Steering Large Language Models between Code Execution and Textual Reasoning

While a lot of recent research focuses on enhancing the textual reasoning capabilities of Large Language Models (LLMs) by optimizing the multi-agent framework or reasoning chains, several benchmark tasks can be solved with 100% success through direct coding, which is more scalable and avoids the computational overhead associated with textual iterating and searching. Textual reasoning has inherent limitations in solving tasks with challenges in math, logics, optimization, and searching, which is unlikely to be solved by simply scaling up the model and data size. The recently released OpenAI GPT Code Interpreter and multi-agent frameworks such as AutoGen have demonstrated remarkable proficiency of integrating code generation and execution to solve complex tasks using LLMs. However, based on our experiments on 7 existing popular methods for steering code/text generation in both single- and multi-turn settings with 14 tasks and 6 types of LLMs (including the new O1-preview), currently there is no optimal method to correctly steer LLMs to write code when needed. We discover some interesting patterns on when models use code vs. textual reasoning with the evolution to task complexity and model sizes, which even result in an astonishingly inverse scaling law. We also discover that results from LLM written code are not always better than using textual reasoning, even if the task could be solved through code. To mitigate the above issues, we propose three methods to better steer LLM code/text generation and achieve a notable improvement. The costs of token lengths and runtime are thoroughly discussed for all the methods. We believe the problem of steering LLM code/text generation is critical for future research and has much space for further improvement. Project Page, Datasets, and Codes are available at https://yongchao98.github.io/CodeSteer/.

CRAB: Cross-environment Agent Benchmark for Multimodal Language Model Agents

The development of autonomous agents increasingly relies on Multimodal Language Models (MLMs) to perform tasks described in natural language with GUI environments, such as websites, desktop computers, or mobile phones. Existing benchmarks for MLM agents in interactive environments are limited by their focus on a single environment, lack of detailed and generalized evaluation methods, and the complexities of constructing tasks and evaluators. To overcome these limitations, we introduce Crab, the first agent benchmark framework designed to support cross-environment tasks, incorporating a graph-based fine-grained evaluation method and an efficient mechanism for task and evaluator construction. Our framework supports multiple devices and can be easily extended to any environment with a Python interface. Leveraging Crab, we developed a cross-platform Crab Benchmark-v0 comprising 100 tasks in computer desktop and mobile phone environments. We evaluated four advanced MLMs using different single and multi-agent system configurations on this benchmark. The experimental results demonstrate that the single agent with GPT-4o achieves the best completion ratio of 35.26%. All framework code, agent code, and task datasets are publicly available at https://github.com/camel-ai/crab.

Large Language Models Can Plan Your Travels Rigorously with Formal Verification Tools

The recent advancements of Large Language Models (LLMs), with their abundant world knowledge and capabilities of tool-using and reasoning, fostered many LLM planning algorithms. However, LLMs have not shown to be able to accurately solve complex combinatorial optimization problems. In Xie et al. (2024), the authors proposed TravelPlanner, a U.S. domestic travel planning benchmark, and showed that LLMs themselves cannot make travel plans that satisfy user requirements with a best success rate of 0.6%. In this work, we propose a framework that enables LLMs to formally formulate and solve the travel planning problem as a satisfiability modulo theory (SMT) problem and use SMT solvers interactively and automatically solve the combinatorial search problem. The SMT solvers guarantee the satisfiable of input constraints and the LLMs can enable a language-based interaction with our framework. When the input constraints cannot be satisfiable, our LLM-based framework will interactively offer suggestions to users to modify their travel requirements via automatic reasoning using the SMT solvers. We evaluate our framework with TravelPlanner and achieve a success rate of 97%. We also create a separate dataset that contain international travel benchmarks and use both dataset to evaluate the effectiveness of our interactive planning framework when the initial user queries cannot be satisfied. Our framework could generate valid plans with an average success rate of 78.6% for our dataset and 85.0% for TravelPlanner according to diverse humans preferences.

PRompt Optimization in Multi-Step Tasks (PROMST): Integrating Human Feedback and Preference Alignment

Prompt optimization aims to find the best prompt to a large language model (LLM) for a given task. LLMs have been successfully used to help find and improve prompt candidates for single-step tasks. However, realistic tasks for agents are multi-step and introduce new challenges: (1) Prompt content is likely to be more extensive and complex, making it more difficult for LLMs to analyze errors, (2) the impact of an individual step is difficult to evaluate, and (3) different people may have varied preferences about task execution. While humans struggle to optimize prompts, they are good at providing feedback about LLM outputs; we therefore introduce a new LLM-driven discrete prompt optimization framework that incorporates human-designed feedback rules about potential errors to automatically offer direct suggestions for improvement. Our framework is stylized as a genetic algorithm in which an LLM generates new candidate prompts from a parent prompt and its associated feedback; we use a learned heuristic function that predicts prompt performance to efficiently sample from these candidates. This approach significantly outperforms both human-engineered prompts and several other prompt optimization methods across eight representative multi-step tasks (an average 27.7% and 28.2% improvement to current best methods on GPT-3.5 and GPT-4, respectively). We further show that the score function for tasks can be modified to better align with individual preferences. We believe our work can serve as a benchmark for automatic prompt optimization for LLM-driven multi-step tasks. Datasets and Codes are available at https://github.com/yongchao98/PROMST. Project Page is available at https://yongchao98.github.io/MIT-REALM-PROMST.

