitive information in unstructured texts. In Mohammad A. Alsmirat and Yaser Jararweh, editors, Sixth Interna- tional Conference on Social Networks Analysis, Management and Security, SNAMS 2019, Granada, Spain, October 22-25, 2019 , pages 53–60. IEEE, 2019. [47] Manuel To... | https://arxiv.org/abs/2505.20910v1 |
Duan, Kaidi Xu, Yuanfang Cai, Eric Sun, and Yue Zhang. A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly. CoRR , abs/2312.02003, 2023. [55] Da Yu, Saurabh Naik, Arturs Backurs, Sivakanth Gopi, Huseyin A. Inan, Gautam Kamath, Janardhan Kulkarni, Yin Tat Lee, Andre Manoel, Lukas... | https://arxiv.org/abs/2505.20910v1 |
and examples of the constructed dataset. One data sample is consists of a user-initiated query and a response from the assistant. Based on this, we add privacy detection results to each query, in the form of a list of privacy. Each element in the list consists of a phrase extracted from the query and the corresponding ... | https://arxiv.org/abs/2505.20910v1 |
minimum of 1e-4. Training is conducted with a batch size of 4, for 10 epochs, and a maximum token length of 1024. For LoRA, we configure the rank to 16, apply a dropout rate of 0.05, and perform weight decomposition on the query, key, and value matrices, as well as the dense MLP layers in the transformer blocks. Additi... | https://arxiv.org/abs/2505.20910v1 |
both syllables of words. Chinese samples. The model precision is set to FP16 and the results is shown in Table 7 and Table 8. These results reveal the time efficiency of local privacy detection with billion-scale LLMs. K Case Study Current local privacy detection methods also have some limitations. Table 9 presents a b... | https://arxiv.org/abs/2505.20910v1 |
Output: “‘json { "reason": "This query exposes the activity of the user and people related to the user intend to do.", "judgment": true } ”’ ...... (other examples) Prompts for privacy leakage classification (Chinese). 用户向你提出一个请求 "<|INPUT|>" 在用户提出请求之前，你只知道用户的真实的唯一标识符，并且不知道用户的任何其他的隐私敏感 信息。 在收到带有用户真实的唯一标识符的请求后，请你进行逐步的思考（... | https://arxiv.org/abs/2505.20910v1 |
appropriate privacy category in the following pre-defined list, and also note if the extracted phrase is not matched with any given privacy category, it should be excluded. Privacy Categories: <|CATEGORIES|> The result output should be in json format with the following format: { "phrase 1" : "privacy category 3", "phra... | https://arxiv.org/abs/2505.20910v1 |
"We plan to play football with Ross." Phrase: "play football with Ross" JSON Output: “‘json { "reason" : "Playing football is the user’s plan that related to the user. Ross is also related to the user.", "judgment" : true } ”’ ...... (other examples) Prompts for filtering privacy phrase with rule 1 (Chinese). 用户向你提出一个请... | https://arxiv.org/abs/2505.20910v1 |
）。 并总结短语泄露的具体隐私敏感信息。 然后，输出如下格式的JSON： “‘json { "phrase":短语, "privacy information": 隐私信息 } ”’ 以下是一些示例： 用户的请求："我正在准备回家，地址是上海市闵行区东川路800号。" 短语："上海市闵行区东川路800号" JSON输出： “‘json { "phrase": " 上海市闵行区东川路800号" , "privacy information": " 用户的家庭地址是上海市闵行区东川路800号。" } ”’ ...... (other examples) 25 Table 18: Data samples with privacy phr... | https://arxiv.org/abs/2505.20910v1 |
general term. You should think step-by-step. Finally, output your answer in JSON. If there is no phrase extracted, the output JSON list should be empty. Here is the format: “‘json ["phrase1", "phrase2", ...] “‘ The user’s QUERY is "<|QUERY|>" Prompts for zero-shot generation (Chinese). 提取用户请求中，揭示了用户、用户相关人员或事物的隐私敏感信息的短语... | https://arxiv.org/abs/2505.20910v1 |
<|CASE|>End of examples. The user’s current Query is "<|QUERY|>" Prompts for in-context learning (Chinese). 提取用户请求中，揭示了用户、用户相关人员或事物的隐私敏感信息的短语。 隐私敏感信息包括用户的观点、偏好、活动、意图或根据请求上下文可能被认为是私人或敏感的信 息。 提取的短语应当在请求的上下文中体现出与用户或用户相关的人员或事物的相关性。 隐私短语应当有具体的指代，而不是一个泛指的词汇。 请一步一步思考（Think step-by-step ）。 给出理由并在最后，以JSON格式输出答案。如果没有提取到任何短语，则输出的... | https://arxiv.org/abs/2505.20910v1 |
arXiv:2505.20921v1 [cs.CL] 27 May 2025Automatic Transmission for LLM Tiers: Optimizing Cost and Accuracy in Large Language Models Injae Na†, Keonwoong Noh†,and Woohwan Jung Department of Applied Artificial Intelligence, Hanyang University {suhoij47, rohgw011, whjung}@hanyang.ac.kr Abstract LLM providers typically offer... | https://arxiv.org/abs/2505.20921v1 |
