A research team has built an AI agent that can do some reasoning without turning each step into words, then switch to ordinary text to use a tool or answer a person. In an October 13 preprint, the researchers report tests across 11 benchmarks and estimate lower generation costs than text-based agents with the same model sizes.
One model can reason internally and act in text
Text-based language-model agents produce text at each reasoning step. That makes their intermediate work easy to read, but it also means generating and decoding many tokens before the agent can call a search tool or run code. Latent reasoning takes another route: the model reuses its internal numerical representation instead of turning each step into a word. Those hidden states are not ordinary sentences a person can read.
The system, called HybridAgent, alternates between the two modes. Its HybridLink component maps a hidden state into the model’s input space so the next step can continue without first producing a text token. When the agent needs an external action, it returns to text for a tool call; a continuous working memory carries the mixed text, hidden states and tool results across turns. The authors trained the transition with supervised examples, then used reinforcement learning to teach the agent when to reason internally and when to act in text.
Benchmark results for the 14B model
The paper evaluates Qwen3-based versions with 4 billion, 8 billion and 14 billion parameters on 11 benchmarks covering math, science, medicine, search and agent tasks. GAIA asks assistants to answer real-world questions using reasoning, web browsing, information handling and tools; it scores answers by normalized near-exact match against a reference answer. WebWalkerQA tests whether an agent can explore website subpages and combine information from one or more sources; it measures answer accuracy using GPT-4 as a judge. For the 14B model, the paper reports average scores of 52.4% on GAIA and 49.5% on WebWalkerQA, compared with 43.7% and 40.5% for the strongest 14B agentic baseline. These percentages come from separate benchmark sets, not a single combined score. The authors report averaging results over five runs with different random seeds. The results come from benchmark tests, not customers or a deployed service.
Lower estimated generation costs in a custom research setup
For matched Qwen3 backbones, the researchers estimate decoding cost per million questions for a text agent and HybridAgent, respectively: $359 and $98 at 4B, $439 and $121 at 8B, and $470 and $176 at 14B. The calculation assumes an H100 hour costs $2.50 and runs at 10% of peak BF16 throughput. It counts the model’s generation compute, but excludes prompt prefill, attention over the growing context and external tool execution, so the figures are not total operating costs.
The team also says existing serving engines did not support its multi-turn combination of hidden-state reasoning and tool calls. It extended its training and inference stack and ran the experiments on 32 H100 GPUs. The preprint therefore describes a research system with a custom setup; it does not show a ready-to-use agent or an independently replicated cost advantage.
Sources
- Jiaru Zou et al., “Hybrid Reasoning Agents,” arXiv:2610.18861v1, submitted Oct. 13, 2026; full text and evaluation tables
- GAIA benchmark paper, evaluation by normalized near-exact match
- WebWalkerQA paper, answer-accuracy metric with GPT-4 judging
- Illustrative image: CSIRO server-rack photograph, CC BY 3.0 license. Image unchanged and not the study’s hardware.