HoPE proposes a positional encoding without long-term decay for LLMs
burny_tech · x · 2026-07-25
HoPE proposes a new positional encoding without long-term decay
The paper introduces HoPE, a positional encoding method aimed at improving context awareness and extrapolation in large language models.
The authors argue that many existing positional encodings assume long-term decay, but that assumption is increasingly outdated for LLMs. Their empirical analysis suggests models often learn a local-decay pattern with a global U-shaped attention profile, and they connect that behavior to components in rotary position encoding (RoPE).
HoPE replaces the position-specific parts of RoPE with a position-independent formulation that preserves only high-frequency signals. The paper claims two main benefits: better spontaneous attention optimization and extrapolation, plus improved optimization of position/semantic components. The screenshot indicates experiments supporting stronger context awareness and extrapolation.
More from Research
- RL Consistently Improves Imagination Models: Photon-1 Beats Gemini — ycombinator · 2026-07-25
- A user wants one-photo side and back views without changing pose for 3D modeling — LoudPoem9492 · 2026-07-25
- Terence Tao slide argues AI-era papers need better exposition, not just proofs — AlexKontorovich · 2026-07-25
- Lean formalizes a sharp product-free set theorem from Gowers and collaborators — satnam6502 · 2026-07-25
- Audit finds frontier model docs mention multiple perspectives, not pluralism — evijit · 2026-07-25
- ICML interpretability talk asks what mechanistic interpretability is actually for — ericjmichaud_ · 2026-07-25