to integrate new LLMs into the selection process. Third, these methods become unreliable when applied to test domains that differ from the training data distribution, limiting their generalizability in real-world applications. To address this problem, we propose the LLM Automatic Transmission ( LLM-AT ), a novel frame-... | https://arxiv.org/abs/2505.20921v1 |
to enhance the accuracy and reliability of LLM- generated responses. These approaches enable models to evaluate their own outputs, incorporate feedback, and iteratively improve their responses without additional training. Madaan et al. (2023) introduce a method where LLMs generate self- feedback on their initial respon... | https://arxiv.org/abs/2505.20921v1 |
on model- specific prompting methods. For detailed prompts, please refer to the appendix C. Our system, as explained in Section 3.1, per- forms a type of iterative refinement process, where inference is repeated within the given tier system until a valid answer is generated. Although pro- viding the output of lower-tie... | https://arxiv.org/abs/2505.20921v1 |
answers by the LLMs. Then, we present the accuracy estimation methods to calculate the accuracy of each tier. 3.4.1 Pseudo-labeling of Correctness To estimate the accuracy of each tier, we rely on pre- vious inference records and the correctness labels. Ideally, human annotated labels would be used, but they are usuall... | https://arxiv.org/abs/2505.20921v1 |
)(4) where top(q)is the set of top-k similar questions ofqandlj,q′is the correctness label of tier jforq′. 4 Experimental Settings 4.1 Datasets We use two datasets with varying difficulty of the question to evaluate whether LLM-AT effectively selects the LLM tier while optimizing trade-offs between performance, cost, a... | https://arxiv.org/abs/2505.20921v1 |
execution time is measured in minutes as the elapsed time for running inference over the entire dataset. Finally, 1The tier system is as of 2025-1-17 the performance of the model is evaluated using the accuracy metrics originally defined in the datasets that comprise MATH and MCQA. 4.5 Implementation Details of LLM-AT ... | https://arxiv.org/abs/2505.20921v1 |
results of LLM-AT show a higher accuracy than the baselines with a lower cost and execution time. LLM-AT capped at o1 with the oracle judge achieves a more efficient performance–cost trade- off than the single o1. Likewise, LLM-AT capped at o1-mini outperforms both the single and itera- tion baselines with o1-mini. The... | https://arxiv.org/abs/2505.20921v1 |
of the weaker model. Additionally, the correct and incorrect pseudo- labels for each model are determined by the evalua- tion of the judge. Based on the reliable performance of the judge, we consider that the pseudo-labels of correctness can be trustworthy, and believe that they support the robustness of LLM-AT . Furth... | https://arxiv.org/abs/2505.20921v1 |
inLLM-AT . Toward the right side of the table, the accuracy difference narrows, and in the Number Theory, GPT-4o even outperforms o1-mini. Reflect- ing this, GPT-4o is selected more frequently as the performance gap becomes smaller. This selection tendency appears to result from the starter strategy of choosing the low... | https://arxiv.org/abs/2505.20921v1 |
rises. Ad- ditionally, lower-tier models are selected more fre- quently as initial points than as final points, while the opposite is observed for higher-tier models. This supports our proposed method of selecting an appropriate initial tier and, when necessary, tran- sitioning to a higher-performing tier for the final... | https://arxiv.org/abs/2505.20921v1 |
numerical reasoning tasks. Transactions on Machine Learning Research . Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv:1803.05457v1 . DeepSeek. 2025. Deepseek into t... | https://arxiv.org/abs/2505.20921v1 |
Neural Information Processing Sys- tems. Kaushal Kumar Maurya, KV Srivatsa, and Ekaterina Kochmar. 2024. Selectllm: Query-aware efficient selection algorithm for large language models. arXiv preprint arXiv:2408.08545 . Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. Can a suit of armor conduct ele... | https://arxiv.org/abs/2505.20921v1 |
