While RLVR is effective for deterministically verifiable tasks, many vision-language tasks are only *partially verifiable*, demanding multi-criterion supervision. To bridge this gap, we propose *Reinforcement Learning with Robust Rubric Rewards (RLR³), extending RL with criterion-level rewards. To ensure faithful scoring, RLR³ routes each criterion to one of two execution paths: an LLM-as-an-extractor paired with a verifier, or an LLM-as-a-Judge for fuzzy criteria. We also introduce a minimal exposure strategy that masks verifier targets from extractors and source images from the GenRM. Furthermore, RLR³ employs decoupled normalization to mitigate score saturation and hierarchical aggregation to prioritize essential criteria. Evaluated on Qwen3-VL-30B-A3B across 15 benchmarks, RLR³ consistently outperforms RLVR, yielding a 4.7-point improvement over the base model and exceeding the official instruct-to-thinking model gap. Controlled audits show that deterministic verifier and minimal exposure strategy reduce exploitable false positives.

The effectiveness of Direct Preference Optimization (DPO) depends on data pairs that capture task-relevant quality differences. Existing data pipelines often rely on *off-policy* perturbations or *coarse* outcome-based signals, which are not well suited to fine-grained visual reasoning. We propose *rDPO*, a rubric-based data pipeline for DPO-based visual preference optimization. For each image-instruction pair, we create a checklist-style rubric of essential and additional criteria to score responses from any possible policies. The instruction-rubric pool is built *offline* and reused during the construction of *on-policy* preference data. On public reward modeling benchmarks, rubric-based prompting massively improves a 30B-A3B judge and brings it close to GPT-5.4. On public downstream benchmarks, rubric-based filtering raises the macro average to 82.69, whereas outcome-based filtering drops it to 75.82 from 81.14. In scalability evaluation on a comprehensive benchmark, rDPO achieves 61.01, markedly outperforming the style-constrained baseline (52.36) and surpassing the 59.48 base model. Together, these results show that visual preference optimization benefits from combining on-policy data construction with instance-specific criterion-level feedback.

In recent months, substantial progress has been made in complex reasoning of Large Language Models (LLMs), particularly through the application of test-time scaling. Notable examples include, though are not limited to, OpenAI’s o1/o3/o4 series and DeepSeek-R1. When responding to a query, these models generate an extended reasoning trajectory, during which the model explores, reflects, backtracks, and self-verifies before arriving at a conclusion. However, fine-tuning models with such reasoning trajectories may not always be optimal. Our findings indicate that not all components within these reasoning trajectories contribute positively to the reasoning process; in fact, some components may affect the overall performance negatively. In this study, we divide a reasoning trajectory into individual subtrajectories and develop a “5+2” framework to: (1) systematically identify suboptimal subtrajectories within the reasoning trajectory based on five human-established criteria; (2) assess the independence of the suboptimal subtrajectories identified in (1) from the subsequent content, ensuring that their elimination does not compromise overall flow and coherence of the reasoning process. Additionally, a sampling algorithm, built upon the “5+2” framework, is employed to select data whose reasoning process is free from suboptimal subtrajectories to the highest degree. Experimental results demonstrate that our method can reduce the number of suboptimal subtrajectories by 25.9% during the inference. Furthermore, our method achieves an average accuracy of 58.92% on highly challenging AIME24, AIME25, AMC24 and MATH500 benchmarks with only two thirds of training data, surpassing the average accuracy of 58.06% achieved with the entire data, and outperforming open-source datasets, including s1K-1.1, Light-R1-SFT-stage-1, OpenR1-Math-94k, and OpenThoughts-114k, when fine-tuning Qwen2.5-Math-7B. Finally, we have validated the efficacy of our method under resource-constrained scenarios, where it exhibits performance improvements across different maximum inference token limits: 2k, 4k, 8k, and 16k tokens.