Large language models (LLMs) inevitably make mistakes when performing step-by-step mathematical reasoning. Process Reward Models (PRMs) have emerged as a promising solution by evaluating each reasoning step. However, existing PRMs typically output evaluation scores directly, limiting both learning efficiency and evaluation accuracy, which is further exacerbated by the scarcity of annotated data. To address these issues, we propose Reasoning-Driven Process Reward Modeling (R-PRM). First, we leverage stronger LLMs to generate seed data from limited annotations, effectively bootstrapping our model's reasoning capabilities and enabling comprehensive step-by-step evaluation. Second, we further enhance performance through preference optimization, without requiring additional annotated data. Third, we introduce inference-time scaling to fully harness the model's reasoning potential. Extensive experiments demonstrate R-PRM's effectiveness: on ProcessBench and PRMBench, it surpasses strong baselines by 11.9 and 8.5 points in F1 scores, respectively. When applied to guide mathematical reasoning, R-PRM achieves consistent accuracy improvements of over 8.5 points across six challenging datasets. Further analysis reveals that R-PRM exhibits more comprehensive evaluation and stronger generalization capabilities, thereby highlighting its significant potential.

Image captioning has been a longstanding challenge in vision-language research. With the rise of LLMs, modern Vision-Language Models (VLMs) generate detailed and comprehensive image descriptions. However, benchmarking the quality of such captions remains unresolved. This paper addresses two key questions: (1) How well do current VLMs actually perform on image captioning, particularly compared to humans? We built CapArena, a platform with over 6000 pairwise caption battles and high-quality human preference votes. Our arena-style evaluation marks a milestone, showing that leading models like GPT-4o achieve or even surpass human performance, while most open-source models lag behind. (2) Can automated metrics reliably assess detailed caption quality? Using human annotations from CapArena, we evaluate traditional and recent captioning metrics, as well as VLM-as-a-Judge. Our analysis reveals that while some metrics (e.g., METEOR) show decent caption-level agreement with humans, their systematic biases lead to inconsistencies in model ranking. In contrast, VLM-as-a-Judge demonstrates robust discernment at both the caption and model levels. Building on these insights, we release CapArena-Auto, an accurate and efficient automated benchmark for detailed captioning, achieving 94.3% correlation with human rankings at just $4 per test. Data and resources will be open-sourced at https://caparena.github.io.

Quality Estimation (QE) models evaluate the quality of machine translations without reference translations, serving as the reward models for the translation task. Due to the data scarcity, synthetic data generation has emerged as a promising solution. However, synthetic QE data often suffers from distribution shift, which can manifest as discrepancies between pseudo and real translations, or in pseudo labels that do not align with human preferences. To tackle this issue, we introduce ADSQE, a novel framework for alleviating distribution shift in synthetic QE data. To reduce the difference between pseudo and real translations, we employ the constrained beam search algorithm and enhance translation diversity through the use of distinct generation models. ADSQE uses references, i.e., translation supervision signals, to guide both the generation and annotation processes, enhancing the quality of word-level labels. ADSE further identifies the shortest phrase covering consecutive error tokens, mimicking human annotation behavior, to assign the final phrase-level labels. Specially, we underscore that the translation model can not annotate translations of itself accurately. Extensive experiments demonstrate that ADSQE outperforms SOTA baselines like COMET in both supervised and unsupervised settings. Further analysis offers insights into synthetic data generation that could benefit reward models for other tasks.

Long-context modelling for large language models (LLMs) has been a key area of recent research because many real world use cases require reasoning over longer inputs such as documents. The focus of research into modelling long context has been on how to model position and there has been little investigation into other important aspects of language modelling such as instruction tuning. Long context training examples are challenging and expensive to create and use. In this paper, we investigate how to design instruction data for the post-training phase of a long context pre-trained model: how much and what type of context is needed for optimal and efficient post-training. Our controlled study reveals that models instruction-tuned on short contexts can effectively generalize to longer ones, while also identifying other critical factors such as instruction difficulty and context composition. Based on these findings, we propose context synthesis, a novel data synthesis framework that leverages off-the-shelf LLMs to generate extended background contexts for high-quality instruction-answer pairs. Experiment results on the document-level benchmark (LongBench) demonstrate that our proposed approach outperforms previous instruction synthesis approaches and comes close to the performance of human-annotated long-context instruction data. The project will be available at: https://github.com/NJUNLP/context-synthesis.

