Tag
AVC-Reward 提出一个事件中心的音视频一致性评估框架，通过可训练奖励模型在全局和谐、时间同步和属性对齐三个层面评估复杂多事件场景，并在与人类判断对齐上达到最先进水平，还可作为在线强化学习信号提升联合生成模型的一致性。
TextVerifier 提出了一种针对中文文本渲染缺陷的逐字符验证方法，利用表意文字描述序列（IDS）作为文本塔输入，并在 TextRenderBench-ZH 基准上达到 0.844 的 F1，显著优于 OCR 置信度、通用 VLM 和 TextPecker，还可作为 Flow-GRPO 后训练的逐字符奖励进一步降低渲染错误。
SyncReward 提出了一套面向音视频生成的人体对齐同步奖励框架，包括 80K 专业同步评分数据集 SyncReward-Data-80K、1K 挑战基准 SyncReward-Bench，以及通过语义-时间对比预训练与人类监督训练的奖励模型，可作为评估指标并用于推理时 best-of-N 选择和强化学习后训练以提升音视频同步质量。
This paper introduces the Thinking Reward Model (TRM) that explicitly formulates case-adaptive rubrics for evaluating visual generation, achieving state-of-the-art performance among open-source reward models.
Dyna Robotics addresses the reliability gap in robotics by combining a large-scale data pyramid, a specialized reward model, and iterative deployment to achieve near-perfect task performance, moving towards commercial-grade autonomy.
A benchmark release comparing 24 LLMs to human writers on 475 creative writing prompts, using a custom reward model to show that frontier models outperform amateurs but not professionals.
F2DR is a fine-grained reward framework designed to evaluate full-pipeline DeepSearch workflows in large language models, addressing limitations of existing reward models by assessing content, trajectory, and answer dimensions, and introducing DeepSearch RM-Bench for benchmarking.
RynnValue introduces an open-source robotic value foundation model that uses temporal distance as a scalable supervision target for reward learning, surpassing preference-supervised state-of-the-art on RBM-EVAL-OOD and improving real-world policy success rates.
This paper introduces a Multi-dimensional Evaluation-Verification Reward (EVR) for reinforcement learning fine-tuning of multi-reference image editing models, improving visual consistency and harmony.
Introduces VlogReward, a reward model for evaluating vlog editing plans across six dimensions, along with a large-scale dataset and benchmark, achieving state-of-the-art results against GPT-5 and Gemini-3-Pro.
This paper introduces PAJAMA, a system that distills LLM-as-a-judge decision logic into programmatic judges, eliminating per-sample API costs while matching the performance of a 13B LLM judge, and providing transparency and efficiency improvements for scalable evaluation.
This paper presents a systems study comparing C++ and PyTorch inference runtimes for reward model scoring in RLHF pipelines, finding that ONNXRuntime provides speedups on CPU while torch.compile leads on GPU, with batching strategy mattering more than language or runtime.
TRACE is a dense credit assignment method for multi-turn agentic reinforcement learning that uses a frozen reference model to compute per-action rewards from log-probability changes at tool boundaries, eliminating the need for a critic or process reward model. It significantly improves long-horizon tool-use performance on benchmarks like BrowseComp-Plus.
This paper introduces SpectraReward, a training-free reward function that leverages pretrained multimodal large language models (MLLMs) as zero-shot reward models for reinforcement learning in text-to-image generation, demonstrating consistent improvements over prior methods.
Introduces PEBS, a per-rater empirical-Bayes shrinkage estimator for calibrating reward models in RLHF, reducing within-user RMSE by over 8.5% on PRISM and over 9.6% on PluriHarms.
TuneJury is an open-source pairwise reward model for text-to-music generation that provides calibrated preference scoring and generalizes across multiple downstream applications.
Neural_avb releases a lightweight Answer-eq Reward Model for RL training on QA tasks, claiming 80% agreement with external judge LM and faster than F1/ROUGE/BertScore.
This paper introduces a framework for time series forecasting that uses importance-aware news compression and process reward model-guided retrieval to incorporate long news articles within fixed context limits, improving prediction accuracy across finance, energy, traffic, and Bitcoin benchmarks.
This paper introduces a method to predict best-of-N inference scaling gains for language models using cheap statistics from a single labeled validation-set sampling pass. A compact predictor with three core features achieves Spearman ρ=0.90 with actual gains, enabling screening of configurations before expensive reward-model scoring.
Introduces Latent Reward Steering (Lrs), an adaptive inference-time framework that uses sparse autoencoder latent states and a learned reward model to implicitly promote cognitive behaviors like verification and backtracking in reasoning LLMs, improving performance across multiple models and benchmarks.