Hybrid Reasoning Agents

An AI agent that thinks silently in vectors but still writes text to use tools.

10.4%

Average performance improvement

Gains while using 35.5%–54.7% fewer tokens than text agents

How it works

Two worlds apart
Text agents can use tools but spell out every thought. Latent reasoners think faster in vectors but cannot call tools.
Switch modes mid-generation
A small adapter, HybridLink, feeds hidden states back as inputs. Special tokens open and close silent latent spans.
Better, cheaper agents
Qwen3 models at 4B to 14B beat text and latent baselines on 11 benchmarks with faster inference.

Which comes out on top on GAIA average accuracy?

try it

Drag the share of latent reasoning up and watch accuracy change.

Latent share30 %

Mixing works best: all-text is slow and verbose, all-latent loses tools and inspection.Illustrative model, not the paper's data

6 terms

Latent reasoning Thinking steps passed as hidden vectors, never turned into words.

Like a chess player who calculates moves silently in their head, then speaks aloud only to announce moves or ask the arbiter.

Where it breaks

Here the silent thoughts can still be read out word by word.

Needs custom serving infrastructure; current engines lack multi-turn hybrid generation.
Latent pseudo-text is a nearest-token readout, possibly not the full meaning.
Tested only on Qwen3-Base models up to 14B with two tools.

From the paper