In this study, we address the issue of enabling an artificial intelligence agent to execute complex language instructions within virtual environments. In our framework, we assume that these instructions involve intricate linguistic structures and multiple interdependent tasks that must be navigated successfully to achieve the desired outcomes. To effectively manage these complexities, we propose a hierarchical framework that combines the deep language comprehension of large language models with the adaptive action-execution capabilities of reinforcement learning agents. The language module (based on LLM) translates the language instruction into a high-level action plan, which is then executed by a pre-trained reinforcement learning agent. We have demonstrated the effectiveness of our approach in two different environments: in IGLU, where agents are instructed to build structures, and in Crafter, where agents perform tasks and interact with objects in the surrounding environment according to language commands.

Large Language Models (LLMs) have demonstrated their capabilities across various tasks, from language translation to complex reasoning. Understanding and predicting human behavior and biases are crucial for artificial intelligence (AI) assisted systems to provide useful assistance, yet it remains an open question whether these models can achieve this. This paper addresses this gap by leveraging the reasoning and generative capabilities of the LLMs to predict human behavior in two sequential decision-making tasks. These tasks involve balancing between exploitative and exploratory actions and handling delayed feedback, both essential for simulating real-life decision processes. We compare the performance of LLMs with a cognitive instance-based learning (IBL) model, which imitates human experiential decision-making. Our findings indicate that LLMs excel at rapidly incorporating feedback to enhance prediction accuracy. In contrast, the cognitive IBL model better accounts for human exploratory behaviors and effectively captures loss aversion bias, i.e., the tendency to choose a sub-optimal goal with fewer step-cost penalties rather than exploring to find the optimal choice, even with limited experience. The results highlight the benefits of integrating LLMs with cognitive architectures, suggesting that this synergy could enhance the modeling and understanding of complex human decision-making patterns.

Semantic role labeling (SRL) enriches many downstream applications, e.g., machine translation, question answering, summarization, and stance/belief detection. However, building multilingual SRL models is challenging due to the scarcity of semantically annotated corpora for multiple languages. Moreover, state-of-the-art SRL projection (XSRL) based on large language models (LLMs) yields output that is riddled with spurious role labels. Remediation of such hallucinations is not straightforward due to the lack of explainability of LLMs. We show that hallucinated role labels are related to naturally occurring divergence types that interfere with initial alignments. We implement Divergence-Aware Hallucination-Remediated SRL projection (DAHRS), leveraging linguistically-informed alignment remediation followed by greedy First-Come First-Assign (FCFA) SRL projection. DAHRS improves the accuracy of SRL projection without additional transformer-based machinery, beating XSRL in both human and automatic comparisons, and advancing beyond headwords to accommodate phrase-level SRL projection (e.g., EN-FR, EN-ES). Using CoNLL-2009 as our ground truth, we achieve a higher word-level F1 over XSRL: 87.6% vs. 77.3% (EN-FR) and 89.0% vs. 82.7% (EN-ES). Human phrase-level assessments yield 89.1% (EN-FR) and 91.0% (EN-ES). We also define a divergence metric to adapt our approach to other language pairs (e.g., English-Tagalog).

The paper presents a study of methods for extracting information about dialogue participants and evaluating their performance in Russian. To train models for this task, the Multi-Session Chat dataset was translated into Russian using multiple translation models, resulting in improved data quality. A metric based on the F-score concept is presented to evaluate the effectiveness of the extraction models. The metric uses a trained classifier to identify the dialogue participant to whom the persona belongs. Experiments were conducted on MBart, FRED-T5, Starling-7B, which is based on the Mistral, and Encoder2Encoder models. The results demonstrated that all models exhibited an insufficient level of recall in the persona extraction task. The incorporation of the NCE Loss improved the model's precision at the expense of its recall. Furthermore, increasing the model's size led to enhanced extraction of personas.

The emergence of Large Language Models (LLMs) has revolutionized how users access information, shifting from traditional search engines to direct question-and-answer interactions with LLMs. However, the widespread adoption of LLMs has revealed a significant challenge known as hallucination, wherein LLMs generate coherent yet factually inaccurate responses. This hallucination phenomenon has led to users' distrust in information retrieval systems based on LLMs. To tackle this challenge, this paper proposes Dynamic Retrieval Augmentation based on hallucination Detection (DRAD) as a novel method to detect and mitigate hallucinations in LLMs. DRAD improves upon traditional retrieval augmentation by dynamically adapting the retrieval process based on real-time hallucination detection. It features two main components: Real-time Hallucination Detection (RHD) for identifying potential hallucinations without external models, and Self-correction based on External Knowledge (SEK) for correcting these errors using external knowledge. Experiment results show that DRAD demonstrates superior performance in both detecting and mitigating hallucinations in LLMs. All of our code and data are open-sourced at https://github.com/oneal2000/EntityHallucination.

Discourse phenomena in existing document-level translation datasets are sparse, which has been a fundamental obstacle in the development of context-aware machine translation models. Moreover, most existing document-level corpora and context-aware machine translation methods rely on an unrealistic assumption on sentence-level alignments. To mitigate these issues, we first curate a novel dataset of Chinese-English literature, which consists of 160 books with intricate discourse structures. Then, we propose a more pragmatic and challenging setting for context-aware translation, termed chapter-to-chapter (Ch2Ch) translation, and investigate the performance of commonly-used machine translation models under this setting. Furthermore, we introduce a potential approach of finetuning large language models (LLMs) within the domain of Ch2Ch literary translation, yielding impressive improvements over baselines. Through our comprehensive analysis, we unveil that literary translation under the Ch2Ch setting is challenging in nature, with respect to both model learning methods and translation decoding algorithms.

Prompt engineering for large language models (LLMs) is often a manual time-intensive process that involves generating, evaluating, and refining prompts iteratively to ensure high-quality outputs. While there has been work on automating prompt engineering, the solutions generally are either tuned to specific tasks with given answers or are quite costly. We introduce GRAD-SUM, a scalable and flexible method for automatic prompt engineering that builds on gradient-based optimization techniques. Our approach incorporates user-defined task descriptions and evaluation criteria, and features a novel gradient summarization module to generalize feedback effectively. Our results demonstrate that GRAD-SUM consistently outperforms existing methods across various benchmarks, highlighting its versatility and effectiveness in automatic prompt optimization.

Transformer-based Large language models (LLMs) have demonstrated their power in various tasks, but their inference incurs significant time and energy costs. To accelerate LLM inference, speculative decoding uses a smaller model to propose one sequence of tokens, which are subsequently validated in batch by the target large model. Compared with autoregressive decoding, speculative decoding generates the same number of tokens with fewer runs of the large model, hence accelerating the overall inference by $1$-$2\times$. However, greedy decoding is not the optimal decoding algorithm in terms of output perplexity, which is a direct measurement of the effectiveness of a decoding algorithm. An algorithm that has better output perplexity and even better efficiency than speculative decoding can be more useful in practice. To achieve this seemingly contradictory goal, we first introduce multi-token joint greedy decoding (MJGD), which greedily generates multiple tokens at each step based on their joint perplexity. We show that it leads to better perplexity for the whole output. But the computation cost of MJGD is infeasible in practice. So we further propose multi-token joint speculative decoding (MJSD), which approximates and accelerates the MJGD from two aspects: it approximates the joint distribution of the large model with that of a small model, and uses a verification step to guarantee the accuracy of approximation; then it uses beam decoding to accelerate the sequence generation from the joint distribution. Compared with vanilla speculative decoding, MJSD has two advantages: (1) it is an approximation of MJGD, thus achieving better output perplexity; (2) verification with joint likelihood allows it to accept the longest prefix sub-sequence of the draft tokens with valid perplexity, leading to better efficiency...

When Large Language Models (LLMs) are compressed using techniques such as quantization, the predominant way to demonstrate the validity of such techniques is by measuring the model's accuracy on various benchmarks.If the accuracies of the baseline model and the compressed model are close, it is assumed that there was negligible degradation in quality.However, even when the accuracy of baseline and compressed model are similar, we observe the phenomenon of flips, wherein answers change from correct to incorrect and vice versa in proportion.We conduct a detailed study of metrics across multiple compression techniques, models and datasets, demonstrating that the behavior of compressed models as visible to end-users is often significantly different from the baseline model, even when accuracy is similar.We further evaluate compressed models qualitatively and quantitatively using MT-Bench and show that compressed models are significantly worse than baseline models in this free-form generative task.Thus, we argue that compression techniques should also be evaluated using distance metrics.We propose two such metrics, KL-Divergence and flips, and show that they are well correlated.

A key limitation of learned robot control policies is their inability to generalize outside their training data. Recent works on vision-language-action models (VLAs) have shown that the use of large, internet pre-trained vision-language models as the backbone of learned robot policies can substantially improve their robustness and generalization ability. Yet, one of the most exciting capabilities of large vision-language models in other domains is their ability to reason iteratively through complex problems. Can that same capability be brought into robotics to allow policies to improve performance by reasoning about a given task before acting? Naive use of "chain-of-thought" (CoT) style prompting is significantly less effective with standard VLAs because of the relatively simple training examples that are available to them. Additionally, purely semantic reasoning about sub-tasks, as is common in regular CoT, is insufficient for robot policies that need to ground their reasoning in sensory observations and the robot state. To this end, we introduce Embodied Chain-of-Thought Reasoning (ECoT) for VLAs, in which we train VLAs to perform multiple steps of reasoning about plans, sub-tasks, motions, and visually grounded features like object bounding boxes and end effector positions, before predicting the robot action. We design a scalable pipeline for generating synthetic training data for ECoT on large robot datasets. We demonstrate, that ECoT increases the absolute success rate of OpenVLA, the current strongest open-source VLA policy, by 28% across challenging generalization tasks, without any additional robot training data. Additionally, ECoT makes it easier for humans to interpret a policy's failures and correct its behavior using natural language.