Research map / Video / Video generation
Middle-Out Decoding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Decode the symmetric sequence from its middle outward | Sequence-to-sequence model · Synthetic de-noising test | MSE: 1.52 \times 1\mathrm{e}{-3} -> 3.51 \times 1\mathrm{e}{-4} | Helped |
| Decode the symmetric sequence from its middle outward | Sequence-to-sequence model · Synthetic de-noising test | Symmetric MSE: 0.0780 -> 1.32 \times 1\mathrm{e}{-4} | Helped |
| Add dual self-attention | Seq2Seq + Attention · MSVD (Youtube2Text) test | METEOR: 34.0 -> 34.4 | Helped |
| Add dual self-attention to middle-out decoding (ablation) | Middle-Out + Attention · MSVD (Youtube2Text) test | METEOR: 30.9 -> 34.1 | Helped |
| Use middle-out decoding instead of left-to-right decoding | Caption generation model · MSVD (Youtube2Text) test | Self METEOR: 49.6 -> 44.8 | Helped |
| Use middle-out decoding instead of left-to-right decoding | Caption generation model · MSVD (Youtube2Text) test | Self BLEU-4: 77.8 -> 70.1 | Helped |
| Provide the oracle middle word rather than the classifier prediction | Middle-Out + Attention + Self-Attention · MSVD (Youtube2Text) test | METEOR: 34.1 -> 40.9 | Helped |
| Use middle-out decoding with the oracle middle word rather than an input-conditioned Seq2Seq decoder | Oracle caption generation · MSVD (Youtube2Text) test | METEOR: 34.5 -> 40.9 | Helped |
| Use output-embedding self-attention instead of dual self-attention (ablation) | Seq2Seq + Attention · MSVD (Youtube2Text) test | METEOR: 34.4 -> 34.2 | Hurt |
| Use output-embedding self-attention instead of dual self-attention (ablation) | Seq2Seq + Attention · MSVD (Youtube2Text) test | CIDEr-D: 81.9 -> 82.8 | Helped |
| Use hidden-state self-attention instead of dual self-attention (ablation) | Seq2Seq + Attention · MSVD (Youtube2Text) test | CIDEr-D: 81.9 -> 80.7 | Hurt |
| Use hidden-state self-attention instead of dual self-attention (ablation) | Seq2Seq + Attention · MSVD (Youtube2Text) test | BLEU-4: 48.3 -> 47.2 | Hurt |
Closest papers by meaning
- SnapKV: LLM Knows What You are Looking for Before Generation (NeurIPS 2024 · similarity 0.84)
- Mimir: Improving Video Diffusion Models for Precise Text Understanding (CVPR 2025 · similarity 0.84)
- When Linear Attention Meets Autoregressive Decoding: Towards More Effective and Efficient Linearized Large Language Models (ICML 2024 · similarity 0.84)
- More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding (NeurIPS 2026 · similarity 0.84)
- A Comprehensive Study of Decoder-Only LLMs for Text-to-Image Generation (CVPR 2025 · similarity 0.84)
- VISTA-LLAMA: Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens (CVPR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.