Transferring policies learned in simulation to the real world is a promising strategy for acquiring robot skills at scale. However, sim-to-real approaches typically rely on manual design and tuning of the task reward function as well as the simulation physics parameters, rendering the process slow and human-labor intensive. In this paper, we investigate using Large Language Models (LLMs) to automate and accelerate sim-to-real design. Our LLM-guided sim-to-real approach, DrEureka, requires only the physics simulation for the target task and automatically constructs suitable reward functions and domain randomization distributions to support real-world transfer. We first demonstrate that our approach can discover sim-to-real configurations that are competitive with existing human-designed ones on quadruped locomotion and dexterous manipulation tasks. Then, we showcase that our approach is capable of solving novel robot tasks, such as quadruped balancing and walking atop a yoga ball, without iterative manual design.

Reinforcement learning from human feedback (RLHF) is a prevalent approach to align AI systems with human values by learning rewards from human preference data. Due to various reasons, however, such data typically takes the form of rankings over pairs of trajectory segments, which fails to capture the varying strengths of preferences across different pairs. In this paper, we propose a novel adaptive preference loss, underpinned by distributionally robust optimization (DRO), designed to address this uncertainty in preference strength. By incorporating an adaptive scaling parameter into the loss for each pair, our method increases the flexibility of the reward function. Specifically, it assigns small scaling parameters to pairs with ambiguous preferences, leading to more comparable rewards, and large scaling parameters to those with clear preferences for more distinct rewards. Computationally, our proposed loss function is strictly convex and univariate with respect to each scaling parameter, enabling its efficient optimization through a simple second-order algorithm. Our method is versatile and can be readily adapted to various preference optimization frameworks, including direct preference optimization (DPO). Our experiments with robotic control and natural language generation with large language models (LLMs) show that our method not only improves policy performance but also aligns reward function selection more closely with policy optimization, simplifying the hyperparameter tuning process.

The ability (and inability) of large language models (LLMs) to perform arithmetic tasks has been the subject of much theoretical and practical debate. We show that LLMs are frequently able to correctly and confidently predict the first digit of n-digit by m-digit multiplication tasks without using chain of thought reasoning, despite these tasks require compounding operations to solve. Simultaneously, LLMs in practice often fail to correctly or confidently predict the last digit of an n-digit by m-digit multiplication, a task equivalent to 1-digit by 1-digit multiplication which can be easily learned or memorized. We show that the latter task can be solved more robustly when the LLM is conditioned on all of the correct higher-order digits, which on average increases the confidence of the correct last digit on 5-digit by 5-digit multiplication tasks using Llama 2-13B by over 230% (0.13 to 0.43) and Mistral-7B by 150% (0.22 to 0.55).

Vehicle motion planning is an essential component of autonomous driving technology. Current rule-based vehicle motion planning methods perform satisfactorily in common scenarios but struggle to generalize to long-tailed situations. Meanwhile, learning-based methods have yet to achieve superior performance over rule-based approaches in large-scale closed-loop scenarios. To address these issues, we propose PlanAgent, the first mid-to-mid planning system based on a Multi-modal Large Language Model (MLLM). MLLM is used as a cognitive agent to introduce human-like knowledge, interpretability, and common-sense reasoning into the closed-loop planning. Specifically, PlanAgent leverages the power of MLLM through three core modules. First, an Environment Transformation module constructs a Bird's Eye View (BEV) map and a lane-graph-based textual description from the environment as inputs. Second, a Reasoning Engine module introduces a hierarchical chain-of-thought from scene understanding to lateral and longitudinal motion instructions, culminating in planner code generation. Last, a Reflection module is integrated to simulate and evaluate the generated planner for reducing MLLM's uncertainty. PlanAgent is endowed with the common-sense reasoning and generalization capability of MLLM, which empowers it to effectively tackle both common and complex long-tailed scenarios. Our proposed PlanAgent is evaluated on the large-scale and challenging nuPlan benchmarks. A comprehensive set of experiments convincingly demonstrates that PlanAgent outperforms the existing state-of-the-art in the closed-loop motion planning task. Codes will be soon released.

Learning multi-task models for jointly detecting stance and verifying rumors poses challenges due to the need for training data of stance at post level and rumor veracity at claim level, which are difficult to obtain. To address this issue, we leverage large language models (LLMs) as the foundation annotators for the joint stance detection (SD) and rumor verification (RV) tasks, dubbed as JSDRV. We introduce a novel reinforcement tuning framework to enhance the joint predictive capabilities of LLM-based SD and RV components. Specifically, we devise a policy for selecting LLM-annotated data at the two levels, employing a hybrid reward mechanism to choose high-quality labels for effective LLM fine-tuning on both tasks. Results demonstrate that JSDRV improves the capabilities of LLMs in the joint tasks, not only outperforming state-of-the-art methods but also generalizing to non-LLMs accommodated as task models.

