Diverse and Effective Red Teaming with Auto-generated Rewards and Multi-step Reinforcement Learning

Autor: Beutel, Alex, Xiao, Kai, Heidecke, Johannes, Weng, Lilian

Automated red teaming can discover rare model failures and generate challenging examples that can be used for training or evaluation. However, a core challenge in automated red teaming is ensuring that the attacks are both diverse and effective. Prio

Externí odkaz: http://arxiv.org/abs/2412.18693

OpenAI o1 System Card

Autor: OpenAI, Jaech, Aaron, Kalai, Adam, Lerer, Adam, Richardson, Adam, El-Kishky, Ahmed, Low, Aiden, Helyar, Alec, Madry, Aleksander, Beutel, Alex, Carney, Alex, Iftimie, Alex, Karpenko, Alex, Passos, Alex Tachard, Neitz, Alexander, Prokofiev, Alexander, Wei, Alexander, Tam, Allison, Bennett, Ally, Kumar, Ananya, Saraiva, Andre, Vallone, Andrea, Duberstein, Andrew, Kondrich, Andrew, Mishchenko, Andrey, Applebaum, Andy, Jiang, Angela, Nair, Ashvin, Zoph, Barret, Ghorbani, Behrooz, Rossen, Ben, Sokolowsky, Benjamin, Barak, Boaz, McGrew, Bob, Minaiev, Borys, Hao, Botao, Baker, Bowen, Houghton, Brandon, McKinzie, Brandon, Eastman, Brydon, Lugaresi, Camillo, Bassin, Cary, Hudson, Cary, Li, Chak Ming, de Bourcy, Charles, Voss, Chelsea, Shen, Chen, Zhang, Chong, Koch, Chris, Orsinger, Chris, Hesse, Christopher, Fischer, Claudia, Chan, Clive, Roberts, Dan, Kappler, Daniel, Levy, Daniel, Selsam, Daniel, Dohan, David, Farhi, David, Mely, David, Robinson, David, Tsipras, Dimitris, Li, Doug, Oprica, Dragos, Freeman, Eben, Zhang, Eddie, Wong, Edmund, Proehl, Elizabeth, Cheung, Enoch, Mitchell, Eric, Wallace, Eric, Ritter, Erik, Mays, Evan, Wang, Fan, Such, Felipe Petroski, Raso, Filippo, Leoni, Florencia, Tsimpourlas, Foivos, Song, Francis, von Lohmann, Fred, Sulit, Freddie, Salmon, Geoff, Parascandolo, Giambattista, Chabot, Gildas, Zhao, Grace, Brockman, Greg, Leclerc, Guillaume, Salman, Hadi, Bao, Haiming, Sheng, Hao, Andrin, Hart, Bagherinezhad, Hessam, Ren, Hongyu, Lightman, Hunter, Chung, Hyung Won, Kivlichan, Ian, O'Connell, Ian, Osband, Ian, Gilaberte, Ignasi Clavera, Akkaya, Ilge, Kostrikov, Ilya, Sutskever, Ilya, Kofman, Irina, Pachocki, Jakub, Lennon, James, Wei, Jason, Harb, Jean, Twore, Jerry, Feng, Jiacheng, Yu, Jiahui, Weng, Jiayi, Tang, Jie, Yu, Jieqi, Candela, Joaquin Quiñonero, Palermo, Joe, Parish, Joel, Heidecke, Johannes, Hallman, John, Rizzo, John, Gordon, Jonathan, Uesato, Jonathan, Ward, Jonathan, Huizinga, Joost, Wang, Julie, Chen, Kai, Xiao, Kai, Singhal, Karan, Nguyen, Karina, Cobbe, Karl, Shi, Katy, Wood, Kayla, Rimbach, Kendra, Gu-Lemberg, Keren, Liu, Kevin, Lu, Kevin, Stone, Kevin, Yu, Kevin, Ahmad, Lama, Yang, Lauren, Liu, Leo, Maksin, Leon, Ho, Leyton, Fedus, Liam, Weng, Lilian, Li, Linden, McCallum, Lindsay, Held, Lindsey, Kuhn, Lorenz, Kondraciuk, Lukas, Kaiser, Lukasz, Metz, Luke, Boyd, Madelaine, Trebacz, Maja, Joglekar, Manas, Chen, Mark, Tintor, Marko, Meyer, Mason, Jones, Matt, Kaufer, Matt, Schwarzer, Max, Shah, Meghan, Yatbaz, Mehmet, Guan, Melody Y., Xu, Mengyuan, Yan, Mengyuan, Glaese, Mia, Chen, Mianna, Lampe, Michael, Malek, Michael, Wang, Michele, Fradin, Michelle, McClay, Mike, Pavlov, Mikhail, Wang, Miles, Wang, Mingxuan, Murati, Mira, Bavarian, Mo, Rohaninejad, Mostafa, McAleese, Nat, Chowdhury, Neil, Ryder, Nick, Tezak, Nikolas, Brown, Noam, Nachum, Ofir, Boiko, Oleg, Murk, Oleg, Watkins, Olivia, Chao, Patrick, Ashbourne, Paul, Izmailov, Pavel, Zhokhov, Peter, Dias, Rachel, Arora, Rahul, Lin, Randall, Lopes, Rapha Gontijo, Gaon, Raz, Miyara, Reah, Leike, Reimar, Hwang, Renny, Garg, Rhythm, Brown, Robin, James, Roshan, Shu, Rui, Cheu, Ryan, Greene, Ryan, Jain, Saachi, Altman, Sam, Toizer, Sam, Toyer, Sam, Miserendino, Samuel, Agarwal, Sandhini, Hernandez, Santiago, Baker, Sasha, McKinney, Scott, Yan, Scottie, Zhao, Shengjia, Hu, Shengli, Santurkar, Shibani, Chaudhuri, Shraman Ray, Zhang, Shuyuan, Fu, Siyuan, Papay, Spencer, Lin, Steph, Balaji, Suchir, Sanjeev, Suvansh, Sidor, Szymon, Broda, Tal, Clark, Aidan, Wang, Tao, Gordon, Taylor, Sanders, Ted, Patwardhan, Tejal, Sottiaux, Thibault, Degry, Thomas, Dimson, Thomas, Zheng, Tianhao, Garipov, Timur, Stasi, Tom, Bansal, Trapit, Creech, Trevor, Peterson, Troy, Eloundou, Tyna, Qi, Valerie, Kosaraju, Vineet, Monaco, Vinnie, Pong, Vitchyr, Fomenko, Vlad, Zheng, Weiyi, Zhou, Wenda, McCabe, Wes, Zaremba, Wojciech, Dubois, Yann, Lu, Yinghai, Chen, Yining, Cha, Young, Bai, Yu, He, Yuchen, Zhang, Yuchen, Wang, Yunyun, Shao, Zheng, Li, Zhuohan

