Carl von Ossietzky Universität Oldenburg

2 related articles

Subscribe to updates

arXiv Oct 12, 2026
Visual-head discovery scores shift with where attention is read: necessity tracks the query span, while steering at default strength is matched by random heads
The internal choices inside a visual-head discovery score reshape which heads it returns in vision-language models. On Qwen3-VL-8B-Instruct, fully crossing query type, discovery corpus and the token positions whose attention is read, and testing all 12 head-sets with steerability and necessity interventions, only necessity follows the attention source: query-span head-sets are more necessary than last-token ones, and the two referring plus query-span sets lie 17–19 standard deviations above 250 layer-matched random sets. At the default strength, no discovered head-set steers better than its layer-matched random sets.
The internal choices inside a visual-head discovery score reshape which heads it returns in vision-language models. On Qwen3-VL-8B-Instruct, fully crossing query type, discovery corpus and the token positions whose attention is read, and testing all 12 head-sets with steerability and necessity interventions, only necessity follows the attention source: query-span head-sets are more necessary than last-token ones, and the two referring plus query-span sets lie 17–19 standard deviations above 250 layer-matched random sets. At the default strength, no discovered head-set steers better than its layer-matched random sets.
German Research Centre for Artificial Intelligence Carl von Ossietzky Universität Oldenburg
Md Abdul Kadir（First author）
Visual Attention Heads CERS (Contrastive Evidence-Region Steering score)NCS (Necessity Contrast Score)Attention Steering Qwen3-VL
arXiv Sep 26, 2026
EAGER boosts generative event extraction with verifiable-reward reinforcement learning, beating prompting, supervised fine-tuning, and prior RL baselines across seven benchmarks
The work presents EAGER, a reinforcement learning framework for generative event extraction that combines fine-grained verifiable rewards with Schema-Contrastive Advantage Estimation to alleviate advantage collapse under sparse binary rewards; its reward design explicitly targets structural validity, extraction accuracy, groundedness, coverage, over-generation, and span precision, and across seven benchmark datasets it consistently outperforms prompting, supervised fine-tuning, and prior reinforcement learning baselines, achieving a substantial improvement over the strongest prior method.
The work presents EAGER, a reinforcement learning framework for generative event extraction that combines fine-grained verifiable rewards with Schema-Contrastive Advantage Estimation to alleviate advantage collapse under sparse binary rewards; its reward design explicitly targets structural validity, extraction accuracy, groundedness, coverage, over-generation, and span precision, and across seven benchmark datasets it consistently outperforms prompting, supervised fine-tuning, and prior reinforcement learning baselines, achieving a substantial improvement over the strongest prior method.
German Research Centre for Artificial Intelligence Carl von Ossietzky Universität Oldenburg
Omar Adjali（First author）
EAGER Event Extraction Reinforcement Learning Verifiable Rewards Schema-Contrastive Advantage Estimation

You’re all caught up

Visual-head discovery scores shift with where attention is read: necessity tracks the query span, while steering at default strength is matched by random heads

EAGER boosts generative event extraction with verifiable-reward reinforcement learning, beating prompting, supervised fine-tuning, and prior RL baselines across seven benchmarks