arxiv: 2604.08000 · v1 · submitted 2026-04-09 · 💻 cs.AI · cs.CL· cs.CV· cs.HC· cs.MA

Recognition: unknown

PASK: Toward Intent-Aware Proactive Agents with Long-Term Memory

Zhifei Xie , Zongzheng Hu , Fangda Ye , Xin Zhang , Haobo Chai , Zihang Liu , Pengcheng Wu , Guibin Zhang

show 5 more authors

Yue Liao Xiaobin Hu Deheng Ye Chunyan Miao Shuicheng Yan

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:01 UTC · model grok-4.3

classification 💻 cs.AI cs.CLcs.CVcs.HCcs.MA

keywords proactive agentslatent needsintent detectionlong-term memoryclosed-loop systemsstreaming modelsdemand detectionAI benchmarks

0 comments

The pith

A streaming model for demand detection combined with hybrid long-term memory lets proactive agents infer latent user needs and intervene under real-time constraints.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to close the gap between laboratory proactivity and real-world demands by showing how agents can detect unspoken needs from ongoing context, store and retrieve evolving user memory across multiple horizons, and act without violating latency or ambiguity limits. It introduces the DD-MM-PAS structure as a reusable way to organize demand detection, memory modeling, and closed-loop action, then builds Pask around a streaming IntentFlow model for the detection step and a three-part memory store. A new benchmark drawn from consented user traces and human refinement is used to test the full loop, with results indicating that the detection component keeps pace with top fast models while surfacing deeper intent.

Core claim

The authors argue that a closed-loop proactive system becomes feasible once demand detection runs in a streaming fashion, memory is maintained as a hybrid of workspace, user-specific, and global stores, and the three elements feed one another continuously. In their Pask instantiation, the IntentFlow model performs demand detection while the memory component grounds longer-horizon actions, and the overall loop is evaluated on LatentNeeds-Bench, a dataset constructed from real consented interactions and refined by thousands of human edits. Under this setup the detection model matches the speed of leading fast language models while identifying more latent needs.

What carries the argument

The DD-MM-PAS paradigm, a three-part structure in which streaming demand detection infers latent needs, hybrid memory maintains context across time scales, and the proactive agent system executes grounded interventions in a closed loop.

If this is right

Agents built this way can maintain continuous awareness of user context without requiring explicit commands at every step.
Hybrid memory stores allow actions to be conditioned on both short-term workspace state and longer-term user patterns.
The closed loop supports ongoing refinement because detected needs and executed actions update the memory stores in turn.
The same paradigm can be re-instantiated with different detection models or memory back-ends while preserving the overall flow.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the approach scales, personal assistants could shift from responding to stated requests toward preempting routine friction points in daily workflows.
The memory modeling component could be extended to handle shared or multi-user contexts where needs are distributed across participants.
Because the benchmark emphasizes real-time constraints, the design implies that similar systems could be embedded in always-on devices without draining resources.

Load-bearing premise

The benchmark built from user-consented data and repeated human editing is representative enough of real-world depth, ambiguity, and timing pressures to confirm that the closed-loop system works outside the lab.

What would settle it

Deployment logs from the system running live with actual users showing either higher latency than reactive baselines or no measurable improvement in anticipating needs that users later confirm as relevant.

read the original abstract

Proactivity is a core expectation for AGI. Prior work remains largely confined to laboratory settings, leaving a clear gap in real-world proactive agent: depth, complexity, ambiguity, precision and real-time constraints. We study this setting, where useful intervention requires inferring latent needs from ongoing context and grounding actions in evolving user memory under latency and long-horizon constraints. We first propose DD-MM-PAS (Demand Detection, Memory Modeling, Proactive Agent System) as a general paradigm for streaming proactive AI agent. We instantiate this paradigm in Pask, with streaming IntentFlow model for DD, a hybrid memory (workspace, user, global) for long-term MM, PAS infra framework and introduce how these components form a closed loop. We also introduce LatentNeeds-Bench, a real-world benchmark built from user-consented data and refined through thousands of rounds of human editing. Experiments show that IntentFlow matches leading Gemini3-Flash models under latency constraints, while identifying deeper user intent.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PASK gives a clean framework for memory-aware proactive agents but the experiments only validate one component and leave the full closed-loop claims untested.

