ACM MobiCom 2026·CCF-A
Zhaomeng Zhou, Lan Zhang, Junyang Wang, Mu Yuan, Junda Lin, Jinke Song
A verify-before-commit spatial trajectory graph that grounds LLMs in deciding what to sense and when across large-scale sensor networks.
Also at the ACM ANAI Workshop @ MobiCom 2025. The TopoSense benchmark is part of Microsoft's System Intelligence Benchmark.
Intelligent systems powered by large-scale sensor networks are shifting from predefined monitoring to intent-driven operation, revealing a critical Semantic-to-Physical Mapping Gap. While large language models (LLMs) excel at semantic understanding, existing perception-centric pipelines operate retrospectively, overlooking the fundamental decision of what to sense and when. We formalize this proactive decision as Semantic-Spatial Sensor Scheduling (S3) and demonstrate that direct LLM planning is unreliable due to inherent gaps in representation, reasoning, and optimization. To bridge these gaps, we introduce the Spatial Trajectory Graph (STG), a neuro-symbolic paradigm governed by a verify-before-commit discipline that transforms open-ended planning into a verifiable graph optimization problem. Based on STG, we implement IoT-Brain, a concrete system embodiment, and construct TopoSense-Bench, a campus-scale benchmark with 5,250 natural-language queries across 2,510 cameras. Evaluations show that IoT-Brain boosts task success rate by 37.6% over the strongest search-intensive methods while running nearly 2 times faster and using 6.6 times fewer prompt tokens. In real-world deployment, it approaches the reliability upper bound while reducing 4.1 times network bandwidth, providing a foundational framework for LLMs to interact with the physical world with unprecedented reliability and efficiency.
@misc{zhou2026iotbraingroundingllmssemanticspatial,
title={IoT-Brain: Grounding LLMs for Semantic-Spatial Sensor Scheduling},
author={Zhaomeng Zhou and Lan Zhang and Junyang Wang and Mu Yuan and Junda Lin and Jinke Song},
year={2026},
eprint={2604.08033},
archivePrefix={arXiv},
primaryClass={cs.AI},
doi={https://doi.org/10.1145/3795866.3796695},
url={https://arxiv.org/abs/2604.08033},
}
ACL 2026 Main Conference·CCF-A
Junda Lin*, Zhaomeng Zhou*, Zhi Zheng, Shuochen Liu, Tong Xu, Yong Chen, Enhong Chen
A verify-before-commit protocol that defends LLM agents against tool-stream injection while preserving reasoning flexibility.
LLM agents operating in open environments face escalating risks from indirect prompt injection, particularly within the tool stream where manipulated metadata and runtime feedback hijack execution flow. Existing defenses encounter a critical dilemma as advanced models prioritize injected rules due to strict alignment while static protection mechanisms sever the feedback loop required for adaptive reasoning. To reconcile this conflict, we propose VIGIL, a framework that shifts the paradigm from restrictive isolation to a verify-before-commit protocol. By facilitating speculative hypothesis generation and enforcing safety through intent-grounded verification, VIGIL preserves reasoning flexibility while ensuring robust control. We further introduce SIREN, a benchmark comprising 959 tool stream injection cases designed to simulate pervasive threats characterized by dynamic dependencies. Extensive experiments demonstrate that VIGIL outperforms state-of-the-art dynamic defenses by reducing the attack success rate by over 22% while more than doubling the utility under attack compared to static baselines, thereby achieving an optimal balance between security and utility.
@inproceedings{lin-etal-2026-vigil,
title = "{VIGIL}: Defending {LLM} Agents Against Tool-Stream Injection via Verify-Before-Commit",
author = "Lin, Junda and
Zhou, Zhaomeng and
Zheng, Zhi and
Liu, Shuochen and
Xu, Tong and
Chen, Yong and
Chen, Enhong",
editor = "Liakata, Maria and
Moreira, Viviane P. and
Zhang, Jiajun and
Jurgens, David",
booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.acl-long.443/",
doi = "10.18653/v1/2026.acl-long.443",
pages = "9764--9785",
ISBN = "979-8-89176-390-6"
}
ACL 2026 Main Conference·CCF-A
Jiawei Cao, Jie Ouyang, Mingyue Cheng, Zhaomeng Zhou, Chunli Liu, Yupeng Li, Zirui Liu, Shijin Wang
Balancing relevance and recency for temporal information retrieval to curb temporal hallucination.
