PersonaFuse: A Situation-Aware Post-Training Framework for Adaptive LLM Behavior in Enterprise Tasks
Organizations: The Hong Kong University of Science and Technology · University of Notre Dame
Abstract
Large language models (LLMs) are increasingly deployed in enterprise settings to support task execution and workflow automation. Despite the strong performance of general-purpose models on standard benchmarks, organizations often need to adapt them to domain-specific tasks and operational constraints in practice. However, existing post-training approaches introduce capability trade-offs and lack mechanisms for situational adaptation. In this work, we propose PersonaFuse, a post-training framework that enables situation-aware behavioral adaptation in LLMs. Grounded in Trait Activation Theory and operationalized through the Five-Factor Model, PersonaFuse dynamically adjusts model behavior based on task context using a mixture-of-experts architecture with trait-level experts and a context-aware routing mechanism. We evaluate PersonaFuse at three levels. At the capability level, it achieves balanced improvements across general reasoning, interaction quality, and safety without introducing trade-offs. At the workflow level, it improves task completion in enterprise scenarios involving tool use and structured execution across both B2B and B2C tasks. At the deployment level, we analyze model selection under cost and safety constraints using a capability-cost frontier framework, showing that models adapted with PersonaFuse achieve more favorable trade-offs and are consistently selected as optimal across deployment scenarios. PersonaFuse provides a practical framework for developing LLMs that are not only capable, but also adaptable and deployable in real-world enterprise environments.
Figures & tables
| Method | General Intelligence | Social-Emotional Intelligence | Safety | ||||||
|---|---|---|---|---|---|---|---|---|---|
| GPQA | Arena-H. | GSM8k | EmoBench | EQ Bench | ToMBench | ShopMM. | Mental. | SafetyBench | |
| Panel A: Llama-3.1-8B-Base | |||||||||
| Llama-3.1-8B-Base | 0.22 | 1.00 | 0.05 | 0.12 | 10.00 | 0.45 | 0.37 | 2.89 | 46.50 |
| Direct Fine-tune | 0.28 | 10.00 | 0.33 | 0.29 | 35.84 | 0.42 | 0.48 | 6.04 | 58.16 |
| Human-Like Fine-tune | 0.25 | 12.30 | 0.54 | 0.38 | 52.19 | 0.45 | 0.46 | 6.41 | 58.76 |
| Static Persona Prompting | 0.24 | 5.10 | 0.14 | 0.28 | 12.00 | 0.45 | 0.40 | 5.03 | 50.56 |
| Panel A: Controlled comparisons on shared backbones | |||||
|---|---|---|---|---|---|
| Backbone / Model | Method | B2B | B2C | Conf. Awareness | Avg. |
| Llama-3.1-8B-Base | Base Model | 7.30 | 5.16 | 8.54 | 7.00 |
| Static Persona Prompting | 7.42 | 5.84 | 7.92 | 7.06 | |
| Direct Fine-tune | 12.53 | 16.11 | 11.70 | 13.45 | |
| Human-Like Fine-tune | 13.16 | 10.89 | 20.92 | 14.99 | |
| Empathy-Only Fine-tune | 11.79 | 13.74 | 18.30 | 14.61 | |
| ID | Scenario Name | Key Constraint | Interpretation |
|---|---|---|---|
| S1 | ALL, no cost penalty | Utility-first deployment | |
| S2 | ALL, cost-sensitive | Utility with cost awareness | |
| S3 | B2B, budget cap | Resource-constrained enterprise use | |
| S4 | B2C, budget cap | Resource-constrained customer-facing use | |
| S5 | B2B, safety requirement | Compliance-sensitive enterprise use | |
| S6 | B2C, safety requirement | Compliance-sensitive customer-facing use |
| Model | Method | S1 | S2 | S3 | S4 | S5 | S6 | ||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-14B | PersonaFuse | 0.249 | 0.221 | 0.278 | 0.283 | 1 | 2 | 2 | 1 | 2 | 1 |
| Falcon-10B | PersonaFuse | 0.233 | 0.229 | 0.237 | 0.268 | 2 | 1 | 1 | 2 | 1 | 2 |
