AI That Thinks Without Words Is 2x Faster, But There’s a Catch

HybridAgent, from researchers at Stanford and Together AI, lets AI agents reason in their internal numerical representations instead of writing every step out in English. Text is only used for tool calls and final answers. It cut text tokens by 35 to 55%, ran about 2x faster and lowered estimated decoding costs by up to 3.7x, with a 10.4% average gain across 11 benchmarks. It runs on a 10M-parameter adapter, and the code is open source. Results so far are on Qwen3 models from 4B to 14B, compared against text-based reasoning agents. The catch: if reasoning happens inside the model, developers can no longer read it to debug decisions. #AIAgents #LatentReasoning #LLM #AIResearch #AIEngineering