Retrieval-Augmented Generation (RAG) is a mainstream approach to mitigating hallucinations in Large Language Models (LLMs), yet in dynamic real-world scenarios, such as weather forecasting or evolving news events, existing retrievers suffer from both temporal-semantic misalignment and outdated-document interference. To address this, we propose Relevance Recency Retrieval (Re³), a novel framework that mitigates temporal hallucinations via two core components: a Time-Aware Dual Relevance Encoder that embeds heterogeneous temporal signals into the semantic space to ensure retrieval fidelity, and a Conflict-Aware Recency Filter that performs listwise arbitration to identify and suppress obsolete factual versions. To rigorously evaluate this setting, we introduce Re² Bench, a large-scale benchmark comprising over 1.3 million instances designed to assess system robustness in realistic environments where temporal constraints and conflicting factual versions coexist. Experiments on three public benchmarks and Re² Bench demonstrate that Re³ consistently outperforms the strongest baselines by an average of 9.7% in generation accuracy, with gains of up to 25.2% on challenging dynamic tasks, while demonstrating robustness across diverse RAG settings.
@inproceedings{cao-etal-2026-re3,
title = "{R}e$^3$: Relevance {\&} Recency Retrieval for Mitigating Temporal Hallucination",
author = "Cao, Jiawei and
Ouyang, Jie and
Cheng, Mingyue and
Zhou, Zhaomeng and
Liu, Chunli and
Li, Yupeng and
Liu, Zirui and
Wang, Shijin",
editor = "Liakata, Maria and
Moreira, Viviane P. and
Zhang, Jiajun and
Jurgens, David",
booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.acl-long.1180/",
doi = "10.18653/v1/2026.acl-long.1180",
pages = "25735--25760",
ISBN = "979-8-89176-390-6"
}
SCIENTIA SINICA Informationis 2026·Position Paper
Yunhao Liu, Mu Yuan, Lan Zhang, Zhaomeng Zhou, Li Liu
A foresight-oriented decision-making paradigm that brings action-conditioned counterfactual prediction into the agent decision loop.
Agent systems have achieved substantial progress in perception, planning, and execution. From Sensing-in-the-Loop (SITL) to Human-in-the-Loop (HITL), existing decision paradigms have successively opened up the perception channel and the instruction channel of agents, enabling agents to respond to the physical world and follow human instructions. In long-horizon, irreversible, and implicitly constrained multi-step decisions, human instructions cannot fully specify all the rules and constraints in advance; an agent that merely samples passively within the given instruction space is highly prone to falling into short-horizon traps that only satisfy immediate needs and to triggering latent long-term risks. The underlying structural gap is that the agent lacks a mechanism to actively construct and compare multiple candidate futures before acting. To address the issue, we introduce Prediction-in-the-Loop (PITL), a foresight-oriented decision-making paradigm for the agent decision loop, which brings action-conditioned counterfactual prediction into the loop as a first-class component—the third in-loop constituent after perception and human feedback—advancing the agent from “deciding based on the present” to “deciding based on the prediction”. We propose the Future-Belief MDP framework, formalizing the prediction–decision coupling in the agent decision loop as a Markov decision process over a future trajectory belief space, and prove core properties including degeneration conditions, the calibration error–decision regret correlation, and existence conditions for the value of foresight. Centered around three progressive scientific questions—how to predict, evaluate, and utilize the prediction—we systematically articulate the challenges that the PITL paradigm faces and use the Future-Belief MDP as a unified yardstick to position related techniques as partial PITL realizations. On the system and empirical front, we develop and open-source a multi-model collaborative PITL prototype (ForesightLM) together with the Foresight Decision Benchmark; experimental results show that, relative to Chain-of-Thought and Tree-of-Thoughts, PITL yields consistent improvements on foresight-critical decisions across model families and decision domains, and offers more cost-efficient operating points on the cost–quality trade-off. Our study seeks to advance autonomous agents from passive instruction-following to a decision style endowed with foresight, providing a unified conceptual framework and reference for research on intelligent decision systems.
@article{LIU_2026,
title = {Prediction-in-the-Loop: a decision-making paradigm endowing agents with foresight},
volume = {56},
ISSN = {1674-7267},
url = {http://dx.doi.org/10.1360/SSI-2026-0130},
DOI = {10.1360/ssi-2026-0130},
number = {8},
journal = {SCIENTIA SINICA Informationis},
publisher = {Science China Press., Co. Ltd.},
author = {Liu, Yunhao and Yuan, Mu and Zhang, Lan and Zhou, Zhaomeng and Liu, Li},
year = {2026},
month = aug,
pages = {1867}
}
