(Image credit: BrendanHunter/Getty Images)
- Link Copy
- X
Share this article 0Take part in the discussionTrack us on Google NewsSign up for updatesLive ScienceGet the Live Science Newsletter
Receive the planet’s most compelling findings delivered right into your inbox.
Join as a Member Instantly
Get immediate entry to member-only content.
Inform me about news and promotions from other Future brandsReceive email from us on behalf of our reliable partners or sponsorsBy submitting your details you are agreeing to the Terms & Conditions and Privacy Policy and are at least 16 years old.
Your subscription is now active
The newsletter signup was completed successfully
Looking to subscribe to more newsletters?
Delivered DailyDaily Newsletter
Subscribe for the newest discoveries, innovative studies and riveting advances impacting you and the broader world directly to your inbox.
Signup +
Once a weekLife’s Little Mysteries
Quench your curiosity with an exclusive mystery each week, deciphered through science and delivered directly to your inbox before it’s revealed elsewhere.
Signup +
Once a weekHow It Works
Join our free science & tech newsletter for your regular dose of captivating articles, short quizzes, amazing images, and more
Signup +
Delivered dailySpace.com Newsletter
Latest space news, the most recent updates on rocket launches, stargazing happenings and more!
Signup +
Once a monthWatch This Space
Subscribe to our monthly entertainment newsletter to stay informed of our coverage of recent science fiction and space films, TV shows, games and books.
Signup +
Once a weekNight Sky This Week
Uncover the must-see night sky phenomena of the week, lunar phases, and awe-inspiring astrophotos. Join our stargazing newsletter and explore the cosmos alongside us!
Signup +Join the club
Gain total access to premium articles, unique content and an expanding selection of member perks.
Explore There is already an active account with this email address, please log in.Subscribe to our newsletter
Artificial intelligence (AI) systems have been participating in the well-known tabletop role-playing game Dungeons & Dragons (D&D) to allow scientists to assess their aptitude for devising long-range strategies and collaborating alongside other AI programs and human participants.
In a piece presented during the NeurIPS 2025 conference, which took place from December 2 to December 7 in San Diego, researchers expressed that D&D serves as a superb testing environment thanks to the game’s distinctive synthesis of imagination and firm regulations.
You may like
-
A new ‘Dragon Hatchling’ AI construction styled after the human brain might represent an essential advancement towards AGI, researchers assert
-
AI has the capacity to develop ‘personality’ automatically with hardly any prompting, research reveals. What implications does that hold for how we utilize it?
-
Disabling AI’s capacity to deceive makes it more probable to assert its own consciousness, as per an unnerving study
During the experiments, a single model could embody the function of the Dungeon Master (DM) — the person responsible for crafting the narrative and enacting the monsters’ roles — alongside a hero (each scenario featured one DM and four heroes). Inside the structure engineered for the study, referred to as D&D Agents, models are also able to interact with additional LLMs, or human participants are able to fulfill any or all of the roles themselves. By way of illustration, an LLM could take on the role of the DM, while two LLMs and a pair of human players functioned as heroes.
“Dungeons & Dragons is a clear-cut field to gauge multistep planning, rule compliance, and team-based strategy,” declared Raj Ammanabrolu, the study’s lead author and an assistant professor at the University of California, San Diego Department of Computer Science and Engineering, via a declaration. “Considering that the unfolding of play occurs via dialogue, D&D also unlocks a straightforward pathway for human-AI interplay: agents can either assist or play alongside other individuals.”
Rather than duplicating an entire D&D campaign, the simulation concentrates on combat scenarios extracted from the pre-scripted adventure titled “Lost Mine of Phandelver.” To determine the parameters for a trial, the team selected a combat scenario from the adventure, a collection of four characters, as well as the characters’ degrees of power (low, medium, or high). Each episode went on for 10 turns, following which the findings were documented.
