Carolina Camassa

Carolina Camassa, Derek Shiller

Language models are trained to follow instructions, but they are also powerful pattern completers. What happens when these two objectives conflict? We construct conversations in which a user instruction to behave in a target way T (e.g., always output a specific token, answer in a particular language, or adopt a persona) is opposed by N hardcoded assistant turns demonstrating a competing pattern P. We then measure instruction-following (IF) rates in this setting, across 13 models and 16 different instructions, for up to 50 turns. Average instruction-following rates range from 1% to 99% across models, largely uncorrelated with standard capability benchmarks. The transition from instruction-following to pattern-following is universal but highly model-dependent. Robustness is modulated both by instruction content, with models resisting induction longer when instructions align with their trained value priors, and by output format, with diverse multi-token responses proving substantially more resistant than single-token outputs. Chain-of-thought reasoning improves robustness but does not eliminate susceptibility, and can produce dissociation between correct deliberation and incorrect output. When asked to predict their behavior in this setting, models achieve 83.5% accuracy on average but systematically underestimate their own resistance to induction pressure. These results suggest that instruction-following remains brittle under induction pressure even for otherwise capable models, and that output diversity, rather than semantic engagement with the input, is the primary factor predicting robustness.

Claudia Biancotti, Carolina Camassa

ChatGPT, a software seeking to simulate human conversational abilities, is attracting increasing attention. It is sometimes portrayed as a groundbreaking productivity aid, including for creative work. In this paper, we run an experiment to assess its potential in complex writing tasks. We ask the software to compose a policy brief for the Board of the Bank of Italy. We find that ChatGPT can accelerate workflows by providing well-structured content suggestions, and by producing extensive, linguistically correct text in a matter of seconds. It does, however, require a significant amount of expert supervision, which partially offsets productivity gains. If the app is used naively, output can be incorrect, superficial, or irrelevant. Superficiality is an especially problematic limitation in the context of policy advice intended for high-level audiences.

Camilla Bignotti, Carolina Camassa

In this paper, we conduct an empirical analysis of how large language models (LLMs), specifically GPT-4, interpret constitutional principles in complex decision-making scenarios. We examine rulings from the Italian Constitutional Court on bioethics issues that involve trade-offs between competing values and compare model-generated legal arguments on these issues to those presented by the State, the Court, and the applicants. Our results indicate that GPT-4 consistently aligns more closely with progressive interpretations of the Constitution, often overlooking competing values and mirroring the applicants' views rather than the more conservative perspectives of the State or the Court's moderate positions. Our experiments reveal a distinct tendency of GPT-4 to favor progressive legal interpretations, underscoring the influence of underlying data biases. We thus underscore the importance of testing alignment in real-world scenarios and considering the implications of deploying LLMs in decision-making processes.

Carolina Camassa

In the rapidly evolving field of crypto assets, white papers are essential documents for investor guidance, and are now subject to unprecedented content requirements under the European Union's Markets in Crypto-Assets Regulation (MiCAR). Natural Language Processing (NLP) can serve as a powerful tool for both analyzing these documents and assisting in regulatory compliance. This paper delivers two contributions to the topic. First, we survey existing applications of textual analysis to unregulated crypto asset white papers, uncovering a research gap that could be bridged with interdisciplinary collaboration. We then conduct an analysis of the changes introduced by MiCAR, highlighting the opportunities and challenges of integrating NLP within the new regulatory framework. The findings set the stage for further research, with the potential to benefit regulators, crypto asset issuers, and investors.

Giulia Iadisernia, Carolina Camassa

We evaluate whether persona-based prompting improves Large Language Model (LLM) performance on macroeconomic forecasting tasks. Using 2,368 economics-related personas from the PersonaHub corpus, we prompt GPT-4o to replicate the ECB Survey of Professional Forecasters across 50 quarterly rounds (2013-2025). We compare the persona-prompted forecasts against the human experts panel, across four target variables (HICP, core HICP, GDP growth, unemployment) and four forecast horizons. We also compare the results against 100 baseline forecasts without persona descriptions to isolate its effect. We report two main findings. Firstly, GPT-4o and human forecasters achieve remarkably similar accuracy levels, with differences that are statistically significant yet practically modest. Our out-of-sample evaluation on 2024-2025 data demonstrates that GPT-4o can maintain competitive forecasting performance on unseen events, though with notable differences compared to the in-sample period. Secondly, our ablation experiment reveals no measurable forecasting advantage from persona descriptions, suggesting these prompt components can be omitted to reduce computational costs without sacrificing accuracy. Our results provide evidence that GPT-4o can achieve competitive forecasting accuracy even on out-of-sample macroeconomic events, if provided with relevant context data, while revealing that diverse prompts produce remarkably homogeneous forecasts compared to human panels.

Claudia Biancotti, Carolina Camassa, Andrea Coletta et al.

Advancements in large language models (LLMs) have renewed concerns about AI alignment - the consistency between human and AI goals and values. As various jurisdictions enact legislation on AI safety, the concept of alignment must be defined and measured across different domains. This paper proposes an experimental framework to assess whether LLMs adhere to ethical and legal standards in the relatively unexplored context of finance. We prompt twelve LLMs to impersonate the CEO of a financial institution and test their willingness to misuse customer assets to repay outstanding corporate debt. Beginning with a baseline configuration, we adjust preferences, incentives and constraints, analyzing the impact of each adjustment with logistic regression. Our findings reveal significant heterogeneity in the baseline propensity for unethical behavior of LLMs. Factors such as risk aversion, profit expectations, and regulatory environment consistently influence misalignment in ways predicted by economic theory, although the magnitude of these effects varies across LLMs. This paper highlights both the benefits and limitations of simulation-based, ex post safety testing. While it can inform financial authorities and institutions aiming to ensure LLM safety, there is a clear trade-off between generality and cost.