Why ChatGPT Gets Stuck Solving Riddles—and How a Simple Trick Gets It Unstuck
Have you ever played one of those mysterious "lateral thinking" puzzles where a bizarre situation is described in just a sentence, and you have to ask a host yes-or-no questions to solve it? Picture this:
A man walks into a bar and orders a glass of water. The bartender pulls out a gun and points it at the man. The man says “Thank you” and leaves. What happened?
To solve it, you need creativity, patience, and ... a bit of zigzagging logic. Humans are pretty good at this kind of thinking. Turns out, large language models like ChatGPT? Not so much.
But what if we could teach ChatGPT how to zigzag better?
That’s exactly what a team of researchers from various universities set out to do. In their study, they found a clever, low-cost way to dramatically boost ChatGPT’s ability to solve these sorts of puzzles — by having it “take a break,” rethink the problem, and come back at it with a fresh perspective.
And the best part? It works.
In this blog post, we’re going to walk you through what they discovered, how they pulled it off, and what it means for the rest of us — especially if you’re using language models to tackle tricky problems or create better prompts.
Wait, What Are “Lateral Thinking” or "Situation Puzzles"?
Let’s start with the basics.
Situation puzzles, a.k.a. lateral thinking puzzles, are brain teasers that require creative reasoning to solve. You're usually given a mysterious or sparse scenario and need to ask only yes/no/irrelevant questions to uncover the full story.
For example:
A man dies in a room with 53 bicycles. What happened?
(Hint: "bicycles" doesn’t mean what you think.)
These puzzles aren't just about decoding facts — they test your ability to make leaps in logic, ask relevant questions, and connect dots in ways you weren’t expecting.
Humans do this naturally — we jump between ideas, make wild guesses, learn from mistakes, and redirect when we’re off track. Language models? Not so much. When they get stuck asking the same type of questions, they hit a wall.
Why ChatGPT Gets Lost in Situation Puzzles
Large Language Models (LLMs) like ChatGPT are great at a lot of tasks — writing stories, summarizing text, even solving math to some degree. But when it comes to situation puzzles, they seem to run in circles.
The research team noticed that ChatGPT often:
- Asks very similar or repetitive questions round after round.
- Focuses too much on one detail and ignores bigger contextual clues.
- Misses the forest for the trees, essentially.
This is a big problem when you’re trying to solve a puzzle that demands out-of-the-box thinking. Instead of getting closer to a solution, the model spirals into question-fatigue — without making meaningful progress.
So how do we help the machines think less like machines?
The Simple Fix: Break, Rethink, Reframe
Here’s the brilliant — and surprisingly simple — idea the researchers tried: Give the AI a break.
More specifically, after a few rounds of back-and-forth guesses, restart the conversation, but — and this is the key — use the information already uncovered to rewrite the puzzle with added hints.
Think of it like this: You’re trying to solve a mystery. After a while, your notes are a mess. You get stuck. So you pause, organize your notes, and summarize what you know. Then you start fresh, but this time with the benefit of a clearer, more concise setup.
That’s what the researchers did. Here's how it worked step-by-step:
1. Let ChatGPT Play the Standard Game
The model plays the situation puzzle game. It asks yes/no questions and occasionally makes a guess.
2. Detect When It’s Stuck
After a few questions (say, 5 rounds) or immediately after a wrong guess, the system says: “Alright, time to stop here.”
3. Reframe the Puzzle
Take the original puzzle + pick the most useful questions and their answers (especially "Yes" ones) and turn them into additional "hints."
For example, if a previous question was:
Q: Was the man lost in a desert?
A: Yes
Then we add a sentence like:
“Hint: The man was lost in the desert.”
Now the reformulated puzzle is clearer and easier to tackle.
4. Restart a New Chat Session
Start a new conversation with the language model, using the rephrased version of the puzzle. The AI can now skip over previously covered ground and aim its questioning in a more productive direction.
