Reclamation of saline-alkali land requires management plans that account for soil and water conditions, treatment evidence, and financial constraints. This study develops Wenyao, an agricultural decision-support system integrating domain adaptation with retrieval-augmented generation. A bilingual corpus of 8,664 research papers was processed using a 110-indicator schema to construct SA-Bench, comprising 4,860 paper-level planning instances. A QwQ-32B model was adapted through low-rank supervised fine-tuning and combined with case retrieval for budget-conditioned planning. Evidence extraction achieved an F1-score of 0.83 on 300 expert-annotated papers. Without retrieval, Wenyao achieved 95.83% accuracy on 936 source-disjoint multiple-choice questions, compared with 87.93% for the strongest general-purpose baseline; removing explanatory supervision reduced accuracy to 89.42%. Matched-evidence comparisons across 69 region-budget cases supported complementary contributions from adaptation and retrieval. In a blind assessment by 31 experts, Wenyao with retrieval obtained the highest mean scores for objectives and projected outcomes, whereas GPT-5 scored highest for cost estimates. These findings support literature-informed reclamation planning with professional review, while ongoing field monitoring will assess soil and crop responses.
Figures & tables
Figure 1: System workflow connecting literature-derived evidence, model adaptation, site-specific analysis, case retrieval, and budget-aware planning. Multiple-choice evaluation uses the adapted model without the online retrieval branch. Only the training partition of SA-Bench is used for supervised adaptation.
Group
Entries
Representative content
Experimental information
5
Location, coordinates, experiment type, meteorological conditions, and duration.
Initial land indicators
35
EC, salt content, pH and measurement method, ESP, SAR, ions, nutrients, texture, water status, and groundwater.
Measures and costs
4
Measure or amendment, application rate, timing, and cost.
Crop growth
24
Crop and cultivar, salt tolerance, physiology, canopy and root traits, water use, and nutrient or salt content.
Post-treatment effects
37
Biomass, yield, and post-treatment soil, water, ion, and nutrient indicators.
Comprehensive benefits
5
Yield increase, salinity or alkalinity reduction, input–output ratio, soil carbon, and microbial responses.
Table 1: The six groups and 110 entries in the extraction schema. Counts refer to schema entries, including distinct pre- and post-treatment fields.
Figure 2: Literature-to-evidence workflow. Documents shorter than 30,000 tokens supply full-text context; longer documents use heading-based hybrid retrieval and cross-encoder reranking. Six indicator groups guide structured extraction. Unit harmonization and expert checks support comparison with the original papers. The expert-checked sample contains 300 papers, and the retained resource contains 4,860 paper-level instances.
Statistic
Value
Initial source papers
8,664
English / Chinese source papers
4,486 / 4,178
Retained papers / planning instances
4,860 / 4,860
Training instances / held-out MCQs
3,924 / 936
Input tokens, minimum / mean / maximum
66 / 453 / 1,605
Output tokens, minimum / mean / maximum
291 / 1,194 / 2,862
Table 2: Paper-level corpus, training/test counts, and instance-length statistics. Each retained source paper contributes one instance.
Figure 3: Assigned SA-Bench labels across salinity severity, soil type, land type, and intervention category. Undetermined labels are not plotted, so the totals vary across dimensions. These are full-resource counts; held-out category counts appear in Appendix C .
Configuration
Precision
Recall
F1
95% CI
Full pipeline
0.76
0.92
0.83
[0.80, 0.85]
pdfplumber only
0.60
0.62
0.61
[0.58, 0.63]
BM25 retrieval only
0.72
0.85
0.78
[0.75, 0.80]
Dense retrieval only
0.70
0.83
0.77
[0.74, 0.79]
Without multi-round batching
0.74
0.76
0.75
[0.73, 0.78]
Simplified extraction prompts
0.66
0.90
0.76
[0.75, 0.78]
Table 3: Evidence-extraction ablations on 300 expert-annotated papers. Intervals are the reported 95% bootstrap intervals for F1.