read the letter

The main thing to know is that this paper lays out DD-MM-PAS as a paradigm for streaming proactive agents and builds PASK around it, but the results shown do not yet prove the system works end-to-end under the real-world conditions it claims to target. The new elements are the overall paradigm, the IntentFlow model for low-latency intent detection, the hybrid memory split into workspace/user/global layers, the PAS infrastructure that ties them into a loop, and LatentNeeds-Bench constructed from consented user data with heavy human editing. These pieces are packaged together in a way that could help people move from lab-style agents to something that runs continuously with memory grounding. The paper does a reasonable job sketching how demand detection feeds memory modeling and then drives actions without breaking latency bounds. That integration is the part worth looking at if you work on agent memory or proactivity. The soft spots sit in the evaluation. The abstract reports that IntentFlow matches Gemini-3-Flash on latency while catching deeper intent, yet gives no baselines, metrics, data splits, or error analysis for that claim, and nothing at all on the full loop such as memory retrieval accuracy over multi-hour sessions or intervention success rates. The benchmark description mentions thousands of editing rounds but supplies no task distribution, inter-annotator numbers, or explicit tests for ambiguity and streaming constraints, so it is hard to judge whether the reported match actually supports the broader DD-MM-PAS story. The stress-test note is accurate on this point. This is for readers who build or study practical agent systems and want concrete ideas on memory structures and latency handling. It is not ready for citation yet because the evidence is incomplete, but the framework itself is coherent enough that a serious editor should send it to referees rather than desk-reject; the reviewers can ask for the missing end-to-end numbers and benchmark details.

Referee Report

3 major / 1 minor

Summary. The paper proposes the DD-MM-PAS paradigm (Demand Detection, Memory Modeling, Proactive Agent System) for streaming proactive AI agents that infer latent needs from context and ground actions in long-term user memory under latency and long-horizon constraints. It instantiates the paradigm in the PASK system using the IntentFlow model for demand detection, a hybrid memory architecture (workspace, user, global), and PAS infrastructure to form a closed loop. The work also introduces LatentNeeds-Bench, a benchmark constructed from user-consented data refined via thousands of rounds of human editing, and reports that IntentFlow matches leading Gemini-3-Flash models under latency constraints while identifying deeper user intent.

Significance. If the experimental claims are substantiated with full methodology and end-to-end metrics, the work could meaningfully advance proactive agents beyond laboratory settings by addressing real-world requirements for depth, ambiguity, precision, and real-time performance through intent-aware systems with persistent memory. The introduction of a general paradigm and a real-world benchmark represents a constructive contribution, though the current manuscript provides limited verifiable evidence for these advances.

major comments (3)

[Experiments] The experimental evaluation section provides no methodology details, baselines, metrics (e.g., exact latency thresholds, accuracy or F1 scores), error analysis, or data splits to support the claim that IntentFlow matches Gemini-3-Flash models under latency constraints while identifying deeper intent. This absence makes it impossible to assess or reproduce the reported performance.
[LatentNeeds-Bench] The LatentNeeds-Bench description lacks any disclosure of task distribution, inter-annotator agreement statistics, or concrete test cases exercising workspace/user/global memory retrieval under streaming real-time constraints. Without these, it is unclear whether the benchmark validates the closed-loop DD-MM-PAS claims.
[Closed-loop system] No end-to-end metrics for the full closed-loop system—such as intervention precision, memory-retrieval accuracy over multi-hour sessions, or user-level success rates—are reported. The central claim that IntentFlow + hybrid memory + PAS infrastructure handles streaming latent-need inference therefore rests on an untested assumption.

