Stanford's Hybrid Reasoning Agents think in latent space, gain 10.4% across 11 benchmarks

alex_verem · x · 2026-11-05

Researchers from Stanford, Together AI, UIUC, and Meta published Hybrid Reasoning Agents (arXiv:2610.18861), an agent that dynamically switches between latent-space reasoning and text-based interaction.

Design

A lightweight HybridLink module toggles between text and latent modes; continuous working memory preserves lossless hybrid context across turns
Routine steps reason in hidden states; tool calls and major decisions return to plain text for inspectability
Post-training pipeline: HybridSFT stabilizes tool-integrated hybrid generation, HybridRL learns when to think latently vs. act in text

Results

+10.4% average across 3 model scales and 11 benchmarks (math, science, medicine, search, agentic reasoning)
35.5%-54.7% token reduction, 1.8x-2.6x faster inference, 2.7x-3.7x cost savings
Largest build on Qwen3-14B scores 49.5 on WebWalker vs. 11.0 for OpenAI's cheapest GPT-5.6 Luna, though it loses 2 of 5 agent benchmarks (HLE 16.0 vs 29.0)
Over a third of math steps skip text generation; per-question time drops from 134s to 63s; generation compute cost per million questions falls from $470 to $176

Interpretability

Swapping hidden thoughts with those from another problem drops math accuracy from 88.2% to 48.6%, confirming answers depend on latent reasoning. Converting hidden steps to rough text, an external monitor judged answers on 560 problems about as well as for a text-only agent. The paper notes that in closed products, only the provider could read the hidden steps — raising trust questions about who gets to see what an AI is thinking.

Original post →

Stanford's Hybrid Reasoning Agents think in latent space, gain 10.4% across 11 benchmarks

More from coding & agent