Causal Evidence Extraction and Triangulation in Crisis Reports using Large Language Models: A ReliefWeb-based Study

# Causal Evidence Extraction and Triangulation in Crisis Reports using Large Language Models: A ReliefWeb-based Study
Source: [https://arxiv.org/html/2608.04576](https://arxiv.org/html/2608.04576)
Yuanjun Zhang University of Oulu, Finland LUT University, Finland yuanjun\.zhang@lut\.fi &Mourad Oussalah University of Oulu, Finland mourad\.oussalah@oulu\.fi

###### Abstract

Humanitarian reports are long, noisy, and multi\-topic, making it difficult to consolidate decision\-relevant causal evidence\. We present a ReliefWeb study \(2000–2024\) and a two\-stage Large Language Model \(LLM\) pipeline that extracts structured intervention\-outcome records with direction and strength attributes\. Query\-conditioned extraction restricts output to a specified intervention class, reducing retrieval\-induced over\-extraction, while snippet grounding links each relation to supporting text for auditability and classification\. In an expert\-annotated dataset of 100 reports, the best closed\-source LLM achieved a weighted F1 score of 90\.73% with strong cost\-efficiency, while Llama\-3\.1\-8B with supervised fine\-tuning reached 94\.15% weighted F1 score\. We further propose context\-preserving triangulation that aggregates strength\-weighted evidence within disaster×\\timessource cells, applies Laplace smoothing and equally weights cells to quantify cross\-context convergence via a Level\-of\-Evidence score\. Applied to cash assistance, food\-related outcomes show strong positive convergence \(LoE=0\.865\) and stable long\-horizon trajectories\.

Causal Evidence Extraction and Triangulation in Crisis Reports using Large Language Models: A ReliefWeb\-based Study

Yuanjun ZhangUniversity of Oulu, FinlandLUT University, Finlandyuanjun\.zhang@lut\.fiMourad Oussalah††thanks:Corresponding authorUniversity of Oulu, Finlandmourad\.oussalah@oulu\.fi

## 1Introduction

Humanitarian crises generate large volumes of narrative situation reports describing interventions and evolving outcomes, yet decision\-relevant evidence about*what works, where, and under which constraints*remains hard to consolidate at scale\. Cash assistance illustrates the challenge: it is widely deployed, but reported impacts vary with market conditions, targeting, delivery mechanisms, and protection risksCash Learning Partnership \(CaLP\) \([2015](https://arxiv.org/html/2608.04576#bib.bib16)\); Doocy and Tappis \([2017](https://arxiv.org/html/2608.04576#bib.bib3)\); Frecceroet al\.\([2019](https://arxiv.org/html/2608.04576#bib.bib17)\); Burton \([2020](https://arxiv.org/html/2608.04576#bib.bib18)\)\. A bottleneck is translating long, multi\-topic narratives into structured, auditable intervention, and outcome evidence that can be compared across contexts\.

Because keyword retrieval prioritizes recall, it often surfaces documents where the queried intervention is marginally mentioned\. Besides, unconstrained extraction can then amplify this noise by producing irrelevant relations\. Recent LLMs make prompt\-based information extraction practical at scaleBrownet al\.\([2020](https://arxiv.org/html/2608.04576#bib.bib4)\), and grounding conclusions in explicit textual evidence is increasingly emphasized for auditabilityDeYounget al\.\([2020](https://arxiv.org/html/2608.04576#bib.bib33)\); Lewiset al\.\([2020](https://arxiv.org/html/2608.04576#bib.bib5)\)\. However, evidence\-synthesis extraction frameworks are typically developed for more topically clean inputShiet al\.\([2025](https://arxiv.org/html/2608.04576#bib.bib6)\), motivating report\-centric and question\-aware designs\.

![Refer to caption](https://arxiv.org/html/2608.04576v1/x1.png)Figure 1:Overall framework for causal evidence extraction and context\-preserving triangulation\. ReliefWeb reports and metadata \(disaster type, source type, time\) are processed by a two\-stage LLM pipeline with query conditioning and snippet grounding to produce structured causal records \(Section[3\.1](https://arxiv.org/html/2608.04576#S3.SS1)\)\. We then curate outcomes and normalize polarity for consistent directionality, and aggregate evidence via cell\-based triangulation over disaster×\\timessource contexts using strength\-weighted counts, Laplace smoothing, and equal cell weighting, yielding global probabilities \(PPP\_\{P\},PZP\_\{Z\},PNP\_\{N\}\), a Level\-of\-Evidence \(LoE\) score, and temporal trajectories \(Section[3\.2](https://arxiv.org/html/2608.04576#S3.SS2)\)\.We address this gap with a two\-stage LLM extraction pipeline for*causal evidence extraction*from ReliefWeb reportsReliefWeb \([2025](https://arxiv.org/html/2608.04576#bib.bib36)\)in Fig\.[1](https://arxiv.org/html/2608.04576#S1.F1)\. InStage 1,*query\-conditioned extraction*restricts outputs to interventions in a query\-defined class and extracts intervention–outcome candidates together with supporting snippets, reducing over\-extraction from low\-precision retrieval\. InStage 2,*snippet\-grounded relation classification*predicts direction and strength for each candidate pair using the supporting snippets, improving auditability and inference\. For synthesis, we then introducecontext\-preserving triangulation, which aggregates strength\-weighted directional evidence within disaster×\\timessource cells and measures cross\-context convergence rather than pooling raw countsJick \([1979](https://arxiv.org/html/2608.04576#bib.bib7)\)\. In addition, for lightweight deployment, we study anoptional distillation step: using the best closed\-source pipeline outputs as supervision, we apply Low\-Rank Adaptation \(LoRA\)Huet al\.\([2022](https://arxiv.org/html/2608.04576#bib.bib58)\)supervised fine\-tuning to Llama\-3\.1\-8B\-InstructGrattafioriet al\.\([2024](https://arxiv.org/html/2608.04576#bib.bib56)\)\.

Using ReliefWeb corpus \(2000–2024\), we compare five strategies across three API LLMs and an open\-weight Llama\-3\.1\-8B\-Instruct model with LoRA fine\-tuning\. Across settings, query conditioning yields large, consistent gains, and snippet grounding further improves faithfulness\. We also demonstrate long\-horizon triangulation for cash assistance, producing interpretable global probabilities and temporal evidence trajectories\.

## 2Related Work

Prior humanitarian natural language processing \(NLP\) research highlights domain shift and the need for trustworthy decision supportRoccaet al\.\([2023](https://arxiv.org/html/2608.04576#bib.bib9)\)\. We study*report\-centric*extraction of intervention–outcome records from long, multi\-topic ReliefWeb reportsReliefWeb \([2025](https://arxiv.org/html/2608.04576#bib.bib36)\), leveraging document\-level IE/REYaoet al\.\([2019](https://arxiv.org/html/2608.04576#bib.bib10)\)and evidence groundingDeYounget al\.\([2020](https://arxiv.org/html/2608.04576#bib.bib33)\)\.

Complementary to text\-based evidence extraction, recent work has also emphasized reliability and explainability in disaster\-related multimodal decision support, showing that preference optimization and explainable vision\-language reasoning can improve both assessment performance and interpretability in disaster severity assessmentZhanget al\.\([2026c](https://arxiv.org/html/2608.04576#bib.bib83)\)\. LLMs are increasingly used for evidence extraction and synthesis, including directional triangulationShiet al\.\([2025](https://arxiv.org/html/2608.04576#bib.bib6)\)and end\-to\-end evidence\-synthesis pipelinesWanget al\.\([2025b](https://arxiv.org/html/2608.04576#bib.bib2)\)\. We adapt this line to humanitarian reports with query\-conditioned extraction and context\-preserving synthesis within disaster×\\timessource cells, while attaching short supporting spans for traceability\. LLM factuality and hallucination detection have also been studied, e\.g\., via consistency\-based checksManakulet al\.\([2023](https://arxiv.org/html/2608.04576#bib.bib1)\)\.

More specifically, our work differs from prior evidence extraction studies in three ways\. First, we target long, noisy, multi\-topic humanitarian reports\. Second, we use query\-conditioned extraction to control retrieval\-induced over\-extraction\. Third, beyond extraction, we introduce context\-preserving triangulation over disaster×\\timessource cells, together with snippet grounding for auditability, instead of relying only on pooled document\-level evidence\.

## 3Methodology

Given a ReliefWeb reportxxand a queryqqspecifying an intervention*class*\(e\.g\.,cash assistance\), we extract a set of directional causal records

e=\(a,o,δ,σ,z\)e=\(a,o,\\delta,\\sigma,z\)\(1\)whereaais an intervention mention \(restricted to classqq\),oois an outcome mention,δ∈\{Inc,Dec,No\}\\delta\\in\\\{\\textsc\{Inc\},\\textsc\{Dec\},\\textsc\{No\}\\\}denotes whether the reported outcome increases, decreases, or shows no clear change,σ∈\{Weak,Mod,Strong\}\\sigma\\in\\\{\\textsc\{Weak\},\\textsc\{Mod\},\\textsc\{Strong\}\\\}is the evidence strength, andzzis a supporting text snippet for auditability\.

#### Data\.

We query ReliefWeb with the keyword’cash assistance’and retain English reports with accessible text \(2000–2024\)\. We restricted to ReliefWeb\-defined disaster and source types: four disasters \(Flood,Drought,Epidemic,Earthquake\) and five sources \(Academic,Government,Media,NGO,Red Cross/Red Crescent Movement\), yielding 8,029 reports;International OrganizationandOtherare excluded due to imbalance/ambiguity \(Appendix[A](https://arxiv.org/html/2608.04576#A1)and Appendix[B](https://arxiv.org/html/2608.04576#A2)\)\. For evaluation, experts annotate 100 sampled reports \(220 relations\); see Appendix[C](https://arxiv.org/html/2608.04576#A3)for details\.