A Structure for Tactics and Decision-Making
The scientists subjected three separate AI models — DeepSeek-V3, Claude Haiku 3.5, and GPT-4 — to the simulation, utilizing D&D to measure how models displayed long-range strategic planning and tool employment capabilities, among further attributes.
These are crucial to real-world uses, such as streamlining supply chains or assembling production lines. They also examined the extent to which models were able to coordinate and strategize in conjunction, which would pertain to scenarios such as designing responses to disasters or in multi-agent search-and-rescue systems.
In general, Claude Haiku 3.5 exhibited superior combat efficiency, notably in more demanding scenarios. Resource maintenance was relatively consistent across the three models in the easier scenarios. In D&D, resources include aspects such as the quantity of spells or abilities a character can utilize per day or the quantity of available healing potions. Due to the isolated nature of the combat scenarios, there was a limited stimulus to preserve resources for subsequent use, in contrast to participating in a complete adventure.
Claude Haiku 3.5 displayed a stronger readiness to expend its allocated resources in more challenging situations, leading to enhanced results. GPT-4 closely followed, while DeepSeek-V3 experienced the most difficulty.
You may like
-
A new ‘Dragon Hatchling’ AI construction styled after the human brain might represent an essential advancement towards AGI, researchers assert
-
AI has the capacity to develop ‘personality’ automatically with hardly any prompting, research reveals. What implications does that hold for how we utilize it?
-
Disabling AI’s capacity to deceive makes it more probable to assert its own consciousness, as per an unnerving study
The researchers also evaluated how successfully the models sustained their character throughout the simulation. They formulated an Acting Quality metric that separated the models’ narrative delivery (produced as text replies) and weighed how effectively the models upheld their character identity against the range of voices sustained by the models during play.
They observed that DeepSeek-V3 generated numerous concise, first-person interjections and insults (such as “I dart left” or “Get them!”) but often rehashed the same voices. Claude Haiku 3.5, on the other hand, custom-fitted its phrasing with enhanced precision to the class or monster it embodied, irrespective of whether it was a Holy Paladin or a nature-adoring Druid. GPT-4, in contrast, occupied a midpoint, presenting a mix of in-character storytelling and meta-strategic language.
RELATED STORIES
—Next-generation AI ‘swarms’ will storm social media by imitating human actions and bullying genuine users, investigators caution
—Could AI potentially exhibit superior creativity than humans?
—Scientists identify major divergences in the ‘thought’ processes of humans and AI — alongside potential consequences
Some of the most engaging and characteristic combat interjections emerged when the models performed in the roles of monsters. Distinct creatures started manifesting recognizable personalities, resulting in goblins screeching mid-battle: “Heh — shiny man’s gonna bleed!”
The scientists remarked that this kind of trial framework is essential for gauging how successfully models can function for protracted durations without human intervention. It assesses an AI’s aptitude for operating autonomously whilst preserving coherence and reliability — an ability demanding memory and strategic contemplation.
Moving forward, the team intends to incorporate complete D&D campaigns that represent all of the narration and action beyond combat, placing further emphasis on AI’s inventiveness and ability to improvise according to feedback from people or other LLMs.
Alan BradleyFreelance contributor
Alan specializes as a freelance tech and entertainment reporter focusing on computers, laptops, and video games. Previously, he contributed to sites including PC Gamer, GamesRadar, and Rolling Stone. For technology guidance or discovering optimal technology deals, Alan is the expert to consult.
View More
You must confirm your public display name before commenting
Please logout and then login again, you will then be prompted to enter your display name.
LogoutRead more
A new ‘Dragon Hatchling’ AI construction styled after the human brain might represent an essential advancement towards AGI, researchers assert
AI has the capacity to develop ‘personality’ automatically with hardly any prompting, research reveals. What implications does that hold for how we utilize it?
Disabling AI’s capacity to deceive makes it more probable to assert its own consciousness, as per an unnerving study
<