When Financial Fine-tuning Fails: A Three-Level Detectability Analysis of Numerical Hallucination in Domain-Adapted Language Models
cs.AI
Quick verdict
Claims financial fine-tuning destroys numerical restraint (L1 5.4%→90.8%); controlled, statistically careful, but single 7B model and synthetic-data-bound.
Summary
The paper asks whether domain fine-tuning improves or degrades numerical reliability in financial summarization. Using Mistral-7B-Instruct-v0.2, the authors compare a Base instruction-tuned model, a finance-language-adapted variant (FT-A), and a numeracy-enhanced variant (FT-A+B+C), all trained with QLoRA on 1,000-2,000 synthetic samples. They introduce a three-level detectability taxonomy (L1 overt currency fabrication, L2 covert-explicit bare magnitudes, L3 covert-implicit quantitative claims) crossed with two grounding conditions (S0 weak, S1 strong) over 240 samples (160 synthetic, 80 real 10-K excerpts). Headline results: L1 hallucination rises from 5.4% (Base) to 82.5% (FT-A) and 90.8% (FT-A+B+C); under weak grounding the fine-tuned models reach 90-100% L1 on synthetic inputs while Base stays at 0.8%. A volume-matched ablation (FT-A×3.5, 64.6% L1) rules out data volume as the driver. Statistical treatment is unusually careful: Wilson CIs, Fisher exact and McNemar tests with Bonferroni correction, plus 50-sample human validation (κ = 0.883 for L1), and detection artifacts inflating Base rates are candidly documented. Key limitations: a single 7B model and single training run with no seed variance; heuristic L2/L3 detectors with ~46% false-negative rates, making those levels lower bounds; and unexamined provenance of the memorized template values ('USD 3.3 billion', 'USD 1011 million') relative to the authors' own synthetic training corpus.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Domain fine-tuning on financial text destroys numerical restraint: L1 hallucination rises from 5.4% (Base) to 82.5% (FT-A) and 90.8% (FT-A+B+C) across 240 summarization samples.
- Numeracy supervision amplifies rather than mitigates hallucination at every detectability level, even though a numeracy-only model (FT-C) improves constrained arithmetic accuracy (32% tolerance vs 0% Base).
- Template injection — insertion of memorized canonical values regardless of input — is a primary hallucination mechanism, with specific fabricated phrases appearing in up to 50% of fine-tuned outputs.
- Under weak grounding (S0), fine-tuned models hallucinate in 90–100% of synthetic-input cases while the Base model stays at 0.8%, isolating restraint rather than reasoning as the failure point.
- A volume-matched ablation (FT-A×3.5, 64.6% L1) shows the discipline collapse is caused by training-data type, not data volume.
Strengths
Controlled factorial design with volume-matched ablation
Three QLoRA variants × three detectability levels × two grounding conditions × two data sources (Tables 2, C.16), plus FT-A×3.5 (Table 11) trained on 3,500 samples to show restraint collapse (64.6% L1) is driven by data type, not volume — the paper's most persuasive control.
Statistical treatment unusual for diagnostic studies
Wilson CIs (Table C.17), Fisher exact plus McNemar tests with Bonferroni correction for the paired design (Table 4), including an honestly reported non-significant L3 comparison (p=0.215) tied to a documented detector artifact (Appendix D.5).
Candid error accounting of its own detector
Concrete, quantified mechanism
Competence/discipline dissociation shown empirically
Weaknesses
Single backbone, single run, no seed variance
Template provenance never verified against training data
L2/L3 measurement fragility
Internal numeric inconsistencies
Undescribed automatic scorer
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.04806, cs.AI) asks whether domain fine-tuning improves or degrades numerical reliability in financial summarization. Its thesis is that 'numerical competence' and 'numerical discipline' are orthogonal — the 'restraint gap' — and that fine-tuning destroys the latter. Central claims: (1) L1 hallucination rises from 5.4% (Base) to 82.5% (FT-A) and 90.8% (FT-A+B+C) over 240 summarization samples (§4.4, Table 7); (2) numeracy supervision amplifies hallucination at all levels (§4.8); (3) 'template injection' — insertion of memorized canonical values ('USD 3.3 billion' at 37% of FT-A outputs, 'USD 1011 million' at 50% of FT-A+B+C outputs; §5.2) — is a primary mechanism; (4) S0 weak-grounding inputs are a catastrophic vulnerability (90–100% L1 on synthetic inputs for fine-tuned models vs 0.8% for Base). Note immediately: the abstract's 'FT-A+B+C reaches 98%' does not match Table 7's 90.8% — 98.3% is FT-A's template frequency, not a hallucination rate.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is numerical hallucination in financial LLM summarization?
It is the generation of numbers or quantitative claims that do not appear in the source document — e.g., inventing a revenue figure, growth rate, or cash flow value when summarizing a financial report. This paper grades such fabrications into three levels: L1 (obvious currency-denominated fakes like 'USD 1.3 billion'), L2 (professionally formatted numbers without currency markers like '3.3 billion'), and L3 (implicit quantitative claims like 'revenue increased').
Does fine-tuning an LLM on financial data make it hallucinate numbers more?
According to this study, dramatically yes — for 7B models fine-tuned with QLoRA. L1 (overt currency) hallucination rose from 5.4% in the base Mistral-7B-Instruct model to 82.5% after financial language fine-tuning and 90.8% after additional numeracy training. The result is demonstrated on a single model family and single training run, so generalization to other models is untested.
What is 'template injection' in this paper?
The insertion of memorized, canonical numeric values into summaries regardless of the input content. The fine-tuned models repeatedly emitted the same fabricated phrases — 'USD 3.3 billion' in 37% of FT-A outputs and 'USD 1011 million' in 50% of FT-A+B+C outputs — even when the source text contained no such figures, indicating reliance on memorized financial boilerplate over input-grounded generation.
Does adding numerical or math training data reduce hallucination?
Not in this study. The numeracy-enhanced model (FT-A+B+C) hallucinated more than the language-only fine-tuned model (FT-A) at every detectability level (+7.9pp L1, +9.5pp L3), even though a numeracy-only model (FT-C) scored better on constrained arithmetic tests (32% vs 0% tolerance accuracy). The authors conclude numerical competence and numerical restraint are separate capabilities.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt: middle sections (parts of §3.2–3.3, §4.4.2–4.6, §5.3–5.8, part of §6.5, Appendix D.1–D.3) were skipped; judgments touching those sections rely on the surrounding visible text.
- The public repository URL (Section 7) falls inside a sampling gap and could not be verified from the excerpt.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.