HybridAgent

Running HybridAgent requires a CUDA GPU, and the Qwen3 tokenizer has built-in anchor tokens.

HybridAgent's tool call execution rate remains at a reliable level of 99.7%.

HybridAgent uses two special tokens, `<latent>` and `</latent>`, to mark entering and exiting the latent space.

HybridLink is a small adapter that maps the model's last-layer hidden state directly back into its own input embedding space, thereby bypassing the round-trip overhead of decoding tokens and re-embedding.

In an ablation experiment, after replacing the model's latent thinking segments with segments from unrelated questions, accuracy dropped by 45%–51%, indicating that latent thinking carries substantive reasoning.

On 11 benchmarks covering math, science, medicine, search and Agent tasks, across three model scales, HybridAgent averages 10.4% higher than frontier pure-text Agents and pure-latent Agents.

The reward signals in the HybridRL stage include answer correctness, output format compliance, and whether tool calls succeed.

HybridAgent's base model is based on Qwen3, in two sizes, 8B and 14B.

HybridAgent provides an "interpretable latent mode" that can render each latent step back into pseudo-text so the internal thinking process can be observed.

Stanford, UIUC and Together AI jointly released the HybridAgent paper and open-source project.

The core mechanism of HybridAgent is to let the model switch back and forth between vector space (latent space) and text space within the same generation process.

The authors note that verl and vLLM do not directly support multi-turn hybrid generation, so they extended their own stack.

The README recommends training HybridRL with 32 GPUs.

Latent-space steps bypass the textual reasoning trace, which is equivalent to moving part of the thinking process out of the logs, causing reduced detection coverage for audit methods that rely on readable intermediate text (reasoning-chain replay, text-based jailbreak and harmful-behavior detection, failure attribution analysis), and requiring alternative means such as hidden-state probes and activation monitoring to fill the gap.

For search-intensive tasks (where only the answer is checked and the path is not audited), cheaper latent-space agents are attractive; for any scenario requiring audit, a readable textual reasoning chain remains irreplaceable.

Appendix B.1 of the paper describes the problem of small errors accumulating progressively over long latent-space segments, gradually eroding the semantic coherence of the latent trajectory.

According to Table 1 of the paper, HybridAgent 8B scores 43.7 on the GAIA benchmark, surpassing GPT-5.6 Luna's 37.9.

After detecting the special `<latent>` token, HybridAgent stops decoding and feeds the last-layer hidden state directly back as the next-step input embedding through the approximately 10-million-parameter HybridLink adapter (Eq. 4 of the paper).

The termination condition of the latent-space segment is decided autonomously by the model: it exits when the output is closest to the `</latent>` token embedding.

HybridAgent retains the full KV cache (including latent-space steps) across turns, avoiding re-encoding the trajectory into tokens.