The o1 model series is trained with large-scale reinforcement learning to reason using chain of thought. These advanced reasoning capabilities provide new avenues for improving the safety and robustness of our models. In particular, our models can re

Externí odkaz: http://arxiv.org/abs/2412.16720

Stellar Loci. VIII. Photometric Metallicities for 100 Million Stars Based on Synthetic Gaia Colors

Autor: Huang, Bowen, Yuan, Haibo, Xu, Shuai, Xiao, Kai, Xiang, Maosheng, Huang, Yang, Beers, Timothy C.

We apply the stellar locus method to synthetic $(BP-RP)_{XPSP}$ and $(BP-G)_{XPSP}$ colors derived from corrected Gaia BP/RP (XP) spectra to obtain accurate and precise estimates of metallicity for about 100 million stars in the Milky Way (34 million

Externí odkaz: http://arxiv.org/abs/2410.19895

A Spatial Uniformity Check of Gaia DR3 Photometry and BP/RP Spectra

Autor: Huang, Bowen, Yuan, Haibo, Xiao, Kai

Gaia DR3 photometry and BP/RP spectra have been widely used as reference in photometric calibrations. In this work, we check the spatial uniformity of Gaia DR3 photometry and BP/RP spectra by comparing the BP, RP and G bands photometry with the synth

Externí odkaz: http://arxiv.org/abs/2408.09779

A Versatile Side Entry Laser System for Scanning Transmission Electron Microscopy

Autor: Dyck, Ondrej, Olunloyo, Olugbenga, Xiao, Kai, Wolf, Benjamin, Moore, Thomas M., Lupini, Andrew R., Jesse, Stephen

We present the design and implementation of a side entry laser system designed for an ultra-high vacuum scanning transmission electron microscope. This system uses a versatile probe design enclosed in a vacuum envelope such that parts can be easily a

Externí odkaz: http://arxiv.org/abs/2407.09461