text string | source string |
|---|---|
Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332 , 2021. Andi Nika, Debmalya Mandal, Parameswaran Kamalaruban, Georgios Tzannetos, Goran Radanovic, and Adish Singla. Reward model learning vs. direct policy optimization: A com- parative analysis of learning from human pref... | https://arxiv.org/abs/2505.19770v1 |
Satinder Singh, and Y . Mansour. Policy gradient meth- ods for reinforcement learning with function approximation. In Neural Information Processing Systems , 1999. Gokul Swamy, Christoph Dann, Rahul Kidambi, Zhiwei Steven Wu, and Alekh Agarwal. A minimaximalist approach to reinforcement learning from human feedback. ar... | https://arxiv.org/abs/2505.19770v1 |
k-wise comparisons. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, Proceedings of the 40th International Conference on Machine Learning , volume 202 of Proceedings of Machine Learning Research , pages 43037–43067. PMLR, 23–29 Jul 2023. Banghua Zhu, Mi... | https://arxiv.org/abs/2505.19770v1 |
. . . . . 26 C.12 Omitted Calculations in Section 3 and appendix B . . . . . . . . . . . . . . . . 26 C.13 Omitted Calculations in Section 4 . . . . . . . . . . . . . . . . . . . . . . . . . 28 D Implementation Details 29 A Supplementary Related Works Reinforcement learning from human feedback (RLHF). Seminal contribut... | https://arxiv.org/abs/2505.19770v1 |
with Equation (3). And this will induce an additional coefficient σ1prϕpyq´rϕpy1qqon the data distribution µpy, y1q. And this by-product explains that PILAF sampler (a variant of online sampler, see Definition 1) is introduced to align the distorted distribution ˜µpy, y1q9µpy, y1q¨σ1prϕpyq´rϕpy1qqwithπθ‹prϕq. If the re... | https://arxiv.org/abs/2505.19770v1 |
the ground-truth reward function satisfy: r“r‹pa1q“r‹pa2qěr‹pa3q“0. 17 Assume the linear feature mapping ψ:YÑRdsatisfies: ψpa1q‰ψpa2q, ψpa3q“1 2ψpa1q`1 2ψpa2q. Define the log-linear policy class Π“ tπθ:θPRdubyπθpaq9πrefpaqexppθJψpaqq, where πref“UnifpYq. Since r‹is realizable, RLHF exactly recovers it and solves: πRLHF... | https://arxiv.org/abs/2505.19770v1 |
conclude that πRLHF“πDPO and hence VπRLHF r‹“VπDPO r‹. C.6 Proof of Proposition 8 Construction 1: VπRLHF r‹ăVπDPO r‹.We first construct an environment satisfying Condition 6 such thatVπRLHF r‹ăVπDPO r‹. Consider the same setup as in Appendix C.4, but define the policy class as Π“∆pYqztπ‹u, where π‹is the optimal policy... | https://arxiv.org/abs/2505.19770v1 |
log-linear policy class: Π“" πθ:θPR2, θJ„ 1 ´1ȷ ě20* , πθpaq9πrefpaqexppθJψpaqq, where πref“UnifpYq. The corresponding surrogate reward class is: FΠ“" ˆrθ: ˆrθpaq“βθJψpaq, θJ„ 1 ´1ȷ ě20* . Now define a strictly larger reward model class: F“FΠYt¯ru,where ¯rpa1q“¯rpa2q“2,¯rpa3q“0. We set the regularization parameter to β... | https://arxiv.org/abs/2505.19770v1 |
5, online DPO can produce a solution πonline DPO, such that VπRLHF r‹ăVπonline DPO r‹. 4 3 2 1 0 1 2 3 4 Parameter: x() 4 2 0246810Numerical Simulation Value function Gradient of DPO loss Gradient of online DPO loss Gradient of RL loss 4 3 2 1 0 1 2 3 4 Parameter: x() 5 05101520Numerical Simulation Value function Gradi... | https://arxiv.org/abs/2505.19770v1 |
Theorem 10. We first prepare the settings and assumptions. For the feature mapping, we set ψpy, ωq “ ψpωq. For DTSP task, we first randomly collect a one-token dataset D:“ typiq 1, ypiq 2uN i“1, and then duplicate it as D;“ typiq 1ypiq 1, ypiq 2ypiq 2uN i“1, and ask human annotators to label these pairs. Now we have co... | https://arxiv.org/abs/2505.19770v1 |
while θ‹ p,0“2rdense, due to the non-singularity of the Gram matrix. We can have a ℓ1-regularized estimator for the reward model: ˆθr,relℓ1“ argmin θ0`τ1PΘB,θ1“e1´1 NNÿ i“1logσprθpypiq wypiq wq´rθpypiq lypiq lqq`γ}θ0}1, ùñ ˆθr,relℓ1,0“argmin θ0`τ1PΘB´1 NNÿ i“1logσpβpθ0`τ1qJpψpypiq wq´ψpypiq lqq`γ}θ0`τ1´τ1}1, where τ1:“... | https://arxiv.org/abs/2505.19770v1 |
the value of σ1prϕpyq´rϕpy1qqto the sampling probability µpy, y1q. Thus we expect µpy, y1q9πθ‹prϕq{σ1prϕpyq´rϕpy1qq. And under the context of surrogate reward learning, we have πθ‹prϕq“πθandrϕ“ˆrθ, and thus µ9πθ‹prϕq{σ1pˆrθpyq´ˆrθpy1qq, which is exactly PILAF sampler (see Definition 1). Calculation of online IPO. For o... | https://arxiv.org/abs/2505.19770v1 |
