Transformers [Vaswani et al., 2017] are versatile models used in various applications, including vision [Yuan et al., 2021, Khan et al., 2022, Dehghani et al., 2023] and natural language processing [Wei et al., 2022b, Touvron et al., 2023]. Their effectiveness in complex tasks is particularly notable in Large Language Models (LLMs) [Wang et al., 2018, Hendrycks et al., 2021], where they excel at generating coherent text and understanding context. This strong performance has led to an increased interest in understanding the Transformer architecture as a computational model capable of executing instructions and solving algorithmic reasoning problems. In this context, Pérez et al. [2021], Wei et al. [2022a] show that Transformers are Turing Complete, and Giannou et al. [2023], Back De Luca and Fountoulakis [2024], Yang et al. [2024] demonstrate that Transformers can effectively encode instructions to solve linear algebra and graphs problems. Additionally, it has been shown that Transformers can perform reasoning tasks using far fewer layers than the number of reasoning steps [Liu et al., 2023], indicating a connection between Transformers and parallel algorithms. To this end, Sanford et al. [2024] further demonstrates that Transformers can simulate the Massively Parallel Computation (MPC) model [Andoni et al., 2018], which is based on the MapReduce framework for large-scale data processing [Dean and Ghemawat, 2008]. Complementing this theoretical framework, empirical studies have demonstrated the capabilities of Transformers, among other models, in executing algorithms [Veličković and Blundell, 2021]. Notable applications include basic arithmetic [Lee et al., 2024], sorting [Tay et al., 2020, Yan et al., 2020], dynamic programming

Long-Form Question Answering (LFQA) refers to generating in-depth, paragraph-level responses to open-ended questions. Although lots of LFQA methods are developed, evaluating LFQA effectively and efficiently remains challenging due to its high complexity and cost. Therefore, there is no standard benchmark for LFQA evaluation till now. To address this gap, we make the first attempt by proposing a well-constructed, reference-based benchmark named Chinese exAmination for LFQA Evaluation (CALF), aiming to rigorously assess the performance of automatic evaluation metrics for LFQA. The CALF benchmark is derived from Chinese examination questions that have been translated into English. It includes up to 1476 examples consisting of knowledge-intensive and nuanced responses. Our evaluation comprises three different settings to ana lyze the behavior of automatic metrics comprehensively. We conducted extensive experiments on 7 traditional evaluation metrics, 3 prompt-based metrics, and 3 trained evaluation metrics, and tested on agent systems for the LFQA evaluation. The results reveal that none of the current automatic evaluation metrics shows comparable performances with humans, indicating that they cannot capture dense information contained in long-form responses well. In addition, we provide a detailed analysis of the reasons why automatic evaluation metrics fail when evaluating LFQA, offering valuable insights to advance LFQA evaluation systems. Dataset and associated codes can be accessed at our GitHub repository.

Deductive coding is a common discourse analysis method widely used by learning science and learning analytics researchers for understanding teaching and learning interactions. It often requires researchers to manually label all discourses to be analyzed according to a theoretically guided coding scheme, which is time-consuming and labor-intensive. The emergence of large language models such as GPT has opened a new avenue for automatic deductive coding to overcome the limitations of traditional deductive coding. To evaluate the usefulness of large language models in automatic deductive coding, we employed three different classification methods driven by different artificial intelligence technologies, including the traditional text classification method with text feature engineering, BERT-like pretrained language model and GPT-like pretrained large language model (LLM). We applied these methods to two different datasets and explored the potential of GPT and prompt engineering in automatic deductive coding. By analyzing and comparing the accuracy and Kappa values of these three classification methods, we found that GPT with prompt engineering outperformed the other two methods on both datasets with limited number of training samples. By providing detailed prompt structures, the reported work demonstrated how large language models can be used in the implementation of automatic deductive coding.

In-context learning, the ability to adapt based on a few examples in the input prompt, is a ubiquitous feature of large language models (LLMs). However, as LLMs' in-context learning abilities continue to improve, understanding this phenomenon mechanistically becomes increasingly important. In particular, it is not well-understood how LLMs learn to solve specific classes of problems, such as reinforcement learning (RL) problems, in-context. Through three different tasks, we first show that Llama $3$ $70$B can solve simple RL problems in-context. We then analyze the residual stream of Llama using Sparse Autoencoders (SAEs) and find representations that closely match temporal difference (TD) errors. Notably, these representations emerge despite the model only being trained to predict the next token. We verify that these representations are indeed causally involved in the computation of TD errors and $Q$-values by performing carefully designed interventions on them. Taken together, our work establishes a methodology for studying and manipulating in-context learning with SAEs, paving the way for a more mechanistic understanding.

Medical imaging is crucial for diagnosing a patient's health condition, and accurate segmentation of these images is essential for isolating regions of interest to ensure precise diagnosis and treatment planning. Existing methods primarily rely on bounding boxes or point-based prompts, while few have explored text-related prompts, despite clinicians often describing their observations and instructions in natural language. To address this gap, we first propose a RAG-based free-form text prompt generator, that leverages the domain corpus to generate diverse and realistic descriptions. Then, we introduce FLanS, a novel medical image segmentation model that handles various free-form text prompts, including professional anatomy-informed queries, anatomy-agnostic position-driven queries, and anatomy-agnostic size-driven queries. Additionally, our model also incorporates a symmetry-aware canonicalization module to ensure consistent, accurate segmentations across varying scan orientations and reduce confusion between the anatomical position of an organ and its appearance in the scan. FLanS is trained on a large-scale dataset of over 100k medical images from 7 public datasets. Comprehensive experiments demonstrate the model's superior language understanding and segmentation precision, along with a deep comprehension of the relationship between them, outperforming SOTA baselines on both in-domain and out-of-domain datasets.

