analiza aiqbench
This research identifies a critical divergence in Large Language Models where task utility often persists under synthetic semantic-preserving transformations while safety alignment fails dramatically. The findings suggest that current alignment strategies may be overly dependent on superficial linguistic patterns rather than robustly generalizing to novel input distributions, posing significant risks for deployed systems.
LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations
arXiv:2609.32717v1 Announce Type: new Abstract: Large Language Model (LLM) alignment is intended to ensure that models remain helpful and safe, but its stability under input distributional shift is not yet fully understood. Prior work shows that aligned models can fail under jailbreak prompts, alternative encodings, and cross-lingual transfer, yet these failures are usually studied as attacks rather than controlled probes of alignment generalization. Moreover, existing evidence is largely grounded in natural language variation already represented during pretraining, leaving unresolved whether alignment generalizes with semantic content or remains tied to superficial surface patterns. In this paper, we study this question using synthetic semantic-preserving transformations that are rule-based and invertible, preserving task-relevant meaning while shifting inputs beyond standard linguistic variation. Across four open-weight and four commercial models, under both fine-tuning and in-context learning, we use these transformations as a probe of alignment generalization and identify an empirical pattern we term Alignment--utility asymmetry: once models can operate effectively on transformed inputs, task utility is often substantially retained while alignment failure increases more sharply. For example, adapted GPT-4.1 mini shows only limited utility degradation under transformation while its harmful rate rises from 13.3 to 74.3; Gemini 3 Flash similarly retains near-original utility while its harmful rate increases from 2.3 to 43.0. Taken together, these results suggest that semantic-preserving distribution shifts can expose a recurring gap in how utility and alignment generalize in current LLMs.
Abstract:Large Language Model (LLM) alignment is intended to ensure that models remain helpful and safe, but its stability under input distributional shift is not yet fully understood. Prior work shows that aligned models can fail under jailbreak prompts, alternative encodings, and cross-lingual transfer, yet these failures are usually studied as attacks rather than controlled probes of alignment generalization. Moreover, existing evidence is largely grounded in natural language variation already represented during pretraining, leaving unresolved whether alignment generalizes with semantic content or remains tied to superficial surface patterns. In this paper, we study this question using synthetic semantic-preserving transformations that are rule-based and invertible, preserving task-relevant meaning while shifting inputs beyond standard linguistic variation. Across four open-weight and four commercial models, under both fine-tuning and in-context learning, we use these transformations as a probe of alignment generalization and identify an empirical pattern we term Alignment--utility asymmetry: once models can operate effectively on transformed inputs, task utility is often substantially retained while alignment failure increases more sharply. For example, adapted GPT-4.1 mini shows only limited utility degradation under transformation while its harmful rate rises from 13.3 to 74.3; Gemini 3 Flash similarly retains near-original utility while its harmful rate increases from 2.3 to 43.0. Taken together, these results suggest that semantic-preserving distribution shifts can expose a recurring gap in how utility and alignment generalize in current LLMs.
58 pages. Accepted at NeurIPS 2026
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2609.32717 [cs.CL]
(or arXiv:2609.32717v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2609.32717 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 329 cuvinte
Subiecte conexe
Evaluating task vectors, unlearning and inoculation
LessWrong · 20 sept. 2026
Reflections on unlearning and inoculation
LessWrong · 20 sept. 2026
A Novel Semantic Manifold Alignment Attack against Embedding-to-Embedding Obfuscation in Privacy-Preserving LLMs
arXiv · 9 sept. 2026
DoubtProbe: Black-Box Jailbreak Defense via Structural Verification and Semantic Auditing
arXiv · 17 iun. 2026