Deprecated: The each() function is deprecated. This message will be suppressed on further calls in /home/zhenxiangba/zhenxiangba.com/public_html/phproxy-improved-master/index.php on line 456 Paper page - ParallelMuse: Agentic Parallel Thinking for Deep Information Seeking
https://github.com/Alibaba-NLP/DeepResearch\n","updatedAt":"2025-10-29T02:41:03.318Z","author":{"_id":"644a4fbc2166258fccc664bc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/8k3b44MbhQiWuo6i8BnYl.jpeg","fullname":"Jialong Wu","name":"callanwu","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":35,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7824311852455139},"editors":["callanwu"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/8k3b44MbhQiWuo6i8BnYl.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2510.24698","authors":[{"_id":"69017d30646208eac0d1f3e7","name":"Baixuan Li","hidden":false},{"_id":"69017d30646208eac0d1f3e8","name":"Dingchu Zhang","hidden":false},{"_id":"69017d30646208eac0d1f3e9","user":{"_id":"644a4fbc2166258fccc664bc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/8k3b44MbhQiWuo6i8BnYl.jpeg","isPro":false,"fullname":"Jialong Wu","user":"callanwu","type":"user"},"name":"Jialong Wu","status":"claimed_verified","statusLastChangedAt":"2026-01-09T08:37:59.196Z","hidden":false},{"_id":"69017d30646208eac0d1f3ea","name":"Wenbiao Yin","hidden":false},{"_id":"69017d30646208eac0d1f3eb","name":"Zhengwei Tao","hidden":false},{"_id":"69017d30646208eac0d1f3ec","user":{"_id":"66e4019518a1920fb7ca19d7","avatarUrl":"/avatars/4f8be2f3244239510235c2ec6e8d38a6.svg","isPro":false,"fullname":"Yida Zhao","user":"zhaoyd","type":"user"},"name":"Yida Zhao","status":"claimed_verified","statusLastChangedAt":"2025-10-29T12:44:30.204Z","hidden":false},{"_id":"69017d30646208eac0d1f3ed","name":"Liwen Zhang","hidden":false},{"_id":"69017d30646208eac0d1f3ee","name":"Haiyang Shen","hidden":false},{"_id":"69017d30646208eac0d1f3ef","name":"Runnan Fang","hidden":false},{"_id":"69017d30646208eac0d1f3f0","name":"Pengjun Xie","hidden":false},{"_id":"69017d30646208eac0d1f3f1","name":"Jingren Zhou","hidden":false},{"_id":"69017d30646208eac0d1f3f2","name":"Yong Jiang","hidden":false}],"publishedAt":"2025-10-28T17:51:50.000Z","submittedOnDailyAt":"2025-10-29T01:04:44.983Z","title":"ParallelMuse: Agentic Parallel Thinking for Deep Information Seeking","submittedOnDailyBy":{"_id":"644a4fbc2166258fccc664bc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/8k3b44MbhQiWuo6i8BnYl.jpeg","isPro":false,"fullname":"Jialong Wu","user":"callanwu","type":"user"},"summary":"Parallel thinking expands exploration breadth, complementing the deep\nexploration of information-seeking (IS) agents to further enhance\nproblem-solving capability. However, conventional parallel thinking faces two\nkey challenges in this setting: inefficiency from repeatedly rolling out from\nscratch, and difficulty in integrating long-horizon reasoning trajectories\nduring answer generation, as limited context capacity prevents full\nconsideration of the reasoning process. To address these issues, we propose\nParallelMuse, a two-stage paradigm designed for deep IS agents. The first\nstage, Functionality-Specified Partial Rollout, partitions generated sequences\ninto functional regions and performs uncertainty-guided path reuse and\nbranching to enhance exploration efficiency. The second stage, Compressed\nReasoning Aggregation, exploits reasoning redundancy to losslessly compress\ninformation relevant to answer derivation and synthesize a coherent final\nanswer. Experiments across multiple open-source agents and benchmarks\ndemonstrate up to 62% performance improvement with a 10--30% reduction in\nexploratory token consumption.","upvotes":21,"discussionId":"69017d31646208eac0d1f3f3","ai_summary":"ParallelMuse enhances problem-solving by efficiently integrating parallel thinking with deep exploration through uncertainty-guided path reuse and reasoning compression.","ai_keywords":["Functionality-Specified Partial Rollout","Compressed Reasoning Aggregation","uncertainty-guided path reuse","reasoning redundancy","information-seeking agents"],"organization":{"_id":"67d15cca6e2cf0e062dbfb54","name":"AlibabaTongyiLab","fullname":"TongyiLab","avatar":"https://cdn-uploads.huggingface.co/production/uploads/67d1502bfabfe9974d1f77bb/XdUSVf6HqBzE7zFBfSDQP.