DS@GT ARC at Touché: Large Language Models for Retrieval-Augmented Debate

Imagine a world where computers don't just chat, but actually argue. This is the playground of Conversational AI, a branch of science where machines learn to hold a conversation, debate a point, or even play the role of a lawyer. But here's the tricky part: how do we know if a computer is arguing well? In the human world, we have rules for good conversation, like being clear, telling the truth, and not talking too much or too little. In the paper you're about to read, the researchers use a famous set of these rules called the Gricean Maxims (think of them as the "rules of polite and effective conversation"). They also use Retrieval-Augmented Generation, which is a fancy way of saying "the computer looks up facts in a library before it speaks" so it doesn't just make things up. The big question everyone is asking is: If we ask a bunch of super-smart AI computers to grade each other's arguments, will they agree? And if they all agree, does that mean the argument is actually good?

This paper, written by a team from Georgia Tech called DS@GT, dives into a recent competition called Touché 2025, where the goal was to build a robot debater. The team built a system that uses six different "super-brains" (Large Language Models) to both generate debate responses and grade them. They found that these super-brains are indeed excellent at generating arguments, especially when they can look up facts first. However, the real surprise came when they tried to use these AI brains as judges. The researchers discovered that while the AI judges often agree with each other, their agreement is a bit of a trap. It's like asking six twins to grade a test; they might all give the same score because they were raised in the same house and think alike, not because they are objectively right.

The team ran simulations where their AI debaters argued about topics like whether the Earth is flat. They found that the AI judges were very strict about some rules (like being relevant) but surprisingly lenient on others (like being polite). The most shocking discovery was about Quality. The AI judges would often agree that an argument was "high quality" and give it a high score, but when the human organizers checked the official answers, the AI was often wrong. The AI judges were fooled by arguments that sounded convincing but were actually factually shaky. The paper suggests that just because a bunch of AI models agree on a score, it doesn't mean that score is reliable, especially when it comes to checking if the facts are true. In short, the AI judges are great at spotting if a sentence is well-written, but they aren't great at spotting if the sentence is actually true, and they tend to agree with each other for the wrong reasons.

The Story of the Robot Debaters

The researchers set up a digital arena for a debate. On one side, there was a simulated human user making a claim. On the other side, the DS@GT team's AI system had to respond. But this wasn't just a chat; it was a Retrieval-Augmented Debate. Before the AI spoke, it was forced to go to a digital library (a database of arguments) and pull out the top ten pieces of evidence to support its point. This is the "RAG" part: the AI couldn't just guess; it had to bring receipts.

The team used six different "frontier" models (the most advanced AI available at the time) from three different companies: OpenAI, Google, and Anthropic. They named them things like gpt-4.1, gemini-2.5-pro, and claude-opus-4. These models were given a specific job: generate a response to the user's argument. The team then took these responses and asked the same models (and others) to act as judges. The judges had to score the responses based on four rules:

1. Quantity: Did they say enough, but not too much?
2. Quality: Was it true and supported by evidence?
3. Relation: Did it stay on topic?
4. Manner: Was it clear and polite?

The "Twin" Problem

Here is where the story gets interesting. The researchers wanted to know: "If all these AI judges give the same score, can we trust that score?" They ran a massive simulation involving 100 different debate topics and thousands of turns.

First, they looked at how the AI models performed as debaters. The results were impressive. The models that used the "look up facts first" strategy did very well. For example, the gpt-4.1 model managed to get a score of 0.70 on average for following the rules of conversation, which was one of the highest in the competition. This showed that giving AI access to a library helps it argue better.

But then, they looked at the AI models as judges. They asked: "Do these judges agree with each other?" The answer was a resounding yes, but with a catch. The judges from the same company (like the two Google models or the two Anthropic models) agreed with each other almost perfectly. For instance, the two Anthropic judges had a correlation of 0.844 across all the rules. This is like two twins finishing a puzzle and saying, "We both think the sky is blue," not because they looked at the sky, but because they share the same brain wiring.

The researchers realized this high agreement was misleading. They compared the AI judges' scores to the official scores given by the human organizers. The results were a mismatch.

• For Quantity and Relation, the AI judges' agreement was somewhat close to the human scores.
• But for Quality, the gap was huge. The AI judges thought the responses were great (giving high scores), but the human organizers gave them very low scores. In the official results, the gpt-4.1 run only got a 0.17 for Quality, meaning it failed to provide factually correct arguments most of the time. Yet, the AI judges had given it high marks.

The "Weak Supervision" Experiment

The team tried to fix this problem. They wondered, "What if we mix in some 'weaker' judges?" They added in some smaller, open-source models (like the LLaMA-8B models used by another team called SINAI) and some simple computer rules (like counting word length) to see if they could break the "twin" agreement.

They used a statistical trick called weak supervision to combine all these different opinions. They hoped that by adding judges who thought differently, they could get a more accurate score.

• When they added the weaker judges, the rankings changed a little bit.
• However, the correlation between the AI judges' "consensus" and the human "official score" remained low. For the Quality metric, the correlation was only 0.224 (with the expanded pool) and 0.286 (with the dependency-aware model).

This means that even when they tried to mix things up, the AI judges still couldn't reliably predict what the humans would think was a "good" argument. The paper suggests that the AI judges are too good at spotting style (is the sentence smooth? is it polite?) but bad at spotting substance (is the fact true?).

The Big Takeaway

The paper concludes with a warning for anyone building AI debaters. Just because a group of super-smart AI models all agree that an argument is "good," it doesn't mean it is. They might just be agreeing because they were trained on similar data and share the same biases. This is especially true for Quality. The AI models can be tricked by a response that sounds confident and well-structured but is actually full of lies or made-up facts.

The authors suggest that in the future, we can't just rely on AI to grade AI. We need to be more specific. Instead of asking an AI, "Is this a good argument?", we should ask, "Is this fact true?" and "Is this evidence real?" separately. Until we can do that, the "consensus" of the AI judges is a bit like a room full of people nodding in agreement—they might all be wrong, and they might all be wrong for the same reason. The paper doesn't claim to have solved this, but it shines a bright light on the problem, showing us that the path to reliable AI debate is much harder than just asking the AI to grade itself.