frontier models
Diffusion Decoding Measures Which Tokens Interfere
Cheap in-pass probes replace a similarity proxy with the actual confidence change between positions.
Published Updated Story ID: mp-2026-11-08-018
Summary
Cheap in-pass probes replace a similarity proxy with the actual confidence change between positions.
Across three models and four benchmarks, measured coupling reduced decoding steps and was faster in most settings, including with a key-value cache, at an average 1.1-point accuracy cost. Shuffled or uniform couplings lost the gain.
Why it matters
Cheap in-pass probes replace a similarity proxy with the actual confidence change between positions.
Limits and context
No additional limitation was separately recorded.
Key claims
Cheap in-pass probes replace a similarity proxy with the actual confidence change between positions.
Evidence: source-2026-11-08-020
Sources
- arXiv preprint 2611.08307, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.