a batch size 32, a learning rate 5e´7, and a gradient accumulation step 2. We train for 3iterations, each for 2epochs. We set rmargin“0.4,1,4for verifications of Condition 1, and set rmargin“1for verifications of Conditions 2 to 4. Offline training: The hyper-parameter setting is based on Zhou et al. [2024]. We use a b... | https://arxiv.org/abs/2505.19770v1 |
arXiv:2505.19773v1 [cs.CL] 26 May 2025What Really Matters in Many-Shot Attacks? An Empirical Study of Long-Context Vulnerabilities in LLMs Sangyeop Kim1,2∗, Yohan Lee1∗, Yongwoo Song1,3*, Kimin Lee4† 1Coxwave,2Seoul National University,3Kyung Hee University,4KAIST {sangyeop.kim, yohan.lee}@coxwave.com, syw5141@khu.ac.k... | https://arxiv.org/abs/2505.19773v1 |
degra- dations suggest that MSJ exploits mechanisms be- yond ICL and may be linked to model design. Therefore, further investigation is needed to un- derstand the factors behind these vulnerabilities. Building on these observations, we conduct ex- tensive experiments to explore key aspects that af- fect these vulnerabi... | https://arxiv.org/abs/2505.19773v1 |
Figure 2: Impact of Instruction Types on ASR across Models. Our experiments confirm the existence of three distinct phases: an initial weakness point, a degradation phase, and a rebound phase. These phases are prominently observed in Secret Role andLove Pliny instructions (middle and right) , while Safe instruction (le... | https://arxiv.org/abs/2505.19773v1 |
Target Query Selection To ensure reliable evaluation of model safety, we carefully filter ambiguous queries from our test dataset. Some queries can be interpreted in non- harmful contexts, making them unsuitable for eval- uating model safety. For instance, queries request- ing constructive discussions about sensitive t... | https://arxiv.org/abs/2505.19773v1 |
analyze how long-context vulnerabilities vary with different harmful topics, we employ Harmful- Adult ,Harmful-Criminal ,Harmful-Cybersecurity , Harmful-Psychology ,Harmful-Discrimination and Harmful-Privacy datasets. Figure 5: ASR Comparison across Different Topic Categories. ASR patterns remain consistent across diff... | https://arxiv.org/abs/2505.19773v1 |
like Llama-3.1 families and Qwen- 2.5-72B, achieving higher ASR. Figure 7 reveals distinct vulnerabilities when using text to fill context compared to QA shots. Ex- cept for Qwen-2.5-7B which shows vulnerability only at 128K context like MSJ, models typically show weakness points at 29to210tokens with QA shots, but the... | https://arxiv.org/abs/2505.19773v1 |
to these at- tacks reinforces the concerning trade-off between model capabilities and safety. 5.2 Shot Repetition Attack Another approach to fill context is simply repeating the same shot multiple times. Unlike fake content which cannot utilize harmfulness characteristics, this repetition strategy further exploits the ... | https://arxiv.org/abs/2505.19773v1 |
increased, suggesting an inverse relationship that warrants further investiga- tion into current safety training procedures. Given these observations, current defense strate- gies may not adequately address the unique chal- lenges of long-context interactions. Our results suggest that safety alignment varies across dif... | https://arxiv.org/abs/2505.19773v1 |
alignment across context lengths, im- proving context processing mechanisms, and de- veloping position-aware safety mechanisms remain important directions for future research. Limitations Model Accessibility and Coverage Limitations Our study focused on open-source models (e.g., Llama, Qwen) accessible under permissibl... | https://arxiv.org/abs/2505.19773v1 |
Bat- son, Meg Tong, Jesse Mu, Daniel J Ford, Francesco Mosconi, Rajashree Agrawal, Rylan Schaeffer, Naomi Bashkansky, Samuel Svenningsen, Mike Lam- bert, Ansh Radhakrishnan, Carson Denison, Evan J Hubinger, Yuntao Bai, Trenton Bricken, Timothy Maxwell, Nicholas Schiefer, James Sully, Alex Tamkin, Tamera Lanham, Karina ... | https://arxiv.org/abs/2505.19773v1 |
Alexander Robey, Maksym Andriushchenko, Francesco Croce, Vikash Sehwag, Edgar Dobriban, Nicolas Flammarion, George J. Pappas, Florian Tramèr, Hamed Hassani, and Eric Wong. 2024. Jailbreakbench: An open ro- bustness benchmark for jailbreaking large language models. In The Thirty-eight Conference on Neural Information Pr... | https://arxiv.org/abs/2505.19773v1 |
defenses for ad- versarial attacks against aligned language models. arXiv preprint arXiv:2309.00614 . Neel Jain, Avi Schwarzschild, Yuxin Wen, Gowthami Somepalli, John Kirchenbauer, Ping yeh Chiang, Micah Goldblum, Aniruddha Saha, Jonas Geiping, and Tom Goldstein. 2024. Baseline defenses for ad- versarial attacks again... | https://arxiv.org/abs/2505.19773v1 |
