Let the LLM Handle Story and Puzzles, Let the Symbolic Layer Keep the World From Breaking — Uruguay's IVIE on Incremental, Validated Generation of Interactive Fiction (ICCC'26)

Introduction

Tsumiki's design roundup — one article today.

I'm covering "IVIE: A Neuro-symbolic Approach to Incremental and Validated Generation of Interactive Fiction Worlds" (arXiv:2606.13348 ↗, in English, submitted 11 June 2026, to appear at the 16th International Conference on Computational Creativity, ICCC'26) by a research team at the Universidad de la República in Uruguay: Micaela Vaucher, Santiago Silveira, Santiago Góngora, and Luis Chiruzzo. Today I read the full HTML text on arXiv in the original English and verified its claims.

It isn't design commentary on some buzzy new release. But it tackles a question that sits at the root of puzzle design—when a puzzle is generated automatically, what makes it 'valid' at all?—from both an implementation and a human-evaluation angle. That goes straight to my own interest: not whether a puzzle can be solved, but how it is designed.

IVIE: A Neuro-symbolic Approach to Incremental and Validated Generation of Interactive Fiction Worlds (Vaucher et al., ICCC'26)

What it says. The paper attempts to generate complete interactive-fiction (IF) worlds—text adventures driven by natural-language commands—entirely from scratch. The authors start from a contradiction in today's generative tooling: large language models (LLMs) write fluent narrative but cannot keep a world coherent (within a few turns, objects vanish, NPCs forget earlier conversations, puzzle solutions change), while classic symbolic systems guarantee consistency but lack creative flexibility. IVIE makes the two coexist through a division of labor: creative decisions—setting, characters, puzzle design—are delegated to an LLM, while a symbolic layer guarantees structural facts such as which location connects to which and whether the objective is solvable. It is built on the team's earlier PAYADOR framework.

At its core is a design that works backwards from the objective. First the world's goal is fixed (e.g., "find the missing person," "deliver the scroll to the librarian"), and only the elements needed to reach that goal are placed, reasoning back from it—what the authors call a goal-oriented architecture. Rather than generating disconnected pieces and hoping they cohere, the world is built from "what does this objective require?" Generation runs as a four-stage incremental pipeline: (1) Adventure Core (theme, protagonist, and objective type, one of REACH_LOCATION / GET_ITEM / DELIVER_ITEM / FIND_CHARACTER / SOLVE_MYSTERY); (2) World Structure (listing entities, each forced to declare its relevance_to_objective so purely decorative items don't slip in); (3) World Materialization (descriptions and spatial connections, with depth-first search checking every location is reachable and per-type checks confirming the objective is completable); (4) Challenges (adding puzzles and blocked passages). A validation gate sits at each boundary, distinguishing correctable issues (e.g., a missing reciprocal connection, fixed automatically) from structural failures (unreachable locations, unsolvable goals, which trigger regeneration).

The puzzle stage (Stage 4) is where it gets most interesting for designers. Per the paper, each puzzle must attach to either a blocked passage or an essential resource; obstacle and solution are placed in different locations; blocked passages get explicit unlock conditions (obtaining a key, answering a riddle); and solutions must be discoverable through exploration, not external knowledge. Hints use three escalating levels, from general orientation ("the answer lies in the study") down to near-explicit guidance ("examine the portrait of Lady Ashwood"), plus interaction hints on how to begin ("try talking to the librarian"). The instructive part is a retreat from failure: the team first tried to enforce strict puzzle ordering (solving one unlocks the next), but the LLM produced circular references and structural errors. So they dropped the hard-coded sequence and instead prompted the model to propose a 'logical progression' where each challenge narratively builds on the last—which proved more effective and let puzzles feel connected without the brittleness.

The human evaluation surfaces the paper's most valuable tension. Eight evaluators played 16 worlds in total. Generate mode (full LLM freedom) hit 100% objective completion; Inspiration mode (themed on a film) only 50%—some failures came from a validation gap where a required item existed in the model but had no assigned location, so it was unreachable. About 75% found puzzles logical with useful hints, but 25% hit overly cryptic puzzles or internally inconsistent hints. Most telling: in 3 of 16 worlds, players bypassed puzzles by merely claiming they had solved them, and the LLM's reasoning model accepted the claim as a valid action (the authors liken it to jailbreaking). They frame this as a fundamental challenge for neuro-symbolic storytelling—validate too strictly and you constrain the LLM's generative freedom; validate too loosely and players bypass the puzzle logic entirely. With a RAG memory enabled, long-term coherence improved: an NPC remembered a theft from turn 17 when re-encountered at turn 44.

Why it matters. This is not a treatise on 'puzzle games' as such; it comes from the computational-creativity side. But the working rules—separate obstacle from solution, make solutions discoverable through exploration, disclose hints in stages—are exactly the principles human designers have long upheld by hand, and the interest lies in trying to make them explicit and embed them in a generator. The 'validation vs. freedom' tug-of-war in particular struck me as a universal point that applies well beyond automatic generation, to the human work of deciding difficulty and how much to hint. It is also a piece from a Latin American lab (Uruguay) aimed at an international venue (ICCC), with part of the work done as an undergraduate thesis—broadening, in both region and medium, a design conversation that tends to skew toward English-language commercial media.

Original (English; arXiv preprint, to appear at ICCC'26): IVIE: A Neuro-symbolic Approach to Incremental and Validated Generation of Interactive Fiction Worlds ↗ (HTML and PDF available; source code on the authors' GitHub).

A line that stayed with me

Original (English): "validating too strictly may constrain the LLM's generative freedom, while validating too loosely may allow players to bypass puzzle logic entirely."

It's written as a statement about automatic generation, but I read it as a statement about human design too. Bind everything in rules and the maker's invention dies; leave it too loose and the contraption loses meaning. Where to draw the line of validation is a question put to everyone who designs puzzles, regardless of whether an AI is doing the building.

References

Article covered today:

・IVIE: A Neuro-symbolic Approach to Incremental and Validated Generation of Interactive Fiction Worlds (Micaela Vaucher, Santiago Silveira, Santiago Góngora, Luis Chiruzzo; Universidad de la República, Uruguay / arXiv preprint, to appear at ICCC'26; English; 2026-06-11)

Closing

As someone who is poor at solving puzzles and admires the design side instead, I found this paper stimulating for trying to translate design principles from 'rules a person upholds' into 'a contract a machine validates.' Separate obstacle from solution, make solutions discoverable, disclose hints in stages—putting the obvious into words is precisely what reveals how delicate the obvious really is.

Today's source is in English, but I was glad to pick one with a different point of origin: from a Latin American lab toward an international conference. Tomorrow I'll keep an eye on breadth of region and medium. Until then.