#### Evaluation\.

We evaluate intervention/outcome extraction via semantic set matching using BERTScore \(threshold 0\.8\)Zhanget al\.\([2020](https://arxiv.org/html/2608.04576#bib.bib15)\)\. Direction and strength are scored on matched intervention–outcome pairs; strength is scored only when direction is correct\. Since strength is ordinal \(Weak<<Mod<<Strong\), adjacent confusions receive partial credit \(0\.5\)\.

### 3\.1Query\-conditioned and snippet\-grounded extraction

As shown in Fig\.[1](https://arxiv.org/html/2608.04576#S1.F1), our final method is atwo\-stage extraction pipelinewith two LLM calls\.Stage 1 \(QC entity & evidence extraction\):the LLM extracts only interventions in classqq\(query\-conditioned, QC\), the associated*outcome mentions*expressed in the report as raw text spans / free\-text phrases, and supporting snippetszz\. The output of Stage 1 therefore consists of candidate intervention–outcome pairs together with their evidence snippets\.Stage 2 \(SG relation classification\):given these candidate pairs and snippets, the LLM predicts the relation labels, i\.e\., directionδ\\deltaand strengthσ\\sigma\. Snippets provide explicit grounding \(snippet\-grounded, SG\) while the full report remains available\.

Separately, we comparesingle\-promptversustwo\-promptprompting as an ablation, and we also ablate QC and SG\. Here,*stage*refers to the two LLM calls in the final extraction pipeline, whereas*single\-prompt*and*two\-prompt*refer only to prompting variants\.

Using the besttwo\-stage QC\+SG extraction pipelineas the teacher, we select the best\-performing proprietary closed\-source LLM\. We then decompose its outputs intoStage\-1 supervision\(QC intervention–outcome pairs with grounded snippets\) andStage\-2 supervision\(direction/strength conditioned on Stage 1\), and train two parameter\-efficient LoRA adapters \(one per stage\) on an open\-weight Llama\-3\.1\-8B\-Instruct student while preserving the same two\-stage inference structure\. To prevent leakage, we exclude the 100 evaluation reports from distillation; to reduce model\-family relatedness bias, we choose a student outside the teacher’s familyWanget al\.\([2025a](https://arxiv.org/html/2608.04576#bib.bib59)\)\.

Table 1:Extraction volume comparison across domains and models\. Values are average \#relations/document\. QC is applied only for humanitarian reports\.
### 3\.2Context\-preserving triangulation

We synthesize records across contexts while preventing high\-volume contexts from dominating\.

#### Polarity normalization\.

Outcomes may be positively framed \(e\.g\.,food security\) or negatively framed \(e\.g\.,food insecurity\)\. We assign polarityπ​\(o\)∈\{\+1,−1\}\\pi\(o\)\\in\\\{\+1,\-1\\\}and flipInc↔Dec\\textsc\{Inc\}\\leftrightarrow\\textsc\{Dec\}whenπ​\(o\)=−1\\pi\(o\)=\-1\(leavingNounchanged\), so a*positive*relation always denotes improvement\. \(For the case study, polarity is labeled for frequent outcomes, e\.g\., frequency≥3\\geq 3\. See Appendix[D](https://arxiv.org/html/2608.04576#A4)for details\.\)

Table 2:Overall performance comparison across models and extraction strategies\. All F1 scores are in %\. Weighted F1 uses weights 0\.30/0\.30/0\.20/0\.20\. Base denotes the baseline extraction strategy fromShiet al\.\([2025](https://arxiv.org/html/2608.04576#bib.bib6)\)\.
#### Cell\-based evidence aggregation\.

Let𝒟\\mathcal\{D\}denote the set of disaster types and𝒮\\mathcal\{S\}denote the set of source types\. Each record inherits disaster typed∈𝒟d\\in\\mathcal\{D\}and source types∈𝒮s\\in\\mathcal\{S\}\. We define triangulation cellsc=\(d,s\)∈𝒟×𝒮c=\(d,s\)\\in\\mathcal\{D\}\\times\\mathcal\{S\}and map normalized directions tor∈\{P,Z,N\}r\\in\\\{P,Z,N\\\}\(positive / no\-change / negative\)\. We convert strength to an ordinal weightw​\(Weak\)=0,w​\(Mod\)=1,w​\(Strong\)=2w\(\\textsc\{Weak\}\)\{=\}0,\\;w\(\\textsc\{Mod\}\)\{=\}1,\\;w\(\\textsc\{Strong\}\)\{=\}2:

Cr\(c\)=∑i∈cw​\(σi\)​𝕀​\(ri=r\)\.C\_\{r\}^\{\(c\)\}=\\sum\_\{i\\in c\}w\(\\sigma\_\{i\}\)\\,\\mathbb\{I\}\(r\_\{i\}=r\)\.\(2\)

#### Smoothed cell probabilities\.

To avoid degenerate estimates in sparse cells, we apply Laplace smoothing \(α=0\.1\\alpha\{=\}0\.1\):

Pr\(c\)=Cr\(c\)\+α∑r′∈\{P,Z,N\}Cr′\(c\)\+3​αP\_\{r\}^\{\(c\)\}=\\frac\{C\_\{r\}^\{\(c\)\}\+\\alpha\}\{\\sum\_\{r^\{\\prime\}\\in\\\{P,Z,N\\\}\}C\_\{r^\{\\prime\}\}^\{\(c\)\}\+3\\alpha\}\(3\)Cells with∑rCr\(c\)=0\\sum\_\{r\}C\_\{r\}^\{\(c\)\}=0are discarded\.

#### Equal cell weighting and Level of Evidence\.

Let𝒞\\mathcal\{C\}denote the set of retained non\-empty cells\. We average cell probabilities equally:

Pr=1\|𝒞\|​∑c∈𝒞Pr\(c\)P\_\{r\}=\\frac\{1\}\{\|\\mathcal\{C\}\|\}\\sum\_\{c\\in\\mathcal\{C\}\}P\_\{r\}^\{\(c\)\}\(4\)Cross\-context convergence is measured by a Level of Evidence score:

LoE=max⁡\(PP,PZ,PN\)−131−13∈\[0,1\]\.\\mathrm\{LoE\}=\\frac\{\\max\(P\_\{P\},P\_\{Z\},P\_\{N\}\)\-\\frac\{1\}\{3\}\}\{1\-\\frac\{1\}\{3\}\}\\in\[0,1\]\.\(5\)Here,PPP\_\{P\},PZP\_\{Z\}, andPNP\_\{N\}denote the global cell\-averaged probabilities of positive / no\-change / negative relationships\. By construction,PPP\_\{P\},PZP\_\{Z\}, andPNP\_\{N\}are non\-negative and sum to 1, somax⁡\(PP,PZ,PN\)∈\[1/3,1\]\\max\(P\_\{P\},P\_\{Z\},P\_\{N\}\)\\in\[1/3,1\]\. Therefore,LoE∈\[0,1\]\\mathrm\{LoE\}\\in\[0,1\]and cannot be negative\.

#### Temporal triangulation\.

For evidence trajectories, we recompute Eqs\. \([2](https://arxiv.org/html/2608.04576#S3.E2)\)–\([5](https://arxiv.org/html/2608.04576#S3.E5)\) cumulatively by yearttusing records withyeari≤t\\mathrm\{year\}\_\{i\}\\leq t, producing\(PP​\(t\),PZ​\(t\),PN​\(t\),LoE​\(t\)\)\(P\_\{P\}\(t\),P\_\{Z\}\(t\),P\_\{N\}\(t\),\\mathrm\{LoE\}\(t\)\)\.

## 4Experiments and Results

We evaluate on four LLMs:Qwen\-PlusBaiet al\.\([2023](https://arxiv.org/html/2608.04576#bib.bib12)\),GPT\-4o\-miniHurstet al\.\([2024](https://arxiv.org/html/2608.04576#bib.bib13)\),DeepSeek\-V3DeepSeek\-AIet al\.\([2025](https://arxiv.org/html/2608.04576#bib.bib14)\), and Llama\-3\.1\-8B\-Instruct\.

### 4\.1QC reduces over\-extraction

Keyword retrieval is recall\-oriented and often surfaces reports where the queried intervention is only marginally mentioned; baseline prompting then over\-extracts query\-irrelevant relations\. Tab\.[1](https://arxiv.org/html/2608.04576#S3.T1)shows severe over\-extraction on humanitarian reports, while query\-conditioned extraction \(QC\) brings the volume close to the expected ground\-truth level across closed\-source models\. Medical numbers are fromShiet al\.\([2025](https://arxiv.org/html/2608.04576#bib.bib6)\)\.

### 4\.2Extraction performance: ablations over QC, SG, and prompting variants

Tab\.[2](https://arxiv.org/html/2608.04576#S3.T2)reports extraction performance on the expert\-annotated benchmark\. Across models, QC provides the largest gains, while prompt decomposition and snippet grounding \(SG\) yield additional consistent improvements\. Among closed\-source models, Qwen\-Plus with two\-step QC\+SG achieves the best weighted F1 \(90\.73%\) and is used for downstream triangulation\. Efficiency is computed asEfficiency=\(Weighted F1/100\)/unit price\\text\{Efficiency\}=\(\\text\{Weighted F1\}/100\)\\,/\\,\\text\{unit price\}\(prices in Appendix Tab\.[5](https://arxiv.org/html/2608.04576#A2.T5)\)\. Our pipeline also provides supervision for distillation: LoRA fine\-tuning Llama\-3\.1\-8B\-Instruct improves weighted F1 from 75\.12% to 94\.15% \(Appendix[F](https://arxiv.org/html/2608.04576#A6)\)\.

