GRU
A streamlined LSTM — a recurrent network that remembers across sequences with fewer moving parts, trading a little capacity for speed and simplicity.
Reviewed July 16, 2026Stable
When not to use it
- For most large-scale sequence tasks in 2026 — a transformer usually outperforms it if you have the data and compute.
- When you specifically need maximum memory capacity — an LSTM occasionally edges it out.
- For very long-range dependencies where attention or state-space models handle the range better.
Reach for something else instead
- LSTM — the more elaborate sibling, occasionally higher-capacity.
- Transformer — the modern default for most sequence tasks, at higher compute cost.
- State-space models (Mamba) — the efficient-recurrence revival for long sequences.
Sources & further reading
- Cho et al. (2014), Learning Phrase Representations using RNN Encoder-Decoder — introduced the GRU.
- Chung et al. (2014), Empirical Evaluation of Gated Recurrent Neural Networks — the GRU-vs-LSTM comparison.
- Goodfellow, Bengio & Courville, Deep Learning — gated recurrent architectures.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Agonising over GRU vs. LSTM when they usually perform within noise of each other — just try both.
- Reaching for recurrence at all when a transformer would clearly win and resources allow.
- Forgetting that the update gate's carry-forward is what prevents vanishing gradients — the whole point of gating.
At a glance
FieldDeep Learning
What it isa simplified LSTM
Gatesupdate + reset (vs. LSTM's three)
Tradea little capacity for speed
DifficultyIntermediate