Language agents powered by large language models (LLMs) have seen exploding development. Their capability of using language as a vehicle for thought and communication lends an incredible level of flexibility and versatility. People have quickly capitalized on this capability to connect LLMs to a wide range of external components and environments: databases, tools, the Internet, robotic embodiment, etc. Many believe an unprecedentedly powerful automation technology is emerging. However, new automation technologies come with new safety risks, especially for intricate systems like language agents. There is a surprisingly large gap between the speed and scale of their development and deployment and our understanding of their safety risks. Are we building a house of cards? In this position paper, we present the first systematic effort in mapping adversarial attacks against language agents. We first present a unified conceptual framework for agents with three major components: Perception, Brain, and Action. Under this framework, we present a comprehensive discussion and propose 12 potential attack scenarios against different components of an agent, covering different attack strategies (e.g., input manipulation, adversarial demonstrations, jailbreaking, backdoors). We also draw connections to successful attack strategies previously applied to LLMs. We emphasize the urgency to gain a thorough understanding of language agent risks before their widespread deployment.

In-context learning has emerged as a groundbreaking ability of Large Language Models (LLMs) and revolutionized various fields by providing a few task-relevant demonstrations in the prompt. However, trustworthy issues with LLM's response, such as hallucination, have also been actively discussed. Existing works have been devoted to quantifying the uncertainty in LLM's response, but they often overlook the complex nature of LLMs and the uniqueness of in-context learning. In this work, we delve into the predictive uncertainty of LLMs associated with in-context learning, highlighting that such uncertainties may stem from both the provided demonstrations (aleatoric uncertainty) and ambiguities tied to the model's configurations (epistemic uncertainty). We propose a novel formulation and corresponding estimation method to quantify both types of uncertainties. The proposed method offers an unsupervised way to understand the prediction of in-context learning in a plug-and-play fashion. Extensive experiments are conducted to demonstrate the effectiveness of the decomposition. The code and data are available at: \url{https://github.com/lingchen0331/UQ_ICL}.

The emergence of Large Language Models (LLMs) like ChatGPT has inspired the development of LLM-based agents capable of addressing complex, real-world tasks. However, these agents often struggle during task execution due to methodological constraints, such as error propagation and limited adaptability. To address this issue, we propose a multi-agent framework based on dynamic Task Decomposition and Agent Generation (TDAG). This framework dynamically decomposes complex tasks into smaller subtasks and assigns each to a specifically generated subagent, thereby enhancing adaptability in diverse and unpredictable real-world tasks. Simultaneously, existing benchmarks often lack the granularity needed to evaluate incremental progress in complex, multi-step tasks. In response, we introduce ItineraryBench in the context of travel planning, featuring interconnected, progressively complex tasks with a fine-grained evaluation system. ItineraryBench is designed to assess agents' abilities in memory, planning, and tool usage across tasks of varying complexity. Our experimental results reveal that TDAG significantly outperforms established baselines, showcasing its superior adaptability and context awareness in complex task scenarios.

Recent work has shown the immense potential of synthetically generated datasets for training large language models (LLMs), especially for acquiring targeted skills. Current large-scale math instruction tuning datasets such as MetaMathQA (Yu et al., 2024) and MAmmoTH (Yue et al., 2024) are constructed using outputs from closed-source LLMs with commercially restrictive licenses. A key reason limiting the use of open-source LLMs in these data generation pipelines has been the wide gap between the mathematical skills of the best closed-source LLMs, such as GPT-4, and the best open-source LLMs. Building on the recent progress in open-source LLMs, our proposed prompting novelty, and some brute-force scaling, we construct OpenMathInstruct-1, a math instruction tuning dataset with 1.8M problem-solution pairs. The dataset is constructed by synthesizing code-interpreter solutions for GSM8K and MATH, two popular math reasoning benchmarks, using the recently released and permissively licensed Mixtral model. Our best model, OpenMath-CodeLlama-70B, trained on a subset of OpenMathInstruct-1, achieves a score of 84.6% on GSM8K and 50.7% on MATH, which is competitive with the best gpt-distilled models. We release our code, models, and the OpenMathInstruct-1 dataset under a commercially permissive license.

Recent large language models (LLMs) have shown remarkable performance in aligning generated text with user intentions across various tasks. When it comes to long-form text generation, there has been a growing interest in generation from a discourse coherence perspective. However, existing lexical or semantic metrics such as BLEU, ROUGE, BertScore cannot effectively capture the discourse coherence. The development of discourse-specific automatic evaluation methods for assessing the output of LLMs warrants greater focus and exploration. In this paper, we present a novel automatic metric designed to quantify the discourse divergence between two long-form articles. Extensive experiments on three datasets from representative domains demonstrate that our metric aligns more closely with human preferences and GPT-4 coherence evaluation, outperforming existing evaluation methods.