![Refer to caption](https://arxiv.org/html/2608.04576v1/Figures/final_output.png)Figure 2:Cumulative temporal triangulation for polarity\-normalized food\-related outcomes: cell\-averagedPPP\_\{P\},PZP\_\{Z\},PNP\_\{N\}and the number of contributing disaster×\\timessource cells over time\.
### 4\.3Triangulation: cross\-context convergence

Using Qwen\-Plus \(two\-step QC\+SG\), we extract 19,568 causal records across four disaster types \(5,999 drought; 3,696 earthquake; 4,412 epidemic; 5,461 flood\)\. We focus on food\-related outcomes due to high evidence density \(Appendix[B](https://arxiv.org/html/2608.04576#A2)\)\. After polarity normalization \(87 positive vs\. 65 negative outcome strings with frequency≥3\\geq 3\), triangulation over4×5=204\\times 5=20disaster×\\timessource cells retains 19 non\-empty cells and yieldsPP=0\.91P\_\{P\}=0\.91andLoE=0\.865\\mathrm\{LoE\}=0\.865\.

Fig\.[2](https://arxiv.org/html/2608.04576#S4.F2)shows cumulative trajectories\. Early years \(2000–2005\) have sparse coverage and volatile estimates; from 2006–2017, expanding cell coverage yields strong positive convergence, peaking in 2017 \(PP=0\.953P\_\{P\}=0\.953,LoE=0\.929\\mathrm\{LoE\}=0\.929\)\. From 2018–2024, the signal remains highly positive with mild softening, consistent with residual context heterogeneity and operational “do no harm” constraintsCash Learning Partnership \(CaLP\) \([2015](https://arxiv.org/html/2608.04576#bib.bib16)\); Frecceroet al\.\([2019](https://arxiv.org/html/2608.04576#bib.bib17)\); Burton \([2020](https://arxiv.org/html/2608.04576#bib.bib18)\)\.

## 5Conclusion

We propose a query\-conditioned, snippet\-grounded two\-stage extraction pipeline to extract auditable intervention–outcome evidence from long humanitarian reports, and a context\-preserving cell\-based triangulation method for cross\-context synthesis\. On ReliefWeb, the best closed\-source setting reaches 90\.73% weighted F1, LoRA SFT on Llama\-3\.1\-8B\-Instruct reaches 94\.15%, and triangulation shows strong positive convergence of cash assistance on food outcomes \(LoE=0\.865\)\.

## Limitations

Due to limitations in funding and manpower, our corpus and case study are constrained by a keyword query \(“cash assistance”\), English\-only accessible text, a limited set of disaster types, and a coarse source taxonomy, which may reduce generalizability to other interventions, languages, hazards, and reporting contexts\. The expert\-annotated evaluation set is also small; while it allows us to capture overall trends, performance estimates may be sensitive to sampling and annotation guidelines\. Our triangulation and aggregation reflect convergence of reported evidence and rely on design choices \(e\.g\., polarity normalization, weighting, smoothing\) that can influence the resulting scores\. We note that the strength labels in this work reflect the strength of reported evidence rather than effect magnitude or population\-scale impact, which we leave for future work\. Therefore, the outputs should be treated as decision support with human review rather than as definitive conclusions\.

## Ethical Considerations

This work analyzes publicly available humanitarian reports and does not intentionally collect or infer sensitive personal information\. Nevertheless, reports may contain incidental mentions of individuals or vulnerable groups; we therefore use the data only for research purposes and recommend caution when releasing derived artifacts\. Because ReliefWeb serves as an upstream source, our method may inherit biases in the reporting ecosystem \(e\.g\., uneven geographic coverage, organizational incentives, and language accessibility\), which can affect extracted evidence and aggregated scores\. In addition, assessments of humanitarian measures in text may themselves be subjective or contested, and LLM\-based extraction may further reflect hidden or unknown model biases\. To mitigate this, we attach supporting snippets for auditability and triangulate across disaster×\\timessource contexts to reduce reliance on any single report or context\. Therefore, the outputs should not be used as a stand\-alone basis for operational choices\. We position the system as an assistive tool to surface and summarize reported evidence, and we recommend human review—especially for contested or ambiguous claims and for any downstream decisions that may impact affected populations\.

## Acknowledgements

This work is funded by the European Union’s Horizon 2020 research and innovation programme under grant agreement No\. 101004509, and the Collaboration of Humanities and Social Sciences in Europe \(CHANSE\) research project Digital Emergency Communication \(DIGeMERGE\)\.

