Paper: arXiv 2510.15993
Authors: Fernando Spadea, Oshani Seneviratne
Abstract
Most financial recommendation systems often fail to account for key behavioral and regulatory factors, leading to advice that is misaligned with user preferences, difficult to interpret, or unlikely to be followed. We present FLARKO (Financial Language-model for Asset Recommendation with Knowledge-graph Optimization), a novel framework that integrates Large Language Models (LLMs), Knowledge Graphs (KGs), and Kahneman-Tversky Optimization (KTO) to generate asset recommendations that are both profitable and behaviorally aligned. FLARKO encodes users’ transaction histories and asset trends as structured KGs, providing interpretable and controllable context for the LLM. To demonstrate the adaptability of our approach, we develop and evaluate both a centralized architecture (CenFLARKO) and a federated variant (FedFLARKO). To our knowledge, this is the first demonstration of combining KTO for fine-tuning of LLMs for financial asset recommendation. We also present the first use of structured KGs to ground LLM reasoning over behavioral financial data in a federated learning (FL) setting. Evaluated on the FAR-Trans dataset, FLARKO consistently outperforms state-of-the-art recommendation baselines on behavioral alignment and joint profitability, while remaining interpretable and resource-efficient.
Complexity vs Empirical Score
- Math Complexity: 5.5/10
- Empirical Rigor: 7.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper combines advanced concepts (KTO, knowledge graphs, federated learning, LLM fine-tuning) with substantial empirical evaluation on a specific dataset, clear performance metrics, and multi-model comparisons.
Research Flowchart
flowchart TD
A["Research Goal"] --> B["Develop FLARKO Framework"]
B --> C["Input Data"]
C --> D["KG & LLM Integration"]
D --> E["Apply KTO for Alignment"]
E --> F["Centralized vs. Federated"]
F --> G["Key Outcomes & Findings"]
subgraph C ["Input Data"]
C1["FAR-Trans Dataset"]
C2["User Transaction History"]
C3["Asset Trends"]
end
subgraph G ["Outcomes"]
G1["Superior Behavioral Alignment"]
G2["Joint Profitability"]
G3["Interpretable Recommendations"]
G4["Resource Efficient"]
end
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes FLARKO, a novel framework that combines large language models, knowledge graphs, and Kahneman‑Tversky Optimization to produce financially profitable and behaviorally aligned asset recommendations.
Source sentence
We present FLARKO (Financial Language-model for Asset Recommendation with Knowledge-graph Optimization), a novel framework that integrates Large Language Models (LLMs), Knowledge Graphs (KGs), and Kahneman-Tversky Optimization (KTO) to generate asset recommendations that are both profitable and behaviorally aligned.
Assumes 4
- Investor intent and preferences can be captured by encoding past transaction behavior in a Personal Knowledge Graph. data
Source sentence
A Personal Knowledge Graph (PKG): Encodes an investor’s past transaction behavior, capturing asset interactions over time and serving as a proxy for user intent and preferences.
- Mid‑sized language models (1.7B‑4B parameters) are sufficient to achieve state‑of‑the‑art performance, and larger models do not necessarily improve results. model
Source sentence
performance does not strictly increase with model size; in fact, models in the 1.7B to 4B parameter range often delivered the best results, even outperforming the larger 8B model.
- Federated learning can be performed without sharing raw sensitive data, using LoRA‑based parameter‑efficient tuning and 4‑bit quantization to keep communication costs low. execution
Source sentence
employs communication-efficient federated learning via low rank adaptation of LLMs (LoRA)-based parameter-efficient tuning.
- Behavioral alignment can be achieved by fine‑tuning the LLM with Kahneman‑Tversky Optimization using only binary feedback on recommendation desirability. model
Source sentence
To ensure that FLARKO’s recommendations align with actual user preferences and investment behavior, we fine-tune the LLM using KTO [2], a lightweight, behaviorally motivated alignment method. KTO requires only binary feedback on whether a model’s recommendation is desirable or not.
Data
- FAR-Trans market-prices unknown
Source sentence
Evaluated on the FAR-Trans dataset, FLARKO consistently outperforms state-of-the-art recommendation baselines on behavioral alignment and joint profitability.
Costs
Not applicable
Finds 4
- FLARKO consistently outperforms state‑of‑the‑art recommendation baselines on behavioral alignment and joint profitability. performance
Source sentence
FLARKO consistently outperforms state-of-the-art recommendation baselines on behavioral alignment and joint profitability.