Alfred Hunt, Justin Tienken-Harder, Kevin Y . Shih, Kemper Talley, John Guan, Ian Steneker, David Campbell, Brad Jokubaitis, Steven Basart, Stephen Fitz, Pon- nurangam Kumaraguru, Kallol Krishna Karmakar, Uday Tupakula, Vijay Varadharajan, Yan Shoshi- taishvili, Jimmy Ba, Kevin M. Esvelt, Alexandr Wang, and Dan Hendryc... | https://arxiv.org/abs/2505.19773v1 |
Anand Siththaranjan, Cassidy Laidlaw, and Dylan Hadfield-Menell. 2024. Distributional preference learning: Understanding and accounting for hidden context in RLHF. In The Twelfth International Con- ference on Learning Representations . Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. Rof... | https://arxiv.org/abs/2505.19773v1 |
Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang R... | https://arxiv.org/abs/2505.19773v1 |
Pliny instruction (Figure 13) is an explicit jailbreaking technique designed to override model safeguards and generate harmful outputs. Through these three categories, we examine how different instruction types influence model safety and vulnerability to attacks. You are a helpful chatbot dedicated to assisting users w... | https://arxiv.org/abs/2505.19773v1 |
the responses carefully and let me know your answer. User Query: {{query}} AI Response: {{response}} Analyze the content above and respond with a single word: - If safe: "SAFE" - If harmful: "UNSAFE" Figure 14: Judge Model Prompt. models tested, GPT-4o demonstrated the highest performance as our judge model, achieving ... | https://arxiv.org/abs/2505.19773v1 |
queries and responses using their existing safety labels (such as is_safe=false in BeaverTails). To maintain manageable dataset sizes, we limited our collection to a maximum of 5,000 samples per dataset, yielding 22,467 candidate samples. To create Harmful-128, 512, 2048 datasets, we first calculated the total token co... | https://arxiv.org/abs/2505.19773v1 |
meet the 128K context length requirement. For the Safe-512 dataset, we drew from the OpenOrca dataset, which contains over 1M sam- ples designed for Instruction Tuning. We specifi- cally excluded prompts containing specialized in- structions like "Think like you are answering to a five year old" or "While performing th... | https://arxiv.org/abs/2505.19773v1 |
intuitive and reliable metric for assessing attack effectiveness. D Effectiveness of Attack Methods on Qwen2.5 Models The effectiveness of Fake Data Attack and Shot Repetition Attack was also evaluated on the Qwen model. The Fake Data Attack generally demon- strates higher ASR, except for specific spike re- gions near ... | https://arxiv.org/abs/2505.19773v1 |
and additional open-source models to verify the generalizability of our findings. Table 6 presents the corresponding ASR result, confirming similar vulnerability patterns across different model archi- tectures. I Detailed Experimental Results We provide the detailed results of our experiments, including mean Attack Suc... | https://arxiv.org/abs/2505.19773v1 |
0.026 (0.010) 0.000 (0.000) 0.142 (0.015) Qwen2.5-72B 0.071 (0.016) 0.049 (0.011) 0.260 (0.020) 0.226 (0.010) 0.043 (0.008) 0.017 (0.008) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000) 0.042 (0.037) Llama-3.2-3B 0.060 (0.000) 0.140 (0.000) 0.360 (0.000) 0.080 (0.000) 0.100 (0.000) 0.080 (0.000) 0.000 (0.000) 0... | https://arxiv.org/abs/2505.19773v1 |
0.000 (0.000) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000) – – Llama-3.1-70B 0.031 (0.011) 0.469 (0.027) 0.487 (0.010) 0.342 (0.016) 0.040 (0.000) 0.000 (0.000) 0.020 (0.000) 0.013 (0.010) 0.042 (0.007) 0.040 (0.000) 0.848 (0.039) – – Harmful-2048Qwen2.5-7B 0.100 (0.020) 0.272 (0.018) 0.204 (0.... | https://arxiv.org/abs/2505.19773v1 |
(0.000) 0.000 (0.000) 0.225 (0.010) 0.000 (0.000) 0.000 (0.000) 0.060 (0.000) 0.120 (0.000) 1.000 (0.000) CybersecurityQwen2.5-7B 0.108 (0.011) 0.200 (0.000) 0.136 (0.009) 0.064 (0.009) 0.100 (0.000) 0.120 (0.000) 0.060 (0.000) 0.020 (0.000) 0.100 (0.000) 0.124 (0.009) 0.376 (0.009) 0.956 (0.009) Qwen2.5-14B 0.356 (0.0... | https://arxiv.org/abs/2505.19773v1 |
(0.011) 0.720 (0.000) Qwen2.5-72B 0.075 (0.010) 0.070 (0.012) 0.020 (0.000) 0.165 (0.010) 0.020 (0.000) 0.060 (0.000) 0.060 (0.000) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000) Llama-3.2-3B 0.060 (0.000) 0.108 (0.011) 0.212 (0.011) 0.208 (0.011) 0.372 (0.011) 0.300 (0.014) 0.160 (0.000) 0.436 (... | https://arxiv.org/abs/2505.19773v1 |
