Jarosław A. Chudziak

Research profile

Research summary

Jarosław A. Chudziak is a researcher affiliated with Warsaw University of Technology whose OpenAlex record pairs a long, sparse early trail with an unusually intense recent phase. The snapshot attributes single works to the 1990s, 2000s and 2014, but output has accelerated sharply: 5 works in 2024, 29 in 2025 and 55 already in 2026, totalling 89 works in the trailing three years. Citation-level metrics remain modest (h-index 9, i10-index 6), consistent with how recent the surge is. The portfolio is deliberately broad: field distribution splits evenly, one third each across legal-reasoning NLP, quantitative finance (financial econometrics / volatility forecasting) and computational argumentation, while leading topics include Multi-Agent Systems and Negotiation (33), Stock Market Forecasting Methods (21), Topic Modeling (14), Logic, Reasoning, and Knowledge (13) and Financial Markets and Investment Strategies (12). In Evolve's catalog he appears as last author on three 2026 arXiv preprints — benchmarking LLM defences against character attacks, loss choice in cryptocurrency volatility forecasting, and catching contradictions in tax reasoning systems — none flagged corresponding. Assessed reproducibility averages 70 across three papers, with the tax-reasoning study rated high. Collaboration is concentrated: 5 distinct coauthors, including Albert Sadowski, Andrzej Tokajuk, Ewelina Gajewska and Katarzyna Budzynska.

Research keywords

Highlights

• LLMs compute through contradictions yet self-verify them

  In the SARA-based tax reasoning study, Evolve's review reports that top LLMs keep computing past contradictions while still being able to self-verify them — a credible, cheap gating signal from a small, well-controlled study, though scope is narrow and confounded. Reproducibility readiness: high.

• In crypto volatility forecasting, forecast level drives raw loss effects

  The reviewed study argues that raw loss-function effects in cryptocurrency volatility forecasting are mostly an artifact of forecast level; after alignment, model choice dominates, while the shared-input check remains inconclusive. Reproducibility readiness: medium.

• Benchmarking LLM defences against character attacks

  The review finds it plausible that LLMs underuse ethotic counterattacks relative to human debaters in political-debate dialogue games, but the claim rests on N=97 descriptive statistics with LLM co-raters and no significance tests. Reproducibility readiness: medium.

Timeline

1. 1990–1996The OpenAlex record attributes its earliest works to this period — 1 work in 1990, 1 in 1992, 2 in 1995 and 1 in 1996 — a sparse early layer predating the current AI agenda; the identity match for these years is unverified.
2. 2005–2014Isolated single works are logged in 2005, 2008, 2009 and 2014, with no multi-work years in between, keeping the pre-2024 record thin in the provided countsByYear data.
3. 2020A single work is recorded for 2020 in the provided bibliometric counts, the only activity logged between 2014 and 2024.
4. 2024–2025Output accelerates: 5 works are recorded in 2024 and 29 in 2025, contributing to the 89 works logged for the trailing three years.
5. 2026The most intense year in the record with 55 logged works; all three Evolve-cataloged arXiv preprints fall here, spanning legal-reasoning NLP (tax reasoning), cryptocurrency volatility forecasting and computational argumentation, each with Chudziak as last author.

Papers in the Evolve catalog

• 2026-09-23

  Benchmarking Argumentative Behaviour of LLMs: A Study of Defences Against Character Attacks

• 2026-09-22

  Loss Choice or Model Choice? The Role of Forecast Level in Cryptocurrency Volatility Forecasting

• 2026-09-05

  Solving versus Verifying: Catching Contradictions in Tax Reasoning Systems

Frequent co-authors