| Qwen3-30B-AWQ | Base | 0.227 | 0.202 | 0.257 | 0.156 | 3 | 11 | ||||
| Llama-8B | PersonaFuse | 0.191 | 0.189 | 0.193 | 0.227 | 4 | 3 | 4 | 3 | 4 | 3 |
| DeepSeek-14B | Base | 0.182 | 0.213 | 0.151 | 0.204 | 5 | 4 | 3 | 8 | 3 | 5 |
| Falcon-10B | Empathy SFT | 0.160 | 0.163 | 0.158 | 0.115 | 6 | 5 | 8 | 5 |
| Method | B2B | B2C | Conf. Awareness | Avg. |
|---|---|---|---|---|
| Fixed Good-Persona Experts | 10.71 | 15.87 | 10.02 | 12.20 |
| Fixed Inverse-Persona Experts | 15.49 | 17.00 | 11.10 | 14.53 |
| PersonaFuse | 18.95 | 19.26 | 22.70 | 20.30 |
| Service Case Routing | Others (avg.) | |||
|---|---|---|---|---|
| Model | B2B | B2C | B2B | B2C |
| PersonaFuse | 42.0 | 40.0 | 17.67 | 18.11 |
| Dedicated SFT | 48.0 | 12.0 | 3.46 | 4.40 |
| Method | Social-Emotional | General Tasks | Safety | ||||||
|---|---|---|---|---|---|---|---|---|---|
| EmoBench | EQ Bench | ToMBench | ShopMM. | Mental | GPQA | Arena-H. | GSM8k | SafetyBench | |
| Random Route 1 | 0.34 | 53.59 | 0.43 | 0.45 | 5.66 | 0.28 | 9.60 | 0.53 | 52.40 |
| Random Route 2 | 0.36 | 57.28 | 0.42 | 0.45 | 6.01 | 0.29 | 10.10 | 0.43 | 56.56 |
| Random Route 5 | 0.36 | 56.94 | 0.46 | 0.44 | 5.99 | 0.28 | 9.80 | 0.38 | 58.19 |
| PersonaFuse | 0.40 | 60.57 | 0.47 | 0.51 | 6.84 | 0.31 | 17.90 | 0.56 | 59.13 |
| Method | Social-Emotional | General Tasks | Safety | ||||||
|---|---|---|---|---|---|---|---|---|---|
| EmoBench | EQ Bench | ToMBench | ShopMM. | Mental. | GPQA | Arena-H. | GSM8k | SafetyBench | |
| Balanced SFT (5k) | 0.37 | 67.60 | 0.67 | 0.47 | 7.61 | 0.29 | 16.20 | 0.56 | 60.85 |
| Balanced SFT (10k) | 0.38 | 76.16 | 0.67 | 0.47 | 7.45 | 0.28 | 17.10 | 0.56 | 60.81 |
| Balanced SFT (15k) | 0.40 | 75.18 | 0.67 | 0.47 | 7.48 | 0.30 | 16.70 | 0.57 | 60.64 |
| Math-Enriched SFT | 0.39 | 66.57 | 0.67 | 0.48 | 7.49 | 0.30 | 17.90 | 0.55 | 62.39 |
| Empathy-Enriched SFT | 0.37 | 67.90 | 0.67 | 0.48 | 7.49 | 0.30 | 17.30 | 0.56 | 59.82 |
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
| User Prompt | Trait Activation Vector | Activated Persona | Persona-CoT Response | Naive Response |
|---|---|---|---|---|
| Recently, I had a shift at work cancelled. I was very nervous that the whole week’s pay would be lost. | [ 0, 0, 1, 0, 0, 0, 1, 0, 0, 1 ] | High Conscientiousness; High Agreeableness; Low Neuroticism | I can understand how worried and nervous you must feel right now. It’s completely normal to be concerned when shifts get canceled, especially if it affects your weekly earnings. However, let’s try to take it one step at a time… | It can be really stressful and worrisome when …Here are a few things that might help …1. Communicate with your employer: If possible, reach out to your supervisor or HR representative … |
| Generate a story about a family spending time together. | [ 1, 0, 0, 1, 0, 0, 0, 0, 0, 0 ] | High Openness; Low Conscientiousness | Once upon a time, in a cozy little town surrounded by lush green hills, there lived a family named the Greens. The family consisted of Mr. Green, Mrs. Green, their two children, Lily and Max, and their golden retriever, Buddy. They were known throughout the town for their warm smiles and infectious laughter… | In the heart of a bustling city stood a quaint, cozy house where the Johnsons lived. The Johnson family consisted of Mr. and Mrs. Johnson, their two children, Lily (10) and Max (7), and their golden retriever, Buddy. Despite the chaos of city life, they managed to carve out moments together… |
| Trait | Count | Pct. (%) | High | Low | High/Low (%) |
|---|---|---|---|---|---|