- Mid‑sized models (1.7B‑4B parameters) achieve the best scores on preference and combined metrics, while the largest 8B model does not improve performance. performance
Source sentence
Qwen3-1.7B achieves the best scores on Pref@3 and Comb@3, while Qwen3-4B achieves the best Prof@3. The largest model, Qwen3-8B, does not outperform its smaller counterparts on any metric.
- In federated settings, Qwen3‑4B improves its peak preference and combined scores compared to its centralized counterpart, showing resilience to federation‑induced degradation. performance
Source sentence
Qwen3-4B manages to improve its results in the federated testing compared to its centralized counterpart, specifically in its peak Pref@3 and Comb@3 scores.
- Ablation studies confirm that both the Market Knowledge Graph and Personal Knowledge Graph contribute meaningfully to recommendation quality, with their combination yielding the strongest performance. performance
Source sentence
Ablation studies further validate FLARKO’s design: both historical market context (MKG) and personalized behavioral signals (PKG) contribute meaningfully to recommendation quality, with their combination yielding the strongest performance.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 21 related papers). Coverage is partial, so this is neither replication nor refutation.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (104)
- Federated fine-tuning of large language models: Kahneman-Tversky vs. direct preference optimization (2025)
- Bursting the Filter Bubble with Knowledge Graph Inversion (2025)
- Avoiding Over-Personalization with Rule-Guided Knowledge Graph Adaptation for LLM Recommendations (2025) — link
- On the opportunities and risks of foundation models (2021) — link
- Universal language model fine-tuning for text classification (2018) — link
- Federated multi-task learning (2017)
- Advances and open problems in federated learning (2021)
- Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision (2023)
and 96 more
- Illustrating Reinforcement Learning from Human Feedback (RLHF) (2022)
- Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations (2022)
- Policy shaping: Integrating human feedback with reinforcement learning (2013)
- Direct preference optimization: your language model is secretly a reward model (2023)
- LLaMA: Open and Efficient Foundation Language Models (2023)
- Language models are few-shot learners (2020)
- What Would Elsa Do? Freezing Layers During Transformer Fine-Tuning (2019)
- Parameter-efficient fine-tuning of large-scale pre-trained language models (2023)
- L o RA +: Efficient Low Rank Adaptation of Large Models (2024) — link
- KTO: Model Alignment as Prospect Theoretic Optimization (2024)
- Improving Weak-to-Strong Generalization with Scalable Oversight and Ensemble Learning (2024)
- Empirical Analysis of the Strengths and Weaknesses of PEFT Techniques for LLMs (2023)
- WRENCH: A Comprehensive Benchmark for Weak Supervision (2021)
- Deep reinforcement learning from human preferences (2017)
- Fine-Tuning Language Models from Human Preferences (2019)
- Sequence tutor: Conservative fine-tuning of sequence generation models with kl-control (2017)
- Direct preference optimization (DPO) (2023) — link
- Freezing layers of your deep learning model - the proper way of doing it (2023) — link
- BitFit: Simple Parameter-efficient Fine-tuning for Transformer-based Masked Language-models (2022)
- Compacter: Efficient low-rank hypercomplex adapter layers (2021)
- Openfedllm: Training large language models on decentralized private data via federated learning (2024)
- FedLLM-Bench: Realistic Benchmarks for Federated Learning of Large Language Models (2024) — link
- Universal and transferable adversarial attacks on aligned language models (2023) — link
- Judging llm-as-a-judge with mt-bench and chatbot arena (2023)
- Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality (2023)
- Communication-efficient learning of deep networks from decentralized data (2017)
- Federated optimization in heterogeneous networks (2020)
- Scaffold: Stochastic controlled averaging for federated learning (2020)
- Measuring the effects of non-identical data distribution for federated visual classification (2019) — link
- Adaptive Federated Optimization
- JudgeLM: Fine-tuned Large Language Models are Scalable Judges (2023) — link
- Stanford Alpaca: An Instruction-following LLaMA model (2023)
- TRL: Transformer Reinforcement Learning (2020)
- Measuring massive multitask language understanding (2020) — link
- Towards deep conversational recommendations (2018)
- Learning hierarchy-aware knowledge graph embeddings for link prediction (2020)
- Learning attention-based embeddings for relation prediction in knowledge graphs (2019) — link