Optimization problems are pervasive in sectors from manufacturing and distribution to healthcare. However, most such problems are still solved heuristically by hand rather than optimally by state-of-the-art solvers because the expertise required to formulate and solve these problems limits the widespread adoption of optimization tools and techniques. This paper introduces OptiMUS, a Large Language Model (LLM)-based agent designed to formulate and solve (mixed integer) linear programming problems from their natural language descriptions. OptiMUS can develop mathematical models, write and debug solver code, evaluate the generated solutions, and improve its model and code based on these evaluations. OptiMUS utilizes a modular structure to process problems, allowing it to handle problems with long descriptions and complex data without long prompts. Experiments demonstrate that OptiMUS outperforms existing state-of-the-art methods on easy datasets by more than $20\%$ and on hard datasets (including a new dataset, NLP4LP, released with this paper that features long and complex problems) by more than $30\%$.

We study the continual pretraining recipe for scaling language models' context lengths to 128K, with a focus on data engineering. We hypothesize that long context modeling, in particular \textit{the ability to utilize information at arbitrary input locations}, is a capability that is mostly already acquired through large-scale pretraining, and that this capability can be readily extended to contexts substantially longer than seen during training~(e.g., 4K to 128K) through lightweight continual pretraining on appropriate data mixture. We investigate the \textit{quantity} and \textit{quality} of the data for continual pretraining: (1) for quantity, we show that 500 million to 5 billion tokens are enough to enable the model to retrieve information anywhere within the 128K context; (2) for quality, our results equally emphasize \textit{domain balance} and \textit{length upsampling}. Concretely, we find that naively upsampling longer data on certain domains like books, a common practice of existing work, gives suboptimal performance, and that a balanced domain mixture is important. We demonstrate that continual pretraining of the full model on 1B-5B tokens of such data is an effective and affordable strategy for scaling the context length of language models to 128K. Our recipe outperforms strong open-source long-context models and closes the gap to frontier models like GPT-4 128K.

As language models continue to scale in size and capability, they display an array of emerging behaviors, both beneficial and concerning. This heightens the need to control model behaviors. We hope to be able to control the personality traits of language models at the inference-time so as to have various character features, on top of which the requirements of different types of tasks can be met. Personality is a higher-level and more abstract behavioral representation for language models. We introduce ControlLM, which leverages differential activation patterns, derived from contrasting behavioral prompts in the model's latent space, to influence the model's personality traits at inference. This approach allows for the precise, real-time adjustment of model behavior. First, we demonstrate ControlLM's capacity to elicit diverse persona behaviors without any training, while precision control allows personality traits to closely match average human values. Subsequently, we showcase improved reasoning and question answering through selective amplification of beneficial attributes like conscientiousness and friendliness. We hope that this work will inspire research on controlling human-like behaviors of language models and provide insights for future research. Our code is publicly available at: https://github.com/wengsyx/ControlLM.

Procedural content generation uses algorithmic techniques to create large amounts of new content for games at much lower production costs. In newer approaches, procedural content generation utilizes machine learning. However, these methods usually require expensive collection of large amounts of data, as well as the development and training of fairly complex learning models, which can be both extremely time-consuming and expensive. The core of our research is to explore whether we can lower the barrier to the use of personalized procedural content generation through a more practical and generalizable approach with large language models. Matching game content with player preferences benefits both players, who enjoy the game more, and developers, who increasingly depend on players enjoying the game before being able to monetize it. Therefore, this paper presents a novel approach to achieving personalization by using large language models to propose levels based on the gameplay data continuously collected from individual players. We compared the levels generated using our approach with levels generated with more traditional procedural generation techniques. Our easily reproducible method has proven viable in a production setting and outperformed levels generated by traditional methods in the probability that a player will not quit the game mid-level.

Instruction tuning is critical to large language models (LLMs) for achieving better instruction following and task adaptation capabilities but its success heavily relies on the training data quality. Many recent methods focus on improving the data quality but often overlook the compatibility of the data with the student model being finetuned. This paper introduces Selective Reflection-Tuning, a novel paradigm that synergizes a teacher LLM's reflection and introspection for improving existing data quality with the data selection capability of the student LLM, to automatically refine existing instruction-tuning data. This teacher-student collaboration produces high-quality and student-compatible instruction-response pairs, resulting in sample-efficient instruction tuning and LLMs of superior performance. Selective Reflection-Tuning is a data augmentation and synthesis that generally improves LLM finetuning and self-improvement without collecting brand-new data. We apply our method to Alpaca and WizardLM data and achieve much stronger and top-tier 7B and 13B LLMs. Our codes, models, and data will be released at https://github.com/tianyi-lab/Reflection_Tuning.