Physics-Enhanced Multi-fidelity Learning for Optical Surface Imprint

Human fingerprints serve as one unique and powerful characteristic for each person, from which policemen can recognize the identity. Similar to humans, many natural bodies and intrinsic mechanical qualities can also be uniquely identified from surface characteristics. To measure the elasto-plastic properties of one material, one formally sharp indenter is pushed into the measured body under constant force and retracted, leaving a unique residual imprint of the minute size from several micrometers to nanometers. However, one great challenge is how to map the optical image of this residual imprint into the real wanted mechanical properties, i.e., the tensile force curve. In this paper, we propose a novel method to use multi-fidelity neural networks (MFNN) to solve this inverse problem. We first actively train the NN model via pure simulation data, and then bridge the sim-to-real gap via transfer learning. The most innovative part is that we use NN to dig out the unknown physics and also implant the known physics into the transfer learning framework, thus highly improving the model stability and decreasing the data requirement. This work serves as one great example of applying machine learning into the real experimental research, especially under the constraints of data limitation and fidelity variance.

Scalable Multi-Robot Collaboration with Large Language Models: Centralized or Decentralized Systems?

A flurry of recent work has demonstrated that pre-trained large language models (LLMs) can be effective task planners for a variety of single-robot tasks. The planning performance of LLMs is significantly improved via prompting techniques, such as in-context learning or re-prompting with state feedback, placing new importance on the token budget for the context window. An under-explored but natural next direction is to investigate LLMs as multi-robot task planners. However, long-horizon, heterogeneous multi-robot planning introduces new challenges of coordination while also pushing up against the limits of context window length. It is therefore critical to find token-efficient LLM planning frameworks that are also able to reason about the complexities of multi-robot coordination. In this work, we compare the task success rate and token efficiency of four multi-agent communication frameworks (centralized, decentralized, and two hybrid) as applied to four coordination-dependent multi-agent 2D task scenarios for increasing numbers of agents. We find that a hybrid framework achieves better task success rates across all four tasks and scales better to more agents. We further demonstrate the hybrid frameworks in 3D simulations where the vision-to-text problem and dynamical errors are considered. See our project website https://yongchao98.github.io/MIT-REALM-Multi-Robot/ for prompts, videos, and code.

AutoTAMP: Autoregressive Task and Motion Planning with LLMs as Translators and Checkers

For effective human-robot interaction, robots need to understand, plan, and execute complex, long-horizon tasks described by natural language. The recent and remarkable advances in large language models (LLMs) have shown promise for translating natural language into robot action sequences for complex tasks. However, many existing approaches either translate the natural language directly into robot trajectories, or factor the inference process by decomposing language into task sub-goals, then relying on a motion planner to execute each sub-goal. When complex environmental and temporal constraints are involved, inference over planning tasks must be performed jointly with motion plans using traditional task-and-motion planning (TAMP) algorithms, making such factorization untenable. Rather than using LLMs to directly plan task sub-goals, we instead perform few-shot translation from natural language task descriptions to an intermediary task representation that can then be consumed by a TAMP algorithm to jointly solve the task and motion plan. To improve translation, we automatically detect and correct both syntactic and semantic errors via autoregressive re-prompting, resulting in significant improvements in task completion. We show that our approach outperforms several methods using LLMs as planners in complex task domains.

NL2TL: Transforming Natural Languages to Temporal Logics using Large Language Models

Temporal Logic (TL) can be used to rigorously specify complex high-level specification for systems in many engineering applications. The translation between natural language (NL) and TL has been under-explored due to the lack of dataset and generalizable model across different application domains. In this paper, we propose an accurate and generalizable transformation framework of English instructions from NL to TL, exploring the use of Large Language Models (LLMs) at multiple stages. Our contributions are twofold. First, we develop a framework to create a dataset of NL-TL pairs combining LLMs and human annotation. We publish a dataset with 28K NL-TL pairs. Then, we finetune T5 models on the lifted versions (i.e., the specific Atomic Propositions (AP) are hidden) of the NL and TL. The enhanced generalizability originates from two aspects: 1) Usage of lifted NL-TL characterizes common logical structures, without constraints of specific domains. 2) Application of LLMs in dataset creation largely enhances corpus richness. We test the generalization of trained models on five varied domains. To achieve full NL-TL transformation, we either combine the lifted model with AP recognition task or do the further finetuning on each specific domain. During the further finetuning, our model achieves higher accuracy (>95%) using only <10% training data, compared with the baseline sequence to sequence (Seq2Seq) model.

FedSSC: Shared Supervised-Contrastive Federated Learning

Federated learning is widely used to perform decentralized training of a global model on multiple devices while preserving the data privacy of each device. However, it suffers from heterogeneous local data on each training device which increases the difficulty to reach the same level of accuracy as the centralized training. Supervised Contrastive Learning which outperform cross-entropy tries to minimizes the difference between feature space of points belongs to the same class and pushes away points from different classes. We propose Supervised Contrastive Federated Learning in which devices can share the learned class-wise feature spaces with each other and add the supervised-contrastive learning loss as a regularization term to foster the feature space learning. The loss tries to minimize the cosine similarity distance between the feature map and the averaged feature map from another device in the same class and maximizes the distance between the feature map and that in a different class. This new regularization term when added on top of the moon regularization term is found to outperform the other state-of-the-art regularization terms in solving the heterogeneous data distribution problem.