Fiduciary-Aware Language Models

• Fiduciary-aware language models are designed to embed legal and ethical fiduciary duties, ensuring trust, accountability, and compliance in sectors like finance and law.
• They employ multi-layered trustworthiness mechanisms including vetted data synthesis, rigorous validation, and dynamic attribution to maintain output traceability.
• These models use difficulty-adaptive training and legal feedback optimization to balance high task proficiency with regulatory adherence in complex, risk-sensitive applications.

Fiduciarily-aware LLMs (FALMs) are LLMs architected, trained, and evaluated with explicit mechanisms to embed and enforce fiduciary obligations—legal and ethical standards requiring agents to act with loyalty, care, and in the best interests of their principals. The central aim is to ensure that LLMs deployed in domains such as finance, law, and autonomous decision-making systems not only demonstrate task proficiency, but also operate with verifiable trustworthiness, accountability, and compliance to legal or regulatory fiduciary standards. This entails combining advancements in trust-oriented knowledge engineering, data synthesis, attribution, and evaluation, with legal-standard-informed training and prompt design (Zheng et al., 22 Jul 2025, Nay, 2023).

1. Theoretical Foundations of Fiduciary Standards in LLMs

Fiduciary duties—specifically the duty of loyalty (avoiding conflicts of interest) and the duty of care (acting with the prudence of a reasonable expert)—form the core of what FALMs must operationalize. In formal terms, a fiduciary relationship is defined as $R = (P, A, \{\text{Loyalty}, \text{Care}\})$, where $P$ is the principal, $A$ the agent (here, the LLM), and the duties are enforced over all $s \in S$ (states):

• Duty of Loyalty: $\arg\max_{a\in A} \mathbb{E}[u_P(s,a)]$ must coincide with the agent’s action.
• Duty of Care: $$\mathbb{E}[u_P(s,\pi^*(s))] \ge \max_{a'}\mathbb{E}[u_P(s,a')] - \varepsilon$$, with $\pi^*$ the fiduciary policy and $\varepsilon$ a prudence margin.

These standards are essential because explicit, hard-coded contracts or rules for all possible state-action pairs are neither feasible nor robust for high-dimensional, open-ended domains where LLMs operate. Therefore, domain-specific legal standards—expressed as "the spirit of the directive"—become the practical means for ensuring robust, aligned, and compliant model behaviors in real-world scenarios (Nay, 2023).

2. Multi-Layered Trustworthiness Assurance

The enforcement of fiduciary obligations in FALMs, as exemplified by the Agentar-Fin-R1 family, is structured through a multi-layered trustworthiness assurance framework:

• Knowledge Engineering (Source Trustworthiness): Construction of a vetted knowledge base $K = \{ k_1, ..., k_n \}$ exclusively from authoritative, audited disclosures and regulatory publications, processed via NER parsing, detoxification, and semantic validation.
• Multi-Agent Data Synthesis (Synthesis Trustworthiness): Domain-specialized agents $\{A_{t_1}, ..., A_{t_m}\}$ generate $P$0 triplets $P$1, which are diversified and refined by a self-evolution agent $P$2. The final synthetic corpus $P$3 supports broad task coverage and data quality.
• Data Validation & Governance (Governance Trustworthiness): Ensemble model consistency is quantified as

$P$4

with classifier agreement on outputs. A reasoning validity verifier $P$5 and a learned rating model $P$6 assign quality scores, producing the final dataset:

$P$7

This multi-layered design ensures that training data is contextually reliable, logically consistent, and adherent to fiduciary and regulatory mandates (Zheng et al., 22 Jul 2025).

3. Dynamic Attribution and Accountability Mechanisms

FALMs enforce traceability and provenance of their outputs through a dynamic attribution system. For every generated answer $P$8, an attribution record $P$9 encapsulates input, authoritative source $A$0, the generating agent $A$1, chain-of-thought, timestamp, and system-assigned confidence $A$2. Execution is governed by a logging and verification loop; during answer generation, any consistency deficit or failed reasoning check triggers a warning for possible hallucination. Only outputs that satisfy both factual and reasoning standards are delivered, and every response can be unambiguously traced back to its data lineage and generating process, supporting auditability and regulatory assurance (Zheng et al., 22 Jul 2025).

