Code & Showcase | CeresAgents

CeresAgents — Code & Showcase

Repository: github.com/GZU-SAMLab/CeresAgents

CeresBench: Benchmark data and annotations

Task-Adaptive Multi-Agent Workflow

A coordinator standardizes the user query, image, location, environmental context, and management history into a structured consultation state. It then selects either a fast generalist route or an expert route. On the expert route, pathology, physiology, and chemistry specialists address complementary parts of the consultation and exchange findings through a shared blackboard.

User consultation
  ├── Intent-based coordinator
  ├── Pathologist → disease hypotheses and visual evidence
  ├── Physiologist → environmental plausibility and abiotic alternatives
  ├── Chemist → registration, dosage, and label-use verification
  ├── Shared blackboard → evidence exchange and conflict reassessment
  └── Final synthesis → diagnosis, management, safety, and uncertainty

Evidence Resources

The evidence layer combines Qwen3-VL-plus for structured image description, EfficientNet-V2-S for disease classification, separate SegFormer models for visible-leaf and lesion segmentation, hybrid BGE-M3/FAISS/BM25 literature retrieval, and Cypher queries over a Neo4j pesticide knowledge graph.

**Figure 4:** Evidence-resource construction: crop–disease image distribution, image collection and annotation, literature preparation, and pesticide knowledge-graph construction.

Overall Results on CeresBench

267manually curated cases

94.89%diagnostic accuracy

85.13safety score

84.60end-to-end quality

System	Diag. acc. (%)	Diag. score	Prescription	Safety	E2E	Conflict
GPT-5	48.91	47.00	61.57	49.01	54.51	78.00
Claude-Sonnet-4.5	47.45	45.23	53.62	42.77	50.56	80.00
Qwen3.5-Plus	61.31	58.61	64.73	54.68	62.78	77.50
Gemini-3-Pro	70.07	64.60	69.82	54.53	65.19	81.00
CeresAgents	94.89	91.67	79.93	85.13	84.60	89.00

Architecture and Evidence Ablations

Configuration	Diagnosis	Prescription	Safety	End-to-end
Full system	91.67	79.93	85.13	84.60
w/o RAG	79.27	75.00	78.85	75.53
w/o KG	88.59	30.43	30.92	70.27
w/o CV	64.61	69.46	66.21	63.54
w/o multi-agent	73.87	74.45	68.44	75.18

Consultation and Human Evaluation

**Figure 5:** Representative end-to-end consultation, from field inputs and specialist assessments to pesticide verification and final advice.

**Figure 6:** Blinded human evaluation. Among 516 preference records, 74% favored CeresAgents, 22% favored Gemini-3-Pro, and 4% indicated no clear preference.

Code & Artifacts

Code: github.com/GZU-SAMLab/CeresAgents

Benchmark: CeresBench benchmark