HybridAgents switch between latent reasoning and text tool calls, cutting tokens 35% to 55%
- HybridAgents (HybridAgent) moves between latent reasoning and text-based tool calls inside one generation process, with a lightweight HybridLink module handling the transitions between the two modes.
- Across 11 benchmarks in math, science, medicine, search, and agentic reasoning, tested at 3 model scales, the authors report a 10.4% performance gain over frontier LLM agents and agent frameworks.
- Token use falls 35.5% to 54.7%, inference runs 1.8x to 2.6x faster, and cost savings reach 2.7x to 3.7x compared with the baselines.
- Training runs in two stages: HybridSFT teaches the backbone to coordinate with HybridLink, and HybridRL learns when to think in latent space and when to act through text.
- Latent thoughts stay directly inspectable, and the code and data are published at github.com/togethercomputer/HybridAgent.
Hacker News opinions
The benchmarks are right there starting around page 9. The abstract claims a 10.4% accuracy gain, plus the token and speed savings.
How is this different from continuous chain of thought? The paper's only justification is that latent vectors lack native tool-call paths and aren't human-readable. But continuous CoT models already make tool calls by leaving thinking mode and emitting ordinary chat tokens, so that argument looks weak to me.
Someone pointed out that the readable text is just another decoder, same as current LLMs. Fair, but that's the part I'd want to see tested.
Racing to remove the only practical way we have to monitor model intent for a few percent more tokens? Less readable reasoning traces means less to audit.
The bigger safety risk to me is training models to override users 'for their own good'. Stop training that and a lot of the trouble goes away.
How does this differ from Astra's looped transformer? That one also reasons in latent space.
Answer is that Astra's model has open weights, a paper explaining it, and code. That's the difference you can actually check.
I wonder about widening the input so each step feeds both the text token and the latent vector. You already pay most of the latent-mode cost while generating text, so why not reason latently at the same time? Monitorability would probably suffer, though.
This is continuous chain of thought, and it's been an active research area for at least two years. Frontier labs' thinking summaries may already look like this in practice.
Calling this thinking is marketing. LLMs produce statistical token correlations, not reasoning.
Plane vs bird: whether something is biological doesn't decide whether it reasons. Read Thinking, Fast and Slow.