Reinforcement learning fine-tuning of large language models increasingly adopts multiple reward dimensions, including verifiable rules, task-specific evaluators, and learned reward models, to provide richer supervision across diverse capabilities. These dimensions are commonly scalarized with fixed aggregation weights. We identify a failure mode in which aggregation itself induces reward hacking: static projection aliases qualitatively different reward profiles into a single scalar, steering optimization toward whichever dimensions are easiest, densest, or systematically favored by the reward signal. Over training, this traps the policy in suboptimal profiles and prevents convergence to better-balanced ones that would yield higher task performance. To address this, we propose **Adaptive Multi-Reward Projection (AMRP)**, a lightweight online method that reallocates aggregation weights using three signals—relative shortfall, reward volatility, and recent progress—increasing pressure on lagging, unstable, or stagnant dimensions while relieving saturated ones. Across structured reasoning, citation-grounded generation, and open-ended alignment under GRPO, AMRP consistently improves reward-profile balance and downstream performance over fixed and dynamic weighting baselines; it also remains effective with GDPO and PPO, supporting compatibility across RL algorithms. Our code is available at https://github.com/yyhappier/AMRP.git.

Defense strategies of large language models besides alignment are introduced to defend against jailbreak attacks, and they have managed to decrease the success rate of jailbreak attacks. However, these defense strategies weakened the helpfulness of large language models. In this work, we propose a universal framework, LlmFixer, acting on large language models equipped with any defense strategy to recover their original helpfulness. LlmFixer consists of an input prompt re-writer and a logic patch. The prompt re-writer is a pre-model for clarifying the intention of input prompts, which promotes large language models to be more helpful to benign inputs and more rejective to malicious inputs. The logic patch is a lightweight structure that enhances large language models’ comprehension capacity by supplementing certain logical relationships. Without updating the parameters of a defensive large language model, LlmFixer fixes its helpfulness while preserving safety. Experiments on three large language models, five jailbreak attacks, and four defense strategies show the effectiveness of LlmFixer.

Large Language Models (LLMs) have shown remarkable capabilities in various natural language processing tasks. However, LLMs may rely on dataset biases as shortcuts for prediction, which can significantly impair their robustness and generalization capabilities. This paper presents Shortcut Suite, a comprehensive test suite designed to evaluate the impact of shortcuts on LLMs’ performance, incorporating six shortcut types, five evaluation metrics, and four prompting strategies. Our extensive experiments yield several key findings: 1) LLMs demonstrate varying reliance on shortcuts for downstream tasks, which significantly impairs their performance. 2) Larger LLMs are more likely to utilize shortcuts under zero-shot and few-shot in-context learning prompts. 3) Chain-of-thought prompting notably reduces shortcut reliance and outperforms other prompting strategies, while few-shot prompts generally underperform compared to zero-shot prompts. 4) LLMs often exhibit overconfidence in their predictions, especially when dealing with datasets that contain shortcuts. 5) LLMs generally have a lower explanation quality in shortcut-laden datasets, with errors falling into three types: distraction, disguised comprehension, and logical fallacy. Our findings offer new insights for evaluating robustness and generalization in LLMs and suggest potential directions for mitigating the reliance on shortcuts.

This paper explores the use of Deep Learning methods for automatic estimation of quality of human translations. Automatic estimation can provide useful feedback for translation teaching, examination and quality control. Conventional methods for solving this task rely on manually engineered features and external knowledge. This paper presents an end-to-end neural model without feature engineering, incorporating a cross attention mechanism to detect which parts in sentence pairs are most relevant for assessing quality. Another contribution concerns oprediction of fine-grained scores for measuring different aspects of translation quality, such as terminological accuracy or idiomatic writing. Empirical results on a large human annotated dataset show that the neural model outperforms feature-based methods significantly. The dataset and the tools are available.

In this paper we introduce MoBiL, a hybrid Monolingual, Bilingual and Language modelling feature set and feature selection and evaluation framework. The set includes translation quality indicators that can be utilized to automatically predict the quality of human translations in terms of content adequacy and language fluency. We compare MoBiL with the QuEst baseline set by using them in classifiers trained with support vector machine and relevance vector machine learning algorithms on the same data set. We also report an experiment on feature selection to opt for fewer but more informative features from MoBiL. Our experiments show that classifiers trained on our feature set perform consistently better in predicting both adequacy and fluency than the classifiers trained on the baseline feature set. MoBiL also performs well when used with both support vector machine and relevance vector machine algorithms.