💬 NLP

Personalized Help for Optimizing Low-Skilled Users' Strategy

This paper demonstrates that augmenting the superhuman AI agent CICERO to provide personalized move and message advice in the game Diplomacy significantly helps novice players compete with or even surpass experienced players, with the mere presence of such advice offering strategic advantages regardless of whether it is followed.

Original authors: Feng Gu, Wichayaporn Wongkamjan, Jonathan K. Kummerfeld, Denis Peskoff, Jonathan May, Jordan Boyd-Graber

Published 2026-02-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Feng Gu, Wichayaporn Wongkamjan, Jonathan K. Kummerfeld, Denis Peskoff, Jonathan May, Jordan Boyd-Graber

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are playing a complex board game called Diplomacy. It's like a high-stakes game of "Seven Deadly Sins" mixed with chess, but with a twist: you can't just move your pieces; you have to talk to the other six players to form alliances, make promises, and occasionally lie to your friends. The goal is to control more than half the map.

Now, imagine you are a total beginner at this game. You don't know the rules well, and the experienced players are about to crush you.

Enter PHOLUS (pronounced like "Pho-lus").

Who is PHOLUS?

Think of PHOLUS as a super-smart, invisible co-pilot sitting next to you. He isn't a robot that plays the game for you. Instead, he's like a sports coach or a wise old uncle whispering in your ear.

PHOLUS is built on top of an AI named CICERO, which is so good at this game it can beat almost any human. But while CICERO is the "player," PHOLUS is the "advisor."

What does PHOLUS actually do?

PHOLUS watches the game and gives you two types of advice, like a GPS for a board game:

The "Move" GPS: "Hey, instead of moving your army to the North, try moving it to the East. It's a safer spot."
The "Chat" GPS: "You should text Italy right now. Tell them, 'Hey, let's team up against Turkey,' but phrase it nicely."

The Big Experiment

The researchers wanted to see: Does having this AI coach actually help?

They set up 12 games with a mix of newbies (people who had never played) and veterans (people who had played for years). They gave some players the advice, some only move advice, some only chat advice, and some no advice at all.

Here is what they found, using some fun analogies:

The "Newbie Superpower": Usually, newbies get crushed immediately. But with PHOLUS, the newbies didn't just survive; they started winning! In fact, with the AI's help, a beginner could beat a veteran. It's like giving a kid a set of training wheels that are so advanced they can race against a professional cyclist.
The "Veteran's Skepticism": The experienced players were less likely to listen. They thought, "I know what I'm doing." They only followed the AI's advice about 6% of the time. The newbies, however, followed the advice about 33% of the time.
The "Magic of Just Listening": Here is the coolest part. Even when players didn't follow the advice, just reading it helped them. It was like reading a map before you drive; even if you take a different route, knowing the terrain makes you a better driver. The advice sparked new ideas in their heads.

Why didn't everyone just do exactly what the AI said?

Because Diplomacy is about human relationships, not just math.

Sometimes the AI would say, "Tell Turkey you are scared of Russia." But the human player thought, "No, I want to trick Turkey into thinking I'm strong." The player would take the AI's idea (talk to Turkey) but change the words to fit their own personality.

Think of it like a ghostwriter. The AI writes a draft of a letter, and the human player edits it to sound more like them. Sometimes they keep the whole letter; sometimes they just keep one sentence.

The Takeaway

This paper proves that Human + AI is a powerful team, even if the human doesn't blindly obey the robot.

For beginners: The AI acts as a safety net, stopping them from making huge mistakes and teaching them how to think strategically.
For experts: The AI acts as a "second opinion," offering a fresh perspective they might have missed.

In a nutshell: PHOLUS isn't there to play the game for you. It's there to be the best study buddy you've ever had, helping you learn the game faster and play it smarter, whether you are a total rookie or a seasoned pro.

1. Problem Statement

While AI agents like CICERO have achieved superhuman performance in the board game Diplomacy (a game requiring complex negotiation, strategy, and simultaneous moves), the utility of these agents as collaborative advisors for humans remains understudied.

The Gap: Current research focuses on AI playing against humans or other AIs. There is a lack of understanding regarding how AI can assist humans, particularly novices, in learning complex strategies and communication without simply replacing human agency.
The Challenge: Diplomacy involves simultaneous secret moves and open communication. Novices often struggle with both strategic reasoning (move optimization) and social negotiation (message crafting). The challenge is to design an AI advisor that provides personalized, real-time guidance that enhances human decision-making without causing over-reliance or blind obedience.

2. Methodology

2.1 System Architecture: PHOLUS

The authors introduce PHOLUS (Personalized Help for Optimizing Low-Skilled Users' Strategy), a natural language agent that acts as a passive advisor rather than an active player.