4. Difficulty-Adaptive Training and Optimization Strategies

FALM training maximizes both task proficiency and domain-aligned compliance by integrating label-guided, curriculum-like mechanisms:

• Difficulty-Aware Weighting: For each financial task label $A$3, relative difficulty is quantified by cross-model pass@$A$4 rates, leading to instance and label weighting:

$A$5

• Weighted Loss Functions: Supervised fine-tuning (SFT) uses

$A$6

and reinforcement-style optimization mixes GRPO with SFT.

• Two-Stage Pipeline: An initial weighted SFT on $A$7 is followed by multi-objective GRPO on hard tasks and targeted SFT to close residual gaps. Curriculum scheduling is implicit, with dynamic reweighting shifting training focus as per-progress on pass@$A$8 (Zheng et al., 22 Jul 2025).

5. Evaluation of Fiduciary Competence and Compliance

Rigorous assessment of FALMs’ fiduciary awareness employs both legal and domain-focused benchmarks. The Finova benchmark, introduced with Agentar-Fin-R1, targets real-world financial reasoning alongside compliance verification across agent capabilities, complex reasoning, and regulatory safety:

Category	Task	# Samples
Agent Capabilities	Intent Detection	150
	Slot Recognition	360
	Tool Planning	258
	Expression Generation	100
Complex Reasoning	Multi-step finance+code	306
Safety & Compliance	Regulatory/risk mitigation	200

Performance metrics include compliance accuracy, reasoning depth (chain-of-thought steps), and overall Finova score, e.g.,

$A$9

Empirical results show Agentar-Fin-R1-32B achieves a Finova average of 69.82%, with compliance at 87.00%, outperforming the previous best (DeepSeek-R1 at 81.00%) by 6 points. Source engineering and synthesis increase factual consistency ($s \in S$0), and 99.2% of statements are traceable to authoritative sources. Weighted and staged training close hard-task pass@1 gaps by 7–12 points (Zheng et al., 22 Jul 2025).

6. Legal Standard Understanding and Reinforcement Learning With Legal Feedback

LLMs’ fiduciary understanding can be systematically evaluated by distilling legal text into $s \in S$1 tuples from case law, where $s \in S$2 describes the factual state, $s \in S$3 the alleged fiduciary action, and $s \in S$4 the court’s legal reward (Positive, Negative, Unsure). Zero-shot classification accuracy on such tuples has been used to benchmark various LLMs: curie (27%), text-davinci-002 (73%), text-davinci-003 (78%). The approach supports future reinforcement learning with legal feedback (RLLF), where a reward model $s \in S$5, trained on legal tuples, guides policy $s \in S$6 optimization:

$s \in S$7

This framework enables the iterative alignment of LLMs to nuanced legal standards, foundational for scalable fiduciarily-aware systems (Nay, 2023).

7. Implications and Limitations

The adoption of fiduciarily-aware LLMs introduces heightened standards for data provenance, output auditability, and compliance to regulatory and legal standards, particularly in sensitive domains. Prompt-level and model-level incorporations of fiduciary instructions, pre-training on legal corpora, and supervised/RL fine-tuning on standard-driven datasets underpin practical implementation. However, current limitations include potential hallucination in LLM-generated labels, loss of nuance due to binarization of legal judgments, and underestimation of zero-shot evaluation capabilities relative to more advanced prompting techniques. Future directions include the expansion to finer-grained, multi-label datasets; more sophisticated chain-of-thought and multi-shot prompt pipelines; and end-to-end benchmarking of reinforcement learning with legal feedback scenarios across decision-critical applications (Zheng et al., 22 Jul 2025, Nay, 2023).