7dbe
ITÖ®¼Ò 5 Ô 1 ÈÕÐÂÎÅ£¬Éî¶ÈÇóË÷£¨DeepSeek£©×òÈÕ£¨4 Ô 30 ÈÕ£©ÔÚ AI ¿ªÔ´ÉçÇø Hugging Face ÉÏ£¬Ðû²¼ÃûΪ DeepSeek-Prover-V2-671B µÄÐÂÄ£×Ó£¬ËæºóÔÚ GitHub µÈƽ̨ÉÏÐû²¼ÁËÂÛÎÄÐÅÏ¢¡£ ITÖ®¼ÒÔ®ÒýÂÛÎĽéÉÜ£¬DeepSeek-Prover-V2 ÊÇÒ»¿îרעÓÚÐÎʽ»¯ÊýÑ§ÍÆÀíµÄ¿ªÔ´´óÐÍÓïÑÔÄ£×Ó£¬»ùÓÚ DeepSeek-V3-0324£¬Í¨¹ýµÝ¹é¶¨Àí֤ʵ¹ÜµÀÌìÉú³õʼÊý¾Ý¡£ Deepseek ÍÆ³öÁË DeepSeek-Prover-V2-671B £¨ÍŽá V3 »ù´¡´óÄ£×Ó£©¡¢ DeepSeek-Prover-V2-7B £¨Ôöǿģ×Ó£©Á½¸öÄ£×Ó£¬ÒÔ¼° DeepSeek-ProverBench Êý¾Ý¼¯¡£ DeepSeek-Prover-V2-671B ½ÓÄÉºÍ DeepSeek V3-0324 ÏàͬµÄ¼Ü¹¹£¬²¢·ÇÓÃÓÚͨÀý¶Ô»°»òÕßÍÆÀí£¬¶øÊÇÓÃÓÚÐÎʽ»¯¶¨Àí֤ʵ¡¢×¨ÃÅÔöÇ¿ÊýѧÄÜÁ¦µÄÄ£×Ó¡£ DeepSeek ÍŶÓÊ×ÏÈÖ¸µ¼ DeepSeek-V3 Ä£×Ó½«ÖØ´ó¶¨ÀíÆÊÎöΪһϵÁÐ×ÓÄ¿µÄ£¨subgoals£©£¬ÕûºÏ·ÇÐÎʽÓëÐÎʽ»¯ÊýÑ§ÍÆÀí£¬ÔÚ Lean 4 ƽ̨ÉÏÐÎʽ»¯Ö¤Êµ°ì·¨¡£ ½Ó×Å£¬Ê¹ÓÃÒ»¸ö½ÏСµÄ 7B ²ÎÊýÄ£×Ó´¦Öóͷ£×ÓÄ¿µÄµÄ֤ʵËÑË÷£¬¼õÇáÅÌËã¼ç¸º¡£×îÖÕ£¬ÍŽáÍêÕûµÄÖð²½Ö¤ÊµÓë DeepSeek-V3 µÄÍ·ÄÔÁ´£¨chain-of-thought£©£¬ÐγÉÇ¿»¯Ñ§Ï°µÄ¡°ÀäÆô¶¯¡±Êý¾Ý¡£ ÔÚѵÁ·ÖУ¬ÍŶÓɸѡ³öÒ»Åú 7B Ä£×ÓÎÞ·¨Ö±½Ó½â¾öµ«×ÓÄ¿µÄÒѱ»Ö¤ÊµµÄÄÑÌ⡣ͨ¹ýÕûºÏ×ÓÄ¿µÄ֤ʵ£¬ÐγÉÍêÕûµÄÐÎʽ»¯Ö¤Êµ£¬²¢Óë DeepSeek-V3 µÄÍÆÀíÀú³Ì¶Ô½Ó£¬ÌìÉúºÏ³ÉÊý¾Ý¡£ Ëæºó£¬Ä£×Ó΢µ÷ÕâЩÊý¾Ý£¬²¢Í¨¹ýÇ¿»¯Ñ§Ï°½øÒ»²½ÌáÉýÄÜÁ¦£¬ÒÔ¶þÔª·´À¡£¨×¼È·»ò¹ýʧ£©×÷Ϊ½±Àø»úÖÆ¡£×îÖÕ£¬DeepSeek-Prover-V2-671B ÔÚÉñ¾¶¨Àí֤ʵÁìÓò´´ÏÂиߣ¬ ÔÚ MiniF2F-test Êý¾Ý¼¯ÉÏͨ¹ýÂÊ´ï 88.9% £¬ÔÚ PutnamBench Êý¾Ý¼¯Öнâ¾ö 658 ¸öÎÊÌâÖÐµÄ 49 ¸ö¡£ ÍŶӻ¹Ðû²¼ÁË ProverBench »ù×¼Êý¾Ý¼¯£¬°üÀ¨ 325 ¸öÐÎʽ»¯ÊýѧÎÊÌâ¡£ÆäÖУ¬15 ¸öÎÊÌâÔ´×Ô½üÆÚ AIME ¾ºÈü£¨AIME 24 ºÍ 25£©£¬Éæ¼°ÊýÂÛÓë´úÊý£¬´ú±í¸ßÖоºÈüÄѶȡ£ ÆäÓà 310 ¸öÎÊÌâÔòÀ´×Ô¾«Ñ¡¿Î±¾ºÍ½ÌѧÄÚÈÝ£¬º¸ÇÏßÐÔ´úÊý¡¢Î¢»ý·Ö¡¢¸ÅÂʵȶà¸öÁìÓò¡£ÕâÒ»Êý¾Ý¼¯Ö¼ÔÚΪ¸ßÖоºÈüºÍ±¾¿ÆÊýѧÌṩÖÜÈ«ÆÀ¹À±ê×¼£¬Íƶ¯Ä£×ÓÔÚ¶àÑù»¯³¡¾°ÏµIJâÊÔÓëÓ¦Óá£