## References

- A\. Asai, Z\. Wu, Y\. Wang, A\. Sil, and H\. Hajishirzi \(2024\)Self\-rag: learning to retrieve, generate, and critique through self\-reflection\.External Links:[Link](https://openreview.net/forum?id=hSyW5go0v8)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p2.1)\.
- J\. Bai, S\. Bai, Y\. Chu, Z\. Cui, K\. Dang, X\. Deng, Y\. Fan, W\. Ge, Y\. Han, F\. Huang, B\. Hui, L\. Ji, M\. Li, J\. Lin, R\. Lin, D\. Liu, G\. Liu, C\. Lu, K\. Lu, J\. Ma, R\. Men, X\. Ren, X\. Ren, C\. Tan, S\. Tan, J\. Tu, P\. Wang, S\. Wang, W\. Wang, S\. Wu, B\. Xu, J\. Xu, A\. Yang, H\. Yang, J\. Yang, S\. Yang, Y\. Yao, B\. Yu, H\. Yuan, Z\. Yuan, J\. Zhang, X\. Zhang, Y\. Zhang, Z\. Zhang, C\. Zhou, J\. Zhou, X\. Zhou, and T\. Zhu \(2023\)Qwen technical report\.External Links:2309\.16609,[Link](https://arxiv.org/abs/2309.16609)Cited by:[§4](https://arxiv.org/html/2608.04576#S4.p1.1)\.
- T\. B\. Brown, B\. Mann, N\. Ryder, M\. Subbiah, J\. Kaplan, P\. Dhariwal, A\. Neelakantan, P\. Shyam, G\. Sastry, A\. Askell, S\. Agarwal, A\. Herbert\-Voss, G\. Krueger, T\. Henighan, R\. Child, A\. Ramesh, D\. M\. Ziegler, J\. Wu, C\. Winter, C\. Hesse, M\. Chen, E\. Sigler, M\. Litwin, S\. Gray, B\. Chess, J\. Clark, C\. Berner, S\. McCandlish, A\. Radford, I\. Sutskever, and D\. Amodei \(2020\)Language models are few\-shot learners\.InProceedings of the 34th International Conference on Neural Information Processing Systems,NIPS ’20,Red Hook, NY, USA\.External Links:ISBN 9781713829546Cited by:[§1](https://arxiv.org/html/2608.04576#S1.p2.1)\.
- J\. Burton \(2020\)“Doing no harm” in the digital age: what the digitalization of cash means for humanitarian action\.International Review of the Red Cross102\(913\),pp\. 43–73\.Cited by:[§1](https://arxiv.org/html/2608.04576#S1.p1.1),[§4\.3](https://arxiv.org/html/2608.04576#S4.SS3.p2.2)\.
- Cash Learning Partnership \(CaLP\) \(2015\)Operational guidance and toolkit for multipurpose cash grants\.Cash Learning Partnership\.Note:Enhanced Response Capacity Project 2014–2015External Links:[Link](https://rohingyaresponse.org/wp-content/uploads/2023/05/Operational-Guidance-and-Toolkit-for-Multipurpose-Cash-Grants-web.pdf)Cited by:[§1](https://arxiv.org/html/2608.04576#S1.p1.1),[§4\.3](https://arxiv.org/html/2608.04576#S4.SS3.p2.2)\.
- Z\. Chen, Y\. Hu, Z\. Fu, Z\. Li, J\. Huang, Q\. Huang, and Y\. Wei \(2026\)INTENT: invariance and discrimination\-aware noise mitigation for robust composed image retrieval\.Proceedings of the AAAI Conference on Artificial Intelligence40\(25\),pp\. 20463–20471\.External Links:[Link](https://ojs.aaai.org/index.php/AAAI/article/view/39181),[Document](https://dx.doi.org/10.1609/aaai.v40i25.39181)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p3.1)\.
- Z\. Chen, Y\. Hu, Z\. Li, Z\. Fu, X\. Song, and L\. Nie \(2025\)OFFSET: segmentation\-based focus shift revision for composed image retrieval\.InProceedings of the 33rd ACM International Conference on Multimedia,MM ’25,New York, NY, USA,pp\. 6113–6122\.External Links:ISBN 9798400720352,[Link](https://doi.org/10.1145/3746027.3755366),[Document](https://dx.doi.org/10.1145/3746027.3755366)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p3.1)\.
- DeepSeek\-AI, A\. Liu, B\. Feng, B\. Xue, B\. Wang, B\. Wu, C\. Lu, C\. Zhao, C\. Deng, C\. Zhang, C\. Ruan, D\. Dai, D\. Guo, D\. Yang, D\. Chen, D\. Ji, E\. Li, F\. Lin, F\. Dai, F\. Luo, G\. Hao, G\. Chen, G\. Li, H\. Zhang, H\. Bao, H\. Xu, H\. Wang, H\. Zhang, H\. Ding, H\. Xin, H\. Gao, H\. Li, H\. Qu, J\. L\. Cai, J\. Liang, J\. Guo, J\. Ni, J\. Li, J\. Wang, J\. Chen, J\. Chen, J\. Yuan, J\. Qiu, J\. Li, J\. Song, K\. Dong, K\. Hu, K\. Gao, K\. Guan, K\. Huang, K\. Yu, L\. Wang, L\. Zhang, L\. Xu, L\. Xia, L\. Zhao, L\. Wang, L\. Zhang, M\. Li, M\. Wang, M\. Zhang, M\. Zhang, M\. Tang, M\. Li, N\. Tian, P\. Huang, P\. Wang, P\. Zhang, Q\. Wang, Q\. Zhu, Q\. Chen, Q\. Du, R\. J\. Chen, R\. L\. Jin, R\. Ge, R\. Zhang, R\. Pan, R\. Wang, R\. Xu, R\. Zhang, R\. Chen, S\. S\. Li, S\. Lu, S\. Zhou, S\. Chen, S\. Wu, S\. Ye, S\. Ye, S\. Ma, S\. Wang, S\. Zhou, S\. Yu, S\. Zhou, S\. Pan, T\. Wang, T\. Yun, T\. Pei, T\. Sun, W\. L\. Xiao, W\. Zeng, W\. Zhao, W\. An, W\. Liu, W\. Liang, W\. Gao, W\. Yu, W\. Zhang, X\. Q\. Li, X\. Jin, X\. Wang, X\. Bi, X\. Liu, X\. Wang, X\. Shen, X\. Chen, X\. Zhang, X\. Chen, X\. Nie, X\. Sun, X\. Wang, X\. Cheng, X\. Liu, X\. Xie, X\. Liu, X\. Yu, X\. Song, X\. Shan, X\. Zhou, X\. Yang, X\. Li, X\. Su, X\. Lin, Y\. K\. Li, Y\. Q\. Wang, Y\. X\. Wei, Y\. X\. Zhu, Y\. Zhang, Y\. Xu, Y\. Xu, Y\. Huang, Y\. Li, Y\. Zhao, Y\. Sun, Y\. Li, Y\. Wang, Y\. Yu, Y\. Zheng, Y\. Zhang, Y\. Shi, Y\. Xiong, Y\. He, Y\. Tang, Y\. Piao, Y\. Wang, Y\. Tan, Y\. Ma, Y\. Liu, Y\. Guo, Y\. Wu, Y\. Ou, Y\. Zhu, Y\. Wang, Y\. Gong, Y\. Zou, Y\. He, Y\. Zha, Y\. Xiong, Y\. Ma, Y\. Yan, Y\. Luo, Y\. You, Y\. Liu, Y\. Zhou, Z\. F\. Wu, Z\. Z\. Ren, Z\. Ren, Z\. Sha, Z\. Fu, Z\. Xu, Z\. Huang, Z\. Zhang, Z\. Xie, Z\. Zhang, Z\. Hao, Z\. Gou, Z\. Ma, Z\. Yan, Z\. Shao, Z\. Xu, Z\. Wu, Z\. Zhang, Z\. Li, Z\. Gu, Z\. Zhu, Z\. Liu, Z\. Li, Z\. Xie, Z\. Song, Z\. Gao, and Z\. Pan \(2025\)DeepSeek\-v3 technical report\.External Links:2412\.19437,[Link](https://arxiv.org/abs/2412.19437)Cited by:[§4](https://arxiv.org/html/2608.04576#S4.p1.1)\.
- J\. DeYoung, S\. Jain, N\. F\. Rajani, E\. Lehman, C\. Xiong, R\. Socher, and B\. C\. Wallace \(2020\)ERASER: a benchmark to evaluate rationalized nlp models\.InProceedings of the 58th annual meeting of the association for computational linguistics,pp\. 4443–4458\.Cited by:[§1](https://arxiv.org/html/2608.04576#S1.p2.1),[§2](https://arxiv.org/html/2608.04576#S2.p1.1)\.
- S\. Doocy and H\. Tappis \(2017\)Cash\-based approaches in humanitarian emergencies: a systematic review\.Campbell Systematic Reviews13\(1\),pp\. 1–200\.Cited by:[§1](https://arxiv.org/html/2608.04576#S1.p1.1)\.
- J\. Freccero, A\. Taylor, J\. Ortega, Z\. Buda, P\. K\. Awah, A\. Blackwell, R\. P\. Cordero, and E\. Stover \(2019\)Safer cash in conflict: exploring protection risks and barriers in cash programming for internally displaced persons in cameroon and afghanistan\.International Review of the Red Cross101\(911\),pp\. 685–713\.Cited by:[§1](https://arxiv.org/html/2608.04576#S1.p1.1),[§4\.3](https://arxiv.org/html/2608.04576#S4.SS3.p2.2)\.
- X\. Ge, S\. Murtaza, A\. Cortez, and H\. Alemzadeh \(2026\)EMSDialog: synthetic multi\-person emergency medical service dialogue generation from electronic patient care reports via multi\-llm agents\.External Links:2604\.07549,[Link](https://arxiv.org/abs/2604.07549)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p4.1)\.
- A\. Grattafiori, A\. Dubey, A\. Jauhri, A\. Pandey, A\. Kadian, A\. Al\-Dahle, A\. Letman, A\. Mathur, A\. Schelten, A\. Vaughan, A\. Yang, A\. Fan, A\. Goyal, A\. Hartshorn, A\. Yang, A\. Mitra, A\. Sravankumar, A\. Korenev, A\. Hinsvark, A\. Rao, A\. Zhang, A\. Rodriguez, A\. Gregerson, A\. Spataru, B\. Roziere, B\. Biron, B\. Tang, B\. Chern, C\. Caucheteux, C\. Nayak, C\. Bi, C\. Marra, C\. McConnell, C\. Keller, C\. Touret, C\. Wu, C\. Wong, C\. C\. Ferrer, C\. Nikolaidis, D\. Allonsius, D\. Song, D\. Pintz, D\. Livshits, D\. Wyatt, D\. Esiobu, D\. Choudhary, D\. Mahajan, D\. Garcia\-Olano, D\. Perino, D\. Hupkes, E\. Lakomkin, E\. AlBadawy, E\. Lobanova, E\. Dinan, E\. M\. Smith, F\. Radenovic, F\. Guzmán, F\. Zhang, G\. Synnaeve, G\. Lee, G\. L\. Anderson, G\. Thattai, G\. Nail, G\. Mialon, G\. Pang, G\. Cucurell, H\. Nguyen, H\. Korevaar, H\. Xu, H\. Touvron, I\. Zarov, I\. A\. Ibarra, I\. Kloumann, I\. Misra, I\. Evtimov, J\. Zhang, J\. Copet, J\. Lee, J\. Geffert, J\. Vranes, J\. Park, J\. Mahadeokar, J\. Shah, J\. van der Linde, J\. Billock, J\. Hong, J\. Lee, J\. Fu, J\. Chi, J\. Huang, J\. Liu, J\. Wang, J\. Yu, J\. Bitton, J\. Spisak, J\. Park, J\. Rocca, J\. Johnstun, J\. Saxe, J\. Jia, K\. V\. Alwala, K\. Prasad, K\. Upasani, K\. Plawiak, K\. Li, K\. Heafield, K\. Stone, K\. El\-Arini, K\. Iyer, K\. Malik, K\. Chiu, K\. Bhalla, K\. Lakhotia, L\. Rantala\-Yeary, L\. van der Maaten, L\. Chen, L\. Tan, L\. Jenkins, L\. Martin, L\. Madaan, L\. Malo, L\. Blecher, L\. Landzaat, L\. de Oliveira, M\. Muzzi, M\. Pasupuleti, M\. Singh, M\. Paluri, M\. Kardas, M\. Tsimpoukelli, M\. Oldham, M\. Rita, M\. Pavlova, M\. Kambadur, M\. Lewis, M\. Si, M\. K\. Singh, M\. Hassan, N\. Goyal, N\. Torabi, N\. Bashlykov, N\. Bogoychev, N\. Chatterji, N\. Zhang, O\. Duchenne, O\. Çelebi, P\. Alrassy, P\. Zhang, P\. Li, P\. Vasic, P\. Weng, P\. Bhargava, P\. Dubal, P\. Krishnan, P\. S\. Koura, P\. Xu, Q\. He, Q\. Dong, R\. Srinivasan, R\. Ganapathy, R\. Calderer, R\. S\. Cabral, R\. Stojnic, R\. Raileanu, R\. Maheswari, R\. Girdhar, R\. Patel, R\. Sauvestre, R\. Polidoro, R\. Sumbaly, R\. Taylor, R\. Silva, R\. Hou, R\. Wang, S\. Hosseini, S\. Chennabasappa, S\. Singh, S\. Bell, S\. S\. Kim, S\. Edunov, S\. Nie, S\. Narang, S\. Raparthy, S\. Shen, S\. Wan, S\. Bhosale, S\. Zhang, S\. Vandenhende, S\. Batra, S\. Whitman, S\. Sootla, S\. Collot, S\. Gururangan, S\. Borodinsky, T\. Herman, T\. Fowler, T\. Sheasha, T\. Georgiou, T\. Scialom, T\. Speckbacher, T\. Mihaylov, T\. Xiao, U\. Karn, V\. Goswami, V\. Gupta, V\. Ramanathan, V\. Kerkez, V\. Gonguet, V\. Do, V\. Vogeti, V\. Albiero, V\. Petrovic, W\. Chu, W\. Xiong, W\. Fu, W\. Meers, X\. Martinet, X\. Wang, X\. Wang, X\. E\. Tan, X\. Xia, X\. Xie, X\. Jia, X\. Wang, Y\. Goldschlag, Y\. Gaur, Y\. Babaei, Y\. Wen, Y\. Song, Y\. Zhang, Y\. Li, Y\. Mao, Z\. D\. Coudert, Z\. Yan, Z\. Chen, Z\. Papakipos, A\. Singh, A\. Srivastava, A\. Jain, A\. Kelsey, A\. Shajnfeld, A\. Gangidi, A\. Victoria, A\. Goldstand, A\. Menon, A\. Sharma, A\. Boesenberg, A\. Baevski, A\. Feinstein, A\. Kallet, A\. Sangani, A\. Teo, A\. Yunus, A\. Lupu, A\. Alvarado, A\. Caples, A\. Gu, A\. Ho, A\. Poulton, A\. Ryan, A\. Ramchandani, A\. Dong, A\. Franco, A\. Goyal, A\. Saraf, A\. Chowdhury, A\. Gabriel, A\. Bharambe, A\. Eisenman, A\. Yazdan, B\. James, B\. Maurer, B\. Leonhardi, B\. Huang, B\. Loyd, B\. D\. Paola, B\. Paranjape, B