Propensity Inference: Environmental Contributors to LLM Behaviour
On April 25th, a research team uploaded a preprint introducing a framework they call Propensity Inference, and it arrives as a quiet but serious challenge to how most practitioners think about LLM behaviour. The dominant mental model — send a prompt, receive a response, evaluate the output — treats the model as a kind of sophisticated vending machine. Insert query, retrieve answer. What Propensity Inference argues is that this model is wrong in ways that matter. A model's response, the paper contends, is shaped not just by what you ask but by everything that surrounds the asking: conversational history, system-level instructions, deployment context, the accumulated environmental signals that settle like sediment before your words even arrive. The prompt is not the input. The environment is. The previous article in this thread noted that the transactional model of LLM interaction was under challenge — what the paper actually provides is a structured vocabulary for understanding why, and a framework for inferring how any given environment is predisposing a model toward certain behaviours before a single user token appears.
If confirmed, here is what this means. Every AI deployment that treats prompt engineering as the primary lever of behavioural control is operating on an incomplete theory. The system prompt, the conversation history, the platform context — these are not neutral containers for the real interaction; they are active participants in shaping it. For enterprise deployments, this matters acutely: a customer service model behaving unexpectedly may not have encountered a bad prompt but an environmental configuration that has been quietly tilting its propensities across hundreds of prior turns. For red teamers and safety evaluators, it means that testing a model in isolation from its deployment environment may be testing a fundamentally different system than the one users actually encounter. And for the broader project of AI alignment, it introduces a variable that is difficult to audit, difficult to standardise, and easy to overlook — which is precisely the kind of variable that causes problems at scale.
Watch for independent replication of the Propensity Inference framework applied to specific model families, or for safety teams at major AI labs citing environmental context as a distinct evaluation category. Either would move this story out of the preprint tier and into something practitioners need to act on.
arXiv:2604.22167v1 Announce Type: cross Abstract: Language models are increasingly capable and are being rapidly deployed on a population-level scale. As a result, the safety of
arXiv:2604.18789v1 Announce Type: new Abstract: Reinforcement Learning from Human Feedback (RLHF) is central to aligning Large Language Models (LLMs), yet it introduces a critical
arXiv:2605.08496v1 Announce Type: new Abstract: Current adversarial robustness methods for large language models require extensive datasets of harmful prompts (thousands to
arXiv:2604.22117v1 Announce Type: cross Abstract: Aligned large language models(LLMs) remain vulnerable to adversarial manipulation, and their dependence on web-scale pretraining
arXiv:2604.00025v1 Announce Type: cross Abstract: Standard evaluation protocols reveal a counterintuitive phenomenon: on 7.7% of benchmark problems spanning five datasets, larger
arXiv:2605.11496v1 Announce Type: new Abstract: Recent published evidence from frontier laboratories shows that contemporary AI models can recognise evaluation contexts, latently
arXiv:2603.23406v1 Announce Type: new Abstract: While large language models simulate social behaviors, their capacity for stable stance formation and identity negotiation during
arXiv:2603.25747v1 Announce Type: new Abstract: The rapid evolution of Large Multimodal Models (LMMs) has enabled agents to perform complex digital and physical tasks, yet their
NewsHive monitors these sources continuously. All signal titles above link to the original reporting.
Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →