The Abstraction and Reasoning Corpus (ARC) is a popular benchmark focused on visual reasoning in the evaluation of Artificial Intelligence systems. In its original framing, an ARC task requires solving a program synthesis problem over small 2D images using a few input-output training pairs. In this work, we adopt the recently popular data-driven approach to the ARC and ask whether a Vision Transformer (ViT) can learn the implicit mapping, from input image to output image, that underlies the task. We show that a ViT -- otherwise a state-of-the-art model for images -- fails dramatically on most ARC tasks even when trained on one million examples per task. This points to an inherent representational deficiency of the ViT architecture that makes it incapable of uncovering the simple structured mappings underlying the ARC tasks. Building on these insights, we propose ViTARC, a ViT-style architecture that unlocks some of the visual reasoning capabilities required by the ARC. Specifically, we use a pixel-level input representation, design a spatially-aware tokenization scheme, and introduce a novel object-based positional encoding that leverages automatic segmentation, among other enhancements. Our task-specific ViTARC models achieve a test solve rate close to 100% on more than half of the 400 public ARC tasks strictly through supervised learning from input-output grids. This calls attention to the importance of imbuing the powerful (Vision) Transformer with the correct inductive biases for abstract visual reasoning that are critical even when the training data is plentiful and the mapping is noise-free. Hence, ViTARC provides a strong foundation for future research in visual reasoning using transformer-based architectures.

Language models are now capable of solving tasks that require dealing with long sequences consisting of hundreds of thousands of tokens. However, they often fail on tasks that require repetitive use of simple rules, even on sequences that are much shorter than those seen during training. For example, state-of-the-art LLMs can find common items in two lists with up to 20 items but fail when lists have 80 items. In this paper, we introduce MLissard, a multilingual benchmark designed to evaluate models' abilities to process and generate texts of varied lengths and offers a mechanism for controlling sequence complexity. Our evaluation of open-source and proprietary models show a consistent decline in performance across all models and languages as the complexity of the sequence increases. Surprisingly, the use of in-context examples in languages other than English helps increase extrapolation performance significantly. The datasets and code are available at https://github.com/unicamp-dl/Lissard

Chemputation is the process of programming chemical robots to do experiments using a universal symbolic language, but the literature can be error prone and hard to read due to ambiguities. Large Language Models (LLMs) have demonstrated remarkable capabilities in various domains, including natural language processing, robotic control, and more recently, chemistry. Despite significant advancements in standardizing the reporting and collection of synthetic chemistry data, the automatic reproduction of reported syntheses remains a labour-intensive task. In this work, we introduce an LLM-based chemical research agent workflow designed for the automatic validation of synthetic literature procedures. Our workflow can autonomously extract synthetic procedures and analytical data from extensive documents, translate these procedures into universal XDL code, simulate the execution of the procedure in a hardware-specific setup, and ultimately execute the procedure on an XDL-controlled robotic system for synthetic chemistry. This demonstrates the potential of LLM-based workflows for autonomous chemical synthesis with Chemputers. Due to the abstraction of XDL this approach is safe, secure, and scalable since hallucinations will not be chemputable and the XDL can be both verified and encrypted. Unlike previous efforts, which either addressed only a limited portion of the workflow, relied on inflexible hard-coded rules, or lacked validation in physical systems, our approach provides four realistic examples of syntheses directly executed from synthetic literature. We anticipate that our workflow will significantly enhance automation in robotically driven synthetic chemistry research, streamline data extraction, improve the reproducibility, scalability, and safety of synthetic and experimental chemistry.

Compressive adaptation approaches, such as QLoRA, are widely popular alternatives for reducing memory requirements during fine-tuning of large language models (LLMs) while producing models capable of handling various downstream tasks. The key idea is to employ a "two-tower" architecture: compressing pretrained LLM parameters into compact representations and fine-tuning the additive full-precision adapter, which typically has few tunable parameters in low-rank format. However, the strict algebraic assumptions, such as low-rank assumption, and the complexity of composing two-tower architectures are some of the known shortcomings, resulting in a poor accuracy-efficiency trade-off. In response to these known limitations, we propose SpaLLM (Sketched Parameter Adaptation of LLMs), a novel compressive adaptation approach for LLMs. This method is also the first to illustrate parameter-sharing compression methods for LLM finetuning, which, unlike QLoRA, are free from strict low-rank algebraic assumptions on adapters. This approach simplifies LLMs' compressive adaptation workflow, potentially improves multi-user serving efficiency, and delivers significantly better accuracy for both natural language understanding and generation tasks. Moreover, by avoiding the "two-tower" architecture, our framework only requires one compressed matrix multiplication per layer during inference, demonstrating superior inference efficiency compared to previous methods. Recent advancements in Large Language Models (LLMs) have demonstrated exceptional performance in Natural Language Processing (NLP), enabling a broad spectrum of downstream applications. LLMs have demonstrated impressive generalization abilities across many downstream tasks in a zero-shot manner. However, compared to training-free methods such as in-context learning (Dong et al., 2022; Rubin et al., 2021) and few-shot prompting (Brown, 2020; Song et al., 2023), fine-tuning on these LLMs is often the ideal method to achieve optimal performance on a specific downstream task (Ding et al., 2023). Clearly, full-precision fine-tuning on these LLMs are often impractical due to the massive requirement of high-performance computing devices such as GPUs. As a result, Parameter-Efficient Fine-Tuning methods (PEFT), such as Low-Rank Adaptation (LoRA) (Hu et al., 2022), emerged as a less resource-intensive approach to fine-tuning while achieving reasonable Clearly, there is a trade-off between accuracy and efficiency.

This paper investigates whether large language models (LLMs) are state-of-the-art quality estimators for machine translation of user-generated content (UGC) that contains emotional expressions, without the use of reference translations. To achieve this, we employ an existing emotion-related dataset with human-annotated errors and calculate quality evaluation scores based on the Multi-dimensional Quality Metrics. We compare the accuracy of several LLMs with that of our fine-tuned baseline models, under in-context learning and parameter-efficient fine-tuning (PEFT) scenarios. We find that PEFT of LLMs leads to better performance in score prediction with human interpretable explanations than fine-tuned models. However, a manual analysis of LLM outputs reveals that they still have problems such as refusal to reply to a prompt and unstable output while evaluating machine translation of UGC.

Reinforcement Learning from Human Feedback (RLHF) has emerged as a pivotal tool for aligning large language models (LLMs) with human preferences. Direct Preference Optimization (DPO), one of the most popular approaches, formulates RLHF as a policy optimization problem without explicitly estimating the reward function. It overcomes the stability and efficiency issues of two-step approaches, which typically involve first estimating the reward function and then optimizing the policy via proximal policy optimization (PPO). Since RLHF is essentially an optimization problem, and it is well-known that momentum techniques can accelerate optimization both theoretically and empirically, a natural question arises: Can RLHF be accelerated by momentum? This paper answers this question in the affirmative. In detail, we first show that the iterative preference optimization method can be viewed as a proximal point method. Based on this observation, we propose a general Accelerated Preference Optimization (APO) framework, which unifies many existing preference optimization algorithms and employs Nesterov's momentum technique to speed up the alignment of LLMs. Theoretically, we demonstrate that APO can achieve a faster convergence rate than the standard iterative preference optimization methods, including DPO and Self-Play Preference Optimization (SPPO). Empirically, we show the superiority of APO over DPO, iterative DPO, and other strong baselines for RLHF on the AlpacaEval 2.0 benchmark.

We introduce a new vulnerability that exploits fixed points in autoregressive models and use it to craft queries that never halt, i.e. an LLM output that does not terminate. More precisely, for what we call non-halting queries, the LLM never samples the end-of-string token (). We rigorously analyze the conditions under which the non-halting anomaly presents itself. In particular, at temperature zero, we prove that if a repeating (cyclic) sequence of tokens is observed at the output beyond the context size, then the LLM does not halt. We demonstrate the non-halting anomaly in a number of experiments performed in base (unaligned) models where repeating tokens immediately lead to a non-halting cyclic behavior as predicted by the analysis. Further, we develop a simple recipe that takes the same fixed points observed in the base model and creates a prompt structure to target aligned models. We study the recipe behavior in bypassing alignment in a number of LLMs including GPT-4o, llama-3-8b-instruct, and gemma-2-9b-it where all models are forced into a non-halting state. Further, we demonstrate the recipe's success in sending most major models released over the past year into a non-halting state with the same simple prompt even at higher temperatures. Further, we study direct inversion based techniques to craft new short prompts to induce the non-halting state. Our experiments with the gradient search based inversion technique ARCA show that non-halting is prevalent across models and may be easily induced with a few input tokens. While its impact on the reliability of hosted systems can be mitigated by configuring a hard maximum token limit in the sampler, the non-halting anomaly still manages to break alignment. This underlines the need for further studies and stronger forms of alignment against non-halting anomalies.

Accommodating human preferences is essential for creating AI agents that deliver personalized and effective interactions. Recent work has shown the potential for LLMs to infer preferences from user interactions, but they often produce broad and generic preferences, failing to capture the unique and individualized nature of human preferences. This paper introduces PREDICT, a method designed to enhance the precision and adaptability of inferring preferences. PREDICT incorporates three key elements: (1) iterative refinement of inferred preferences, (2) decomposition of preferences into constituent components, and (3) validation of preferences across multiple trajectories. We evaluate PREDICT on two distinct environments: a gridworld setting and a new text-domain environment (PLUME).

While LLMs have emerged as performant architectures for reasoning tasks, their compositional generalization capabilities have been questioned. In this work, we introduce a Compositional Generalization Challenge for Graph-based Commonsense Reasoning (CGGC) that goes beyond previous evaluations that are based on sequences or tree structures - and instead involves a reasoning graph: It requires models to generate a natural sentence based on given concepts and a corresponding reasoning graph, where the presented graph involves a previously unseen combination of relation types. To master this challenge, models need to learn how to reason over relation tupels within the graph, and how to compose them when conceptualizing a verbalization. We evaluate seven well-known LLMs using in-context learning and find that performant LLMs still struggle in compositional generalization. We investigate potential causes of this gap by analyzing the structures of reasoning graphs, and find that different structures present varying levels of difficulty for compositional generalization. Arranging the order of demonstrations according to the structures' difficulty shows that organizing samples in an easy-to-hard schema enhances the compositional generalization ability of LLMs.

With the success of ChatGPT and other similarly sized SotA LLMs, claims of emergent human like social reasoning capabilities, especially Theory of Mind (ToM), in these models have appeared in the scientific literature. On the one hand those ToM-capabilities have been successfully tested using tasks styled similar to those used in psychology (Kosinski, 2023). On the other hand, follow up studies showed that those capabilities vanished when the tasks were slightly altered (Ullman, 2023). In this work we introduce a novel dataset of 68 tasks for probing ToM in LLMs, including potentially challenging variations which are assigned to 10 complexity classes. This way it is providing novel insights into the challenges LLMs face with those task variations. We evaluate the ToM performance of four SotA open source LLMs on our dataset and the dataset introduced by (Kosinski, 2023). The overall low goal accuracy across all evaluated models indicates only a limited degree of ToM capabilities. The LLMs' performance on simple complexity class tasks from both datasets are similar. Whereas we find a consistent tendency in all tested LLMs to perform poorly on tasks that require the realization that an agent has knowledge of automatic state changes in its environment, even when those are spelled out to the model. For task complications that change the relationship between objects by replacing prepositions, we notice a performance drop in all models, with the strongest impact on the mixture-of-experts model. With our dataset of tasks grouped by complexity we offer directions for further research on how to stabilize and advance ToM capabilities in LLM.