\. Liu, B\. Wu, B\. Ni, B\. Hancock, B\. Wasti, B\. Spence, B\. Stojkovic, B\. Gamido, B\. Montalvo, C\. Parker, C\. Burton, C\. Mejia, C\. Liu, C\. Wang, C\. Kim, C\. Zhou, C\. Hu, C\. Chu, C\. Cai, C\. Tindal, C\. Feichtenhofer, C\. Gao, D\. Civin, D\. Beaty, D\. Kreymer, D\. Li, D\. Adkins, D\. Xu, D\. Testuggine, D\. David, D\. Parikh, D\. Liskovich, D\. Foss, D\. Wang, D\. Le, D\. Holland, E\. Dowling, E\. Jamil, E\. Montgomery, E\. Presani, E\. Hahn, E\. Wood, E\. Le, E\. Brinkman, E\. Arcaute, E\. Dunbar, E\. Smothers, F\. Sun, F\. Kreuk, F\. Tian, F\. Kokkinos, F\. Ozgenel, F\. Caggioni, F\. Kanayet, F\. Seide, G\. M\. Florez, G\. Schwarz, G\. Badeer, G\. Swee, G\. Halpern, G\. Herman, G\. Sizov, Guangyi, Zhang, G\. Lakshminarayanan, H\. Inan, H\. Shojanazeri, H\. Zou, H\. Wang, H\. Zha, H\. Habeeb, H\. Rudolph, H\. Suk, H\. Aspegren, H\. Goldman, H\. Zhan, I\. Damlaj, I\. Molybog, I\. Tufanov, I\. Leontiadis, I\. Veliche, I\. Gat, J\. Weissman, J\. Geboski, J\. Kohli, J\. Lam, J\. Asher, J\. Gaya, J\. Marcus, J\. Tang, J\. Chan, J\. Zhen, J\. Reizenstein, J\. Teboul, J\. Zhong, J\. Jin, J\. Yang, J\. Cummings, J\. Carvill, J\. Shepard, J\. McPhie, J\. Torres, J\. Ginsburg, J\. Wang, K\. Wu, K\. H\. U, K\. Saxena, K\. Khandelwal, K\. Zand, K\. Matosich, K\. Veeraraghavan, K\. Michelena, K\. Li, K\. Jagadeesh, K\. Huang, K\. Chawla, K\. Huang, L\. Chen, L\. Garg, L\. A, L\. Silva, L\. Bell, L\. Zhang, L\. Guo, L\. Yu, L\. Moshkovich, L\. Wehrstedt, M\. Khabsa, M\. Avalani, M\. Bhatt, M\. Mankus, M\. Hasson, M\. Lennie, M\. Reso, M\. Groshev, M\. Naumov, M\. Lathi, M\. Keneally, M\. Liu, M\. L\. Seltzer, M\. Valko, M\. Restrepo, M\. Patel, M\. Vyatskov, M\. Samvelyan, M\. Clark, M\. Macey, M\. Wang, M\. J\. Hermoso, M\. Metanat, M\. Rastegari, M\. Bansal, N\. Santhanam, N\. Parks, N\. White, N\. Bawa, N\. Singhal, N\. Egebo, N\. Usunier, N\. Mehta, N\. P\. Laptev, N\. Dong, N\. Cheng, O\. Chernoguz, O\. Hart, O\. Salpekar, O\. Kalinli, P\. Kent, P\. Parekh, P\. Saab, P\. Balaji, P\. Rittner, P\. Bontrager, P\. Roux, P\. Dollar, P\. Zvyagina, P\. Ratanchandani, P\. Yuvraj, Q\. Liang, R\. Alao, R\. Rodriguez, R\. Ayub, R\. Murthy, R\. Nayani, R\. Mitra, R\. Parthasarathy, R\. Li, R\. Hogan, R\. Battey, R\. Wang, R\. Howes, R\. Rinott, S\. Mehta, S\. Siby, S\. J\. Bondu, S\. Datta, S\. Chugh, S\. Hunt, S\. Dhillon, S\. Sidorov, S\. Pan, S\. Mahajan, S\. Verma, S\. Yamamoto, S\. Ramaswamy, S\. Lindsay, S\. Lindsay, S\. Feng, S\. Lin, S\. C\. Zha, S\. Patil, S\. Shankar, S\. Zhang, S\. Zhang, S\. Wang, S\. Agarwal, S\. Sajuyigbe, S\. Chintala, S\. Max, S\. Chen, S\. Kehoe, S\. Satterfield, S\. Govindaprasad, S\. Gupta, S\. Deng, S\. Cho, S\. Virk, S\. Subramanian, S\. Choudhury, S\. Goldman, T\. Remez, T\. Glaser, T\. Best, T\. Koehler, T\. Robinson, T\. Li, T\. Zhang, T\. Matthews, T\. Chou, T\. Shaked, V\. Vontimitta, V\. Ajayi, V\. Montanez, V\. Mohan, V\. S\. Kumar, V\. Mangla, V\. Ionescu, V\. Poenaru, V\. T\. Mihailescu, V\. Ivanov, W\. Li, W\. Wang, W\. Jiang, W\. Bouaziz, W\. Constable, X\. Tang, X\. Wu, X\. Wang, X\. Wu, X\. Gao, Y\. Kleinman, Y\. Chen, Y\. Hu, Y\. Jia, Y\. Qi, Y\. Li, Y\. Zhang, Y\. Zhang, Y\. Adi, Y\. Nam, Yu, Wang, Y\. Zhao, Y\. Hao, Y\. Qian, Y\. Li, Y\. He, Z\. Rait, Z\. DeVito, Z\. Rosnbrick, Z\. Wen, Z\. Yang, Z\. Zhao, and Z\. Ma \(2024\)The llama 3 herd of models\.External Links:2407\.21783,[Link](https://arxiv.org/abs/2407.21783)Cited by:[§1](https://arxiv.org/html/2608.04576#S1.p3.1)\.
- K\. Guu, K\. Lee, Z\. Tung, P\. Pasupat, and M\. Chang \(2020\)Retrieval augmented language model pre\-training\.InProceedings of the 37th International Conference on Machine LearningThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7\-11, 2024The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7\-11, 2024Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),H\. D\. III, A\. Singh, L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Proceedings of Machine Learning Research, Vol\.119,pp\. 3929–3938\.External Links:[Link](https://proceedings.mlr.press/v119/guu20a.html)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p2.1)\.
- E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. Chen \(2022\)LoRA: low\-rank adaptation of large language models\.InThe Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25\-29, 2022,External Links:[Link](https://openreview.net/forum?id=nZeVKeeFYf9)Cited by:[Appendix F](https://arxiv.org/html/2608.04576#A6.SS0.SSS0.Px1.p1.2),[§1](https://arxiv.org/html/2608.04576#S1.p3.1)\.
- Y\. Hu, Z\. Li, Z\. Chen, Q\. Huang, Z\. Fu, M\. Xu, and L\. Nie \(2026\)REFINE: composed video retrieval via shared and differential semantics enhancement\.ACM Trans\. Multimedia Comput\. Commun\. Appl\.\.Note:Just AcceptedExternal Links:ISSN 1551\-6857,[Link](https://doi.org/10.1145/3796712),[Document](https://dx.doi.org/10.1145/3796712)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p3.1)\.
- A\. Hurst, A\. Lerer, A\. P\. Goucher, A\. Perelman, A\. Ramesh, A\. Clark, A\. Ostrow, A\. Welihinda, A\. Hayes, A\. Radford, A\. Mądry, A\. Baker\-Whitcomb, A\. Beutel, A\. Borzunov, A\. Carney, A\. Chow, A\. Kirillov, A\. Nichol, A\. Paino, A\. Renzin, A\. T\. Passos, A\. Kirillov, A\. Christakis, A\. Conneau, A\. Kamali, A\. Jabri, A\. Moyer, A\. Tam, A\. Crookes, A\. Tootoochian, A\. Tootoonchian, A\. Kumar, A\. Vallone, A\. Karpathy, A\. Braunstein, A\. Cann, A\. Codispoti, A\. Galu, A\. Kondrich, A\. Tulloch, A\. Mishchenko, A\. Baek, A\. Jiang, A\. Pelisse, A\. Woodford, A\. Gosalia, A\. Dhar, A\. Pantuliano, A\. Nayak, A\. Oliver, B\. Zoph, B\. Ghorbani, B\. Leimberger, B\. Rossen, B\. Sokolowsky, B\. Wang, B\. Zweig, B\. Hoover, B\. Samic, B\. McGrew, B\. Spero, B\. Giertler, B\. Cheng, B\. Lightcap, B\. Walkin, B\. Quinn, B\. Guarraci, B\. Hsu, B\. Kellogg, B\. Eastman, C\. Lugaresi, C\. Wainwright, C\. Bassin, C\. Hudson, C\. Chu, C\. Nelson, C\. Li, C\. J\. Shern, C\. Conger, C\. Barette, C\. Voss, C\. Ding, C\. Lu, C\. Zhang, C\. Beaumont, C\. Hallacy, C\. Koch, C\. Gibson, C\. Kim, C\. Choi, C\. McLeavey, C\. Hesse, C\. Fischer, C\. Winter, C\. Czarnecki, C\. Jarvis, C\. Wei, C\. Koumouzelis, D\. Sherburn, D\. Kappler, D\. Levin, D\. Levy, D\. Carr, D\. Farhi, D\. Mely, D\. Robinson, D\. Sasaki, D\. Jin, D\. Valladares, D\. Tsipras, D\. Li, D\. P\. Nguyen, D\. Findlay, E\. Oiwoh, E\. Wong, E\. Asdar, E\. Proehl, E\. Yang, E\. Antonow, E\. Kramer, E\. Peterson, E\. Sigler, E\. Wallace, E\. Brevdo, E\. Mays, F\. Khorasani, F\. P\. Such, F\. Raso, F\. Zhang, F\. von Lohmann, F\. Sulit, G\. Goh, G\. Oden, G\. Salmon, G\. Starace, G\. Brockman, H\. Salman, H\. Bao, H\. Hu, H\. Wong, H\. Wang, H\. Schmidt, H\. Whitney, H\. Jun, H\. Kirchner, H\. P\. de Oliveira Pinto, H\. Ren, H\. Chang, H\. W\. Chung, I\. Kivlichan, I\. O’Connell, I\. O’Connell, I\. Osband, I\. Silber, I\. Sohl, I\. Okuyucu, I\. Lan, I\. Kostrikov, I\. Sutskever, I\. Kanitscheider, I\. Gulrajani, J\. Coxon, J\. Menick, J\. Pachocki, J\. Aung, J\. Betker, J\. Crooks, J\. Lennon, J\. Kiros, J\. Leike, J\. Park, J\. Kwon, J\. Phang, J\. Teplitz, J\. Wei, J\. Wolfe, J\. Chen, J\. Harris, J\. Varavva, J\. G\. Lee, J\. Shieh, J\. Lin, J\. Yu, J\. Weng, J\. Tang, J\. Yu, J\. Jang, J\. Q\. Candela, J\. Beutler, J\. Landers, J\. Parish, J\. Heidecke, J\. Schulman, J\. Lachman, J\. McKay, J\. Uesato, J\. Ward, J\. W\. Kim, J\. Huizinga, J\. Sitkin, J\. Kraaijeveld, J\. Gross, J\. Kaplan, J\. Snyder, J\. Achiam, J\. Jiao, J\. Lee, J\. Zhuang, J\. Harriman, K\. Fricke, K\. Hayashi, K\. Singhal, K\. Shi, K\. Karthik, K\. Wood, K\. Rimbach, K\. Hsu, K\. Nguyen, K\. Gu\-Lemberg, K\. Button, K\. Liu, K\. Howe, K\. Muthukumar, K\. Luther, L\. Ahmad, L\. Kai, L\. Itow, L\. Workman, L\. Pathak, L\. Chen, L\. Jing, L\. Guy, L\. Fedus, L\. Zhou, L\. Mamitsuka, L\. Weng, L\. McCallum, L\. Held, L\. Ouyang, L\. Feuvrier, L\. Zhang, L\. Kondraciuk, L\. Kaiser, L\. Hewitt, L\. Metz, L\. Doshi, M\. Aflak, M\. Simens, M\. Boyd, M\. Thompson, M\. Dukhan, M\. Chen, M\. Gray, M\. Hudnall, M\. Zhang, M\. Aljubeh, M\. Litwin, M\. Zeng, M\. Johnson, M\. Shetty, M\. Gupta, M\. Shah, M\. Yatbaz, M\. J\. Yang, M\. Zhong, M\. Glaese, M\. Chen, M\. Janner, M\. Lampe, M\. Petrov, M\. Wu, M\. Wang, M\. Fradin, M\. Pokrass, M\. Castro, M\. O\. T\. de Castro, M\. Pavlov, M\. Brundage, M\. Wang, M\. Khan, M\. Murati, M\. Bavarian, M\. Lin, M\. Yesildal, N\. Soto, N\. Gimelshein, N\. Cone, N\. Staudacher, N\. Summers, N\. LaFontaine, N\. Chowdhury, N\. Ryder, N\. Stathas, N\. Turley, N\. Tezak, N\. Felix, N\. Kudige, N\. Keskar, N\. Deutsch, N\. Bundick, N\. Puckett, O\. Nachum, O\. Okelola, O\. Boiko, O\. Murk, O\. Jaffe, O\. Watkins, O\. Godement, O\. Campbell\-Moore, P\. Chao, P\. McMillan, P\. Belov, P\. Su, P\. Bak, P\. Bakkum, P\. Deng, P\. Dolan, P\. Hoeschele, P\. Welinder, P\. Tillet, P\. Pronin, P\. Tillet, P\. Dhariwal, Q\. Yuan, R\. Dias, R\. Lim, R\. Arora, R\. Troll, R\. Lin, R\. G\. Lopes, R\. Puri, R\. Miyara, R\. Leike, R\. Gaubert, R\. Zamani, R\. Wang, R\. Donnelly, R\. Honsby, R\. Smith, R\. Sahai, R\. Ramchandani, R\. Huet, R\. Carmichael, R\. Zellers, R\. Chen, R\. Chen, R\. Nigmatullin, R\. Cheu, S\. Jain, S\. Altman, S\. Schoenholz, S\. Toizer, S\. Miserendino, S\. Agarwal, S\. Culver, S\. Ethersmith, S\. Gray, S\. Grove, S\. Metzger, S\. Hermani, S\. Jain, S\. Zhao, S\. Wu, S\. Jomoto, S\. Wu, Shuaiqi, Xia, S\. Phene, S\. Papay, S\. Narayanan, S\. Coffey, S\. Lee, S\. Hall, S\. Balaji, T\. Broda, T\. Stramer, T\. Xu, T\. Gogineni, T\. Christianson, T\. Sanders, T\. Patwardhan, T\. Cunninghman, T\. Degry, T\. Dimson, T\. Raoux, T\. Shadwell, T\. Zheng, T\. Underwood, T\. Markov, T\. Sherbakov, T\. Rubin, T\. Stasi, T\. Kaftan, T\. Heywood, T\. Peterson, T\. Walters, T\. Eloundou, V\. Qi, V\. Moeller, V\. Monaco, V\. Kuo, V\. Fomenko, W\. Chang, W\. Zheng, W\. Zhou, W\. Manassra, W\. Sheu, W\. Zaremba, Y\. Patil, Y\. Qian, Y\. Kim, Y\. Cheng, Y\. Zhang, Y\. He, Y\. Zhang, Y\. Jin, Y\. Dai, and Y\. Malkov \(2024\)GPT\-4o system card\.External Links:2410\.21276,[Link](https://arxiv.org/abs/2410.21276)Cited by:[§4](https://arxiv.org/html/2608.04576#S4.p1.1)\.
