LLM Output Drift: Cross-Provider Validation & Mitigation for Financial Workflows
Authors: Raffi Khatchadourian, Rolando Franco
Venue: AI4F @ ACM ICAIF 2025 (Nov 15 in Singapore)
Paper: https://arxiv.org/abs/2511.07585
Code: https://github.com/ibm-client-engineering/output-drift-financial-llms
This repository is a Hugging Face landing page for the paper and its open-source implementation. It focuses on deterministic test harnesses, cross-provider validation, and risk-tiered deployment for financial LLM workflows (SEC 10-Ks, RAG over filings, JSON/SQL tasks).
🔑 Key finding
Well-engineered 7–8B models achieve 100% output consistency at T=0.0, while a 120B model reaches only 12.5% consistency, regardless of configuration.
Across 480 runs (5 models, 3 tasks, 2 temperatures, 3 concurrency levels), we show an inverse relationship between model size and determinism and map this to regulatory requirements (FSB, BIS, CFTC).
📊 Model tier classification
| Tier | Models | Consistency @ T=0.0 | Status | Recommended use |
|---|---|---|---|---|
| 1 | Granite-3-8B, Qwen2.5-7B | 100% | ✅ Production-ready | All regulated tasks |
| 2 | Llama-3.3-70B, Mistral-Medium-2505 | 56–100% | ⚠️ Task-specific | SQL / structured only |
| 3 | GPT-OSS-120B | 12.5% | ❌ Non-compliant | Not for compliance |
n = 480 runs (16 per condition), 95% Wilson CIs, p < 0.0001 (Fisher’s exact).
🎯 Why this matters
Financial institutions face a “verification tax”: human review erodes AI productivity gains when outputs are nondeterministic.
This framework shows:
- Audit-ready determinism is achievable with the right model + decoding setup.
- Cross-provider consistency: behavior transfers between local (Ollama) and cloud (IBM watsonx.ai).
- Task-specific drift: SQL and structured summaries remain stable even at T=0.2; RAG is far more sensitive.