A Simple Voting Tweak Makes AI Thinkers 10% More Accurate
Smarter AI answers with no retraining needed — here's how it works.
When an AI chatbot tackles a tricky question — a math problem, a logic puzzle — it doesn't just blurt out one answer. It quietly explores several possible chains of reasoning at the same time, like a student scribbling multiple attempts on scratch paper. Most current systems then delete the weaker attempts and keep only the strongest one. The problem: sometimes the answer that looked weakest early on turns out to be right, and it gets thrown away before anyone notices.
A team of researchers is proposing a fix. Instead of wiping out the low-ranked attempts, their method keeps far more of them alive and simply stops any single attempt from dominating. Think of a meeting where a few loud voices normally drown everyone out: rather than silencing the quiet people, you keep the whole room talking. Then, at the end, the system groups together all the answers that mean the same thing and goes with whichever answer the largest number of separate reasoning paths arrived at — a majority vote, in effect.
The results were solid. Tested across three openly available AI models and five sets of reasoning tasks, their approach found a correct answer somewhere among its attempts more often in 13 of 15 cases. When it came to the final answer, it matched or beat the previous best method in 14 of 15 cases, with improvements of as much as 10.6 percentage points. Notably, none of this required retraining the model — the gain comes purely from how the AI thinks at the moment it answers, which makes it cheap and quick to adopt.
The catch: running many reasoning paths at once uses more computing power, so each answer costs more time and money behind the scenes. This is also a research paper presented at an academic workshop, not a product you can use today. Still, it points to where AI progress is increasingly coming from — smarter thinking on the fly, rather than simply bigger models.
- AI models explore several reasoning paths at once, and old methods deleted the 'weaker' ones — sometimes losing the correct answer in the process.
- The new method keeps more paths alive, then picks the answer that the most separate paths agree on, like a majority vote.
- It improved accuracy by up to 10.6 percentage points across three AI models and five test sets, with no retraining required — but it costs more computing power per answer.
Why It Matters
Better AI reasoning means fewer wrong answers for you — at the cost of more computing power per reply.