Long-form RewardBench: Evaluating Reward Models for Long-form Generation

The widespread adoption of reinforcement learning-based alignment highlights the growing importance of reward models. Various benchmarks have been built to evaluate reward models in various domains and scenarios. However, a significant gap remains in assessing reward models for long-form generation, despite its critical role in real-world applications. To bridge this, we introduce Long-form RewardBench, the first reward modeling testbed specifically designed for long-form generation. Our benchmark encompasses five key subtasks: QA, RAG, Chat, Writing, and Reasoning. We collected instruction and preference data through a meticulously designed multi-stage data collection process, and conducted extensive experiments on 20+ mainstream reward models, including both classifiers and generative models. Our findings reveal that current models still lack long-form reward modeling capabilities. Furthermore, we designed a novel Long-form Needle-in-a-Haystack Test, which revealed a correlation between reward modeling performance and the error's position within a response, as well as the overall response length, with distinct characteristics observed between classification and generative models. Finally, we demonstrate that classifiers exhibit better generalizability compared to generative models trained on the same data. As the first benchmark for long-form reward modeling, this work aims to offer a robust platform for visualizing progress in this crucial area.

Toward Robust LLM-Based Judges: Taxonomic Bias Evaluation and Debiasing Optimization

Large language model (LLM)-based judges are widely adopted for automated evaluation and reward modeling, yet their judgments are often affected by judgment biases. Accurately evaluating these biases is essential for ensuring the reliability of LLM-based judges. However, existing studies typically investigate limited biases under a single judge formulation, either generative or discriminative, lacking a comprehensive evaluation. To bridge this gap, we propose JudgeBiasBench, a benchmark for systematically quantifying biases in LLM-based judges. JudgeBiasBench defines a taxonomy of judgment biases across 4 dimensions, and constructs bias-augmented evaluation instances through a controlled bias injection pipeline, covering 12 representative bias types. We conduct extensive experiments across both generative and discriminative judges, revealing that current judges exhibit significant and diverse bias patterns that often compromise the reliability of automated evaluation. To mitigate judgment bias, we propose bias-aware training that explicitly incorporates bias-related attributes into the training process, encouraging judges to disentangle task-relevant quality from bias-correlated cues. By adopting reinforcement learning for generative judges and contrastive learning for discriminative judges, our methods effectively reduce judgment biases while largely preserving general evaluation capability.

Beyond Token-Level Policy Gradients for Complex Reasoning with Large Language Models

Existing policy-gradient methods for auto-regressive language models typically select subsequent tokens one at a time as actions in the policy. While effective for many generation tasks, such an approach may not fully capture the structure of complex reasoning tasks, where a single semantic decision is often realized across multiple tokens--for example, when defining variables or composing equations. This introduces a potential mismatch between token-level optimization and the inherently block-level nature of reasoning in these settings. To bridge this gap, we propose Multi-token Policy Gradient Optimization (MPO), a framework that treats sequences of K consecutive tokens as unified semantic actions. This block-level perspective enables our method to capture the compositional structure of reasoning trajectories and supports optimization over coherent, higher-level objectives. Experiments on mathematical reasoning and coding benchmarks show that MPO outperforms standard token-level policy gradient baselines, highlight the limitations of token-level policy gradients for complex reasoning, motivating future research to look beyond token-level granularity for reasoning-intensive language tasks.

Thinking with Comics: Enhancing Multimodal Reasoning through Structured Visual Storytelling

Chain-of-Thought reasoning has driven large language models to extend from thinking with text to thinking with images and videos. However, different modalities still have clear limitations: static images struggle to represent temporal structure, while videos introduce substantial redundancy and computational cost. In this work, we propose Thinking with Comics, a visual reasoning paradigm that uses comics as a high information-density medium positioned between images and videos. Comics preserve temporal structure, embedded text, and narrative coherence while requiring significantly lower reasoning cost. We systematically study two reasoning paths based on comics and evaluate them on a range of reasoning tasks and long-context understanding tasks. Experimental results show that Thinking with Comics outperforms Thinking with Images on multi-step temporal and causal reasoning tasks, while remaining substantially more efficient than Thinking with Video. Further analysis indicates that different comic narrative structures and styles consistently affect performance across tasks, suggesting that comics serve as an effective intermediate visual representation for improving multimodal reasoning.

RM-Distiller: Exploiting Generative LLM for Reward Model Distillation

Reward models (RMs) play a pivotal role in aligning large language models (LLMs) with human preferences. Due to the difficulty of obtaining high-quality human preference annotations, distilling preferences from generative LLMs has emerged as a standard practice. However, existing approaches predominantly treat teacher models as simple binary annotators, failing to fully exploit the rich knowledge and capabilities for RM distillation. To address this, we propose RM-Distiller, a framework designed to systematically exploit the multifaceted capabilities of teacher LLMs: (1) Refinement capability, which synthesizes highly correlated response pairs to create fine-grained and contrastive signals. (2) Scoring capability, which guides the RM in capturing precise preference strength via a margin-aware optimization objective. (3) Generation capability, which incorporates the teacher's generative distribution to regularize the RM to preserve its fundamental linguistic knowledge. Extensive experiments demonstrate that RM-Distiller significantly outperforms traditional distillation methods both on RM benchmarks and reinforcement learning-based alignment, proving that exploiting multifaceted teacher capabilities is critical for effective reward modeling. To the best of our knowledge, this is the first systematic research on RM distillation from generative LLMs.