- T\. D\. Jick \(1979\)Mixing qualitative and quantitative methods: triangulation in action\.Administrative science quarterly24\(4\),pp\. 602–611\.Cited by:[§1](https://arxiv.org/html/2608.04576#S1.p3.1)\.
- P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel, S\. Riedel, and D\. Kiela \(2020\)Retrieval\-augmented generation for knowledge\-intensive nlp tasks\.InProceedings of the 34th International Conference on Neural Information Processing Systems,NIPS ’20,Red Hook, NY, USA\.External Links:ISBN 9781713829546Cited by:[§1](https://arxiv.org/html/2608.04576#S1.p2.1)\.
- J\. Li, X\. Xu, S\. Ma, and S\. Li \(2025\)FaithAct: faithfulness planning and acting in mllms\.arXiv preprint arXiv:2511\.08409\.Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p3.1)\.
- Z\. Li, Y\. Hu, Z\. Chen, Q\. Huang, G\. Qiu, Z\. Fu, and M\. Liu \(2026a\)ReTrack: evidence\-driven dual\-stream directional anchor calibration network for composed video retrieval\.Proceedings of the AAAI Conference on Artificial Intelligence40\(28\),pp\. 23373–23381\.External Links:[Link](https://ojs.aaai.org/index.php/AAAI/article/view/39507),[Document](https://dx.doi.org/10.1609/aaai.v40i28.39507)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p3.1)\.
- Z\. Li, Y\. Hu, Z\. Chen, S\. Zhang, Q\. Huang, Z\. Fu, and Y\. Wei \(2026b\)HABIT: chrono\-synergia robust progressive learning framework for composed image retrieval\.Proceedings of the AAAI Conference on Artificial Intelligence40\(8\),pp\. 6762–6770\.External Links:[Link](https://ojs.aaai.org/index.php/AAAI/article/view/37608),[Document](https://dx.doi.org/10.1609/aaai.v40i8.37608)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p4.1)\.
- H\. Lin, Z\. Liu, Y\. Zhu, C\. Qin, J\. Lin, X\. Shang, C\. He, W\. Zhang, and L\. Wu \(2026\)MMFineReason: closing the multimodal reasoning gap via open data\-centric methods\.External Links:2601\.21821,[Link](https://arxiv.org/abs/2601.21821)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p4.1)\.
- P\. Liu, Z\. Cui, D\. Liang, and W\. Ye \(2025a\)Who stole your data? a method for detecting unauthorized rag theft\.External Links:2510\.07728,[Link](https://arxiv.org/abs/2510.07728)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p5.1)\.
- P\. Liu, X\. Wang, Z\. Cui, and W\. Ye \(2025b\)Queries are not alone: clustering text embeddings for video search\.InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval,SIGIR ’25,New York, NY, USA,pp\. 874–883\.External Links:ISBN 9798400715921,[Link](https://doi.org/10.1145/3726302.3730066),[Document](https://dx.doi.org/10.1145/3726302.3730066)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p2.1)\.
- P\. Liu, J\. Yang, L\. Wang, S\. Wang, Y\. Hao, and H\. Bai \(2023\)Retrieval\-based unsupervised noisy label detection on text data\.InProceedings of the 32nd ACM International Conference on Information and Knowledge Management,CIKM ’23,New York, NY, USA,pp\. 4099–4104\.External Links:ISBN 9798400701245,[Link](https://doi.org/10.1145/3583780.3615146),[Document](https://dx.doi.org/10.1145/3583780.3615146)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p4.1)\.
- Z\. Liu, H\. Liang, B\. Li, W\. Xiong, C\. Chen, C\. He, W\. Zhang, and B\. Cui \(2025c\)SynthVLM: towards high\-quality and efficient synthesis of image\-caption datasets for vision\-language models\.External Links:2407\.20756,[Link](https://arxiv.org/abs/2407.20756)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p4.1)\.
- Y\. Long, Y\. Liu, L\. Xu, and A\. Brintrup \(2025\)EmoDebt: bayesian\-optimized emotional intelligence for strategic agent\-to\-agent debt recovery\.External Links:2503\.21080,[Link](https://arxiv.org/abs/2503.21080)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p5.1)\.
- Y\. Long, Y\. Liu, and L\. Xu \(2026\)EmoMAS: emotion\-aware multi\-agent system for high\-stakes edge\-deployable negotiation with bayesian orchestration\.External Links:2604\.07003,[Link](https://arxiv.org/abs/2604.07003)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p5.1)\.
- K\. Ma, R\. Jin, W\. Haotian, W\. Xi, H\. Chen, Y\. Tang, and Q\. Wang \(2024\)Context\-driven index trimming: a data quality perspective to enhancing precision of RALMs\.InFindings of the Association for Computational Linguistics: EMNLP 2024,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 4886–4901\.External Links:[Link](https://aclanthology.org/2024.findings-emnlp.281/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-emnlp.281)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p2.1)\.
- X\. Ma, X\. Zhang, and Z\. Weng \(2026\)Stable and explainable personality trait evaluation in large language models with internal activations\.External Links:2601\.09833,[Link](https://arxiv.org/abs/2601.09833)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p5.1)\.
- P\. Manakul, A\. Liusie, and M\. Gales \(2023\)SelfCheckGPT: zero\-resource black\-box hallucination detection for generative large language models\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 9004–9017\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.557/),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.557)Cited by:[§2](https://arxiv.org/html/2608.04576#S2.p2.1)\.
- C\. Niu, Y\. Wu, J\. Zhu, S\. Xu, K\. Shum, R\. Zhong, J\. Song, and T\. Zhang \(2024\)RAGTruth: a hallucination corpus for developing trustworthy retrieval\-augmented language models\.Bangkok, Thailand,pp\. 10862–10878\.External Links:[Link](https://aclanthology.org/2024.acl-long.585/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.585)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p3.1)\.
- ReliefWeb \(2025\)ReliefWeb api documentation\.Note:[https://apidoc\.reliefweb\.int/](https://apidoc.reliefweb.int/)Service provided by UN OCHA; accessed 2025\-12\-20Cited by:[§1](https://arxiv.org/html/2608.04576#S1.p3.1),[§2](https://arxiv.org/html/2608.04576#S2.p1.1)\.
- R\. Rocca, N\. Tamagnone, S\. Fekih, X\. Contla, and N\. Rekabsaz \(2023\)Natural language processing for humanitarian action: opportunities, challenges, and the path toward humanitarian nlp\.Frontiers in big Data6,pp\. 1082787\.Cited by:[§2](https://arxiv.org/html/2608.04576#S2.p1.1)\.
- X\. Shi, W\. Zhao, T\. Chen, C\. Yang, and J\. Du \(2025\)Evidence triangulator: using large language models to extract and synthesize causal evidence across study designs\.Nature Communications16\(1\),pp\. 7355\.Cited by:[§1](https://arxiv.org/html/2608.04576#S1.p2.1),[§2](https://arxiv.org/html/2608.04576#S2.p2.1),[Table 2](https://arxiv.org/html/2608.04576#S3.T2),[§4\.1](https://arxiv.org/html/2608.04576#S4.SS1.p1.1)\.
- Q\. Wang, V\. B\. Nguyen, N\. Feldhus, L\. F\. Villa\-Arenas, C\. Seifert, S\. Möller, and V\. Schmitt \(2025a\)Truth or twist? optimal model selection for reliable label flipping evaluation in LLM\-based counterfactuals\.InProceedings of the 18th International Natural Language Generation Conference,L\. Flek, S\. Narayan, L\. H\. Phương, and J\. Pei \(Eds\.\),Hanoi, Vietnam,pp\. 80–97\.External Links:[Link](https://aclanthology.org/2025.inlg-main.5/)Cited by:[§3\.1](https://arxiv.org/html/2608.04576#S3.SS1.p3.1)\.
- Z\. Wang, L\. Cao, B\. Danek, Q\. Jin, Z\. Lu, and J\. Sun \(2025b\)Accelerating clinical evidence synthesis with large language models\.npj Digital Medicine8\(1\),pp\. 509\.Cited by:[§2](https://arxiv.org/html/2608.04576#S2.p2.1)\.
- Q\. Yang, Z\. Chen, Y\. Hu, Z\. Li, Z\. Fu, and L\. Nie \(2026\)STABLE: efficient hybrid nearest neighbor search via magnitude\-uniformity and cardinality\-robustness\.External Links:2604\.01617,[Link](https://arxiv.org/abs/2604.01617)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p2.1)\.
- Y\. Yao, D\. Ye, P\. Li, X\. Han, Y\. Lin, Z\. Liu, Z\. Liu, L\. Huang, J\. Zhou, and M\. Sun \(2019\)DocRED: a large\-scale document\-level relation extraction dataset\.InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics,A\. Korhonen, D\. Traum, and L\. Màrquez \(Eds\.\),Florence, Italy,pp\. 764–777\.External Links:[Link](https://aclanthology.org/P19-1074/),[Document](https://dx.doi.org/10.18653/v1/P19-1074)Cited by:[§2](https://arxiv.org/html/2608.04576#S2.p1.1)\.
- O\. Yoran, T\. Wolfson, O\. Ram, and J\. Berant \(2024\)Making retrieval\-augmented language models robust to irrelevant context\.External Links:[Link](https://openreview.net/forum?id=ZS4m74kZpH)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p2.1)\.
- Q\. Zhang, H\. Zhang, L\. Pang, Y\. Tong, H\. Zheng, and Z\. Zheng \(2026a\)Less is more: compact clue selection for efficient retrieval\-augmented generation reasoning\.InProceedings of the ACM Web Conference 2026,WWW ’26,New York, NY, USA,pp\. 1971–1982\.External Links:ISBN 9798400723070,[Link](https://doi.org/10.1145/3774904.3792158),[Document](https://dx.doi.org/10.1145/3774904.3792158)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p2.1)\.
- T\. Zhang, V\. Kishore, F\. Wu, K\. Q\. Weinberger, and Y\. Artzi \(2020\)BERTScore: evaluating text generation with BERT\.In8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26\-30, 2020,External Links:[Link](https://openreview.net/forum?id=SkeHuCVFDr)Cited by:[§3](https://arxiv.org/html/2608.04576#S3.SS0.SSS0.Px2.p1.2)\.
- W\. Zhang, X\. Zhang, H\. Yu, S\. Nie, B\. Wu, J\. Yue, T\. Liu, and Y\. Li \(2026b\)ExpSeek: self\-triggered experience seeking for web agents\.External Links:2601\.08605,[Link](https://arxiv.org/abs/2601.08605)Cited by:[Appendix E](https://arxiv.org/html/2608.04576#A5.p2.1)\.
- Y\. Zhang, F\. A\. Shaik, S\. Acharjee, F\. Khalid, and M\. Oussalah \(2026c\)Towards reliable multimodal disaster severity assessment through preference optimization and explainable vision\-language reasoning\.Reliability Engineering & System Safety,pp\. 112674\.External Links:ISSN 0951\-8320,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.ress.2026.112674),[Link](https://www.sciencedirect.com/science/article/pii/S0951832026004862)Cited by:[§2](https://arxiv.org/html/2608.04576#S2.p2.1)\.

