Agents learn to think in latent space: HybridAgent gains 10.4% on average across 11 benchmarks while cutting tokens by up to 54.7%
Synopsis
HybridAgent lets a single agent move back and forth between continuous latent reasoning and text interaction within one generation, so part of its thinking leaves token-by-token decoding while tool use stays intact. Text agents previously had to write every step as tokens, while latent-reasoning models lacked native tool access and were hard to inspect. Across 11 benchmarks in mathematics, science, medicine, search and agentic reasoning at 4B to 14B scale, it reports a 10.4% average gain over frontier text and latent agents, with 35.5%–54.7% fewer tokens and 1.8–2.6 times faster inference.
Interpretation
An agent can alternate between two media in one generation: continuous latent vectors carry internal reasoning, and text handles tool calls and outward expression. Earlier agents decoded every intermediate reasoning step into text tokens before acting, while latent-reasoning models kept computation inside the model, lacked a native path for multi-turn tool calls, and were hard to supervise or audit. On 11 benchmarks spanning mathematics, science, medicine, knowledge search and agentic reasoning, HybridAgent reports a 10.4% average gain over frontier text and latent agents and agentic frameworks, together with 35.5%–54.7% token reduction, 1.8–2.6 times inference speedup and 2.7–3.7 times cost savings.
Switching between the two modes is handled by a lightweight adapter that maps the last-layer hidden state back into the input embedding space, so a latent step can continue generation after a text step. Feeding the last-layer hidden state directly back as the next input creates a representation mismatch with the pretrained input space, making transitions between continuous latent states and discrete tokens difficult. The adapter has roughly 10M parameters for a 4B model and uses a residual connection to preserve the original hidden-state semantics; on a 14B backbone, the two-layer residual design scores 88.2, 81.7 and 34.5 on MATH500, MedQA and MuSiQue, above direct reuse without alignment at 71.4, 73.7 and 24.5.
Multi-turn hybrid generation runs on continuous working memory, where text, latent vectors and tool observations accumulate as KV entries in positional order and the next turn resumes from that memory. Standard agents re-encode each turn's generated content and tool outputs into a textual context, whereas hybrid trajectories interleave text tokens with latent vectors and cannot be reconstructed by conventional tokenization alone. Working memory preserves the complete hybrid context without converting latent thoughts back to text; training proceeds in two stages, HybridSFT on 54K hybrid trajectories followed by HybridRL on 12K deep-reasoning examples and 1K agentic search examples.
Perspective
Evaluation covers 11 benchmarks in mathematics, science, medicine, knowledge search and agentic reasoning, with Qwen3-Base models at 4B, 8B and 14B scale; tools are limited to a sandboxed Python interpreter and a live search engine, with at most 10 tool calls per trajectory.
The efficiency figures are reported relative to text and latent baselines; cost estimates assume $2.50 per H100-hour and 10% utilization of peak BF16 throughput, and exclude prompt prefill, context-dependent attention and external tool execution. Existing training and inference frameworks do not directly support multi-turn hybrid generation and had to be extended.