Recent advances in test-time alignment methods, such as Best-of-N sampling, offer a simple and effective way to steer language models (LMs) toward preferred behaviors using reward models (RM). However, these approaches can be computationally expensive, especially when applied uniformly across prompts without accounting for differences in alignment difficulty. In this work, we propose a prompt-adaptive strategy for Best-of-N alignment that allocates inference-time compute more efficiently. Motivated by latency concerns, we develop a two-stage algorithm: an initial exploratory phase estimates the reward distribution for each prompt using a small exploration budget, and a second stage adaptively allocates the remaining budget using these estimates. Our method is simple, practical, and compatible with any LM-RM combination. Empirical results on prompts from the AlpacaEval, HH-RLHF, and PKU-SafeRLHF datasets for 12 LM-RM pairs and 50 different batches of prompts show that our adaptive strategy outperforms the uniform allocation with the same inference budget. Moreover, we show that our adaptive strategy remains competitive against uniform allocations with 20% larger inference budgets and improves in performance as the batch size grows. Language Models (LMs) have demonstrated human-like capabilities across a wide range of tasks, including mathematics, coding, and creative writing (Brown et al., 2020; Achiam et al., 2023). While pre-training on massive corpora equips these models with extensive knowledge, it is crucial that their responses at inference-time adhere to ethical standards and safety guidelines. A common approach involves leveraging preference data to steer the model toward more desirable outputs. For example, post-training methods such as Reinforcement Learning with Human Feedback (RLHF) (Christiano et al., 2017; Ouyang et al., 2022), Direct Preference Optimization (DPO) (Rafailov et al., 2023), and its variants (Glaese et al., 2022), fine-tune the model weights, while constraining the updated model to remain close to a reference model. Despite its empirical success, post-training methods are computationally expensive and can introduce unintended and opaque changes to the base model (Ouyang et al., 2022; Bai et al., 2022).

Idioms are defined as a group of words with a figurative meaning not deducible from their individual components. Although modern machine translation systems have made remarkable progress, translating idioms remains a major challenge, especially for speech-to-text systems, where research on this topic is notably sparse. In this paper, we systematically evaluate idiom translation as compared to conventional news translation in both text-to-text machine translation (MT) and speech-to-text translation (SLT) systems across two language pairs (German to English, Russian to English). We compare state-of-the-art end-to-end SLT systems (SeamlessM4T SLT-to-text, Whisper Large v3) with MT systems (SeamlessM4T SLT-to-text, No Language Left Behind), Large Language Models (DeepSeek, LLaMA) and cascaded alternatives. Our results reveal that SLT systems experience a pronounced performance drop on idiomatic data, often reverting to literal translations even in higher layers, whereas MT systems and Large Language Models demonstrate better handling of idioms. These findings underscore the need for idiom-specific strategies and improved internal representations in SLT architectures.

In this study, we investigate whether LLMs can be used to indicate if a study in the behavioural social sciences is replicable. Using a dataset of 14 previously replicated studies (9 successful, 5 unsuccessful), we evaluate the ability of both open-source (Llama 3 8B, Qwen 2 7B, Mistral 7B) and proprietary (GPT-4o) instruction-tuned LLMs to discriminate between replicable and non-replicable findings. We use LLMs to generate synthetic samples of responses from behavioural studies and estimate whether the measured effects support the original findings. When compared with human replication results for these studies, we achieve F1 values of up to $77\%$ with Mistral 7B, $67\%$ with GPT-4o and Llama 3 8B, and $55\%$ with Qwen 2 7B, suggesting their potential for this task. We also analyse how effect size calculations are affected by sampling temperature and find that low variance (due to temperature) leads to biased effect estimates.

This study introduces RelCAT (Relation Concept Annotation Toolkit), an interactive tool, library, and workflow designed to classify relations between entities extracted from clinical narratives. Building upon the CogStack MedCAT framework, RelCAT addresses the challenge of capturing complete clinical relations dispersed within text. The toolkit implements state-of-the-art machine learning models such as BERT and Llama along with proven evaluation and training methods. We demonstrate a dataset annotation tool (built within MedCATTrainer), model training, and evaluate our methodology on both openly available gold-standard and real-world UK National Health Service (NHS) hospital clinical datasets. We perform extensive experimentation and a comparative analysis of the various publicly available models with varied approaches selected for model fine-tuning. Finally, we achieve macro F1-scores of 0.977 on the gold-standard n2c2, surpassing the previous state-of-the-art performance, and achieve performance of >=0.93 F1 on our NHS gathered datasets.

SMILES-based molecular generative models have been pivotal in drug design but face challenges in fragment-constrained tasks. To address this, the Sequential Attachment-based Fragment Embedding (SAFE) representation was recently introduced as an alternative that streamlines those tasks. In this study, we investigate the optimal setups for training SAFE generative models, focusing on dataset size, data augmentation through randomization, model architecture, and bond disconnection algorithms. We found that larger, more diverse datasets improve performance, with the LLaMA architecture using Rotary Positional Embedding proving most robust. SAFE-based models also consistently outperform SMILES-based approaches in scaffold decoration and linker design, particularly with BRICS decomposition yielding the best results.

