Back to feed
arXiv cs.CL
arXiv cs.CL
6/19/2026
New paper finds query position

New paper finds query position

Original: Where to Place the Query? Unveiling and Mitigating Positional Bias in In-Context Learning for Diffusion LLMs via Decoding Dynamics

Short summary

Diffusion Large Language Models (dLLMs) handle in-context learning differently than autoregressive models, making query position a first-order variable affecting generation quality. The paper identifies a 'Recency Effect' in attention and proposes Average Confidence and Auto-ICL, a training-free strategy to dynamically optimize query placement.

  • •Query position is a critical variable in dLLMs unlike AR models
  • •Spatial 'Recency Effect' in attention explains positional sensitivity
  • •Average Confidence metric and Auto-ICL strategy proposed as solutions

Generated with AI, which can make mistakes.

#research-breakthrough#ai-tools
Read full article at arXiv cs.CL

Is this a good recommendation for you?

Explore more

Quantifying Aleatoric Uncertainty of In-Context Learning for Robust Measure of LLM Prediction Confidence
arXiv cs.CL
Adaptive scheduling steers diffusion L
arXiv cs.LG
SproutRAG: Attention-Guided Tree Search with Progressive Embeddings for Long-Document RAG
arXiv cs.CL
LEAP: Unlocking dLLM Parallelism via Lookahead Early-Convergence Token Detection
arXiv cs.LG