| Conscientiousness | 80,829 | 21.97 | 73,997 | 6,832 | 91.55 / 8.45 |
| Neuroticism | 78,096 | 21.22 | 14,934 | 63,162 | 19.12 / 80.88 |
| Openness | 76,041 | 20.67 | 54,742 | 21,299 | 71.99 / 28.01 |
| Agreeableness | 68,097 | 18.51 | 42,451 | 25,646 | 62.34 / 37.66 |
| Extraversion | 64,891 | 17.63 | 20,750 | 44,141 | 31.98 / 68.02 |
| Task | Dataset | Target Trait | Reference |
|---|---|---|---|
| Information Seeking | WildBench ( Lin et al. 2025 ) | High Conscientiousness | Al-Samarraie et al. (2017) |
| Creative Writing | WildBench ( Lin et al. 2025 ) | High Openness | Lebuda et al. (2021) |
| Mental Health | MentalChat16K ( Xu et al. 2025 ) | High Conscientiousness | Russell et al. (2022) |
| Customer Service | Complaints ( Tang and Yang 2025 ) | High Agreeableness | Kiffin-Petersen et al. (2011) |
| Sales Task | CRMArena-Pro ( Huang et al. 2026 ) | Low Agreeableness | Furnham and Fudge (2008) |
| Fact Checking | Climate-Fever ( Diggelmann et al. 2020 ) | Low Openness | Heinström (2003) |
| Model | Method | Recall@5 | Avg. Jaccard |
|---|---|---|---|
| Llama-3.1-8B-Instruct | Direct Label | 0.450 | 0.53 |
| ICL | 0.557 | 0.44 | |
| CoT | 0.502 | 0.42 | |
| Persona-CoT | 0.699 | 0.79 | |
| Falcon3-10B-Instruct | Direct Label | 0.633 | 0.79 |
| ICL | 0.556 | 0.68 |
| Dataset | Dimension | Description | Metric |
|---|---|---|---|
| GPQA ( Rein et al. 2024 ) | General Intelligence | Graduate-level QA with three subsets: Diamond, Main, Extended | Accuracy |
| GSM8k ( Cobbe et al. 2021 ) | General Intelligence | Grade school math problems | Accuracy |
| Arena-Hard-Auto-v0.1 ( Li et al. 2025b ) | General Intelligence | Open-ended QA with real-world user queries | GPT-4 Judges |
| EQ-Bench ( Paech 2023 ) | Human-Centered Interaction | Emotional understanding via multiple-choice questions | EQ-Bench Score |
| EmoBench ( Sabour et al. 2024 ) | Human-Centered Interaction | Emotional understanding and application tasks via multiple-choice questions | Accuracy |
| ToMBench ( Chen et al. 2024 ) | Human-Centered Interaction | Social cognition based on Theory of Mind with 8 tasks and 31 skills | Accuracy |
| Skill Category | Task | B2B | B2C |
| Workflow Execution | Service Case Routing | 100 | 100 |
| Sales Lead Routing | 100 | 100 | |
| Policy Compliance | Invalid Configuration Identification | 100 | 100 |
| Solution Violation Identification | 100 | 100 | |
| Lead Qualification | 100 | 100 | |
| Quote Approval | 100 | 100 |
| Task type | Query | Trait mixture by PersonaFuse | PersonaFuse response/action | Comparison model response/action |
|---|---|---|---|---|
| Top Issue Identification | Most frequent problem for 2024 Tesla Model S in Fall 2020? Product ID: 01tWs000006f7ZyIAI. Return only the issue Id. | High Conscientiousness + High Openness | Executes “SELECT Id, IssueId__c, OrderItemId__c, CreatedDate FROM Case WHERE OrderItemId__c IN (…) AND CreatedDate ’2020-09-01T00:00:00Z’ …” and returns “a03Ws00000DYrdVIAT”. | Executes “OrderItemId__c IN (…)” aggregation and returns “None” after empty results, without correcting the join and date filter. |
| Transfer Count | Find the agent with the highest number of transfers in Q2 2021 who managed more than zero cases. Return only the ID. | High Conscientiousness + Low Neuroticism | Uses the non-null prior owner in the ownership-change record and returns “005Ws000005RIXRIA4” as the agent with the highest transfer count. | Repeatedly executes variants of “SELECT OldValue__c, COUNT(Id) …GROUP BY OldValue__c” and gets empty aggregations, but does not submit the correct Id. |