Figure 4: Multiple-choice accuracy and mean response time on 936 held-out questions without retrieval. Response times describe the reported serving environments; the time axis is logarithmic.
First system versus second system
DeepSeek-R1
GPT-5
Wenyao versus Qwen3-Max
65.22
62.32
Wenyao versus Qwen3-Max + RAG
55.07
52.17
Wenyao + RAG versus Wenyao
66.67
60.87
Wenyao + RAG versus Qwen3-Max
76.81
75.36
Wenyao + RAG versus Qwen3-Max + RAG
60.87
57.97
Qwen3-Max + RAG versus Qwen3-Max
72.46
69.57
Table 4: Preference rates (%) favoring the first system. Each judge evaluates the same 69 region–budget cases; values are 100W/69 , rounded to two decimal places.
System
Objectives
Cost
Conclusions
Wenyao + RAG
6.89
6.95
7.60
Qwen3-Max
6.80
6.98
6.80
DeepSeek-R1
6.55
7.00
7.10
GPT-5
6.79
7.13
7.43
Table 5: Primary assessment by 31 experts from 22 institutes. The 36 questionnaires cover 12 regions sampled from the 23-region automatic evaluation pool, with three budgets per region. Each questionnaire has five or six expert ratings on a 0–10 scale.
System
Objectives
Cost
Conclusions
Wenyao
6.82
6.63
6.90
Wenyao + RAG
7.16
6.92
7.68
Qwen3-Max
6.67
6.94
6.86
Qwen3-Max + RAG
6.96
7.02
7.38
Table 6: Supplementary matched-evidence assessment by six experts over eight regions and three budgets, giving 24 region–budget cases. Means are reported separately from the primary study.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Component
Reported setting
Structured extraction model
DeepSeek-R1-Distill-Qwen-32B
Planning model initialization
QwQ-32B
Framework / adaptation
LLaMA-Factory; LoRA
Learning rate / optimizer
5×10−5 / AdamW
Epochs / LoRA rank
6 / 8
LoRA alpha / dropout
16 / 0
Appendix
Table A1: Adaptation, generation, and extraction-evaluation settings.
Table A2: Retrieval and reranking settings. Library defaults correspond to the reported software versions. The extraction and planning paths have different final retrieval stages.
Field
Source-case content
Location and soil
Xiaotun Township, Linze County, Gansu, China; Heihe River Basin; meadow solonchak.
Initial conditions
pH 8.45; salt content 8.18 g kg -1 ; groundwater depth 1.50 m; bulk density 1.43 g cm -3 ; organic matter 8.50 g kg -1 .
Climate
Mean annual precipitation 109 mm; mean annual evaporation 1,966 mm; mean annual temperature 7.8 ∘ C.
Reported intervention
Three years of alfalfa cultivation with manure, mineral fertilizer, and irrigation. The full source instance specifies treatment quantities and timing.
Reported soil response
At 0–20 cm, pH decreases from 8.45 to 8.20; salt content decreases by 4.63 g kg -1 ; reported desalination rate 56.60%; bulk density decreases by 0.24 g cm -3 .
Subsequent yield
Spring wheat 6.57 t ha -1 and maize 9.84 t ha -1 after ploughing the alfalfa.
Appendix
Table A3: Selected fields from the Linze County evidence-to-plan example. Outcomes were reported for the source treatment and duration.
Dimension
Category
n
Wenyao
QwQ
Qwen3
V3
R1
GPT-5
Gemini
Severity
Mild
171
93.57
74.85
82.46
81.29
80.12
78.36
80.12
Moderate
185
96.22
80.54
84.86
84.86
78.38
82.16
80.54
Severe
108
95.37
81.48
84.26
86.11
81.48
76.85
86.11
Soil
Chloride
212
96.23
85.85
91.04
90.09
86.32
88.68
86.32
Sulphate
48
93.75
70.83
81.25
81.25
75.00
75.00
83.33
Sodic
285
96.14
80.70
83.86
84.56
78.95
81.05
82.46
Appendix
Table A4: Measure-selection accuracy (%) without retrieval. QwQ denotes QwQ-32B; Qwen3 denotes Qwen3-Max; V3 and R1 denote DeepSeek-V3 and DeepSeek-R1; Gemini denotes Gemini-2.5-Pro.