In-context learning (ICL) achieves remarkable performance in various domains such as knowledge acquisition, commonsense reasoning, and semantic understanding. However, its performance significantly deteriorates for emotion detection tasks, especially fine-grained emotion recognition. The underlying reasons for this remain unclear. In this paper, we identify the reasons behind ICL's poor performance from the perspective of prototype theory and propose a method to address this issue. Specifically, we conduct extensive pilot experiments and find that ICL conforms to the prototype theory on fine-grained emotion recognition. Based on this theory, we uncover the following deficiencies in ICL: (1) It relies on prototypes (example-label pairs) that are semantically similar but emotionally inaccurate to predict emotions. (2) It is prone to interference from irrelevant categories, affecting the accuracy and robustness of the predictions. To address these issues, we propose an Emotion Context Learning method (E-ICL) on fine-grained emotion recognition. E-ICL relies on more emotionally accurate prototypes to predict categories by referring to emotionally similar examples with dynamic labels. Simultaneously, E-ICL employs an exclusionary emotion prediction strategy to avoid interference from irrelevant categories, thereby increasing its accuracy and robustness. Note that the entire process is accomplished with the assistance of a plug-and-play emotion auxiliary model, without additional training. Experiments on the fine-grained emotion datasets EDOS, Empathetic-Dialogues, EmpatheticIntent, and GoEmotions show that E-ICL achieves superior emotion prediction performance. Furthermore, even when the emotion auxiliary model used is lower than 10% of the LLMs, E-ICL can still boost the performance of LLMs by over 4% on multiple datasets.

Interpretability and explainability of AI are becoming increasingly important in light of the rapid development of large language models (LLMs). This paper investigates the interpretation of LLMs in the context of the knowledge-based question answering. The main hypothesis of the study is that correct and incorrect model behavior can be distinguished at the level of hidden states. The quantized models LLaMA-2-7B-Chat, Mistral-7B, Vicuna-7B and the MuSeRC question-answering dataset are used to test this hypothesis. The results of the analysis support the proposed hypothesis. We also identify the layers which have a negative effect on the model's behavior. As a prospect of practical application of the hypothesis, we propose to train such "weak" layers additionally in order to improve the quality of the task solution.

We explore uncertainty quantification in large language models (LLMs), with the goal to identify when uncertainty in responses given a query is large. We simultaneously consider both epistemic and aleatoric uncertainties, where the former comes from the lack of knowledge about the ground truth (such as about facts or the language), and the latter comes from irreducible randomness (such as multiple possible answers). In particular, we derive an information-theoretic metric that allows to reliably detect when only epistemic uncertainty is large, in which case the output of the model is unreliable. This condition can be computed based solely on the output of the model obtained simply by some special iterative prompting based on the previous responses. Such quantification, for instance, allows to detect hallucinations (cases when epistemic uncertainty is high) in both single- and multi-answer responses. This is in contrast to many standard uncertainty quantification strategies (such as thresholding the log-likelihood of a response) where hallucinations in the multi-answer case cannot be detected. We conduct a series of experiments which demonstrate the advantage of our formulation. Further, our investigations shed some light on how the probabilities assigned to a given output by an LLM can be amplified by iterative prompting, which might be of independent interest.

We present $\texttt{ACCORD}$, a framework and benchmark suite for disentangling the commonsense grounding and reasoning abilities of large language models (LLMs) through controlled, multi-hop counterfactuals. $\texttt{ACCORD}$ introduces formal elements to commonsense reasoning to explicitly control and quantify reasoning complexity beyond the typical 1 or 2 hops. Uniquely, $\texttt{ACCORD}$ can automatically generate benchmarks of arbitrary reasoning complexity, and so it scales with future LLM improvements. Benchmarking state-of-the-art LLMs -- including GPT-4o (2024-05-13), Llama-3-70B-Instruct, and Mixtral-8x22B-Instruct-v0.1 -- shows performance degrading to random chance with only moderate scaling, leaving substantial headroom for improvement. We release a leaderboard of the benchmark suite tested in this work, as well as code for automatically generating more complex benchmarks.

Although Retrieval-Augmented Large Language Models (RALMs) demonstrate their superiority in terms of factuality, they do not consistently outperform the original retrieval-free Language Models (LMs). Our experiments reveal that this example-level performance inconsistency exists not only between retrieval-augmented and retrieval-free LM but also among different retrievers. To understand this phenomenon, we investigate the degeneration behavior of RALMs and theoretically decompose it into four categories. Further analysis based on our decomposition reveals that the innate difference in knowledge sources and the unpredictable degeneration of the reader model contribute most to the inconsistency. Drawing from our analysis, we introduce Ensemble of Retrievers (EoR), a trainable framework that can adaptively retrieve from different knowledge sources and effectively decrease unpredictable reader errors. Our experiments on Open Domain Question Answering show that EoR substantially improves performance over the RALM with a single retriever by considerably reducing inconsistent behaviors.