LLM Output Drift in Financial Workflows: Validation and Mitigation (arXiv)
16 comments
·November 12, 2025raffisk
Empirical study on LLM output consistency in regulated financial tasks (RAG, JSON, SQL). Governance focus: Smaller models (Qwen2.5-7B, Granite-3-8B) hit 100% determinism at T=0.0, passing audits (FSB/BIS/CFTC), vs. larger like GPT-OSS-120B at 12.5%. Gaps are huge (87.5%, p<0.0001, n=16) and survive multiple-testing corrections.
Caveat: Measures reproducibility (edit distance), not full accuracy—determinism is necessary for compliance but needs semantic checks (e.g., embeddings to ground truth). Includes harness, invariants (±5%), and attestation.
Thoughts on inverse size-reliability? Planning follow-up with accuracy metrics vs. just repro.
throwdbaaway
It is the reasoning. During the reasoning process, the top few tokens have very similar or even same logprobs. With gpt-oss-120b, you should be able to get deterministic output by turning off reasoning, e.g. by appending:
{"role": "assistant", "content": "<think></think>"}
raffisk
Good call—reasoning token variance is likely a factor, esp with logprob clustering at T=0. Your <think></think> workaround would work, but we need reasoning intact for financial QA accuracy.
Also the mistral medium model we tested had ~70% deterministic outputs across the 16 runs for the text to sql gen and summarization in json tasks- and it had reasoning on. Llama 3.3 70b started to degrade and doesn’t have reasoning. But it’s a relevant variable to consider
doctorpangloss
“Determinism is necessary for compliance”
Says who?
The stuff you comply with changes in real time. How’s that for determinism?
raffisk
Author here—fair point, regs are a moving target . But FSB/BIS/CFTC explicitly require reproducible outputs for audits (no random drift in financial reports). Determinism = traceability, even when rules update at the very least
Most groups I work with stick to traditional automation/rules systems, but top-down mandates are pushing them toward frontier models for general tasks—which then get plugged into these workflows. A lot stays in sandbox, but you'd be surprised what's already live in fin services.
The authorities I cited (FSB/BIS/CFTC) literally just said last month AI monitoring is "still at early stage" cc https://www.fsb.org/2024/11/the-financial-stability-implicat...
Curious how you'd tackle that real-time changing reg?
nomel
Also, what happens if you add a space to the end of the prompt? Or write a 12.00 to 12.000?
raffisk
Good q—spacing could mess with tokenization, untested but def plausible. Worth a quick test on the setup - through the code for the fin svcs harness for tinkering / testing diff prompts/model arch’s based on feedback https://github.com/ibm-client-engineering/output-drift-finan...
colechristensen
Outputs not being deterministic with temperature = 0 doesn't match my understanding of what "temperature" meant, I thought the definition of T=0 was determinism.
Is this perhaps inference implementation details somehow introducing randomness?
kakugawa
Defeating Nondeterminism in LLM Inference
https://news.ycombinator.com/item?id=45200925
https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
> As it turns out, our request’s output does depend on the parallel user requests. Not because we’re somehow leaking information across batches — instead, it’s because our forward pass lacks “batch invariance”, causing our request’s output to depend on the batch size of our forward pass.
tl;dr: the way inference is batched introduces non-determinism.
measurablefunc
This is b/c these things are Markov chains. You can not expect consistent results & outputs.
SrslyJosh
Using an LLM for a "financial workflow" makes as much sense as integrating one with Excel. But who needs correct results when you're just working with money, right? ¯\_(ツ)_/¯
mirekrusin
Humans are non deterministic yet they use excel, work with financial workflows and deal with the money.
thfuran
And because one system that aims to achieve deterministic operation can’t quite perfectly do so, we might as well abandon any attempt at determinism?
Terr_
"Humans make math errors, yet they do math anyway, therefore this calculator that makes errors is also OK."
What do you call a fallacy where the universe is imperfect therefore nobody can have higher standards for acceptable flaws?
Mankind has spent literal centuries observing deficiencies and faults in human bookkeeping and calculation, constantly trying to improve it with processes and machinery. There's no good reason to suddenly stop caring about those issues simply because the latest proposal is marketed as "AI".
ACCount37
Did you actually read what the paper was about before leaving a low quality comment?
Don't use LLMs for financial workflows. Use them to create software for financial workflows. Software doesn't "drift".