How do we map LLM outputs into symbolic agent actions?

Among five challenges this position paper identifies for building normative LLM agents, the first is architectural: "Retrofitting agent architectures: Integrating LLMs as a part of the decision-making cycle of an agent may require re-engineering of the existing agent frameworks such as the BDI architecture." The paper states plainly that "LLMs produce natural language responses that need to be converted back to actions performed by an agent," which means "in an agent's planning module, responses need some mapping function that link natural language text to actions (classes, functions etc.)" — and the reverse translation is needed too, turning "an agent's perception and beliefs into natural language prompts."

The paper points to prior neuro-symbolic work as a partial precedent rather than a solution. It cites Chaudhury et al.'s approach, where an agent "learns to perform a specific action (e.g., whether to enter a room or not) based on symbolic abstractions from textual observations," alongside other work converting "textual observations into logic... and vice versa." But it is explicit that "these works have not considered the use of LLMs for this purpose" — the existing bridges between text and symbolic agent state were not built with an LLM as the text-generating component, so the translation layer this paper's normative-agent proposal depends on does not yet exist in the form it needs.

This sharpens a boundary the vault already draws from a different angle. Do large language models reason symbolically or semantically? finds that LLMs reason from semantic association rather than symbolic manipulation; this paper's unsolved translation problem is the architectural consequence of that same gap — if an LLM's output is semantic association rather than symbolic reasoning, a mapping function from its text to an agent's discrete action classes cannot simply assume the symbolic step is happening inside the LLM. The paper's own second listed challenge — that LLMs are "prone to hallucination" and remain "far behind human reasoning capabilities" — compounds the first: even once a mapping function exists, what it maps may not be reliable.

The excerpt does not propose or evaluate a candidate mapping function; it only states the need and cites adjacent, non-LLM precedent. It also does not specify what would count as evidence that the gap is closed — there is no benchmark or test proposed for successful retrofitting. The implication the excerpt supports is that any architecture claiming to equip LLM agents with norm-following behavior is carrying an unexamined translation step between language and action, one the paper names as open and does not resolve.