(0.009) 0.092 (0.011) 0.300 (0.000) 0.200 (0.000) 0.400 (0.000) 0.240 (0.000) 0.280 (0.000) 0.240 (0.000) 0.120 (0.000) 0.080 (0.000) 0.040 (0.000) Llama-3.1-70B 0.025 (0.010) 0.045 (0.010) 0.250 (0.026) 0.450 (0.012) 0.605 (0.010) 0.575 (0.019) 0.470 (0.020) 0.140 (0.016) 0.000 (0.000) 0.000 (0.000) 0.760 (0.073) Tabl... | https://arxiv.org/abs/2505.19773v1 |
0.000 (0.000) 0.000 (0.000) Qwen2.5-72B 0.072 (0.011) 0.080 (0.014) 0.052 (0.027) 0.052 (0.030) 0.048 (0.033) 0.036 (0.036) 0.008 (0.011) 0.012 (0.011) 0.004 (0.009) 0.000 (0.000) 0.010 (0.015) Llama-3.2-3B 0.060 (0.000) 0.124 (0.033) 0.256 (0.105) 0.272 (0.113) 0.256 (0.038) 0.216 (0.043) 0.216 (0.036) 0.240 (0.051) 0... | https://arxiv.org/abs/2505.19773v1 |
arXiv:2505.19776v1 [cs.CL] 26 May 2025Analyzing Political Bias in LLMs via Target-Oriented Sentiment Classification Akram Elbouanani1, Evan Dufraisse1, Adrian Popescu1, 1Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France Abstract Political biases encoded by LLMs might have detrimental effects on downstream ... | https://arxiv.org/abs/2505.19776v1 |
et al., 2024; Huang et al., 2020) prompt LLMs to generate political content and deploy sentiment or stance detection to infer ideological alignments. While interesting, these approaches have signif- icant limitations. First, the number of interac- tions with LLMs is usually reduced, leading to limited statistical power... | https://arxiv.org/abs/2505.19776v1 |
alignment, while (Santurkar et al., 2023) leverage public opin- ion polls to evaluate alignment with demographic groups. (Hartmann et al., 2023) and (Motoki et al., 2024) utilize political compass tests and voting ad- vice applications to assess political leanings. These methods provide easily quantifiable and repro- d... | https://arxiv.org/abs/2505.19776v1 |
a dataset of political entities and sentences, translating them into six languages, and evaluat- ing multiple LLMs. We diversify personal and political attributes to control confounders such as demographics and popularity. We also introduce a control group of synthetic entities to isolate theeffects of entity-oriented ... | https://arxiv.org/abs/2505.19776v1 |
al., 2019) and entity-linking using mGENRE (De Cao et al., 2022) over CC-News (Mackenzie et al., 2020) to count politician mentions in the news and retain the most frequently mentioned politicians. We then employ a hierarchical sampling algorithm to obtain a list of 1319 politicians diversified across countries and pol... | https://arxiv.org/abs/2505.19776v1 |
on the left-to-right spectrum, (2) politician-level similarities, (3) 2-dimensional political compass analysis, (4) cross-language sentiment patterns, and (5) the effect of model size on bias intensity. 4 Results Sentiment analysis across models and languages reveals systematic biases based on political align- ment. Fi... | https://arxiv.org/abs/2505.19776v1 |
trade-offs in multilingual model design and the need to consider how training data and methodolo- gies influence bias amplification carefully. LLMs have a left-libertarian bias. Using Parl- Gov data, which assigns parties a progressive- conservative and authoritarian-libertarian score, we map politicians onto a 10 ×10 ... | https://arxiv.org/abs/2505.19776v1 |
similarity exhibit more aligned senti- ment patterns, reflecting ideological or contextual affinities. For example, Pedro Sánchez andOlaf Scholz , both European socialist leaders, show high similarity (+0.92), likely due to shared political ide- ology and regional context. Similarly, Joe Biden andKamala Harris , who sh... | https://arxiv.org/abs/2505.19776v1 |
templates to obtain a massive TSC predic- tion dataset and aggregate them to analyze differ- ent bias facets. Our results have several important implications. First, we contribute to studying the LLMs’ political biases by introducing a flexibleframework that builds on TSC, a fine-grained NLP task. This modeling enables... | https://arxiv.org/abs/2505.19776v1 |
77 100 77 76 78 78 79 77 100eng spa fra rus zho araFigure 4: Jaccard similarity index between the sentiment predictions in the tested languages obtained with Qwen-7B, and Qwen-72B , and Aya-Expanse-32B. 012345678910GPT-4o-mini (zho) 0 1 2 3 4 5 6 7 8 9 10 Economic Policy 0 1 2 3 4 5 6 7 8 9 10Social Policy GPT-4o-mini ... | https://arxiv.org/abs/2505.19776v1 |
74 73 74 78 80 79 82 85 10084 81 81 79 83 81 80 80 73 73 73 75 78 79 80 84 87 84 1008789 87 91 90 85 86 75 74 73 76 77 80 82 85 86 81 87 100 90 90 89 89 84 85 73 73 72 75 77 79 79 83 85 81 89 90 100 88 90 92 8788 73 72 72 75 76 79 80 84 84 79 87 90 88 100 88 88 85 86 74 74 74 76 79 82 81 86 878391 89 90 88 100 92 8788 ... | https://arxiv.org/abs/2505.19776v1 |
into how biases manifest in usual political discussions in other con- texts. More advanced political bias mitigation tech- niques would require in-depth model adaptations beyond this contribution’s scope. Evolving Nature of Biases. The biases observed reflect the state of political text included in the LLMs’ training d... | https://arxiv.org/abs/2505.19776v1 |
Rasul, Stefan Schweter, and Roland V ollgraf. 2019. Flair: An easy-to-use framework for state-of-the-art nlp. In Proceedings of the 2019 conference of the North American chapter of the association for com- putational linguistics (demonstrations) , pages 54–59. Badr AlKhamissi, Muhammad ElNokrashy, Mai Alkhamissi, and M... | https://arxiv.org/abs/2505.19776v1 |
pro- environmental, left-libertarian orientation. Preprint , arXiv:2301.01768. Andrew Heywood. 2021. Political ideologies: An intro- duction . Bloomsbury Publishing. Yi-Li Hsu, Jui-Ning Chen, Yang Fan Chiang, Shang- Chien Liu, Aiping Xiong, and Lun-Wei Ku. 2024. Enhancing perception: Refining explanations of news claim... | https://arxiv.org/abs/2505.19776v1 |
Feng, Yang Wu, Yongheng Zhang, Yinghui Li, Min Li, Wanxiang Che, and Philip S Yu. 2024. Large language models meet nlp: A survey. arXiv preprint arXiv:2405.12819 . Paul Röttger, Valentin Hofmann, Valentina Pyatkin, Musashi Hinck, Hannah Rose Kirk, Hinrich Schütze, and Dirk Hovy. 2024. Political Compass or Spin- ning Ar... | https://arxiv.org/abs/2505.19776v1 |
nian Harold Li, and Kai-Wei Chang. 2022. Geom- lama: Geo-diverse commonsense probing on multi- lingual pre-trained language models. arXiv preprint arXiv:2205.12247 . Wenxuan Zhang, Yue Deng, Bing Liu, Sinno Pan, and Lidong Bing. 2024. Sentiment Analysis in the Era of Large Language Models: A Reality Check. In Findings ... | https://arxiv.org/abs/2505.19776v1 |
Remove "Big Tent" from alignments 7: ifalignments is empty then 8: return "Big Tent" 9: end if 10:end if 11:sum←0 12:foreach alignment inalignments do 13: sum←sum+mapping[alignment] 14:end for 15:avg←sum/ length( alignments ) 16:rounded ←ROUND AWAY FROM ZERO(avg) 17:foreach ( label ,score ) inmapping do 18: ifscore =ro... | https://arxiv.org/abs/2505.19776v1 |
sampled_politicians 6: end for 7: foreach alignment in {"CL", "CR"} do 8: LetSubset_CL_CR ←filter(entities, where Country = country and Alignment = alignment) 9: Append top k2entities from Sub- set_CL_CR to sampled_politicians 10: end for 11: foreach alignment in {"CC"} do 12: Let Subset_CC ← filter(entities, where Cou... | https://arxiv.org/abs/2505.19776v1 |
provided in the input. This ensures that the control group consists of entirely fictional entities with no overlap or similarity to real-world politicians. The prompt is structured as follows: •Thesystem role defines the task as generating culturally and temporally appropriate names that are unique and avoid repetition... | https://arxiv.org/abs/2505.19776v1 |
from this step, as we adjust for gender and personality traits are not overly specific. •Check for additional entities in the sentence and ensure that their relationship with the main entity is not too specific to generalize. The table 1 provides examples of sentence eval- uation, including the original sentence, the s... | https://arxiv.org/abs/2505.19776v1 |
“At the time, he also recommended that Mussolini and Hitler develop closer ties.”6 Accepted The relationship between Hitler and Mussolini is not contextualized and can be generalized to other enti- ties. in non-Western languages tied to different geopo- litical contexts, we translated the sentences into multiple langua... | https://arxiv.org/abs/2505.19776v1 |
were among the most highly rated entities. In contrast, Adolf Hitler (-0.81), Heinrich Himmler (-0.55), and Joseph Stalin (-0.34) received some of the most negative sentiment scores. Nick Griffin (-0.36), a British far-right politician and for- mer leader of the British National Party, Marine Le Pen(-0.28), leader of F... | https://arxiv.org/abs/2505.19776v1 |