Sparse reward environments in reinforcement learning (RL) pose significant challenges for exploration, often leading to inefficient or incomplete learning processes. To tackle this issue, this work proposes a teacher-student RL framework that leverages Large Language Models (LLMs) as "teachers" to guide the agent's learning process by decomposing complex tasks into subgoals. Due to their inherent capability to understand RL environments based on a textual description of structure and purpose, LLMs can provide subgoals to accomplish the task defined for the environment in a similar fashion to how a human would do. In doing so, three types of subgoals are proposed: positional targets relative to the agent, object representations, and language-based instructions generated directly by the LLM. More importantly, we show that it is possible to query the LLM only during the training phase, enabling agents to operate within the environment without any LLM intervention. We assess the performance of this proposed framework by evaluating three state-of-the-art open-source LLMs (Llama, DeepSeek, Qwen) eliciting subgoals across various procedurally generated environment of the MiniGrid benchmark. Experimental results demonstrate that this curriculum-based approach accelerates learning and enhances exploration in complex tasks, achieving up to 30 to 200 times faster convergence in training steps compared to recent baselines designed for sparse reward environments.

Ensembling multiple models has always been an effective approach to push the limits of existing performance and is widely used in classification tasks by simply averaging the classification probability vectors from multiple classifiers to achieve better accuracy. However, in the thriving open-source Large Language Model (LLM) community, ensembling methods are rare and typically limited to ensembling the full-text outputs of LLMs, such as selecting the best output using a ranker, which leads to underutilization of token-level probability information. In this paper, we treat the Generation of each token by LLMs as a Classification (GaC) for ensembling. This approach fully exploits the probability information at each generation step and better prevents LLMs from producing early incorrect tokens that lead to snowballing errors. In experiments, we ensemble state-of-the-art LLMs on several benchmarks, including exams, mathematics and reasoning, and observe that our method breaks the existing community performance ceiling. Furthermore, we observed that most of the tokens in the answer are simple and do not affect the correctness of the final answer. Therefore, we also experimented with ensembling only key tokens, and the results showed better performance with lower latency across benchmarks.

This paper explores an intriguing observation: fine-tuning a large language model (LLM) with responses generated by a LLM often yields better results than using responses generated by humans. We conduct an in-depth investigation to understand why this occurs. Contrary to the common belief that these instances is simply due to the more detailed nature of LLM-generated content, our study identifies another contributing factor: an LLM is inherently more "familiar" with LLM generated responses. This familiarity is evidenced by lower perplexity before fine-tuning. We design a series of experiments to understand the impact of the "familiarity" and our conclusion reveals that this "familiarity" significantly impacts learning performance. Training with LLM-generated responses not only enhances performance but also helps maintain the model's capabilities in other tasks after fine-tuning on a specific task.

Employing Large Language Models (LLM) in various downstream applications such as classification is crucial, especially for smaller companies lacking the expertise and resources required for fine-tuning a model. Fairness in LLMs helps ensure inclusivity, equal representation based on factors such as race, gender and promotes responsible AI deployment. As the use of LLMs has become increasingly prevalent, it is essential to assess whether LLMs can generate fair outcomes when subjected to considerations of fairness. In this study, we introduce a framework outlining fairness regulations aligned with various fairness definitions, with each definition being modulated by varying degrees of abstraction. We explore the configuration for in-context learning and the procedure for selecting in-context demonstrations using RAG, while incorporating fairness rules into the process. Experiments conducted with different LLMs indicate that GPT-4 delivers superior results in terms of both accuracy and fairness compared to other models. This work is one of the early attempts to achieve fairness in prediction tasks by utilizing LLMs through in-context learning.

The extraordinary performance of large language models (LLMs) heightens the importance of detecting whether the context is generated by an AI system. More importantly, while more and more companies and institutions release their LLMs, the origin can be hard to trace. Since LLMs are heading towards the time of AGI, similar to the origin tracing in anthropology, it is of great importance to trace the origin of LLMs. In this paper, we first raise the concern of the origin tracing of LLMs and propose an effective method to trace and detect AI-generated contexts. We introduce a novel algorithm that leverages the contrastive features between LLMs and extracts model-wise features to trace the text origins. Our proposed method works under both white-box and black-box settings therefore can be widely generalized to detect various LLMs.(e.g. can be generalized to detect GPT-3 models without the GPT-3 models). Also, our proposed method requires only limited data compared with the supervised learning methods and can be extended to trace new-coming model origins. We construct extensive experiments to examine whether we can trace the origins of given texts. We provide valuable observations based on the experimental results, such as the difficulty level of AI origin tracing, and the AI origin similarities, and call for ethical concerns of LLM providers. We are releasing all codes and data as a toolkit and benchmark for future AI origin tracing and detecting studies. \footnote{We are releasing all available resource at \url{https://github.com/OpenLMLab/}.}