This paper investigates the challenges associated with bias, toxicity, unreliability, and lack of robustness in large language models (LLMs) such as ChatGPT. It emphasizes that these issues primarily stem from the quality and diversity of data on which LLMs are trained, rather than the model architectures themselves. As LLMs are increasingly integrated into various real-world applications, their potential to negatively impact society by amplifying existing biases and generating harmful content becomes a pressing concern. The paper calls for interdisciplinary efforts to address these challenges. Additionally, it highlights the need for collaboration between researchers, practitioners, and stakeholders to establish governance frameworks, oversight, and accountability mechanisms to mitigate the harmful consequences of biased LLMs.

Recent advancements in Large Language Models (LLMs) have spurred numerous attempts to apply these technologies to embodied tasks, particularly focusing on high-level task planning and task decomposition. To further explore this area, we introduce a new embodied task planning benchmark, ET-Plan-Bench, which specifically targets embodied task planning using LLMs. It features a controllable and diverse set of embodied tasks varying in different levels of difficulties and complexities, and is designed to evaluate two critical dimensions of LLMs' application in embodied task understanding: spatial (relation constraint, occlusion for target objects) and temporal & causal understanding of the sequence of actions in the environment. By using multi-source simulators as the backend simulator, it can provide immediate environment feedback to LLMs, which enables LLMs to interact dynamically with the environment and re-plan as necessary. We evaluated the state-of-the-art open source and closed source foundation models, including GPT-4, LLAMA and Mistral on our proposed benchmark. While they perform adequately well on simple navigation tasks, their performance can significantly deteriorate when faced with tasks that require a deeper understanding of spatial, temporal, and causal relationships. Thus, our benchmark distinguishes itself as a large-scale, quantifiable, highly automated, and fine-grained diagnostic framework that presents a significant challenge to the latest foundation models. We hope it can spark and drive further research in embodied task planning using foundation models.

Legal intake, the process of finding out if an applicant is eligible for help from a free legal aid program, takes significant time and resources. In part this is because eligibility criteria are nuanced, open-textured, and require frequent revision as grants start and end. In this paper, we investigate the use of large language models (LLMs) to reduce this burden. We describe a digital intake platform that combines logical rules with LLMs to offer eligibility recommendations, and we evaluate the ability of 8 different LLMs to perform this task. We find promising results for this approach to help close the access to justice gap, with the best model reaching an F1 score of .82, while minimizing false negatives.

Reinforcement Learning from Human Feedback (RLHF) has greatly improved the performance of modern Large Language Models (LLMs). The RLHF process is resource-intensive and technically challenging, generally requiring a large collection of human preference labels over model-generated outputs. Reinforcement Learning from AI Feedback (RLAIF) addresses this data collection challenge by leveraging synthetic preferences generated by an LLM. However, recent work has shown that synthetic preferences labels may not align well with human preference judgments (Zeng et al., 2023). To address this, we propose a hybrid approach that unifies RLHF and RLAIF methodologies. We introduce GenRM, an iterative algorithm that trains an LLM on self-generated reasoning traces, leading to synthetic preference labels matching human preference judgments. Empirically, we show that zero-shot LLM-based judgments under-perform compared to Bradley-Terry reward models on in-distribution tasks (between 9-36%). In contrast, GenRM achieves in-distribution accuracy comparable to Bradley-Terry models, while significantly outperforming them on out-of-distribution tasks (between 10-45%). Our results show that combining the strengths of RLHF and RLAIF offers a promising approach for improving the quality of synthetic preference labels. Reinforcement Learning from Human Feedback (RLHF) has significantly improved the performance of modern Large Language Models (LLMs) (see e.g., Reid et al., 2024; OpenAI, 2023). Despite its effectiveness, the RLHF process presents several challenges.

The rapid proliferation of frontier model agents promises significant societal advances but also raises concerns about systemic risks arising from unsafe interactions. Collusion to the disadvantage of others has been identified as a central form of undesirable agent cooperation. The use of information hiding (steganography) in agent communications could render collusion practically undetectable. This underscores the need for evaluation frameworks to monitor and mitigate steganographic collusion capabilities. We address a crucial gap in the literature by demonstrating, for the first time, that robust steganographic collusion in LLMs can arise indirectly from optimization pressure. To investigate this problem we design two approaches -- a gradient-based reinforcement learning (GBRL) method and an in-context reinforcement learning (ICRL) method -- for reliably eliciting sophisticated LLM-generated linguistic text steganography. Importantly, we find that emergent steganographic collusion can be robust to both passive steganalytic oversight of model outputs and active mitigation through communication paraphrasing. We contribute a novel model evaluation framework and discuss limitations and future work. Our findings imply that effective risk mitigation from steganographic collusion post-deployment requires innovation in passive and active oversight techniques.