The standard model for reasoning agents today is chain-of-thought: the model externalizes its thinking as text, then uses tools or produces an answer. The reasoning is inspectable, the token budget is measurable, and every tool call happens in the sequence of output tokens. The cost is that generating intermediate reasoning tokens is expensive in proportion to how much of the computation they represent, and most of those tokens are never read by anyone.
A line of research is pushing in a different direction: do the reasoning in latent space — in the model’s internal activation space — rather than in generated tokens. The idea is that an intermediate representation doesn’t need to be human-readable text to be useful computation. If the model can route reasoning through its own activations rather than through token generation, you save the autoregressive cost of producing those tokens. LatentMAS, accepted to ICML 2026, moved multi-agent collaboration into latent space and reported 50–80% token reduction; agents pass latent thoughts rather than exchanging reasoning traces as text.
The problem with pure latent reasoning is that it has no natural interface for tool calls. Tool use happens at the boundary of model outputs: the model generates a function call, the harness executes it, and the result comes back as tokens. Latent reasoning happens before that boundary, inside the model. You can’t call a shell command from inside an activation vector.
A paper from Stanford and Together AI, Hybrid Reasoning Agents (arXiv:2610.18861, submitted October 13), proposes threading the needle. Their system, HybridAgent, switches dynamically between latent-space reasoning and text generation within a single generation pass. A lightweight component called HybridLink handles the transitions. A continuous working memory module preserves state across turns so the hybrid context survives multi-step interactions without forcing everything back to text. The backbone is Qwen3 at 4B, 8B, and 14B parameter scales.
The training is two-stage. First, HybridSFT teaches the model to coordinate with HybridLink — essentially: how to transition between modes and how to preserve context. Then HybridRL trains the agent when to use each mode. The RL stage is what makes this interesting: the model learns whether latent reasoning or text-mode reasoning is better for the current step, rather than having that decision made statically by the architecture. In practice this means tool-call steps go through text (where they have to), while reasoning steps that don’t require external interaction can stay in latent space.
The reported numbers across 11 benchmarks spanning math, science, medicine, search, and agentic reasoning: 10.4% better average performance than text-based LLM agents, 35.5–54.7% fewer tokens, 1.8–2.6× faster inference, and 2.7–3.7× lower cost. The 10.4% performance gain is interesting because it suggests the efficiency isn’t just coming from skipping reasoning tokens but from doing something computationally different in latent space. Whether that holds at 70B+ scale is not established by this paper; the largest backbone tested is 14B.
The broader question this work raises is whether token-space chain-of-thought is a transitional technique rather than a permanent feature of agent architectures. CoT was developed for models that weren’t trained to reason otherwise, and it works well enough that it became the default. But the cost scales directly with how much reasoning you want, which is why inference costs for complex agentic tasks remain high. If you can move a substantial fraction of that reasoning into the forward pass itself — retaining text for the interfaces that actually require it — you get a different cost curve.
The code for HybridAgent is available at togethercomputer/HybridAgent on GitHub. The Qwen3 backbone keeps the entry barrier manageable: 4B and 8B are within range of consumer hardware, which should make it reasonably easy to replicate the core claims.