The magic here is that ChatGPT avoids getting stuck in an unproductive thought loop — because it gets a reboot with better context.
The Results: It Actually Works (Really Well)
So, does this "reformulation" actually help?
In short: Yes — big time.
When tested on several classic situation puzzles, this method led to:
- Higher success rates (i.e., more puzzles correctly solved)
- Fewer questions needed to reach the correct guess
- Better question quality, with more “Yes” answers and fewer irrelevant ones
Even better, the model stopped repeating itself and asked questions that directly built on what it already knew.
Real-World Case Study
Let’s walk through one example from the research.
Imagine the original puzzle is:
A man is found dead in a phone booth with a broken umbrella.
ChatGPT starts by asking a series of questions — some helpful, some not.
- Q1: Was it raining? → No
- Q2: Was he trying to call someone? → Yes
- Q3: Was he attacked? → No
- Q4: Was there glass on the ground? → Irrelevant
- Q5: Was he ill? → No
After this, the model seems to stumble. It doesn’t know where else to go.
So — reformulation time!
The setup is now rewritten like this:
A man is found dead in a phone booth with a broken umbrella.
Hints:
1. He was trying to call someone.
2. He was not attacked.
3. It was not raining.
Now the AI restarts — and suddenly it asks smarter questions and quickly solves the puzzle.
By boiling down what’s been learned and feeding it back in a cleaner format, LLMs like ChatGPT can dramatically improve their reasoning.
Why This Matters: Beyond Riddles
Sure, it's fun to watch machines solve puzzles — but there’s a bigger picture here.
This method of external reformulation has huge implications for how we interact with AI in real-world reasoning tasks. Think about:
- Customer service bots trying to resolve complicated issues
- Virtual tutors helping students work through problems
- Proposed AI legal assistants handling nuanced cases
In all these scenarios, the AI can get “stuck in a loop” — repeating questions, misreading context, or missing the point. But if we introduce strategic checkpoints, summarize what’s been learned, and reframe the problem, we can improve outcomes significantly.
This approach is almost like helping AI build a memory that it can actually reflect on — without flooding its context window or expecting perfect recall.
🧠 Put another way: Instead of just giving AI more information, we help it think better with the information it already has.
Got a Prompting Hack from This? You Bet.
If you’re using ChatGPT or similar tools to solve complex problems — writing code, planning projects, doing research — you might experience similar chatbot burnout.
Here’s how you can apply this idea yourself:
- When the model starts looping or going nowhere, pause.
- Summarize the key insights from your interaction so far.
- Feed that summary + the original problem back into a new chat.
- Restart — and watch the magic unfold.
This little trick is simple but powerful. Because, as the study shows, a fresh start with the right focus can change everything.
Key Takeaways
✅ ChatGPT struggles with multi-step reasoning tasks like situation puzzles — often asking repetitive or unhelpful questions after a few rounds.
✅ Reformulating the problem mid-way by summarizing the most useful prior Q&As into hints leads to a massive improvement in puzzle-solving success.
✅ Starting a new chat session with this reformulated puzzle prevents the AI from getting stuck in unproductive loops and helps it ask better questions.
✅ In practice, this method makes ChatGPT more efficient, increasing win rates, reducing the number of questions and failed guesses, and improving question quality overall.
✅ This strategy has real-world applications in any domain where complex reasoning is needed — from tutoring to debugging to decision support.
✅ Bonus Prompting Tip: If ChatGPT gets stuck or gives generic responses, stop the convo, reframe what’s been learned, and restart the session with those learnings included. It works.
Solving riddles might seem like a quirky AI challenge, but it speaks to a deeper goal: teaching machines to think creatively, not just logically. And with small tweaks like reformulation, we’re getting a lot closer.
So next time you hit a dead end with your favorite AI assistant, don’t throw your arms up — just say, "Let’s take a step back," and try feeding it what it already knows in a smarter way.
Who knew a fresh start could be so intelligent?