minor comments (1)

[Abstract] The abstract refers to 'Gemini3-Flash' without specifying the exact model version or release; this should be clarified for reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

Thank you for the thorough and constructive review. We appreciate the identification of areas where additional detail is required to substantiate the claims. We will revise the manuscript to include expanded experimental methodology, benchmark statistics, and available end-to-end metrics. Point-by-point responses follow.

read point-by-point responses

Referee: [Experiments] The experimental evaluation section provides no methodology details, baselines, metrics (e.g., exact latency thresholds, accuracy or F1 scores), error analysis, or data splits to support the claim that IntentFlow matches Gemini-3-Flash models under latency constraints while identifying deeper intent. This absence makes it impossible to assess or reproduce the reported performance.

Authors: We agree that the experimental section requires substantially more detail for reproducibility. In the revised manuscript we will add the full methodology, including the exact latency thresholds applied, accuracy and F1 scores for IntentFlow versus Gemini-3-Flash, the complete set of baselines, error analysis, and the train/validation/test splits used. These additions will directly support the reported performance claims. revision: yes
Referee: [LatentNeeds-Bench] The LatentNeeds-Bench description lacks any disclosure of task distribution, inter-annotator agreement statistics, or concrete test cases exercising workspace/user/global memory retrieval under streaming real-time constraints. Without these, it is unclear whether the benchmark validates the closed-loop DD-MM-PAS claims.

Authors: We acknowledge the need for greater transparency on the benchmark. The revision will include task distribution statistics, inter-annotator agreement figures from the multi-round human editing process, and concrete test-case examples that exercise workspace, user, and global memory retrieval under streaming constraints. This will clarify how the benchmark supports the DD-MM-PAS paradigm. revision: yes
Referee: [Closed-loop system] No end-to-end metrics for the full closed-loop system—such as intervention precision, memory-retrieval accuracy over multi-hour sessions, or user-level success rates—are reported. The central claim that IntentFlow + hybrid memory + PAS infrastructure handles streaming latent-need inference therefore rests on an untested assumption.

Authors: The current manuscript emphasizes component-level results and the benchmark; however, we recognize that end-to-end evaluation is essential. In the revision we will report all available end-to-end metrics (intervention precision and memory-retrieval accuracy) from the benchmark runs. For multi-hour session metrics we will add a discussion of current limitations and any preliminary aggregated results we can provide, while noting that full user-level longitudinal studies remain future work. These additions will strengthen the support for the closed-loop claims. revision: partial

Circularity Check

0 steps flagged

No significant circularity; claims rest on empirical description without self-referential reduction

full rationale

The paper introduces the DD-MM-PAS paradigm and its Pask instantiation as a descriptive framework for proactive agents, along with the LatentNeeds-Bench constructed from user data and human editing. No equations, derivations, fitted parameters, or mathematical predictions appear in the abstract or described components. Experimental claims (IntentFlow matching Gemini-3-Flash under latency while identifying deeper intent) are presented as direct results rather than reductions to prior inputs by construction. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing steps in the provided text. The derivation chain is therefore self-contained and non-circular, with central claims depending on external benchmark validation instead of internal redefinition.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 2 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities with supporting details; components like IntentFlow and the benchmark are introduced without upstream grounding.

invented entities (2)

IntentFlow model no independent evidence
purpose: Streaming demand detection under latency constraints
New model proposed for intent inference in the DD component.
LatentNeeds-Bench no independent evidence
purpose: Real-world benchmark for proactive agent evaluation
Constructed from user-consented data with human refinement.

pith-pipeline@v0.9.0 · 5519 in / 1210 out tokens · 35424 ms · 2026-05-10T18:01:52.757861+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 20 canonical work pages · 8 internal anchors

[1]

Funaudiollm: V oice understanding and generation foundation models for natural interaction between humans and llms.arXiv preprint arXiv:2407.04051, 2024

