The Pulse
Together AI’s HybridAgent Claims Better Results With Fewer Tokens
A paper from researchers at Together AI, Stanford, UIUC and Meta introduces HybridAgent, which switches between hidden-state reasoning and text for tool use. Its authors report higher average benchmark performance with fewer tokens, while n
AI.info Team · Written with AI ·
Text gives agents visibility; latent steps promise speed
AI agents face a design tradeoff: text-based reasoning makes intermediate steps and tool calls visible, but generating every thought as text takes tokens. Latent reasoning can keep more computation inside a model, yet it is harder to inspect and does not naturally connect to external tools. A paper submitted to arXiv on October 13, 2026, proposes a system meant to combine both approaches rather than choose between them.
The authors, affiliated with Stanford University, Together AI, the University of Illinois Urbana-Champaign and Meta, call the system HybridAgent. It shifts between latent reasoning and text during generation, using text for tool calls and answers and latent steps for internal reasoning. The paper describes its findings as an average performance improvement alongside reductions in token use, inference time and estimated cost.
“Compared with frontier LLM Agents and agentic frameworks spanning both text-based and latent methods, HybridAgent consistently improves performance by 10.4%, while delivering 35.5%–54.7% token reduction, 1.8×–2.6× faster inference, and 2.7×–3.7× cost savings.”
Jiaru Zou et al., “Hybrid Reasoning Agents”
HybridLink connects hidden states to text
HybridAgent’s switching mechanism, called HybridLink, maps the model’s final-layer hidden states back into its input-embedding space. That lets the agent continue with latent steps, then return to text when it needs to call a tool, provide an answer or expose information for monitoring. A continuous working memory carries the text, latent steps and tool observations between turns.
The authors describe two training stages. HybridSFT teaches a model to coordinate with HybridLink and use tools; HybridRL trains it to choose when to reason in latent space and when to act through text. The paper says the process adapts Qwen3 base models at 4-billion, 8-billion and 14-billion parameter scales, using 54,000 curated hybrid trajectories for supervised fine-tuning, followed by reinforcement learning on reasoning and search examples.
The researchers also offer a way to inspect latent spans by mapping them to their nearest token embeddings, which they render as pseudo-text. That is the paper’s proposed interpretation method; it does not by itself establish that a readout fully or faithfully represents the model’s internal computation.
Eleven benchmarks, with a serving-system caveat
The evaluation covers 11 benchmarks in mathematics, science, medicine, knowledge-intensive search and agentic reasoning. The paper names tests including AIME 2025 and 2026, MATH500, GPQA-Diamond, MedQA, GAIA, Humanity’s Last Exam, WebWalker, SealQA and DeepSearchQA. Baselines include text-based and latent reasoning methods, agent frameworks such as ReAct and WebThinker, and several larger language-model agents.
The authors ran their experiments on H100 GPUs and report five independent seeds. Their headline 10.4% figure is described as an average performance improvement; the abstract does not specify whether that figure is relative or expressed in percentage points. The results are the authors’ benchmark findings, not a guarantee that the same gains will hold across other tasks or deployments.
The paper’s cost comparison is an estimate, not a bill from a production service. It assumes an H100-hour costs $2.50 and GPU use reaches 10% of peak BF16 throughput; it excludes prompt prefill, context-dependent attention and external tool execution. The researchers also state that current serving engines do not support multi-turn hybrid generation, so their experiments require custom training and inference infrastructure.
Code and checkpoints are available
Together AI’s GitHub repository provides the method’s code, training tools and instructions for running batch inference. It lists 8B and 14B agentic checkpoints, along with training datasets and tools for viewing latent spans as pseudo-text. The paper’s broader claim—that one agent can retain text-based tool use while spending fewer generated tokens on internal reasoning—now rests on results that outside researchers can examine and test against the published setup.