- Representing text for joint embedding of text and knowledge bases (2015)
- Llm-assisted knowledge graph engineering: Experiments with chatgpt (2023)
- Unveiling User Preferences: A Knowledge Graph and LLM-Driven Approach for Conversational Recommendation (2024) — link
- KTOTrainer (2024)
- ReDial Dataset (2024)
- Qwen3-0.6B (2024)
- Fede: Embedding knowledge graphs in federated setting (2021)
- FedKGRec: privacy-preserving federated knowledge graph aware recommender system (2024)
- Flower: A friendly federated learning research framework (2020) — link
- The dual echo chamber: Modeling social media polarization for interventional recommending (2021)
- Mitigating Filter Bubbles within Deep Recommender Systems (2022) — link
- Calibrating noise to sensitivity in private data analysis (2006)
- Privacy-preserving cross-domain recommendation with federated graph learning (2024)
- Food.com Recipes and Interactions (2019) — link
- Recommender systems based on neuro-symbolic knowledge graph embeddings encoding first-order logic rules (2024)
- Bursting filter bubble: Enhancing serendipity recommendations with aligned large language models (2025) — link
- The filter bubble: What the Internet is hiding from you (2011)
- Exploring the filter bubble: the effect of using recommender systems on content diversity (2014)
- How the machine ‘thinks’: Understanding opacity in machine learning algorithms (2016)
- Recommender systems and their ethical challenges (2020)
- A survey of serendipity in recommender systems (2016)
- Diversity, serendipity, novelty, and coverage: a survey and empirical analysis of beyond-accuracy objectives in recommender systems (2016)
- Ripplenet: Propagating user preferences on the knowledge graph for recommender systems (2018)
- Kgat: Knowledge graph attention network for recommendation (2019)
- Harnessing Deep Neural Networks with Logic Rules (2016)
- Adaptive Personalized Knowledge Graph Summarization
- Explainable recommendation: A survey and new perspectives (2020)
- Adaptive Constraint Relaxation in Personalized Nutrition Recommendations: An LLM-Driven Knowledge Graph Retrieval Approach (2025)
- FAR-Trans: An Investment Dataset for Financial Asset Recommendation (2024) — link
- Stock Recommendations for Individual Investors: A Temporal Graph Network Approach with Mean-Variance Efficient Sampling — → 2404.07223
- Gemma 3 technical report (2025) — link
- Yarn: Efficient context window extension of large language models (2023) — link
- Knowledge graphs (2021)
- Findkg: Dynamic knowledge graphs with large language models for detecting global trends in financial markets — → 2407.10909
- Finbert: A pre-trained financial language representation model for financial text mining (2021)
- Bloomberggpt: A large language model for finance (2023) — link
- Fingpt: Democratizing internet-scale data for financial large language models (2023) — link
- Investlm: A large language model for investment using financial domain instruction tuning — → 2309.13064
- Finllama: Llm-based financial sentiment analysis for algorithmic trading (2024)
- Recommender Systems meet Finance: A literature review (2016)
- Effects of data sparsity on recommender systems based on collaborative filtering (2018)
- Facing the cold start problem in recommender systems (2014)
- Dynamic Risk Prediction in Financial-Production Systems Using Temporal Self-Attention and Adaptive Autoregressive Models (2025)
- Portfolio theory: as I still see it (2010)
- An Expert System for Financial Planners (1987)
- Insights into alignment: Evaluating dpo and its variants across multiple tasks (2024) — link
- JSON-LD 1.0 (2014)
- Lightgbm: A highly efficient gradient boosting decision tree (2017)
- Lightgcn: Simplifying and powering graph convolution network for recommendation (2020)
- Neural collaborative filtering vs. matrix factorization revisited (2020)
- Trust your neighbors: A comprehensive survey of neighborhood-based methods for recommender systems (2021)
- R. Fast algorithms for mining association rules (1994)
- Retrieval-augmented generation for knowledge-intensive nlp tasks (2020)
- Gemini: a family of highly capable multimodal models (2023) — link
- Lora: Low-rank adaptation of large language models. (2022)
- Artificial intelligence and fintech: An overview of opportunities and risks for banking, investments, and microfinance (2021)
- Blockchain-based Framework for Scalable and Incentivized Federated Learning (2025)
- Fed-RD: Privacy-preserving federated learning for financial crime detection (2024)
- A differentially private blockchain-based approach for vertical federated learning (2024)
3 of 104 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.