From Perception to Reasoning: Deep Thinking Empowers Multimodal Large Language Models

With the remarkable success of Multimodal Large Language Models (MLLMs) in perception tasks, enhancing their complex reasoning capabilities has emerged as a critical research focus. Existing models still suffer from challenges such as opaque reasoning paths and insufficient generalization ability. Chain-of-Thought (CoT) reasoning, which has demonstrated significant efficacy in language models by enhancing reasoning transparency and output interpretability, holds promise for improving model reasoning capabilities when extended to the multimodal domain. This paper provides a systematic review centered on "Multimodal Chain-of-Thought" (MCoT). First, it analyzes the background and theoretical motivations for its inception from the perspectives of technical evolution and task demands. Then, it introduces mainstream MCoT methods from three aspects: CoT paradigms, the post-training stage, and the inference stage, while also analyzing their underlying mechanisms. Furthermore, the paper summarizes existing evaluation benchmarks and metrics, and discusses the application scenarios of MCoT. Finally, it analyzes the challenges currently facing MCoT and provides an outlook on its future research directions.

ToolSample: Dual Dynamic Sampling Methods with Curriculum Learning for RL-based Tool Learning

While reinforcement learning (RL) is increasingly used for LLM-based tool learning, its efficiency is often hampered by an overabundance of simple samples that provide diminishing learning value as training progresses. Existing dynamic sampling techniques are ill-suited for the multi-task structure and fine-grained reward mechanisms inherent to tool learning. This paper introduces Dynamic Sampling with Curriculum Learning (DSCL), a framework specifically designed to address this challenge by targeting the unique characteristics of tool learning: its multiple interdependent sub-tasks and multi-valued reward functions. DSCL features two core components: Reward-Based Dynamic Sampling, which uses multi-dimensional reward statistics (mean and variance) to prioritize valuable data, and Task-Based Dynamic Curriculum Learning, which adaptively focuses training on less-mastered sub-tasks. Through extensive experiments, we demonstrate that DSCL significantly improves training efficiency and model performance over strong baselines, achieving a 3.29\% improvement on the BFCLv3 benchmark. Our method provides a tailored solution that effectively leverages the complex reward signals and sub-task dynamics within tool learning to achieve superior results.

Enhancing Large Language Models'Machine Translation via Dynamic Focus Anchoring

Large language models have demonstrated exceptional performance across multiple crosslingual NLP tasks, including machine translation (MT). However, persistent challenges remain in addressing context-sensitive units (CSUs), such as polysemous words. These CSUs not only affect the local translation accuracy of LLMs, but also affect LLMs' understanding capability for sentences and tasks, and even lead to translation failure. To address this problem, we propose a simple but effective method to enhance LLMs' MT capabilities by acquiring CSUs and applying semantic focus. Specifically, we dynamically analyze and identify translation challenges, then incorporate them into LLMs in a structured manner to mitigate mistranslations or misunderstandings of CSUs caused by information flattening. Efficiently activate LLMs to identify and apply relevant knowledge from its vast data pool in this way, ensuring more accurate translations for translating difficult terms. On a benchmark dataset of MT, our proposed method achieved competitive performance compared to multiple existing open-sourced MT baseline models. It demonstrates effectiveness and robustness across multiple language pairs, including both similar language pairs and distant language pairs. Notably, the proposed method requires no additional model training and enhances LLMs' performance across multiple NLP tasks with minimal resource consumption.

Cross-Domain Bilingual Lexicon Induction via Pretrained Language Models

Bilingual Lexicon Induction (BLI) is generally based on common domain data to obtain monolingual word embedding, and by aligning the monolingual word embeddings to obtain the cross-lingual embeddings which are used to get the word translation pairs. In this paper, we propose a new task of BLI, which is to use the monolingual corpus of the general domain and target domain to extract domain-specific bilingual dictionaries. Motivated by the ability of Pre-trained models, we propose a method to get better word embeddings that build on the recent work on BLI. This way, we introduce the Code Switch(Qin et al., 2020) firstly in the cross-domain BLI task, which can match differit is yet to be seen whether these methods are suitable for bilingual lexicon extraction in professional fields. As we can see in table 1, the classic and efficient BLI approach, Muse and Vecmap, perform much worse on the Medical dataset than on the Wiki dataset. On one hand, the specialized domain data set is relatively smaller compared to the generic domain data set generally, and specialized words have a lower frequency, which will directly affect the translation quality of bilingual dictionaries. On the other hand, static word embeddings are widely used for BLI, however, in some specific fields, the meaning of words is greatly influenced by context, in this case, using only static word embeddings may lead to greater bias. ent strategies in different contexts, making the model more suitable for this task. Experimental results show that our method can improve performances over robust BLI baselines on three specific domains by averagely improving 0.78 points.

Speculative Decoding Meets Quantization: Compatibility Evaluation and Hierarchical Framework Design

Speculative decoding and quantization effectively accelerate memory-bound inference of large language models. Speculative decoding mitigates the memory bandwidth bottleneck by verifying multiple tokens within a single forward pass, which increases computational effort. Quantization achieves this optimization by compressing weights and activations into lower bit-widths and also reduces computations via low-bit matrix multiplications. To further leverage their strengths, we investigate the integration of these two techniques. Surprisingly, experiments applying the advanced speculative decoding method EAGLE-2 to various quantized models reveal that the memory benefits from 4-bit weight quantization are diminished by the computational load from speculative decoding. Specifically, verifying a tree-style draft incurs significantly more time overhead than a single-token forward pass on 4-bit weight quantized models. This finding led to our new speculative decoding design: a hierarchical framework that employs a small model as an intermediate stage to turn tree-style drafts into sequence drafts, leveraging the memory access benefits of the target quantized model. Experimental results show that our hierarchical approach achieves a 2.78$\times$ speedup across various tasks for the 4-bit weight Llama-3-70B model on an A100 GPU, outperforming EAGLE-2 by 1.31$\times$. Code available at https://github.com/AI9Stars/SpecMQuant.