PRISM: Transfer and Limits of Activation Interventions in Tool-Calling Language Models
Quick Answer
The PRISM framework reveals that using lower-resource settings and cross-lingual transfer can significantly enhance model alignment, reducing unsafe tool calls in Llama-3.1-8B by 21.95 percentage points.
Quick Take
This approach effectively transfers safety interventions across languages and larger models, demonstrating the potential of auxiliary supervision from compressed models.
Key Points
- PRISM evaluates activation interventions across precision, languages, and scales.
- Using NF4 labels enables stable intervention estimation with minimal scenarios.
- Five-language NF4 directions reduced unsafe violations in Llama-3.1-8B by 21.95%.
- Safety interventions successfully transfer to models with up to 70B parameters.
- Auxiliary supervision from multilingual data mitigates hazardous tool-calling behaviors.
DeepSignal Analysis
What happened
The PRISM framework investigates the transfer of activation interventions in language models, revealing that lower-resource settings and cross-lingual transfer can enhance model alignment. Specifically, it reports a 21.95 percentage point reduction in unsafe tool calls in the Llama-3.1-8B model. The findings suggest that auxiliary supervision from compressed models can improve safety across different languages and model sizes.
Key evidence
- The PRISM framework evaluates the transfer of activation interventions across precision, languages, and scales, focusing on model alignment.
- Using NF4 behavioral labels allows stable intervention estimation with minimal scenarios, enhancing the effectiveness of safety interventions.
- The study reports a 21.95 percentage point reduction in unsafe tool calls in Llama-3.1-8B through cross-lingual transfer without target-language tuning.
Why it matters
The findings from the PRISM framework highlight a significant advancement in improving the safety of language models, particularly in reducing unsafe tool calls. This is crucial as language models are increasingly integrated into applications where safety is paramount. By demonstrating that lower-resource settings can effectively enhance model alignment, the research opens avenues for more robust and safer AI systems, potentially influencing future model training and deployment strategies.
Paper Resources
📖 Reader Mode
~2 min readAbstract: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, these alignment improvements transfer effectively across languages: five-language NF4-supervised directions successfully transfer to held-out Arabic, reducing unsafe violations in Llama-3.1-8B by 21.95 percentage points without target-language tuning. By mapping paired benign activations, we also successfully transfer these safety interventions to much larger models (up to 70B parameters). Ultimately, PRISM highlights how auxiliary supervision from compressed models and multilingual data can uncover and mitigate hazardous tool-calling behaviors across different model families. We release all experimental code and supporting artifacts at this https URL, alongside an interactive explorer of recorded model responses and experimental results at this https URL.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.20808 [cs.CL] |
| (or arXiv:2610.20808v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.20808 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Giovana Bonás [view email]
[v1]
Wed, 14 Oct 2026 18:55:23 UTC (176 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →The "10th Juror": Open-Set Standpoint Screening for Bureaucratic Bias Detection
MARS-Gov introduces a framework for detecting bureaucratic bias in Dutch government documents, achieving a new state-of-the-art F1 score of 0.880. This model outperforms existing zero-shot detectors by 20.2 points and reduces unnecessary interventions to just 2.5%. The framework's dynamic '10th juror' adapts to emerging biases, enhancing legal language processing.