Llama 3 70B 0-shot FR 0.57 0.50 0.57 0.73 EN 0.65 0.38 0.56 0.74 AR 0.38 0.43 0.64 0.64 6-shot FR 0.61 0.64 0.69 0.72 EN 0.65 0.66 0.54 0.69 AR 0.24 0.41 0.63 0.69 9-shot FR 0.66 0.66 0.67 0.74 EN 0.69 0.59 0.69 0.73 AR 0.53 0.63 0.62 0.71 Anal yze t he sent i m ent t ow ar ds t he t ar get of t he f ol l ow i ng sent ... | https://arxiv.org/abs/2505.19776v1 |
ci ona sol o l a punt uaci ón de sent i m i ent o basada en l a escal a pr opor ci onada. La r espuest a sol o debe cont ener l a pal abr a negat i vo, neut r al o posi t i vo, nada m ás. or aci ón : {sentence} obj et i vo : {target} sent i m i ent o: П роанализируйте отнош ение к цели следую щ его предлож ения и класс... | https://arxiv.org/abs/2505.19776v1 |
names in negative sentiments intro- duces a slight negative bias in others. For instance, Russian names are overrepresented in the FL align- ment while women are overrepresented in the LL alignment. These findings highlight the difficulty of achieving complete neutrality through fictional name generation. J Statistical... | https://arxiv.org/abs/2505.19776v1 |
p-values. Model Language FL LL CL CC CR RR FR BT Aya FrenchFL 0.9972 0.9960 0.8845 0.0012 0.0000 0.0000 0.2945 LL 0.0029 0.4070 0.0641 0.0000 0.0000 0.0000 0.0017 CL 0.0040 0.5934 0.0878 0.0000 0.0000 0.0000 0.0028 CC 0.1157 0.9361 0.9123 0.0000 0.0000 0.0000 0.0520 CR 0.9988 1.0000 1.0000 1.0000 0.0021 0.0000 0.9870 R... | https://arxiv.org/abs/2505.19776v1 |
0.0000 0.1221 0.4831 CR 1.0000 1.0000 0.9999 0.9928 0.0271 0.9146 0.9815 RR 1.0000 1.0000 1.0000 1.0000 0.9730 0.9994 0.9999 FR 0.9985 1.0000 0.9934 0.8781 0.0856 0.0006 0.8210 BT 0.9431 0.9950 0.8822 0.5176 0.0185 0.0001 0.1794 GPT SpanishFL 0.9727 0.0656 0.0565 0.0000 0.0000 0.0000 0.0035 LL 0.0273 0.0001 0.0002 0.00... | https://arxiv.org/abs/2505.19776v1 |
0.6134 0.0075 0.0001 0.2847 LL 0.0270 0.5975 0.3040 0.0433 0.0000 0.0000 0.0094 CL 0.0171 0.4029 0.2201 0.0298 0.0000 0.0000 0.0059 CC 0.0732 0.6963 0.7801 0.1019 0.0000 0.0000 0.0235 CR 0.3871 0.9568 0.9702 0.8983 0.0018 0.0000 0.1804 RR 0.9925 1.0000 1.0000 1.0000 0.9982 0.0719 0.9518 FR 0.9999 1.0000 1.0000 1.0000 1... | https://arxiv.org/abs/2505.19776v1 |
0.9993 0.8363 0.5772 0.9763 FR 0.8822 0.9995 0.9999 0.9984 0.7637 0.4232 0.9623 BT 0.2367 0.8483 0.9032 0.7679 0.1182 0.0238 0.0379 Table 10: Mistral-7B: Mann-Withney test p-values. Model Language FL LL CL CC CR RR FR BT Mistral FrenchFL 1.0000 1.0000 1.0000 0.9995 0.0092 0.0000 0.4689 LL 0.0000 0.8222 0.7063 0.1071 0.... | https://arxiv.org/abs/2505.19776v1 |
0.9872 0.0350 0.0000 0.7508 LL 0.0000 0.9378 0.2113 0.0023 0.0000 0.0000 0.0001 CL 0.0000 0.0623 0.0088 0.0000 0.0000 0.0000 0.0000 CC 0.0000 0.7890 0.9912 0.0246 0.0000 0.0000 0.0012 CR 0.0128 0.9977 1.0000 0.9754 0.0000 0.0000 0.0924 RR 0.9651 1.0000 1.0000 1.0000 1.0000 0.0000 0.9881 FR 1.0000 1.0000 1.0000 1.0000 1... | https://arxiv.org/abs/2505.19776v1 |
0.9510 0.0563 0.9547 FR 0.9999 1.0000 1.0000 1.0000 0.9992 0.9439 0.9984 BT 0.6061 0.9703 0.9588 0.9765 0.3896 0.0454 0.0016 Qwen7b RussianFL 0.9999 0.9998 0.9987 0.9614 0.9479 0.0020 0.9426 LL 0.0001 0.3918 0.1805 0.0189 0.0053 0.0000 0.0483 CL 0.0002 0.6085 0.2855 0.0401 0.0181 0.0000 0.0730 CC 0.0013 0.8198 0.7148 0... | https://arxiv.org/abs/2505.19776v1 |
arXiv:2505.19797v2 [cs.CL] 28 May 2025TheAvengers : A Simple Recipe for Uniting Smaller Language Models to Challenge Proprietary Giants Yiqun Zhang1,2♣‡Hao Li2,3♣‡Chenxu Wang2,4‡Linyao Chen2,5‡ Qiaosheng Zhang2Peng Ye2Shi Feng1†Daling Wang1Zhen Wang3 Xinrun Wang6Jia Xu2Lei Bai2Wanli Ouyang2Shuyue Hu2† 1Northeastern Uni... | https://arxiv.org/abs/2505.19797v2 |
broader community of researchers and practitioners to engage in and contribute to cutting-edge LM research. Second, the open-source ecosystem has already invested significant effort and computational resources in model development.1 Demonstrating strong performance from these models would justify the reuse of existing ... | https://arxiv.org/abs/2505.19797v2 |
clustering-based routing strategy, and empirically, in the notably strong performance it achieves. 1For example, as of May 2025, there are over 1,318 models fine-tuned on LLaMA-3.1-8B-Instruct and 489 on Qwen-2.5-7B-Instruct, despite both being released less than a year ago. 2 Our experiments show that, the Avengers , ... | https://arxiv.org/abs/2505.19797v2 |
a small group of university students). These results not only demonstrate the effectiveness of the Avengers , but also more importantly, highlight the untapped potential of open-source contributions—advancing the vision of a more democratic and inclusive AI ecosystem. 2 Related Work The harnessing of collective intelli... | https://arxiv.org/abs/2505.19797v2 |
