EASE: evidence-anchored spatial attention lifts multimodal RLVR by up to 3.1 points, EMNLP 2026
jiqizhixin · x · 2026-09-11
Harbin Institute of Technology and Zhongguancun College present EASE (Evidence-Anchored Spatial Attention), accepted to EMNLP 2026 Findings. Outcome-only reward in multimodal RLVR raises accuracy but leaves a systematic mismatch between where the model looks and where the evidence is. EASE adds a process-level signal: an annotation pipeline produces bounding boxes for answer-supporting regions (training metadata only), each converted to a 2D Gaussian (2σ covering half the box), equal-weight mixing of multiple boxes, and 10% background smoothing. Gains: +2.9 on Qwen2.5-VL-7B, +3.1 on Qwen3-VL-4B, +2.5 on Qwen3-VL-8B over the DAPO baseline.
More from Research
- ParetoRSI-Gym benchmarks whether AI agents can recursively self-improve LLMs across domains — Haojin Wang · 2026-10-28
- Stanford's Ψ-0.6 is a unified vision-language world model with adaptive reasoning streams — Stanford · 2026-10-28
- Shanghai AI Lab's OpenESM shows energy-steered models can scale as foundation learners — Shanghai-AI-Laboratory · 2026-10-28
- CellTrack-Bench tests VLMs on microscopic cell tracking, exposing reasoning gaps — biohub · 2026-10-28
- Hybrid LLM agent scores 83.2%, beats nine VLA policies on new SO-101 robot arm benchmark — Truman Hickok · 2026-10-28
- DeepMind researcher: all of mathematics will soon be formalizable, with agents hunting proofs — ChrSzegedy · 2026-10-28