A safety decision-making framework for large language models in cotton production integrates multi-source heterogeneous data and architecture-level hallucination mitigation
Existing large language model (LLM) agent frameworks lack semantic reliability verification for multi-source heterogeneous agricultural perception data, leaving them vulnerable to the “garbage in, garbage out”(GIGO) effect and catastrophic hallucinated decisions. This study proposes a four-layer functionally decoupled safety architecture grounded in an “untrusted-by-default” design principle and evaluates its decision-making safety in Xinjiang cotton production scenarios. The architecture reduces the catastrophic hallucination rate from 45% (tightly coupled baseline) to 13%, a markedly greater reduction than that achieved by the soft-validation reflective baseline (29%). Ablation experiments confirm the independent contributions of the perception-layer hard validation and symbolic audit modules, which together constitute a defense-in-depth mechanism. On the Xinjiang Cotton Benchmark, a domain-adapted 8B model attains 72.1% accuracy-surpassing Llama3-70B (61.3%) with approximately one-ninth of the parameters. The architecture provides a trustworthy, efficient, and transferable solution for real-time multi-source heterogeneous data fusion and decision-making under computational constraints in smart cotton production.
S3与S5的对比进一步揭示了两模块的协同价值:二者均配备符号审计与闭环反馈,尽管S3因失去前线拦截导致更多脏数据涌入决策层,迫使符号审计被动触发了更高的拒答率(21% vs. 16%),但其CHR依然大幅回升至27%。更深层次的逻辑在于系统可用性:完整架构(S5)在保持最低幻觉率(13%)的同时,实现了高达71%的安全放行率(即系统在感知层精准、静默地剔除单一污染源后,仍能基于富余的干净上下文做出完全正确的决策);而S3架构的安全放行率仅为52%。上述结果验证了纵深防御设计的有效性:单独依赖任一模块均无法达到完整架构的安全水平,两者协同才能系统性地将CHR从45%压降至13%,对13%残留失败案例(共13例)进行定性分析,可归纳为两类边界情形:1)阈值邻近型(9例),污染数据的关键元指标(如云覆盖率32%)仅略超阈值,时序积分统计的正常浮动导致个别场景中规则未被触发;2)多源协同降级型(4例),气象API轻度延迟与遥感影像轻度污染同时发生,各单项指标均未达到硬拒绝阈值,但综合数据质量已低于可靠决策所需的信息下限。这些失败模式均属规则覆盖盲区,而非软约束意义上的推理局限,为后续多元协同降级联合规则与自适应阈值校准研究指明了方向。
YaoS, ZhaoJ, YuD, DuN, ShafranI, NarasimhanK, CaoY. ReAct: Synergizing reasoning and acting in language models[EB/OL].[--].
[4]
HongS, ZhugeM, ChenJ, ZhengX W, ChengY H, WangJ L, ZhangC Y, YauS, LinZ J, ZhouL Y. MetaGPT: Meta programming for a multi-agent collaborative framework [C]. In: The Twelfth International Conference on Learning Representations. Vienna: International Conference on Learning Representations, 2024: 23247-23275
[5]
LinQ, XuL, WuS, MaoR, WangC, FengH, HuangB, DuZ. GeoAgent: A hierarchical LLM-based multi-agent architecture for autonomous spatial analysis [J]. International Journal of Geographical Information Science, 2026: 1-40
[6]
WeiJ, WangX Z, SchuurmansD, BosmaM, IchterB, XiaF, ChiE H, LeQ V, ZhouD N. Chain-of-thought prompting elicits reasoning in large language models [C]. In: Advances in Neural Information Processing Systems. Red Hook: Curran Associates, 2022, 35: 24824-24837
[7]
LewisP, PerezE, PiktusA, PetroniF, KarpukhinV, GoyalN, KüttlerH, LewisM, YihW T, RocktäschelT, RiedelS, KielaD. Retrieval-augmented generation for knowledge-intensive NLP tasks [C]. In: Advances in Neural Information Processing Systems. Red Hook: Curran Associates, 2020, 33: 9459-9474
[8]
MadaanA, TandonN, GuptaP, HallinanS, GaoL Y, WiegreffeS, AlonU, DziriN, PrabhumoyeS, YangY M, GuptaS, MajumderB P, HermannK, WelleckS, YazdanbakhshA, ClarkP. Self-refine: Iterative refinement with self-feedback [C]. In: Advances in Neural Information Processing Systems. Red Hook: Curran Associates, 2023, 36: 46534-46594
[9]
TonmoyS M I, ZamanS M M, JainV, RaniA, RawteV, ChadhaA, DasA. A comprehensive survey of hallucination mitigation techniques in large language models[EB/OL].[--].
[10]
JiZ W, LeeN, FrieskeR, YuT Z, SuD, XuY, IshiiE, BangY J, MadattoA, FungP. Survey of hallucination in natural language generation[J]. ACM Computing Surveys, 2023, 55(12): 1-38
[11]
PanY, SunJ X, YuH F, LuckJ, BaiG, ChamaraN, GeY F, AwadaT. Building multi-agent copilot towards autonomous agricultural data management and analysis[C]. In: 2024 IEEE International Conference on Big Data. Washington DC: Institute of Electrical and Electronics Engineers, 2024: 4384-4393
[12]
HuE J, ShenY L, WallisP, Allen-ZhuZ, LiY Z, WangS, WangL, ChenW Z. LoRA: Low-rank adaptation of large language models [EB/OL].[--].
[13]
HintonG, VinyalsO, DeanJ. Distilling the knowledge in a neural network[EB/OL].[--].
[14]
ZhuX Y, LiJ, LiuY, MaC, WangW P. A survey on model compression for large language models[J]. Transactions of the Association for Computational Linguistics, 2024, 12: 1556-1577
[15]
ColvinS. Pydantic: Data validation using Python type hints[CP/OL].[--].
[16]
GorelickN, HancherM, DixonM, IlyushchenkoS, ThauD, MooreR. Google Earth Engine: Planetary-scale geospatial analysis for everyone [J]. Remote Sensing of Environment, 2017, 202: 18-27
[17]
WuC Y, GanY K, GeY X, LuZ Y, WangJ H, FengY, ShanY, LuoP. LLaMA Pro: Progressive LLaMA with block expansion[C].In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: )Long Papers. Bangkok: Association for Computational Linguistics, 2024: 6518-6537.
ChenJ L, XiaoS T, ZhangP T, LuoK, LianD F, LiuZ. BGE M3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation [C].In: Findings of the Association for Computational Linguistics: ACL 2024. Bangkok: Association for Computational Linguistics, 2024: 2318-2335
[20]
JohnsonJ, DouzeM, JégouH. Billion-scale similarity search with GPUs[J]. IEEE Transactions on Big Data, 2019, 7(3): 535-547
[21]
BhuyanB P, Ramdane-CherifA, TomarR, SinghT P. Neuro-symbolic artificial intelligence: A survey [J]. Neural Computing and Applications, 2024, 36(21): 12809-12844