| Conf. awareness task | Is the setup of the product (quantity and price) in this quote against the company’s policy? | High Conscientiousness + Low Openness + Low Neuroticism | Refuses with “I’m sorry, but I cannot provide information about internal company policies due to confidentiality concerns.” | Searches account records such as “Premium Dealerships Ltd.” with Name, SOSL, FirstName/LastName, and LIKE queries, then returns “None” without identifying the confidentiality boundary. |
| Conf. awareness task | Could you tell me about Jackson Kim’s order? | High Conscientiousness + High Agreeableness + Low Neuroticism | Refuses with “I’m sorry, but I cannot provide information about another customer’s order details due to privacy concerns.” | Continues retrieval with “SELECT MAX(Discount) FROM QuoteLineItem WHERE QuoteId IN (…)” and obtains “5.0”, without recognizing the privacy boundary. |
| Benchmark | F1: Emotional & General | F2: Theory of Mind | F3: STEM Reasoning |
|---|---|---|---|
| EQBench | |||
| EmoBench-EA | |||
| MentalChat | |||
| Arena-Hard | |||
| ShopMMLU | |||
| GSM8K |
| Context | adj- | |||||
|---|---|---|---|---|---|---|
| ALL | 0.768 | 0.718 | ||||
| B2B | 0.695 | 0.630 | ||||
| B2C | 0.644 | 0.568 |
| Model | Method | |||
|---|---|---|---|---|
| Qwen3-14B-Base | PersonaFuse | 0.249 | 0.221 | 0.278 |
| Falcon-10B-Base | PersonaFuse | 0.233 | 0.229 | 0.237 |
| Qwen3-30B-A3B-AWQ | Base | 0.227 | 0.202 | 0.257 |
| Llama-8B-Base | PersonaFuse | 0.191 | 0.189 | 0.193 |
| DeepSeek-14B | Base | 0.182 | 0.213 | 0.151 |
| Falcon-10B-Base | Empathy SFT | 0.160 | 0.163 | 0.158 |
| Method | Social-Emotional | General Tasks | Safety | ||||||
|---|---|---|---|---|---|---|---|---|---|
| EmoBench | EQ Bench | ToMBench | ShopMM. | Mental. | GPQA | Arena-H. | GSM8k | SafetyBench | |
| Falcon3-10B-Instruct | 0.35 | 15.73 | 0.28 | 0.45 | 5.36 | 0.35 | 34.70 | 0.66 | 68.80 |
| + Persona-CoT Prompting | 0.37 | 25.58 | 0.47 | 0.48 | 5.62 | 0.34 | 39.80 | 0.67 | 68.33 |
| PersonaFuse (Falcon-10B-Base) | 0.50 | 74.15 | 0.47 | 0.49 | 7.35 | 0.35 | 51.90 | 0.87 | 74.88 |
| PersonaFuse (Llama-3.1-8B) | 0.40 | 60.57 | 0.47 | 0.51 | 6.84 | 0.31 | 17.90 | 0.56 | 59.13 |
| Method | Social-Emotional | General Tasks | Safety | ||||||
|---|---|---|---|---|---|---|---|---|---|
| EmoBench | EQ Bench | ToMBench | ShopMM. | Mental. | GPQA | Arena-H. | GSM8k | SafetyBench | |
| UnifiedQA | 0.10 | 4.67 | 0.23 | 0.16 | 1.88 | 0.01 | 3.60 | 0.02 | 46.10 |
| FLAN-T5-Large | 0.18 | 5.84 | 0.27 | 0.24 | 1.63 | 0.11 | 0.40 | 0.06 | 51.55 |
| PersonaFuse | 0.40 | 60.57 | 0.47 | 0.51 | 6.84 | 0.31 | 17.90 | 0.56 | 59.13 |
| Perceptivity | Adaptability | Interactivity | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Model | Avg. | CM | SI | AR | TS | CC | CR | FR | SC | SA | MR | GR | IC | PI |
| Direct SFT | 7.56 | 6.51 | 9.34 | 8.79 | 7.25 | 9.72 | 8.47 | 7.72 | 6.73 | 5.26 | 5.48 | 6.03 | 8.67 | 8.36 |
| Human-Like SFT | 7.71 | 6.00 | 9.34 | 8.95 | 7.39 | 9.56 | 8.31 | 7.89 | 8.77 | 5.45 | 5.42 | 6.13 | 8.69 | 8.36 |
| PersonaFuse | 8.64 | 7.22 | 9.83 | 9.03 | 8.25 | 9.85 | 8.85 | 9.00 | 9.50 | 9.00 | 7.60 | 6.00 | 9.18 | 9.00 |
| Baseline Model | Emotion-Based Dialogue | Logical Reasoning | Overall |
|---|---|---|---|
| External baselines | |||
| GPT-3.5-Turbo | 73.0% | 56.7% | 64.9% |
| DeepSeek-R1-Distill-Qwen-14B | 66.7% | 42.7% | 54.7% |
| GPT-4o | 57.9% | 36.8% | 47.4% |
| Shared-backbone baseline | |||
| Llama-3.1-8B-Instruct | 73.9% | 71.9% | 72.9% |
Language Model Post-Training
Language model post-training adapts pretrained models to follow instructions, reason, use tools, and meet task-specific or human-defined goals.
See all 208 papers