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"646def60df618b303b419323","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/646def60df618b303b419323/JLJGYen4-5M8ivsLsSk0w.jpeg","isPro":false,"fullname":"Lei Wang","user":"demolei","type":"user"},{"_id":"644a4fbc2166258fccc664bc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/8k3b44MbhQiWuo6i8BnYl.jpeg","isPro":false,"fullname":"Jialong Wu","user":"callanwu","type":"user"},{"_id":"6538bdfdf5f5016df35f5faf","avatarUrl":"/avatars/054fc6f8cb46805e66de5c3c856d4fb9.svg","isPro":false,"fullname":"Baixuan Li","user":"MuBai2001","type":"user"},{"_id":"64e44e37f8d8389c1ab8fba6","avatarUrl":"/avatars/d64e67b2c991c0ea4c231b5a996b4216.svg","isPro":false,"fullname":"caishihao","user":"caishihao","type":"user"},{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"},{"_id":"65e6970d135c27ea806526fe","avatarUrl":"/avatars/4aced113d9cab055ae06f3945869a280.svg","isPro":false,"fullname":"Xuanzhong Chen","user":"chenxz","type":"user"},{"_id":"64b73f9317570fdff9b0d1c4","avatarUrl":"/avatars/62124ae3e929b53f99f37e97226a877d.svg","isPro":false,"fullname":"Wang Xinyu","user":"oriuta","type":"user"},{"_id":"63fc4c00a3c067e62899d32b","avatarUrl":"/avatars/b54f2a406afdbbe2cd305d4d9f88ced2.svg","isPro":false,"fullname":"Wenbiao Yin","user":"NLPblue","type":"user"},{"_id":"6680b6c42698e06471df1307","avatarUrl":"/avatars/5570ab6041c70f91828df2f9ddfd6303.svg","isPro":false,"fullname":"Li","user":"BX2001","type":"user"},{"_id":"6867c1ce45b5d529582cfc6d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/bqwK31WooK66CNf1usskF.png","isPro":false,"fullname":"zhangdingchu","user":"zhangdingchu123","type":"user"},{"_id":"66e4019518a1920fb7ca19d7","avatarUrl":"/avatars/4f8be2f3244239510235c2ec6e8d38a6.svg","isPro":false,"fullname":"Yida Zhao","user":"zhaoyd","type":"user"},{"_id":"63ecd42a60ff4b318ad1ef47","avatarUrl":"/avatars/c05001fa9c41f713d1fe31d11212faae.svg","isPro":false,"fullname":"RuiYe","user":"RuiYe1129","type":"user"}],"acceptLanguages":["*"],"dailyPaperRank":0,"organization":{"_id":"67d15cca6e2cf0e062dbfb54","name":"AlibabaTongyiLab","fullname":"TongyiLab","avatar":"https://cdn-uploads.huggingface.co/production/uploads/67d1502bfabfe9974d1f77bb/XdUSVf6HqBzE7zFBfSDQP.png"}}">
ParallelMuse enhances problem-solving by efficiently integrating parallel thinking with deep exploration through uncertainty-guided path reuse and reasoning compression.
AI-generated summary
Parallel thinking expands exploration breadth, complementing the deep
exploration of information-seeking (IS) agents to further enhance
problem-solving capability. However, conventional parallel thinking faces two
key challenges in this setting: inefficiency from repeatedly rolling out from
scratch, and difficulty in integrating long-horizon reasoning trajectories
during answer generation, as limited context capacity prevents full
consideration of the reasoning process. To address these issues, we propose
ParallelMuse, a two-stage paradigm designed for deep IS agents. The first
stage, Functionality-Specified Partial Rollout, partitions generated sequences
into functional regions and performs uncertainty-guided path reuse and
branching to enhance exploration efficiency. The second stage, Compressed
Reasoning Aggregation, exploits reasoning redundancy to losslessly compress
information relevant to answer derivation and synthesize a coherent final
answer. Experiments across multiple open-source agents and benchmarks
demonstrate up to 62% performance improvement with a 10--30% reduction in
exploratory token consumption.
Parallel thinking expands exploration breadth, complementing the deep exploration of information-seeking (IS) agents to further enhance problem-solving capability. However, conventional parallel thinking faces two key challenges in this setting: inefficiency from repeatedly rolling out from scratch, and difficulty in integrating long-horizon reasoning trajectories during answer generation, as limited context capacity prevents full consideration of the reasoning process. To address these issues, we propose ParallelMuse, a two-stage paradigm designed for deep IS agents. The first stage, Functionality-Specified Partial Rollout, partitions generated sequences into functional regions and performs uncertainty-guided path reuse and branching to enhance exploration efficiency. The second stage, Compressed Reasoning Aggregation, exploits reasoning redundancy to losslessly compress information relevant to answer derivation and synthesize a coherent final answer. Experiments across multiple open-source agents and benchmarks demonstrate up to 62% performance improvement with a 10--30% reduction in exploratory token consumption.