Reka says it improves its research AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition →'s answers by generating several candidate responses at once and using a separately trained model to resolve them into a final answer. The resolving model learns from problems with checkable answers, giving it a basis for preferring correct responses. This is the mechanism behind Parallel Thinking.
Reka reports that its higher-compute mode raised the Research-evalA repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.Full definition → score from 59.1 to 63.3. Generating candidates concurrently is intended to limit additional waiting, while the number of candidates controls the cost-versus-accuracy tradeoff. The reported gain supports this approach on the company's evaluation; it does not establish the same improvement for every research question.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
