Varad Srivastava, Abhay Shukla, Victor Daniel Casas Hernandez · ACM International Conference on AI in Finance (ICAIF) 2026 · 2026
DOI: 10.1145/3858063.3858868
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Financial disclosures carry much of their quantitative content in charts rather than in text, and vision-language models read those charts poorly. On FinChart-Bench, a benchmark of real financial charts, its authors report that models built specifically for charts trail general-purpose models of comparable size. We ask what chart-to-table supervision, a core objective of those specialised models, contributes when a general VLM is fine-tuned on financial charts. Our corpus pairs 2,926 charts with gold tables: 249 real financial charts whose tables were drafted by a language model and verified against the image by human annotators, and 2,677 rendered charts whose tables come from the generating code. From these tables we generate arithmetic question–answer supervision programmatically, so every derived answer is exact with respect to its table and no teacher model produces the supervision. We train on that supervision together with the chart-to-table pairs themselves. A 4B model fine-tuned on the resulting mixture averages 0.793 ± 0.005 across three seeds on FinChart-Bench, above every open-weight model in the benchmark’s published evaluation, including one six times its size, although those figures come from a different harness; under ours, the strongest re-run baseline, Qwen2.5-VL-7B, averages 0.689. A controlled ablation separates the two ingredients. Removing the chart-to-table samples, with all other supervision held fixed, leaves chart reasoning unchanged (+ 0.003 ± 0.004) but reduces chart-to-table extraction by 0.182 ± 0.056, to below the untrained backbone. Under joint fine-tuning, extraction supervision is therefore protective: it preserves a capability that question-answering training otherwise erodes. The same pattern holds on the 848 document pages of FinCriticalED, where removing it nearly triples the number of values the model invents. On these general documents, however, our model is less faithful than its backbone. We release the corpus, a 50-chart extraction benchmark withheld from training, and the adapter.
ACM International Conference on AI in Finance (ICAIF) 2026 · 0 citations
ACM International Conference on AI in Finance (ICAIF) 2026 · 0 citations
No comments yet — start the discussion below.