An 8B Agent That Thinks Without Writing Tokens Beats GPT-5.6 Luna on GAIA. Not on HLE

An 8-billion-parameter agent is outperforming GPT-5.6 Luna on the GAIA benchmark, according to a new Hybrid Reasoning Agents paper released on arXiv. The HybridAgent-8B manages a score of 43.7, compared to Luna's 37.9. However, on the Humanity's Last Exam, a more challenging knowledge and reasoning test, the larger GPT-5.6 Luna scores double the 14.4 points achieved by the smaller agent.

The researchers have made the weights, training data, and code publicly available under an MIT license. Both the 8B and 14B variants of the hybrid agent utilize a latent reasoning mode, which allows the model to perform part of its thinking without writing a single token. This saves on token costs and reduces the amount of text that needs to be logged and analyzed.

The latent reasoning is achieved through a small residual adapter called HybridLink, which processes the last-layer hidden state and feeds it back as the next input embedding. While the average improvement over the baseline GPT models is 10.4%, the margin becomes more significant when comparing the 8B and 14B variants to Luna. The researchers also provide a tool for monitoring the latent reasoning steps, although this may reduce the amount of readable traceability for safety and observability purposes.