Generating adversarial examples contributes to mainstream neural machine translation~(NMT) robustness. However, popular adversarial policies are apt for fixed tokenization, hindering its efficacy for common character perturbations involving versatile tokenization. Based on existing adversarial generation via reinforcement learning~(RL), we propose the `DexChar policy' that introduces character perturbations for the existing mainstream adversarial policy based on token substitution. Furthermore, we improve the self-supervised matching that provides feedback in RL to cater to the semantic constraints required during training adversaries. Experiments show that our method is compatible with the scenario where baseline adversaries fail, and can generate high-efficiency adversarial examples for analysis and optimization of the system.

Benchmarks are crucial for evaluating machine learning algorithm performance, facilitating comparison and identifying superior solutions. However, biases within datasets can lead models to learn shortcut patterns, resulting in inaccurate assessments and hindering real-world applicability. This paper addresses the issue of entity bias in relation extraction tasks, where models tend to rely on entity mentions rather than context. We propose a debiased relation extraction benchmark DREB that breaks the pseudo-correlation between entity mentions and relation types through entity replacement. DREB utilizes Bias Evaluator and PPL Evaluator to ensure low bias and high naturalness, providing a reliable and accurate assessment of model generalization in entity bias scenarios. To establish a new baseline on DREB, we introduce MixDebias, a debiasing method combining data-level and model training-level techniques. MixDebias effectively improves model performance on DREB while maintaining performance on the original dataset. Extensive experiments demonstrate the effectiveness and robustness of MixDebias compared to existing methods, highlighting its potential for improving the generalization ability of relation extraction models. We will release DREB and MixDebias publicly.

We propose utilizing background operators for mathematical reasoning in large language models (LLMs). To achieve this, we define a set of fundamental mathematical predicates as the basic building blocks. For each mathematical problem, we develop a Prolog solution that includes problem-specific predicates and intermediate predicates derived from these background operators, ensuring that each solution adheres to the defined operator set. We introduce the MATH-Prolog corpus, which is derived from the counting and probability categories of the MATH corpus. For efficient data augmentation, we apply K-fold cross-validated self-training. This method incrementally generates new Prolog solutions for each fold, incorporating those verified as correct into the training set throughout the model training process. Our experimental results demonstrate that 5-fold crossvalidated self-training effectively identifies new, accurate Prolog solutions, achieving an accuracy of 84.6% on the cross-validated set, and 84.8% on the test set during fine-tuning the Meta-Llama-3.1-8B-Instruct model. This approach successfully uncovers new solutions with fully computable inference steps for previously unseen problems. Additionally, incorporating the background mathematical predicates into the prompt enhances solution coverage.

Having been trained on massive pretraining data, large language models have shown excellent performance on many knowledge-intensive tasks. However, pretraining data tends to contain misleading and even conflicting information, and it is intriguing to understand how LLMs handle these noisy data during training. In this study, we systematically analyze LLMs' learning preferences for data with conflicting knowledge. We find that pretrained LLMs establish learning preferences similar to humans, i.e., preferences towards formal texts and texts with fewer spelling errors, resulting in faster learning and more favorable treatment of knowledge in data with such features when facing conflicts. This finding is generalizable across models and languages and is more evident in larger models. An in-depth analysis reveals that LLMs tend to trust data with features that signify consistency with the majority of data, and it is possible to instill new preferences and erase old ones by manipulating the degree of consistency with the majority data.

Decoding by contrasting layers (DoLa), is designed to improve the generation quality of large language models (LLMs) by contrasting the prediction probabilities between an early exit output (amateur logits) and the final output (expert logits). However, we find that this approach does not work well on non-English tasks. Inspired by previous interpretability work on language transition during the model's forward pass, we discover that this issue arises from a language mismatch between early exit output and final output. In this work, we propose an improved contrastive decoding algorithm that is effective for diverse languages beyond English. To obtain more helpful amateur logits, we devise two strategies to skip a set of bottom, language-agnostic layers based on our preliminary analysis. Experimental results on multilingual reasoning benchmarks demonstrate that our proposed method outperforms previous contrastive decoding baselines and substantially improves LLM's chain-of-thought reasoning accuracy across 11 languages. The project will be available at: https://github.com/NJUNLP/SkipLayerCD.

The process of meaning composition, wherein smaller units like morphemes or words combine to form the meaning of phrases and sentences, is essential for human sentence comprehension. Despite extensive neurolinguistic research into the brain regions involved in meaning composition, a computational metric to quantify the extent of composition is still lacking. Drawing on the key-value memory interpretation of transformer feed-forward network blocks, we introduce the Composition Score, a novel model-based metric designed to quantify the degree of meaning composition during sentence comprehension. Experimental findings show that this metric correlates with brain clusters associated with word frequency, structural processing, and general sensitivity to words, suggesting the multifaceted nature of meaning composition during human sentence comprehension.