Recent reinforcement learning (RL) approaches, such as outcome-supervised GRPO, have advanced reasoning in Large Language Models (LLMs), yet their adaptation to multimodal LLMs (MLLMs) remains underexplored. Progress has been further limited by the lack of evaluation settings that jointly test perception and reasoning under controlled generalization challenges. To enable such analysis, we present **SEED-Bench-R1**, a structured testbed featuring real-world video tasks and hierarchical evaluation across in-distribution, cross-environment, and cross-environment-task scenarios. Our analysis reveals that standard outcome-supervised GRPO often yields "logical incoherence"—achieving correct answers through flawed reasoning—due to its exclusive focus on final-answer rewards and rigid KL penalties. To address this, we propose **GRPO-CARE**, a consistency-aware RL framework that eliminates KL penalties while introducing a two-tiered reward system: a base reward for accuracy and an adaptive bonus for consistency. This bonus, derived from a slowly evolving reference model through group-relative likelihood calibration, rewards reasoning paths that logically support the final answer without requiring expensive process supervision. Experiments on SEED-Bench-R1 show that GRPO-CARE consistently outperforms standard GRPO, achieving a 6.7% gain on the hardest evaluation level and a 24.5% increase in reasoning consistency. Moreover, models trained with GRPO-CARE transfer effectively to diverse video understanding and even language-only reasoning benchmarks, validating its robustness and generality.

Despite the significant success of large vision-language models (LVLMs), some studies have revealed that LVLMs suffer from the hallucination problem, where the LVLMs’ response contains descriptions of non-existent objects. Although various benchmarks have been proposed to investigate this problem, they mostly focus on single-turn evaluation and overlook the hallucination raised by textual inputs. To investigate the hallucination problem of LVLMs when given long-term misleading textual history, we propose a novel visual dialogue hallucination evaluation benchmark VisDiaHalBench. The benchmark consists of samples with five-turn questions about an edited image and its original version. VisDiaHalBench differs from previous hallucination benchmarks in the following three points: 1) The questions and answers are unambiguously grounded by annotated scene graphs. 2) The images are uncommonly edited to inspect the visual model and common-object hallucination in LLMs. 3) The carefully designed dialogue refers a same object in different turns to assess the image consistency and influence of history for LVLMs. The detailed analysis of several state-of-the-art LVLMs across image consistency, visual understanding, history influence, and other dimensions reveals their substantial performance gap with single-turn VQA tasks. The benchmark is released in: https://github.com/qingxingcao/VisDiaHalBench