Paper: More Retrieved Pages Hurt Diffusion LMs; ECF Framework Solves It
_reachsumit · x · 2026-08-10
Traditional visual RAG typically feeds all retrieved pages to the generator, but this paper reveals this assumption fails for Diffusion Language Models (DLMs).
- Counter-intuitive Phenomenon: While retrieving more pages improves answer recall, passing them all to the model decreases accuracy due to semantic conflicts. The root cause lies in parallel denoising, where position-wise proposals combine incompatible visual sources into unsupported answers.
- Proposed Solution: The authors introduce ECF (Entropy-Based Candidate Filter), a training-free evidence-admission framework. It constructs multi-granularity evidence units and uses blank-controlled block confidence and retrieval rank to selectively admit beneficial context before decoding.
- Results: Experiments across three multimodal DLMs show ECF successfully preserves retrieval coverage while limiting harmful visual exposure.
More from Research
- CityU Hong Kong unveils RefineEdit, a training-free image editing framework that tops PIE-Bench — CityU-HongKong · 2026-09-21
- tokenbender shares a link to "some of the nicest material to read" — tokenbender · 2026-09-21
- QwenImage 2.1 adds KV caching, measured 2.55x faster image generation — RisingSayak · 2026-09-21
- The biggest AI-for-science mistakes happen before you pick a model: 11 ideas — bravo_abad · 2026-09-21
- 19 teams, 478 submissions, $50k in prizes: Aletheia's Quest AI lie detection winners announced — gsarti_ · 2026-09-21
- IBM Launches Steerability Challenge: Suppress LLM Dishonesty, Win Cash and NeurIPS Spot — davidbau · 2026-09-21