Keyu An, Qian Chen, Chong Deng, Zhihao Du, Changfeng Gao, Zhifu Gao, Yue Gu, Ting He, Hangrui Hu, Kai Hu, et al. Funaudiollm: Voice understanding and generation foundation models for natural interaction between humans and llms. arXiv preprint arXiv:2407.04051,

work page arXiv
[2]

Qwen3-VL Technical Report

Anthropic. Claude ai.https://www.anthropic.com/claude, 2025a. Accessed: 2026-03-15. Anthropic. Introducing claude haiku 4.5. https://www.anthropic.com/news/claude-haiku-4-5, 2025b. Official release post for Claude Haiku 4.5. Accessed: 2026-03-15. Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[3]

Ruisheng Cao, Mouxiang Chen, Jiawei Chen, Zeyu Cui, Yunlong Feng, Binyuan Hui, Yuheng Jing, Kaixin Li, Mingze Li, Junyang Lin, et al

Ruisheng Cao, Mouxiang Chen, Jiawei Chen, Zeyu Cui, Yunlong Feng, Binyuan Hui, Yuheng Jing, Kaixin Li, Mingze Li, Junyang Lin, et al. Qwen3-coder-next technical report.arXiv preprint arXiv:2603.00729,

work page arXiv
[4]

arXiv preprint arXiv:2310.05915 , year=

Baian Chen, Chang Shu, Ehsan Shareghi, Nigel Collier, Karthik Narasimhan, and Shunyu Yao. Fireact: Toward language agent fine-tuning.arXiv preprint arXiv:2310.05915,

work page arXiv
[5]

Need help? designing proactive ai assistants for programming

Valerie Chen, Alan Zhu, Sebastian Zhao, Hussein Mozannar, David Sontag, and Ameet Talwalkar. Need help? designing proactive ai assistants for programming. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems, pages 1–18,

2025
[6]

Moshi: a speech-text foundation model for real-time dialogue

Accessed: 2026-03-15. Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Hervé Jégou, Edouard Grave, and Neil Zeghidour. Moshi: a speech-text foundation model for real-time dialogue.arXiv preprint arXiv:2410.00037,

work page internal anchor Pith review arXiv 2026
[7]

Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to-end speech recognition.arXiv preprint arXiv:2206.08317, 2022

22 Zhifu Gao, Shiliang Zhang, Ian McLoughlin, and Zhijie Yan. Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to-end speech recognition.arXiv preprint arXiv:2206.08317,

work page arXiv
[8]

Gemini 2.5 flash-lite

Google. Gemini 2.5 flash-lite. https://ai.google.dev/gemini-api/docs/models/gemini-2.5-flash-lite, 2025a. Official Gemini API model documentation. Accessed: 2026-03-15. Google. Gemini 3 flash. https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/ 3-flash, 2025b. Official Vertex AI model documentation. Accessed: 2026-03-15. Google DeepM...

2026
[9]

Zhenpeng Huang, Xinhao Li, Jiaqi Li, Jing Wang, Xiangyu Zeng, Cheng Liang, Tao Wu, Xi Chen, Liang Li, and Limin Wang

Accessed: 2026-03-15. Zhenpeng Huang, Xinhao Li, Jiaqi Li, Jing Wang, Xiangyu Zeng, Cheng Liang, Tao Wu, Xi Chen, Liang Li, and Limin Wang. Online video understanding: Ovbench and videochat-online. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 3328–3338,

2026
[10]

Hello again! llm-powered personalized agent for long-term dialogue

Hao Li, Chenghao Yang, An Zhang, Yang Deng, Xiang Wang, and Tat-Seng Chua. Hello again! llm-powered personalized agent for long-term dialogue. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 5259–5276,

2025
[11]

