Modern robots vary significantly in shape, size, and sensor configurations used to perceive and interact with their environments. However, most navigation policies are embodiment-specific; a policy learned using one robot's configuration does not typically gracefully generalize to another. Even small changes in the body size or camera viewpoint may cause failures. With the recent surge in custom hardware developments, it is necessary to learn a single policy that can be transferred to other embodiments, eliminating the need to (re)train for each specific robot. In this paper, we introduce RING (Robotic Indoor Navigation Generalist), an embodiment-agnostic policy, trained solely in simulation with diverse randomly initialized embodiments at scale. Specifically, we augment the AI2-THOR simulator with the ability to instantiate robot embodiments with controllable configurations, varying across body size, rotation pivot point, and camera configurations. In the visual object-goal navigation task, RING achieves robust performance on real unseen robot platforms (Stretch RE-1, LoCoBot, Unitree's Go1), achieving an average of 72.1% and 78.9% success rate across 5 embodiments in simulation and 4 robot platforms in the real world. (project website: https://one-ring-policy.allen.ai/)

Data contamination hinders fair LLM evaluation by introducing test data into newer models' training sets. Existing studies solve this challenge by updating benchmarks with newly collected data. However, they fail to guarantee contamination-free evaluation as the newly collected data may contain pre-existing knowledge, and their benchmark updates rely on intensive human labor. To address these issues, we in this paper propose AntiLeak-Bench, an automated anti-leakage benchmarking framework. Instead of simply using newly collected data, we construct samples with explicitly new knowledge absent from LLMs' training sets, which thus ensures strictly contamination-free evaluation. We further design a fully automated workflow to build and update our benchmark without human labor. This significantly reduces the cost of benchmark maintenance to accommodate emerging LLMs. Through extensive experiments, we highlight that data contamination likely exists before LLMs' cutoff time and demonstrate AntiLeak-Bench effectively overcomes this challenge.

Causal discovery is essential for understanding complex systems, yet traditional methods often depend on strong, untestable assumptions, making the process challenging. Large Language Models (LLMs) present a promising alternative for extracting causal insights from text-based metadata, which consolidates domain expertise. However, LLMs are prone to unreliability and hallucinations, necessitating strategies that account for their limitations. One such strategy involves leveraging a consistency measure to evaluate reliability. Additionally, most text metadata does not clearly distinguish direct causal relationships from indirect ones, further complicating the inference of causal graphs. As a result, focusing on causal orderings, rather than causal graphs, emerges as a more practical and robust approach. We propose a novel method to derive a distribution of acyclic tournaments (representing plausible causal orders) that maximizes a consistency score. Our approach begins by computing pairwise consistency scores between variables, yielding a cyclic tournament that aggregates these scores. From this structure, we identify optimal acyclic tournaments compatible with the original tournament, prioritizing those that maximize consistency across all configurations. We tested our method on both classical and well-established bechmarks, as well as real-world datasets from epidemiology and public health. Our results demonstrate the effectiveness of our approach in recovering distributions causal orders with minimal error.

The reasoning abilities of Large Language Models (LLMs) are attracting increasing attention. In this work, we focus on causal reasoning and address the task of establishing causal relationships based on correlation information, a highly challenging problem on which several LLMs have shown poor performance. We introduce a prompting strategy for this problem that breaks the original task into fixed subquestions, with each subquestion corresponding to one step of a formal causal discovery algorithm, the PC algorithm. The proposed prompting strategy, PC-SubQ, guides the LLM to follow these algorithmic steps, by sequentially prompting it with one subquestion at a time, augmenting the next subquestion's prompt with the answer to the previous one(s). We evaluate our approach on an existing causal benchmark, Corr2Cause: our experiments indicate a performance improvement across five LLMs when comparing PC-SubQ to baseline prompting strategies. Results are robust to causal query perturbations, when modifying the variable names or paraphrasing the expressions.

Governments typically collect and steward a vast amount of high-quality data on their citizens and institutions, and the UK government is exploring how it can better publish and provision this data to the benefit of the AI landscape. However, the compositions of generative AI training corpora remain closely guarded secrets, making the planning of data sharing initiatives difficult. To address this, we devise two methods to assess UK government data usage for the training of Large Language Models (LLMs) and 'peek behind the curtain' in order to observe the UK government's current contributions as a data provider for AI. The first method, an ablation study that utilises LLM 'unlearning', seeks to examine the importance of the information held on UK government websites for LLMs and their performance in citizen query tasks. The second method, an information leakage study, seeks to ascertain whether LLMs are aware of the information held in the datasets published on the UK government's open data initiative data$.$gov$.$uk. Our findings indicate that UK government websites are important data sources for AI (heterogenously across subject matters) while data$.$gov$.$uk is not. This paper serves as a technical report, explaining in-depth the designs, mechanics, and limitations of the above experiments. It is accompanied by a complementary non-technical report on the ODI website in which we summarise the experiments and key findings, interpret them, and build a set of actionable recommendations for the UK government to take forward as it seeks to design AI policy. While we focus on UK open government data, we believe that the methods introduced in this paper present a reproducible approach to tackle the opaqueness of AI training corpora and provide organisations a framework to evaluate and maximize their contributions to AI development.