## Appendix AData Filtering and Preprocessing

This section describes the dataset before filtering, including the full set of reports retrieved from ReliefWeb, the distribution across disaster types and source categories, and the filtering criteria applied to obtain the final dataset used in our experiments\.

#### Retrieval overview and disaster coverage\.

We first retrieve ReliefWeb reports using the keyword“cash assistance”and keep English reports with accessible plain text \(2000–2024\)\. Tab\.[3](https://arxiv.org/html/2608.04576#A1.T3)summarizes the disaster\-type distribution*before*applying our final selection criteria\. The corpus spans a wide range of hazards, withFlood,Epidemic,Drought, andEarthquakebeing the most frequently covered types\. To ensure sufficient evidence density and stable cross\-context analysis, we focus on these four high\-coverage disaster types in subsequent experiments\.

#### Source\-type composition over time\.

Because reporting volume and institutional participation vary substantially across years, we further examine the annual distribution by source type\. Tab\.[4](https://arxiv.org/html/2608.04576#A1.T4)provides pre\-filter year\-by\-year counts for the source categories, highlighting both long\-term growth in reporting and shifts in which institutions contribute reports\. Fig\.[3](https://arxiv.org/html/2608.04576#A1.F3)visualizes the same trend\.

#### Filtering rationale\.

Based on the above distributions, we apply filtering and selection to balance coverage and interpretability: \(i\) retain only reports with accessible text and valid metadata, \(ii\) restrict to the four most prevalent disaster types for stable cell\-level analysis, and \(iii\) map ReliefWeb sources into the source\-type taxonomy used in the main experiments\. The resulting dataset forms the basis for both extraction evaluation and downstream triangulation\.

Table 3:Distribution \(before applying the final selection criteria\) of ReliefWeb reports by disaster type \(2000–2024\)\.Table 4:Annual number of ReliefWeb reports by source type \(2000–2024\)\. Academic = Academic and Research Institution; Government = Government; Intl\. Org\. = International Organization; Media = Media; NGO = Non\-governmental Organization; Other = Other; RC/RC = Red Cross/Red Crescent Movement\.![Refer to caption](https://arxiv.org/html/2608.04576v1/Figures/00_Overall_Trend.png)Figure 3:Annual number of ReliefWeb reports by source type\.

