Reinforcement learning has substantially improved large language model (LLM) agents in verifiable domains, but remains difficult to apply to open-ended agent tasks, where solutions are diverse and reliable scalar rewards are hard to obtain. Recent pairwise evaluation methods alleviate reward discrimination collapse by replacing pointwise scoring with relative preferences. However, they still compress rich comparative feedback into a single trajectory-level reward, obscuring decisive intermediate steps and preventing successful behaviors from being consolidated into reusable skills. We propose ArenaFlow, a hierarchical credit propagation framework for open-ended agent reinforcement learning. ArenaFlow leverages tournament-based relative ranking to derive trajectory-level reward signals. Each comparison is further equipped with structured reflective evaluation, which reveals three types of supervision: pivotal success steps, reusable strategy skills, and usage attribution of retrieved skills. At the step level, ArenaFlow propagates trajectory-level advantages to high-confidence pivotal steps according to tournament survival depth, enabling more targeted optimization of local reasoning behaviors. At the skill level, ArenaFlow estimates skill utility from group-level usage attribution and maintains a global skill memory through utility-aware updating, pruning, and retrieval. The resulting high-utility skills further serve as policy priors for future exploration. Extensive experiments validate ArenaFlow’s effectiveness on open-ended agent tasks.

Large Language Models (LLMs) continue to grow, reaching hundreds of billions of parameters and making it challenging for Deep Learning practitioners with resource-constrained systems to use them, e.g., fine-tuning these models for a downstream task of their interest. Adapters, such as low-rank adapters (LoRA), have been proposed to reduce the number of trainable parameters in a model, reducing memory requirements and enabling smaller systems to fine-tune these models. Orthogonal to this work, Neural Architecture Search (NAS) has been used to discover compressed and more efficient architectures without sacrificing performance compared to similar base models. This paper introduces a novel approach, LoNAS, to use NAS on language models by exploring a search space of elastic low-rank adapters while reducing memory and compute requirements of full-scale NAS, resulting in high-performing compressed models obtained from weight-sharing super-networks. Compared to models fine-tuned with LoRA, these models contain fewer total parameters, reducing the inference time with only minor decreases in accuracy and, in some cases, even improving accuracy. We discuss the limitations of LoNAS and share observations for the research community regarding its generalization capabilities, which have motivated our follow-up work.

Transformer-based models have demonstrated outstanding performance in natural language processing (NLP) tasks and many other domains, e.g., computer vision. Depending on the size of these models, which have grown exponentially in the past few years, machine learning practitioners might be restricted from deploying them in resource-constrained environments. This paper discusses the compression of transformer-based models for multiple resource budgets. Integrating neural architecture search (NAS) and network pruning techniques, we effectively generate and train weight-sharing super-networks that contain efficient, high-performing, and compressed transformer-based models. A common challenge in NAS is the design of the search space, for which we propose a method to automatically obtain the boundaries of the search space and then derive the rest of the intermediate possible architectures using a first-order weight importance technique. The proposed end-to-end NAS solution, EFTNAS, discovers efficient subnetworks that have been compressed and fine-tuned for downstream NLP tasks. We demonstrate EFTNAS on the General Language Understanding Evaluation (GLUE) benchmark and the Stanford Question Answering Dataset (SQuAD), obtaining high-performing smaller models with a reduction of more than 5x in size without or with little degradation in performance.

Counterspeech has been proposed as a solution to the proliferation of online hate. Research has shown that natural language processing (NLP) approaches could generate such counterspeech automatically, but there are competing ideas for how NLP models might be used for this task and a variety of evaluation metrics whose relationship to one another is unclear. We test three different approaches and collect ratings of the generated counterspeech for 1,740 tweet-participant pairs to systematically compare the counterspeech on three aspects: quality, effectiveness and user preferences. We examine which model performs best at which metric and which aspects of counterspeech predict user preferences. A free-form text generation approach using ChatGPT performs the most consistently well, though its generations are occasionally unspecific and repetitive. In our experiment, participants’ preferences for counterspeech are predicted by the quality of the counterspeech, not its perceived effectiveness. The results can help future research approach counterspeech evaluation more systematically.

Causal chain reasoning (CCR) is an essential ability for many decision-making AI systems, which requires the model to build reliable causal chains by connecting causal pairs. However, CCR suffers from two main transitive problems: threshold effect and scene drift. In other words, the causal pairs to be spliced may have a conflicting threshold boundary or scenario.To address these issues, we propose a novel Reliable Causal chain reasoning framework (ReCo), which introduces exogenous variables to represent the threshold and scene factors of each causal pair within the causal chain, and estimates the threshold and scene contradictions across exogenous variables via structural causal recurrent neural networks (SRNN). Experiments show that ReCo outperforms a series of strong baselines on both Chinese and English CCR datasets. Moreover, by injecting reliable causal chain knowledge distilled by ReCo, BERT can achieve better performances on four downstream causal-related tasks than BERT models enhanced by other kinds of knowledge.

We introduce the task of historical text summarisation, where documents in historical forms of a language are summarised in the corresponding modern language. This is a fundamentally important routine to historians and digital humanities researchers but has never been automated. We compile a high-quality gold-standard text summarisation dataset, which consists of historical German and Chinese news from hundreds of years ago summarised in modern German or Chinese. Based on cross-lingual transfer learning techniques, we propose a summarisation model that can be trained even with no cross-lingual (historical to modern) parallel data, and further benchmark it against state-of-the-art algorithms. We report automatic and human evaluations that distinguish the historic to modern language summarisation task from standard cross-lingual summarisation (i.e., modern to modern language), highlight the distinctness and value of our dataset, and demonstrate that our transfer learning approach outperforms standard cross-lingual benchmarks on this task.