Stanford and Together AI's HybridAgents reason in latent space, gaining 10.4% with 2.7-3.7x cost savings
omarsar0 · x · 2026-10-15
Researchers from Stanford, Together AI, UIUC, and Meta introduce Hybrid Reasoning Agents (HybridAgents), which reason internally in latent space and only switch to text for tool calls or final answers.
Core design:
- HybridLink: a lightweight module mapping last-layer hidden states back into embedding space for seamless text↔latent mode switching, keeping latent thoughts inspectable
- Working memory: preserves hybrid context across turns; tool calls stay plain text
- Two-stage training: HybridFT (SFT) stabilizes tool-integrated hybrid generation; HybridRL learns when to think latently vs act in text
Results: across 11 benchmarks (math, science, medicine, search, agentic reasoning) at 3 model scales, it improves performance by 10.4% over frontier LLM agents and frameworks, with 35.5%-64.7% token reduction, 1.8-2.6x faster inference, and 2.7-3.7x cost savings. Code: github.com/togethercomputer/HybridAgent.
More from coding & agent
- Skip the deck: present with Codex Sites and embed live demos instead — gabrielchua · 2026-10-15
- Geoffrey Huntley: hand-typing code is over—delete your skills, engineer the loop — rseroter · 2026-10-15
- AI coding agent deletes production database in 9 seconds; Keycard calls for task-based agent auth — matt_slotnick · 2026-10-15
- vLLM's bitwise-consistent RL: zero KL on Nemotron NVFP4, 91.3% throughput on DeepSeek-V4-Flash — DynamicWebPaige · 2026-10-15
- vLLM achieves bitwise-consistent RL training: DeepSeek-V4-Flash hits 0 diff over 100 steps — vllm_project · 2026-10-15
- ChatGPT Sites + MCP Events + dots: A Workflow-Triggered Agent Pattern — gabrielchua · 2026-10-15