than those used in our work), yet still failto demonstrate competitive performance. To the best of our knowledge, this is the first work to show that a training-free, router- based framework—when integrated with automatic complementary model selection and voting—can elevate small open-source models to match the overall... | https://arxiv.org/abs/2505.19797v2 |
a vector p= [p1,···, pk], where pidenotes the model’s performance score on cluster ci∈ C. These profiles inform routing decisions at the inference time. Online Inference At the inference time, the Avengers no longer performs clustering or profiling. For each query in the test set Dtest, the Avengers first computes its ... | https://arxiv.org/abs/2505.19797v2 |
42], Knights and Knaves [ 43], and BBH [ 44]),Knowledge (ARC Challenge [ 45], MMLUPro [ 46], GPQA [ 47], FinQA [ 48], and MedQA [ 49]), and Affective (EmoryNLP [ 50] and MELD [ 51]). Additionally, to evaluate the generalization capabilities of the Avengers on out-of- distribution (OOD) tasks, we further include one ext... | https://arxiv.org/abs/2505.19797v2 |
•MODEL -SAT [10]: Employs capability instruction tuning, translating candidate models’ perfor- mance into textual capability descriptions, which are embedded and passed to a trainable LLM to predict model suitability for each query. •Mixture of Agents (MoA) [11]: A hierarchical mixture approach where multiple layers of... | https://arxiv.org/abs/2505.19797v2 |
struggle in Out-of-Distribution (OOD) settings. To compare their generalization ability, we introduce a new dataset for each of the five task categories to evaluate performance. As shown in Table 3, the Avengers exhibits the most robust generalization among all compared methods, achieving an average score of 74.42. Thi... | https://arxiv.org/abs/2505.19797v2 |
66.46 39.60 54.38 48.12 glm-4-chat 5.00 58.40 22.14 62.01 65.85 37.60 21.57 47.59 92.15 51.75 31.25 58.59 64.96 41.32 57.46 47.84 Llama-3.1-it 1.67 49.80 20.71 61.81 71.95 27.60 11.71 65.74 88.48 47.35 25.67 53.97 69.76 35.15 51.62 45.53 Granite-3.1-it 1.67 61.40 20.00 37.17 39.63 31.44 19.29 36.39 85.24 44.06 34.82 65... | https://arxiv.org/abs/2505.19797v2 |
62.50 69.77 70.91 59.20 66.89 MODEL-SAT†97.20 58.22 68.37 70.58 48.40 68.55 EmbedLLM†97.20 63.06 71.11 69.34 47.98 69.77 TheAvengers (ours) 96.67 68.52 77.43 71.79 57.70 74.42 - vs EmbedLLM (%) ↓0.55 ↑4.98 ↑19.17 ↑2.20 ↑24.30 ↑8.14 Model Selection Figure 2 shows the performance as the number of selected models in the A... | https://arxiv.org/abs/2505.19797v2 |
key hyperparameters such as the number of clusters ( k). With an increasing number of integrated models, the Avengers exhibits complementary advantages, especially in knowledge, code and affective tasks. 9 The superior performance of the Avengers highlights the untapped potential of community-driven, open-source models... | https://arxiv.org/abs/2505.19797v2 |
California press, 1967. [17] Joe H Ward Jr. Hierarchical grouping to optimize an objective function. Journal of the American statistical association , 58(301):236–244, 1963. [18] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statis... | https://arxiv.org/abs/2505.19797v2 |
Liang, Yupei Lin, Yandong Chen, Shanshan Zhong, Hefeng Wu, and Liang Lin. Routereval: A comprehensive benchmark for routing llms to explore model-level scaling up in llms. arXiv preprint arXiv:2503.10657 , 2025. [34] Keming Lu, Hongyi Yuan, Runji Lin, Junyang Lin, Zheng Yuan, Chang Zhou, and Jingren Zhou. Routing to th... | https://arxiv.org/abs/2505.19797v2 |
V Le, Ed H Chi, Denny Zhou, , and Jason Wei. Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261 , 2022. 12 [45] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? t... | https://arxiv.org/abs/2505.19797v2 |
preprint arXiv:2308.03281 , 2023. 13 [60] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems , 36:7093–7115, 2023. [61] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language mo... | https://arxiv.org/abs/2505.19797v2 |
glm-4 all tools. arXiv preprint arXiv:2406.12793 , 2024. 14 [74] Microsoft, :, Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Alon Benhaim, Martin Cai, Vishrav Chaudhary, Congcong Chen, Dong Chen, Dongdong Chen, Junkun Chen, Weizhu Chen, Yen-Chun Chen, Yi ling Chen, Qi Da... | https://arxiv.org/abs/2505.19797v2 |
