HybridAgent Paper Claims 10.4% Gain and 2.7x to 3.7x Lower Cost, but Its Own Text Gives 10.4% Two Different Meanings

A preprint from researchers at Stanford, UIUC and Together AI reports that agents trained to reason in hidden states and act in text beat text-only agents by 10.4% on average. The paper's abstract and its results section use that figure differently, and the paper does not say whether it is a relative or absolute gain.

"Hybrid Reasoning Agents" was submitted to arXiv on October 13 by Jiaru Zou and seven co-authors, including James Zou. As of Monday, no independent replication appears in the coverage cited below.

What the paper builds

The system, HybridAgent, lets a model switch between latent reasoning, meaning computation in its hidden states, and ordinary text, inside a single generation. A component called HybridLink maps the last-layer hidden states back into the input-embedding space so the model can keep going without writing words. A continuous working memory carries context across turns.

Training has two stages. HybridSFT, per the paper's HTML version, uses 54K hybrid trajectories created by automatic annotation. HybridRL uses 12K deep-reasoning examples and 1K agentic-search examples.

The 10.4% figure

The abstract says HybridAgent "consistently improves performance by 10.4%" over frontier LLM agents and agentic frameworks. The introduction calls it an average. Section 5.1 gives 18.5% for the SFT stage and 10.4% for the RL stage, each against matched text and latent baselines.

The paper does not say what the 10.4% is averaged over, or whether it is a relative or absolute change. AI Weekly, in its summary, said that the abstract does not name the frontier baselines or the efficiency baselines.

The absolute scores are more modest than the headline suggests.

Benchmark	HybridAgent 8B	HybridAgent 14B
GAIA, average	43.7	52.4
Humanity's Last Exam, average	14.4	16.0
GPQA-Diamond	48.0	55.1
AIME 25	39.2	43.3

All four rows are the authors' own figures, averaged over five seeds. The paper tests 11 benchmarks in total, including MATH500, MedQA, MuSiQue, WebWalker, SealQA and DeepSearchQA.

1. HybridAgent 8B43.7 points
2. HybridAgent 14B52.4 points

Source: Zou et al., Hybrid Reasoning Agents, arXiv:2610.18861, Table 1, accessed 2026-10-19

Efficiency claims

The abstract reports 35.5% to 54.7% fewer tokens, 1.8x to 2.6x faster inference and 2.7x to 3.7x lower cost. Crypto Briefing converted the cost range to roughly 63% to 73% less, which is the same claim in a different unit.

The token figure rests on Table 7, where text-mode tokens per instance are 302 against 596 for the 8B model and 284 against 497 for the 14B model, compared with text agents. Section 5.2 cites 2.1x wall-clock speed against text baselines without a measurement method, and the cost breakdown sits in Appendix G.2.

Crypto Briefing, dated October 14, said the work has not necessarily been through peer review and that scaling past 14B parameters is an open question. It also said pseudo-text, the readable rendering of latent steps, is not the same as reading a model's literal chain of thought. It tied that point to regulated fields such as medicine.

What can be downloaded

The GitHub repository is MIT-licensed and had 267 stars and 30 forks when fetched. Its README names Qwen3-8B-Base and Qwen3-14B-Base as the starting points and lists HybridAgent-8B-Agentic and HybridAgent-14B-Agentic weights on Hugging Face. The README reports no benchmark scores.

The 14B model page shows an MIT licence, 975 downloads in the last month and 8 likes. It lists 15B parameters in BF16 while the name says 14B.

Reproducing the training is expensive. The README recommends 8 or 16 GPUs for HybridSFT and 32 GPUs for HybridRL, and the search tool needs a Tavily API key.

For other open agent code that arrived with large star counts and thin evidence, see Google Research's Harness2. On licences for small open models, see Liquid's LFM2.5-24B-A2B.

The repository had one open issue when fetched. The README links the HybridSFT and HybridRL training datasets but reports no benchmark scores, so the only place the 10.4% can be checked against numbers is the paper itself.