The paper investigates numerical hallucination in financial language models, comparing a base instruction-tuned model, a domain-adapted model (FT‑A), and a numeracy-enhanced domain model (FT‑A+B+C). A three-level detectability taxonomy is introduced, and the study finds that domain fine-tuning dramatically increases hallucination rates—up to 98%—and that numeracy supervision actually amplifies hallucination. Template injection is identified as a key mechanism, leading the authors to conclude that numerical hallucination stems from degraded numerical restraint rather than lack of reasoning, and they recommend comprehensive evaluation and grounding-aware deployment practices.
arXiv:2606.21359v2 Announce Type: replace Abstract: Large language models (LLMs) are increasingly used to communicate and explain scientific concepts, yet their tendency to hallucinate poses signific...
arXiv:2609.32964v2 Announce Type: replace Abstract: Language models (LMs) often hallucinate by committing to confident answers rather than abstaining, even when they do not have enough information to...
arXiv:2606. 07537v1 Announce Type: cross Abstract: Large language models hallucinate--producing fluent, confident, factually wrong outputs--with a consistency that persists across generations and scales.
arXiv:2607. 26368v1 Announce Type: cross Abstract: Financial disclosures contain numerical claims, temporal statements, entity references, policy commitments, and risk descriptions that may conflict in qualitatively different ways.
The study investigates whether large language models (LLMs) can generalize arithmetic reasoning across different input formats, such as numeric and verbal expressions in English, Spanish, and Italian. By using attribution patching, the researchers identified the specific internal circuits each model employs for numeric versus verbal arithmetic tasks. They found that the degree of overlap between a model’s numeric circuit and its verbal circuit predicts how well the model generalizes to verbal formats, explaining variations in performance across languages and formats without requiring labeled data.
Financial disclosures contain numerical claims, temporal statements, entity references, policy commitments, and risk descriptions that may conflict in qualitatively different ways. Detecting a conflict is only the first step: review workflows may also need to determine its type, since numerical, temporal, referential, factual, and normative inconsistencies require different evidence and downstream checks.
arXiv:2603.13911v2 Announce Type: replace Abstract: A language model can give the wrong answer even when the correct answer is decodable from its intermediate states. To study this gap between decoda...
The paper proposes using the temporal volatility of internal attention mechanisms—measured by an unsupervised attention dispersion metric—as a diagnostic signal for hallucinations in large language models. It demonstrates that spikes in attention entropy within intermediate layers correlate with reasoning breakdowns, and shows statistically significant AUC improvements of up to +0.076 over output-based baselines on GSM8K and MATH-500 benchmarks using the Qwen2.5 model family.
arXiv:2607. 17883v1 Announce Type: cross Abstract: Enterprises will not deploy AI agents they cannot trust, and the most-cited reason for distrust is hallucination: confident, fluent output that is simply not true.
arXiv:2607. 19257v1 Announce Type: cross Abstract: Practitioners make three prompt-design decisions with almost no controlled evidence behind them: how to format instructions and context (markdown, plain text, prose, or tabular), how many simultaneous instructions a system prompt can carry before compliance degrades, and how much context a model can hold before recall and honesty degrade.
FinVector-Market-4B adapts the Qwen/Qwen3.5-4B language model using rank‑16 LoRA on a 22,000‑example financial corpus. On a 600‑example benchmark with implicit and explicit JSON‑schema contracts, the adapted model improves JSON validity from 0% to 91.3% and boosts task‑specific metrics—FinQA exact match from 14.7% to 40.0%, calculator‑expression correctness from 48.0% to 82.7%, scenario branch‑label agreement from 20.1% to 89.5%, and implication‑direction agreement from 52.4% to 87.2%. A post‑hoc audit shows that macro‑F1 differences stem from changing label sets, and the benchmark’s generalization claims are limited by filing overlap and calculator‑target inconsistencies.
arXiv:2607. 05861v1 Announce Type: cross Abstract: Large reasoning models (LRMs) improve language model capabilities by generating explicit thinking traces before final answers.