PRISM: Transfer and Limits of Activation Interventions in Tool-Calling Language Models
Giovana Kerche Bonás, Thales Sales Almeida, Marcelo da Silva Reis, Rodrigo Nogueira, Marcos Meideiros Raimundo
Abstract
Recent literature shows that unsafe tool calls can often accompany textual refusals in language models. To investigate this hidden vulnerability, we introduce PRISM, a framework for evaluating the transfer of activation interventions across precision, languages, and scales. Our findings demonstrate that leveraging lower-resource settings (specifically quantized NF4 labels and cross-lingual transfer) offers a robust pathway to expose and improve model alignment. Notably, using behavioral labels from an NF4 copy enables stable intervention estimation even with very few scenarios. Furthermore, th…
Who it is from
Maritaca AIautor: Rodrigo Nogueira (fundador e CEO)
Other papers from the same organizations
- Lumen-bench: Cross-Lingual Safety Gaps and the Say-Do Problem in Tool-Calling Large Language ModelsMaritaca AI
- Eleven Years of BRACIS: A Meta-Scientific Study of the Brazilian Conference on Intelligent SystemsMaritaca AI
- BLUEX v2: Benchmarking LLMs on Open-Ended Questions from Brazilian University Entrance ExamsTropic-AI
- moBERTo: A Modern Encoder for Portuguese via Continued Pretraining of ModernBERTTropic-AI
- LLM-Based Persuasion Enables Guardrail Override in Frontier LLMsMaritaca AI
- Magis-Bench: Evaluating LLMs on Magistrate-Level Legal TasksMaritaca AI