Improving multi-agent debate with sparse communication topology

Yunxuan Li, Yibing Du, Jiageng Zhang, Le Hou, Peter Grabowski, Yeqing Li, and Eugene Ie. Improving multi-agent debate with sparse communication topology. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 7281–7294,

2024
[12]

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, et al. Deepseek-v3. 2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556,

work page internal anchor Pith review Pith/arXiv arXiv
[13]

Proactive agent: Shifting llm agents from re- active responses to active assistance,

Yaxi Lu, Shenzhi Yang, Cheng Qian, Guirong Chen, Qinyu Luo, Yesai Wu, Huadong Wang, Xin Cong, Zhong Zhang, Yankai Lin, et al. Proactive agent: Shifting llm agents from reactive responses to active assistance.arXiv preprint arXiv:2410.12361,

work page arXiv
[14]

Omni-captioner: Data pipeline, models, and benchmark for omni detailed perception,

Ziyang Ma, Ruiyang Xu, Zhenghao Xing, Yunfei Chu, Yuxuan Wang, Jinzheng He, Jin Xu, Pheng-Ann Heng, Kai Yu, Junyang Lin, et al. Omni-captioner: Data pipeline, models, and benchmark for omni detailed perception.arXiv preprint arXiv:2510.12720,

work page arXiv
[15]

Karen Myers and Neil Yorke-Smith

Accessed: 2026-03-15. Karen Myers and Neil Yorke-Smith. Proactive behavior of a personal assistive agent. InProceedings of the AAMAS Workshopon Metareasoning in Agent-Based Systems, Honolulu, HI, pages 31–45,

2026
[16]

Gpt-5 mini

OpenAI. Gpt-5 mini. https://developers.openai.com/api/docs/models/gpt-5-mini, 2025a. OpenAI API model documentation. Accessed: 2026-03-15. OpenAI. Gpt-5 nano. https://developers.openai.com/api/docs/models/gpt-5-nano, 2025b. OpenAI API model documentation. Accessed: 2026-03-15. OpenAI. gpt-oss-120b & gpt-oss-20b model card.https://openai.com/index/gpt-oss-...

2026
[17]

OpenClaw Team

Accessed: 2026-03-15. OpenClaw Team. Openclaw: Open computer-use agent workspace.https://github.com/OpenClaw/OpenClaw,

2026
[18]

Siva Karthik Parimi and Rajesh Cherukuri

Accessed: 2026-03-15. Siva Karthik Parimi and Rajesh Cherukuri. Proactive ai systems: Engineering intelligent platforms that sense, predict, and act.International Journal of Emerging Trendsin Computer Science and Information Technology, 5(3):122–130,

2026
[19]

Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever

Accessed: 2026-03-15. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. InInternational conference on machine learning, pages 28492–28518. PMLR,

2026
[20]

Zep: A Temporal Knowledge Graph Architecture for Agent Memory

23 Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais, Jack Ryan, and Daniel Chalef. Zep: a temporal knowledge graph architecture for agent memory.arXiv preprint arXiv:2501.13956,

work page internal anchor Pith review arXiv
[21]

Qwen3-ASR Technical Report

Xian Shi, Xiong Wang, Zhifang Guo, Yongqi Wang, Pei Zhang, Xinyu Zhang, Zishan Guo, Hongkun Hao, Yu Xi, Baosong Yang, et al. Qwen3-asr technical report.arXiv preprint arXiv:2601.21337,

work page internal anchor Pith review arXiv
[22]

MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling

MiroMind Team, Song Bai, Lidong Bing, Carson Chen, Guanzheng Chen, Yuntao Chen, Zhe Chen, Ziyi Chen, Jifeng Dai, Xuan Dong, et al. Mirothinker: Pushing the performance boundaries of open-source research agents via model, context, and interactive scaling.arXiv preprint arXiv:2511.11793,

work page internal anchor Pith review Pith/arXiv arXiv
[23]

Rema: Learning to meta-think for llms with multi-agent reinforcement learning.CoRR, abs/2503.09501, 2025