performance was used. B Additional Experimental Results B.1 Smoothing Factor Specialized for the Benchmarks Figure 6 presents the results when using bench- mark performance similar to or identical to each dataset as AccBench. The results show that using benchmark performance similar to the target task as a smoothing fa... | https://arxiv.org/abs/2505.20921v1 |
Starter & w/o Abstention 0 2 4 6 8 10 API Cost ($)868788899091 GPT-4o-miniGPT-4oo1-miniMCQA 500 550 600 650 # of Calls68707274767880AccuracyMATH LLM-AT w/o Starter w/o Starter & w/o Abstention 5 10 15 API Cost ($)68707274767880 o1-miniMATHFigure 7: Ablation study. The left graph shows the ac- curacy and number of API c... | https://arxiv.org/abs/2505.20921v1 |
the ab- stention rate also increases and the accuracy onabstained questions gradually decreases. This in- dicates that the model is not abstaining at random, but rather selectively abstaining from questions that are more difficult overall or comparatively harder within the same difficulty level. C Prompt Templates We p... | https://arxiv.org/abs/2505.20921v1 |
sympy import sqrt x1, y1, z1 = 2, 1, -4 x2, y2, z2 = 5, 8, -3 distance_squared = (x2 - x1)**2 + (y2 - y1)**2 + (z2 - z1)**2 distance = sqrt(distance_squared) answer = distance Question: How many zeros are at the end of the product 25 $ times$ 240? Code: def count_trailing_zeros_of_product(num1, num2): product = num1 * ... | https://arxiv.org/abs/2505.20921v1 |
"no" if the output and reasoning process are not logically aligned with the question. Alternatively, if an error occurs in the code execution results, always evaluate it as "no". The input format is as follows: question: The question asked to the LLM. generated output: The answer generated by the LLM in response to the... | https://arxiv.org/abs/2505.20921v1 |
arXiv:2505.20925v1 [cs.CL] 27 May 2025Multi-objective Large Language Model Alignment with Hierarchical Experts Zhuo Li1∗, Guodong Du1∗, Weiyang Guo1, Yigeng Zhou1, Xiucheng Li1, Wenya Wang2,Fangming Liu3,Yequan Wang4,Deheng Ye5,Min Zhang1,Jing Li1/envel⌢pe 1Harbin Institute of Technology, Shenzhen, China 2Nanyang Techn... | https://arxiv.org/abs/2505.20925v1 |
subprob- lems [ 67]. Each subproblem is handled by a specialized parameters, referred to as experts. These experts are each assigned to a distinct preference, focus solely on their corresponding preferences and optimize within their localized subproblem regions. This strategy avoids the pitfalls of a single monolithic ... | https://arxiv.org/abs/2505.20925v1 |
breaks down the multi-objective alignment problem into a series of single-preference subproblems, each handled by a set of specialized experts, enabling fine-grained control and full Pareto coverage. 2 Table 1: Comparison with other alignment methods. M is number of preference and N is number of objectives. Note that M... | https://arxiv.org/abs/2505.20925v1 |
extend this concept to MOA. LoraMoE [ 8], the closest work to ours, is a Mixture-of-Experts (MoE) approach that uses LoRA Adapters [ 18] as experts, integrating LLM knowledge by activating select experts via a router network. However, it requires costly training across all LoRA experts simultaneously and limits knowled... | https://arxiv.org/abs/2505.20925v1 |
diverse LLMs and alignment objectives. Consequently, we replace all linear module in the Transformer architecture with MoE-style plugin modules, incorporating the LoRA experts. When receiving user preference λ∈RN, we simply linearly combining the NLoRA experts’ outputs with weightings λ, yielding: Oλ(x) =Wprex+NX i=1λi... | https://arxiv.org/abs/2505.20925v1 |
(TCH) scalarization, optimizing for the worst-case objective: J(θ|λ) =max θmin i{λi(Ri(x, y)−z∗ i)} (6) where z∗∈RNis a reference point , and λnow encodes objective importance, and Ex∼D, y∼πθ(·|x)[Ri(x, y)]is abbreviated as Ri(θ). We solve this max-min objective via Online Mirror Descent (OMD) [35], yielding: J(θ|λ) =m... | https://arxiv.org/abs/2505.20925v1 |