Foundation: PHOLUS leverages CICERO, a state-of-the-art AI capable of playing Diplomacy with near-optimal strategy and natural language communication.
Function: Instead of submitting moves or sending messages itself, PHOLUS observes the game state and the player's history to generate advice in two forms:
1. Move Advice: Recommended unit orders (e.g., "Move Army to Galicia").
2. Message Advice: Suggested communication content and targets (e.g., "Propose a DMZ in Tyrolia to Germany").
Personalization: The advice is conditioned on the specific player's past moves and messages, ensuring the suggestions align with the current game context and the player's trajectory.

2.2 Experimental Setup

Participants: 41 players recruited from the Diplomacy community (experienced "veterans") and university board game clubs (novices with no prior experience).
Game Structure: 12 games were played, each lasting ~3 hours with 2–5 human players per game.
Conditions: In each turn, players were randomly assigned to one of four settings:
1. No Advice: Control group.
2. Message Advice Only.
3. Move Advice Only.
4. Combined Advice: Both move and message suggestions.
Data Collection: The study generated a dataset of 1,070 player turns, 117 playing hours, over 3,600 move advice instances, and 4,300 message advice instances.

2.3 Evaluation Metrics

Quantitative:
- Point Gain/Loss: A linear regression model with regularization was trained to predict supply center gains based on advice settings, player type, and game state.
- Acceptance Rate: The percentage of advice accepted by players.
- Agreement & Equivalence: Measuring the overlap between PHOLUS's suggested moves and the player's actual moves ( $A = |x \cap y| / |x|$ ) and exact matches ( $E$ ).
Qualitative:
- Semantic Similarity: Used Abstract Meaning Representation (AMR) and SMATCH scores to compare the semantic structure of PHOLUS's suggested messages against actual player messages.
- Surveys: Post-game surveys assessed player perception of advice helpfulness and trust.

3. Key Contributions

Novel Advisor Paradigm: Shifts the focus from AI as a player to AI as a companion/advisor, demonstrating that AI can enhance human learning in complex, multi-agent environments.
Bridging the Skill Gap: Demonstrated that AI advice can allow novices to compete with, and occasionally surpass, experienced players.
Nuanced Understanding of Reliance: Revealed that "acceptance" is not binary. Players often synthesize advice rather than blindly following it. Even when advice is rejected, the mere presence of the suggestion can positively influence a player's strategic thinking.
Dataset Release: Provided a rich dataset of human-AI interaction in Diplomacy, including move/message pairs and acceptance behaviors.

4. Key Results

4.1 Performance Impact

Novice Improvement: Novices receiving both move and message advice performed significantly better than those receiving no advice. In games where novices received combined advice, they were rarely eliminated early and often ended with more supply centers than they started with.
Veteran Benefit: Experienced players also benefited, though they were less likely to accept the advice.
Regression Analysis:
- Receiving no advice had a negative coefficient ( $\approx -0.05$ ), indicating a disadvantage.
- Move advice showed a positive correlation with point gains.
- Combined advice (Move + Message) yielded the highest positive impact.
- Interestingly, Message advice alone had a slight negative effect on outcomes, suggesting that without strategic move guidance, communication advice alone may be insufficient or distracting.

4.2 User Behavior and Reliance

Acceptance Rates:
- Novices accepted 32.6% of move advice and 6.3% of message advice.
- Veterans accepted only 6.4% of move advice and 3.4% of message advice.
Blind Obedience vs. Synthesis:
- While novices accepted move advice frequently, they did not follow it blindly. Agreement dropped from 80% at the start of a turn to 70% by the end, and equivalence dropped from 46% to 38%.
- Players often used the AI's strategic insights to inform their own communication strategies, even if they didn't execute the exact move suggested.
Semantic Alignment: High SMATCH scores indicated that many player messages were minor variations of the AI's suggestions, showing that players internalized the AI's intent even when rephrasing.

5. Significance and Future Directions

Human-AI Collaboration: The study validates the "Centaur" model (human + AI) in complex strategic games, showing that AI can act as a "scaffold" for learning rather than a replacement.
Educational Potential: The findings suggest that personalized AI agents can accelerate learning in unfamiliar, high-stakes environments by providing tailored guidance that adapts to the user's skill level.
Limitations & Future Work:
- Intent Misalignment: PHOLUS sometimes optimized for its own (CICERO's) win conditions rather than the human's specific goals (e.g., suggesting aggressive moves when a player wanted to build an alliance).
- High-Level Intent: Current systems struggle to infer high-level meta-goals (e.g., "I want to defeat Austria") from low-level inputs. Future work aims to incorporate Theory of Mind to better align advice with user intentions.
- Accessibility: The system currently requires high-end GPUs (A100), limiting accessibility. Future research should focus on distilling these models for broader use.

Conclusion: PHOLUS demonstrates that AI agents can effectively augment human intelligence in complex social-strategic games. By providing personalized move and message advice, AI can significantly level the playing field for novices, fostering a collaborative environment where humans learn to make better decisions through guided interaction rather than blind obedience.