Shuo Zhang, Songyang Zhang, Wenjian Zhang, Wenwei Zhang, Xingcheng Zhang, Xinyue Zhang, Hui Zhao, Qian Zhao, Xiaomeng Zhao, Fengzhe Zhou, Zaida Zhou, Jingming Zhuo, Yicheng Zou, Xipeng Qiu, Yu Qiao, and Dahua Lin. Internlm2 technical report, 2024. [79] Ryan Teknium, Jeffrey Quesnelle, and Chen Guang. Hermes 3 technical... | https://arxiv.org/abs/2505.19797v2 |
Gu, Zhiyuan Liu, and Zonghong Dai. Yi: Open foundation models by 01.ai, 2025. URL https://arxiv.org/abs/2403.04652 . A Technical Appendices and Supplementary Material A.1 Further Related Work Mixture-based methods This line of research typically processes each query using multiple models in parallel to generate several... | https://arxiv.org/abs/2505.19797v2 |
models. Subsequently, an automated model selection method is applied to select the 10 models listed in Table 7. All results report in the main text, including those for baseline comparisons, consistently use the same set of models selected by the Avengers . Table 8: Other models. Model Parameters (B) Phi-4-mini-instruc... | https://arxiv.org/abs/2505.19797v2 |
language models, while the projector uses a higher learning rate of 1e-4. Initially, we exclusively fine-tune the projector for approximately 10,000 steps. Following this phase, we continue fine-tuning all model parameters for the remainder of the training process. A warmup ratio of 0.1 is applied to stabilize the earl... | https://arxiv.org/abs/2505.19797v2 |
code on GitHub: https://anonymous.4open.science/r/Avengers-2C0D B.1 Benchmark Dataset Descriptions AIME Olympiad-level math benchmark of American Invitational Mathematics Examination prob- lems (2024 & 2025); numeric-answer tasks covering algebra, geometry, number theory and combinatorics that push high-school contest ... | https://arxiv.org/abs/2505.19797v2 |
tuned first for code reliability rather than broad expert knowledge. DeepSeek-R1-Distill-Qwen-7B Dominates Olympiad-level mathematics and logic riddles, while retaining competitive breadth across finance and encyclopaedic knowledge tests—a rare feat for its parameter count. The main gap is in highly specialised clinica... | https://arxiv.org/abs/2505.19797v2 |
its effect on model performance. As shown in Figure 7, results indicate that the Avengers achieves the best overall performance when the test set proportion is approximately between 0.3 and 0.35. Additionally, it is noteworthy that with only 30% of the data used to fit the clustering model, the Avengers already surpass... | https://arxiv.org/abs/2505.19797v2 |
their generalization performance. Moreover, identifying a simple yet effective early-stopping criterion to prevent overfitting becomes challenging in practice. This explains why, in Tables 2 and 3, we report peak performance on the test set rather than the validation set for these trained methods. In the Appendix A.4, ... | https://arxiv.org/abs/2505.19797v2 |
MOLE: Metadata Extraction and Validation in Scientific Papers Using LLMs Zaid Alyafeai1Maged S. Al-Shaibani2Bernard Ghanem1 1KAUST2SDAIA-KFUPM Joint Research Center for AI, KFUPM Abstract Metadata extraction is essential for cataloging and preserving datasets, enabling effective re- search discovery and reproducibility... | https://arxiv.org/abs/2505.19800v1 |
Figure 2: MOLE pipeline. The paper text and Schema are used as input, and the output is the extracted metadata content. approaches typically extract around 5-10 attributes (Ahmad and Afzal, 2020; Tkaczyk et al., 2015), our work automatically extracts around 30 differ- ent attributes per paper, providing a substantially... | https://arxiv.org/abs/2505.19800v1 |
issues. •Domain we use the following options to de- scribe the attribute: social media, news ar- ticles, commentary, books, Wikipedia, web pages, public datasets, TV channels, captions, LLM, or other. The options are improved by including recent approaches for synthetic data generation using LLMs. •Collection Style sim... | https://arxiv.org/abs/2505.19800v1 |
use three types of validations to make sure the output is consistent with our schema: Table 1: Results of all models on the main categories Arabic (ar), English (en), Japanese (jp), French (fr), Russian (ru), Multilingual (multi) datasets. Average shows the weighted average of all categories. Maximum across category is... | https://arxiv.org/abs/2505.19800v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.