Comparison model
DeepSeek-R1 judge
GPT-5 judge
QwQ-32B
79.71
76.81
Qwen3-Max
76.81
75.36
DeepSeek-V3
63.77
60.87
GPT-5
59.42
—
Gemini-2.5-Pro
60.87
75.36
DeepSeek-R1
—
65.22
Appendix
Table A5: Preference rates (%) favoring Wenyao + RAG across 69 region–budget comparisons. A dash denotes an omitted self-comparison.
Just as Socrates recognized the limits of his own knowledge, Retrieval-Augmented Language Models (RALMs) should learn to abstain when the retrieved evidence cannot support a reliable response. Existing approaches largely rely on monolithic LLMs to handle heterogeneous retrieval failures in a single step, resulting in limited abstention performance and high computational costs. We instead decompose retrieval failures into two distinct states: (i) the unanswerable state, where the required evidence is absent, and (ii) the distracted state, where relevant evidence is mixed with conflicting, negated, or adversarial information. Based on this decomposition, we introduce a lightweight module (Sieve) that screens retrieved document sets for distracting evidence before invoking a costly LLM (Sage) for grounded generation and abstention. Evaluated across both general and high-stakes expert domains, our Sieve and Sage framework preemptively detects distracting noise, improving system accuracy by up to 69.4 percentage points and Macro-F1 by 55.2 percentage points compared to one-stage baselines. Furthermore, it achieves up to a 1.99x speedup, establishing a highly efficient and reliable abstention pipeline for RALM with abstention.
Jongbin Won, Sung Geun An, Jay-yoon Lee
Graduate School of Data Science / Seoul National University
Frontier large language models (LLMs) can generate heuristic functions that guide search to achieve state-of-the-art performance in satisficing planning, where any plan is acceptable. However, these heuristics are not guaranteed to be admissible and can lead to suboptimal plans. We introduce LeanPlan, the first planning system that finds optimal plans with LLM-generated heuristics whose admissibility is machine-checked. Given a domain description and training tasks, an agentic loop uses planner feedback to iteratively improve a reusable domain-specific heuristic, its admissibility proof and the required domain assumptions. LeanPlan implements the heuristic, its proof and an efficient planner with machine-checked grounding and search in Lean 4. We evaluate LeanPlan on ten domains from the International Planning Competition and three new domains, using test tasks with up to 57 times as many objects as the training tasks. With GPT-5.6 Sol in the agentic loop, we successfully generate heuristics and admissibility proofs for all these domains. With the resulting heuristics, LeanPlan usually expands fewer states than the state-of-the-art Scorpion planner and solves more tasks overall.
André G. Pereira, Augusto B. Corrêa, Felipe Meneguzzi +1
Federal University of Rio Grande do Sul Brazil · University of Oxford United Kingdom · University of Aberdeen United Kingdom +1
Unsupervised domain adaptation generalizes neural retrievers to an unseen domain by generating pseudo queries on target domain documents. The quality and efficiency of this adaptation critically depend on which documents are selected for pseudo query generation. The existing document sampling method focuses on diversity but fails to capture model uncertainty. In contrast, we propose Uncertainty-based Iterative Document Sampling (UnIte) addressing these limitations by (1) filtering documents with high aleatoric uncertainty and (2) prioritizing those with high epistemic uncertainty, maximizing the learning utility of the current model. We conducted extensive experiments on a large corpus of BEIR with small and large models, showing significant gains of +2.45 and +3.49 nDCG@10 with a smaller training sample size, 4k on average.
Jongyoon Kim, Minseong Hwang, Seung-won Hwang
Interdisciplinary Program in Artificial Intelligence, Seoul National University