used open-source reward model. Additionally, we report GPT-4-based win rates—comparative against base model—for further evaluation. Specially, for the math benchmark, we report PASS@1 accuracy, and for the over-refusal benchmark [ 7], we report safety and helpfulness ratio. For more details, refer to Appendix E. 6 HoE ... | https://arxiv.org/abs/2505.20925v1 |
0.17 0.25 0.25 RS 66.7 67.8 76.2 38.9 42.6 58.44 RIC 70 67.6 76.5 42.3 46.2 60.52 MOD 68.1 68.9 76.3 40.9 47.1 60.26 HoE(OURS ) 70 71.1 77.7 42.9 48.7 62.08 (+3.6) PREFERENCE 0.11 0.11 0.11 0.33 0.33 RS 66.4 67.5 75.8 40.5 44.3 58.9 RIC 67.7 62.4 73.9 44 49.9 59.58 MOD 67.7 68.2 75.6 42.9 48.1 60.5 HoE(OURS ) 69.8 70.8... | https://arxiv.org/abs/2505.20925v1 |
full PF coverage. This confirms its advantage in multi-objective optimization stability. 6.2 Advantages over Existing Methods While existing methods each excel in specific areas, HoE offers seven notable advantages, with quantitative comparisons provided in Tab. 3. The checklist of advantages are listed in Tab.1. 1)Lig... | https://arxiv.org/abs/2505.20925v1 |
[2]Y . Bai, A. Jones, K. Ndousse, A. Askell, A. Chen, N. DasSarma, D. Drain, S. Fort, D. Ganguli, T. Henighan, N. Joseph, S. Kadavath, J. Kernion, T. Conerly, S. E. Showk, N. Elhage, Z. Hatfield- Dodds, D. Hernandez, T. Hume, S. Johnston, S. Kravec, L. Lovitt, N. Nanda, C. Olsson, D. Amodei, T. B. Brown, J. Clark, S. M... | https://arxiv.org/abs/2505.20925v1 |
alignment for llms through multi-round red-teaming. arXiv preprint arXiv:2505.17147 , 2025. URL https://arxiv.org/abs/2505.17147 . [15] D. Guodong, J. Lee, J. Li, R. Jiang, Y . Guo, S. Yu, H. Liu, S. K. Goh, H.-K. Tang, D. He, et al. Parameter competition balancing for model merging. In Proceedings of the Thirty-eighth... | https://arxiv.org/abs/2505.20925v1 |
K. Goh, W. Wang, Y . Wang, F. Liu, H.-K. Tang, S. Alharbi, D. He, and M. Zhang. Multi-modality expansion and retention for llms through parameter merging and decoupling. arXiv preprint arXiv:2505.17110 , 2025. URL https://arxiv.org/abs/2505. 17110 . [30] K. Li, T. Zhang, and R. Wang. Deep reinforcement learning for mul... | https://arxiv.org/abs/2505.20925v1 |
alignment by interpolating weights fine-tuned on diverse rewards. InProceedings of the Annual Conference on Neural Information Processing Systems (NeurIPS) , 2023. [45] N. Rimsky, N. Gabrieli, J. Schulz, M. Tong, E. Hubinger, and A. M. Turner. Steering llama 2 via contrastive activation addition. In Proceedings of the ... | https://arxiv.org/abs/2505.20925v1 |
Processing Systems (NeurIPS) , 2023. [60] E. Yang, Z. Wang, L. Shen, S. Liu, G. Guo, X. Wang, and D. Tao. Adamerging: Adaptive model merging for multi-task learning. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR) , 2024. [61] K. Yang, Z. Liu, Q. Xie, J. Huang, T. Zhang, and S.... | https://arxiv.org/abs/2505.20925v1 |
Experts R,HoE model Θ ={} uniformly select weightings {λl}l∼[N+L+R]∼∆N fori= 1toNdo τi←extract LoRA from (θi−θpre) end for forl=NtoN+Ldo τi+l←Merging {θi}i∈[N]with weighting λl end for Θ ={τi}i=[N+L] forr=N+LtoN+L+Rdo τr←Train router experts on λrwithΘ end for insert Θ ={τi}i=[N+L+R] output Θ B Additional Results B.1 M... | https://arxiv.org/abs/2505.20925v1 |
conducted across three dimensions: helpfulness, harmlessness, and humor. We take the average win rate across these three metrics as the final result. B.2 Cost Analysis Table 3: Comparison of Training Costs, Storage Costs, and Inference Costs for various baselines, when using Llama2-7B as the base model to align on thre... | https://arxiv.org/abs/2505.20925v1 |
performance by 40%, and MOLoRA expert reducing storage by 30% while retaining performance, confirming its necessity. Q2. Could other approaches such as MOD or RS similarly use LoRA to reduce storage? One may wonder whether alternatives like MOD and RS could benefit equally from LoRA-based compression. While both method... | https://arxiv.org/abs/2505.20925v1 |
(an open-source model described in Appendix E). Next, we prune 40% of the least significant parameters across the entire model based on absolute magnitude, resulting in a sparse matrix. We then apply SVD decomposition, sorting the singular values and selecting the top 128 rank-1 matrices. To further optimize performanc... | https://arxiv.org/abs/2505.20925v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.