With the integration of large language models (LLMs), embodied agents have strong capabilities to execute complicated instructions in natural language, paving a way for the potential deployment of embodied robots. However, a foreseeable issue is that those embodied agents can also flawlessly execute some hazardous tasks, potentially causing damages in real world. To study this issue, we present SafeAgentBench -- a new benchmark for safety-aware task planning of embodied LLM agents. SafeAgentBench includes: (1) a new dataset with 750 tasks, covering 10 potential hazards and 3 task types; (2) SafeAgentEnv, a universal embodied environment with a low-level controller, supporting multi-agent execution with 17 high-level actions for 8 state-of-the-art baselines; and (3) reliable evaluation methods from both execution and semantic perspectives. Experimental results show that the best-performing baseline gets 69% success rate for safe tasks, but only 5% rejection rate for hazardous tasks, indicating significant safety risks. More details and codes are available at https://github.com/shengyin1224/SafeAgentBench.

Since their introduction LLM have gained widespread traction in different domains. They can be used as stand-alone products, but also to augment existing software products such as applications (also called agentic functions) or machine learning agents (also called LLM-based agents) to increase their capabilities. In this section, we show challenges during development and operation of LLM-based applications on three examples. Users interact with LLM-based applications by entering input into the LLM, the so-called prompt. Jang et al. showed in 2023 that the LLM's output is very sensitive to variations of the prompt [1]. Thus, the task of finding the best prompt to generate expected or best output leads to manual, trial-and-error-prompt experimenting - a method well known as prompt-engineering (cf. White et al. in 2023 for ChatGPT [2] or a survey of prompt techniques by Schulhoff et al. in 2024 [3]). Additionally, the outputs of an LLM-based application can not only vary, but also be wrong without telling a user ("hallucination", cf.

Natural Language Processing (NLP) of news articles can play an important role in understanding the dynamics and causes of violent conflict. Despite the availability of datasets categorizing various conflict events, the existing labels often do not cover all of the fine-grained violent conflict event types relevant to areas like the Horn of Africa. In this paper, we introduce a new benchmark dataset Conflict Events in the Horn of Africa region (CEHA) and propose a new task for identifying violent conflict events using online resources with this dataset. The dataset consists of 500 English event descriptions regarding conflict events in the Horn of Africa region with fine-grained event-type definitions that emphasize the cause of the conflict. This dataset categorizes the key types of conflict risk according to specific areas required by stakeholders in the Humanitarian-Peace-Development Nexus. Additionally, we conduct extensive experiments on two tasks supported by this dataset: Event-relevance Classification and Event-type Classification. Our baseline models demonstrate the challenging nature of these tasks and the usefulness of our dataset for model evaluations in low-resource settings with limited number of training data.

Current ophthalmology clinical workflows are plagued by over-referrals, long waits, and complex and heterogeneous medical records. Large language models (LLMs) present a promising solution to automate various procedures such as triaging, preliminary tests like visual acuity assessment, and report summaries. However, LLMs have demonstrated significantly varied performance across different languages in natural language question-answering tasks, potentially exacerbating healthcare disparities in Low and Middle-Income Countries (LMICs). This study introduces the first multilingual ophthalmological question-answering benchmark with manually curated questions parallel across languages, allowing for direct cross-lingual comparisons. Our evaluation of 6 popular LLMs across 7 different languages reveals substantial bias across different languages, highlighting risks for clinical deployment of LLMs in LMICs. Existing debiasing methods such as Translation Chain-of-Thought or Retrieval-augmented generation (RAG) by themselves fall short of closing this performance gap, often failing to improve performance across all languages and lacking specificity for the medical domain. To address this issue, We propose CLARA (Cross-Lingual Reflective Agentic system), a novel inference time de-biasing method leveraging retrieval augmented generation and self-verification. Our approach not only improves performance across all languages but also significantly reduces the multilingual bias gap, facilitating equitable LLM application across the globe.

We present DriveGPT, a scalable behavior model for autonomous driving. We model driving as a sequential decision making task, and learn a transformer model to predict future agent states as tokens in an autoregressive fashion. We scale up our model parameters and training data by multiple orders of magnitude, enabling us to explore the scaling properties in terms of dataset size, model parameters, and compute. We evaluate DriveGPT across different scales in a planning task, through both quantitative metrics and qualitative examples including closed-loop driving in complex real-world scenarios. In a separate prediction task, DriveGPT outperforms a state-of-the-art baseline and exhibits improved performance by pretraining on a large-scale dataset, further validating the benefits of data scaling.