HybridAgent
A hybrid reasoning agent architecture that allows a large language model to switch dynamically between textual reasoning and latent-space reasoning. It enters latent-space mode when deeper thought is required and returns to text mode when tools must be called or output produced, while retaining tool-calling capability and reasoning interpretability.
Core Facts
Timeline (last 90 days)
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.
Table 9 estimates computational cost rather than API bills: at $2.50 per H100 hour and 10% utilization of peak BF16 throughput, each latent-space step is billed as one full forward pass.
Under this cost-estimation basis, the 8B agent's cost per million questions drops from $439 to $121.
Appendix F of the paper provides a pseudo-text window: each latent-space step can be mapped to the nearest token embedding, and the readability score is taken as the mean of the best cosine match over the vocabulary.
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.
The HybridAgent repository releases weights, training data and code under the MIT license.
The model card explicitly states that loading with `transformers` alone yields a pure-text model; latent-space mode and tool calling must go through the HybridAgent code stack, and currently no inference provider hosts this model.
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.
16 more timeline events
All Facts (20)
Compared with the baseline, HybridAgent reduces token consumption by 35.5% to 54.7%, improves inference speed by 1.8x to 2.6x, and lowers cost by 2.7x to 3.7x.
75%UnverifiedFor 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.
50%UnverifiedThe HybridRL training stage is responsible for teaching the model when to think silently and when to put pen to paper and output.
50%UnverifiedThe incremental gains from latent-space mode are: GAIA rises from 38.8 to 43.7, and WebWalker jumps from 32.0 to 42.5.
50%UnverifiedTable 9 estimates computational cost rather than API bills: at $2.50 per H100 hour and 10% utilization of peak BF16 throughput, each latent-space step is billed as one full forward pass.
50%UnverifiedUnder this cost-estimation basis, the 8B agent's cost per million questions drops from $439 to $121.
50%UnverifiedAppendix F of the paper provides a pseudo-text window: each latent-space step can be mapped to the nearest token embedding, and the readability score is taken as the mean of the best cosine match over the vocabulary.
50%UnverifiedAppendix 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.
50%UnverifiedThe model card explicitly states that loading with `transformers` alone yields a pure-text model; latent-space mode and tool calling must go through the HybridAgent code stack, and currently no inference provider hosts this model.
50%UnverifiedThe authors note that verl and vLLM do not directly support multi-turn hybrid generation, so they extended their own stack.
50%UnverifiedThe README recommends training HybridRL with 32 GPUs.
50%UnverifiedLatent-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.
50%UnverifiedAccording to Table 1 of the paper, HybridAgent 8B scores 43.7 on the GAIA benchmark, surpassing GPT-5.6 Luna's 37.9.
50%UnverifiedAfter 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).
50%UnverifiedThe 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.
50%UnverifiedHybridAgent retains the full KV cache (including latent-space steps) across turns, avoiding re-encoding the trajectory into tokens.
50%UnverifiedHybridAgent's dual-mode mechanism enters latent-space reasoning when deep thinking is needed and switches back to text mode when interacting with tools or the environment.
50%UnverifiedHybridAgent adopts a two-stage post-training pipeline: HybridSFT (hybrid supervised fine-tuning) and HybridRL (hybrid reinforcement learning).
50%UnverifiedCompared with frontier LLM Agents and agent frameworks (including pure-text and pure-latent-space methods), HybridAgent achieves a consistent performance improvement of 10.4%.
50%UnverifiedHybridLink is a lightweight translation module in HybridAgent used for the representation between text mode and latent space mode.
50%