## Appendix BAdditional Analyses

### B\.1Outcome distribution

#### Motivation for the food\-related case study\.

To choose an outcome family with sufficient evidence for long\-horizon triangulation, we inspect the frequency distribution of extracted outcome mentions\. Fig\.[4](https://arxiv.org/html/2608.04576#A2.F4)shows a frequency\-based word cloud of outcomes extracted by our best\-performing configuration, revealing that food\-related outcomes appear prominently\. This observation motivates the focus on food\-related outcomes in the main triangulation analysis \(Section[3\.2](https://arxiv.org/html/2608.04576#S3.SS2)\)\.

![Refer to caption](https://arxiv.org/html/2608.04576v1/Figures/outcome_cloud.png)Figure 4:Frequency\-based word cloud of extracted outcomes \(used to motivate the food\-related case study\)\.

### B\.2API pricing for efficiency metric

#### Cost accounting\.

To compare practical deployability across LLMs, we report a cost\-efficiency metric in Tab\.[2](https://arxiv.org/html/2608.04576#S3.T2)based on per\-token API prices\. Tab\.[5](https://arxiv.org/html/2608.04576#A2.T5)lists the input/output prices used in our calculations \(per 1M tokens\), and we convert qwen\-plus and deepseek\-v3 pricing from RMB using a fixed exchange rate stated in the caption\.

Table 5:API pricing used in cost\-efficiency calculations \(prices per 1M tokens\)\. And qwen\-plus/deepseek\-v3 are converted from RMB using 1 RMB = 0\.1414 USD\.

## Appendix CSampling Details of the Expert Evaluation Set

#### Sampling procedure\.

The expert evaluation set consists of 100 full\-length crisis reports randomly sampled from our 8029\-report corpus\. We use simple random sampling at the report level, without stratification by source type, year, or region\. Because each report is typically long and may contain multiple interventions and outcomes, expert annotation in this setting is substantially more time\- and expertise\-intensive than standard single\-label annotation\.

#### Source\-type composition of the sampled reports\.

As a descriptive check, the 100 sampled reports cover multiple major source categories in the corpus, including Red Cross/Red Crescent Movement \(34 reports\), Non\-governmental Organization \(32 reports\), Government \(25 reports\), Media \(5 reports\), and Academic and Research Institution \(4 reports\)\. Since the evaluation set is obtained through simple random sampling rather than stratified sampling, we do not expect its source\-type proportions to exactly match the corpus\-level distribution\. Instead, this breakdown is reported to show that the sampled reports span the major source categories represented in the corpus\.

## Appendix DPolarity Labeling for Frequent Outcomes

#### Rationale for labeling polarity only for frequent outcomes\.

In the triangulation case study, we label polarity only for outcome terms occurring at least 3 times\. This decision is motivated by the frequency distribution of outcome terms\. Specifically, outcome terms with frequency≥3\\geq 3comprise only 941 unique vocabularies, yet they cover 13554 out of 19568 samples \(69\.27%\)\. By contrast, outcome terms occurring only 1–2 times include 5080 unique vocabularies, but together they cover only 6014 out of 19568 samples \(30\.73%\)\. This distribution reflects a typical long\-tail pattern: a relatively small set of frequent outcomes covers the majority of the samples, whereas a large number of rare outcomes contributes limited sample coverage\. Given our effort under limited manpower, we prioritize the high\-frequency outcomes in polarity labeling\. We further clarify that these polarity labels are used only in the triangulation case study and do not affect the main extraction evaluation\.

## Appendix ESupplementary Discussion and Targeted Future Directions

Our results show that query\-conditioned extraction and snippet\-grounded relation classification can make long humanitarian reports more auditable and more amenable to context\-preserving triangulation\. Several broader methodological directions also suggest how this framework could be strengthened in future work\.

Retrieval reliability and clue selection remain key bottlenecks for long\-document humanitarian evidence extraction\. Our current corpus construction relies on keyword\-based retrieval, which is practical but can still surface reports in which the queried intervention is only marginally discussed\. More broadly, retrieval\-augmented modeling has long motivated modular and interpretable use of external evidenceGuuet al\.\([2020](https://arxiv.org/html/2608.04576#bib.bib77)\)\. For our setting, future versions of the pipeline could incorporate retrieval\-quality control and contextual filtering before extraction\. Context\-driven index trimming may help remove query\-inconsistent retrieved contexts, while compact clue selection can reduce reasoning cost by retaining only the minimum sufficient evidenceMaet al\.\([2024](https://arxiv.org/html/2608.04576#bib.bib74)\); Zhanget al\.\([2026a](https://arxiv.org/html/2608.04576#bib.bib61)\)\. Query clustering and hybrid search further suggest richer retrieval over both free text and structured metadata such as source type, disaster type, and timeLiuet al\.\([2025b](https://arxiv.org/html/2608.04576#bib.bib65)\); Yanget al\.\([2026](https://arxiv.org/html/2608.04576#bib.bib72)\)\. Adaptive retrieval and self\-reflective evidence use are also relevant: rather than always applying the same fixed pipeline depth, a future system could decide when extra retrieval or verification is necessary and remain robust when some retrieved content is irrelevantAsaiet al\.\([2024](https://arxiv.org/html/2608.04576#bib.bib78)\); Zhanget al\.\([2026b](https://arxiv.org/html/2608.04576#bib.bib60)\); Yoranet al\.\([2024](https://arxiv.org/html/2608.04576#bib.bib79)\)\.

Ensuring faithful, focused, and robust evidence grounding is equally important in this setting\. Our current method attaches supporting snippets and performs relation classification with access to both snippets and the full report, but future systems could evaluate more explicitly whether intermediate reasoning remains faithful to the cited evidence\. Work on faithfulness\-aware planning and acting is directly relevant here because it emphasizes evidence\-constrained reasoning rather than relying only on final\-answer qualityLiet al\.\([2025](https://arxiv.org/html/2608.04576#bib.bib73)\)\. Related work on hallucination analysis in retrieval\-augmented systems also suggests the value of finer\-grained checks for unsupported or weakly supported outputsNiuet al\.\([2024](https://arxiv.org/html/2608.04576#bib.bib80)\)\. In parallel, robust composed retrieval studies highlight transferable ideas for evidence calibration and focus control, such as aligning heterogeneous signals and reducing the influence of noisy or peripheral contentLiet al\.\([2026a](https://arxiv.org/html/2608.04576#bib.bib67)\); Chenet al\.\([2025](https://arxiv.org/html/2608.04576#bib.bib69)\)\. Related work on noise mitigation and differential semantics is also relevant when multiple snippets or contexts must be combined without collapsing distinct signals into a single oversimplified summaryChenet al\.\([2026](https://arxiv.org/html/2608.04576#bib.bib70)\); Huet al\.\([2026](https://arxiv.org/html/2608.04576#bib.bib71)\)\.

Further improvements are also likely to come from a more data\-centric treatment of both the extraction model and the distillation pipeline\. Recent work shows that gains can come not only from larger models, but also from better supervision design, stronger filtering, and more robust treatment of noisy data\. For our setting, this suggests several practical extensions: synthesizing or augmenting training instances for rare intervention–outcome patterns, curating higher\-quality teacher data, and selecting fine\-tuning examples based on difficulty and utilityLiuet al\.\([2025c](https://arxiv.org/html/2608.04576#bib.bib62)\); Linet al\.\([2026](https://arxiv.org/html/2608.04576#bib.bib63)\); Geet al\.\([2026](https://arxiv.org/html/2608.04576#bib.bib81)\)\. Retrieval\-based noisy\-label detection provides a natural starting point for identifying suspicious labels or low\-quality teacher supervisionLiuet al\.\([2023](https://arxiv.org/html/2608.04576#bib.bib66)\)\. Progressive robust learning under noisy correspondence further suggests that student training may benefit from improved filtering, curriculum design, or confidence\-aware weighting when teacher outputs are imperfectLiet al\.\([2026b](https://arxiv.org/html/2608.04576#bib.bib68)\)\.

As retrieval\-enhanced evidence systems mature, questions of provenance, release control, modular orchestration, and stable model characterization become increasingly importantMaet al\.\([2026](https://arxiv.org/html/2608.04576#bib.bib82)\)\. Methods for detecting unauthorized use of retrieval\-enhanced data suggest that future evidence platforms may need stronger provenance protection for curated corpora, extracted records, and derived knowledge storesLiuet al\.\([2025a](https://arxiv.org/html/2608.04576#bib.bib64)\)\. More broadly, high\-stakes multi\-agent work suggests a natural architectural extension of our current multi\-stage pipeline: separate agents could specialize in retrieval, verification, grounding, normalization, and triangulation, while a higher\-level controller dynamically allocates trust, requests more evidence, or routes difficult cases for human reviewLonget al\.\([2026](https://arxiv.org/html/2608.04576#bib.bib75),[2025](https://arxiv.org/html/2608.04576#bib.bib76)\)\. In our context, such an architecture may be more appropriate than a single monolithic extractor for handling long reports, heterogeneous evidence, and uncertainty\-sensitive humanitarian decision support\.

## Appendix FLoRA Fine\-Tuning Details

#### Implementation details \(LoRA SFT\)\.

We fine\-tune an open\-weightLlama\-3\.1\-8B\-Instructwith LoRAHuet al\.\([2022](https://arxiv.org/html/2608.04576#bib.bib58)\)and train two adapters to match our two\-step pipeline \(for Step 1 and Step 2\)\. For both adapters, we use rankr=8r\{=\}8\(lora\_target=all\), cutoff length 4096, 3 epochs, learning rate1​e−41\\mathrm\{e\}\{\-4\}\.

Training runs on a single node with 3×\\timesV100 GPUs; we evaluate/save every 111 steps and load the best checkpoint by dev loss\. Distillation data are built from a disjoint training split and exclude the 100 expert\-annotated evaluation reports\.

## Appendix GSummary of Notations

For readability, Tab\.[6](https://arxiv.org/html/2608.04576#A7.T6)summarizes the main notations used in this paper\.

Table 6:Summary of the main notations used in the paper\.