Ziyu Wan, Yunxiang Li, Xiaoyu Wen, Yan Song, Hanjing Wang, Linyi Yang, Mark Schmidt, Jun Wang, Weinan Zhang, Shuyue Hu, et al. Rema: Learning to meta-think for llms with multi-agent reinforcement learning.arXiv preprint arXiv:2503.09501,

work page arXiv
[24]

Emotionthinker: Prosody-aware reinforcement learning for explainable speech emotion reasoning.arXiv preprint arXiv:2601.15668,

Dingdong Wang, Shujie Liu, Tianhua Zhang, Youjun Chen, Jinyu Li, and Helen Meng. Emotionthinker: Prosody-aware reinforcement learning for explainable speech emotion reasoning.arXiv preprint arXiv:2601.15668,

work page arXiv
[25]

Videollm knows when to speak: Enhancing time-sensitive video comprehension with video-text duet interaction format

Yueqian Wang, Xiaojun Meng, Yuxuan Wang, Jianxin Liang, Jiansheng Wei, Huishuai Zhang, and Dongyan Zhao. Videollm knows when to speak: Enhancing time-sensitive video comprehension with video-text duet interaction format. arXiv preprint arXiv:2411.17991, 1(3):5,

work page arXiv
[26]

DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understanding

Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302,

work page internal anchor Pith review arXiv
[27]

Xie and C

Zhifei Xie and Changqiao Wu. Mini-omni: Language models can hear, talk while thinking in streaming.arXiv preprint arXiv:2408.16725, 2024a. Zhifei Xie and Changqiao Wu. Mini-omni2: Towards open-source gpt-4o with vision, speech and duplex capabilities. arXiv preprint arXiv:2410.11190, 2024b. Jin Xu, Zhifang Guo, Hangrui Hu, Yunfei Chu, Xiong Wang, Jinzheng...

work page arXiv
[28]

ProAgent: Harnessing On-Demand Sensory Contexts for Proactive LLM Agent Systems in the Wild

Bufang Yang, Lilin Xu, Liekang Zeng, Yunqi Guo, Siyang Jiang, Wenrui Lu, Kaiwei Liu, Hancheng Xiang, Xiaofan Jiang, Guoliang Xing, et al. Proagent: Harnessing on-demand sensory contexts for proactive llm agent systems. arXiv preprint arXiv:2512.06721, 2025a. Bufang Yang, Lilin Xu, Liekang Zeng, Kaiwei Liu, Siyang Jiang, Wenrui Lu, Hongkai Chen, Xiaofan Ji...

work page internal anchor Pith review Pith/arXiv arXiv
[29]

Speakerlm: End-to-end versa- tile speaker diarization and recognition with multimodal large lan- guage models,

Han Yin, Yafeng Chen, Chong Deng, Luyao Cheng, Hui Wang, Chao-Hong Tan, Qian Chen, Wen Wang, and Xiangang Li. Speakerlm: End-to-end versatile speaker diarization and recognition with multimodal large language models. arXiv preprint arXiv:2508.06372,

work page arXiv
[30]

Proagent: building proactive cooperative agents with large language models

Ceyao Zhang, Kaijie Yang, Siyi Hu, Zihao Wang, Guanghe Li, Yihang Sun, Cheng Zhang, Zhaowei Zhang, Anji Liu, Song-Chun Zhu, et al. Proagent: building proactive cooperative agents with large language models. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17591–17599, 2024a. Shaolei Zhang, Shoutao Guo, Qingkai Fang, Yan Zh...

work page arXiv
[31]

Ask-before-plan: Proactive language agents for real-world planning

Xuan Zhang, Yang Deng, Zifeng Ren, See Kiong Ng, and Tat-Seng Chua. Ask-before-plan: Proactive language agents for real-world planning. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 10836–10863, 2024b. Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. Expel: Llm agents are experientiallearne...

2024