We propose a novel framework that leverages LLMs for full causal graph discovery. While previous LLM-based methods have used a pairwise query approach, this requires a quadratic number of queries which quickly becomes impractical for larger causal graphs. In contrast, the proposed framework uses a breadth-first search (BFS) approach which allows it to use only a linear number of queries. We also show that the proposed method can easily incorporate observational data when available, to improve performance. In addition to being more time and data-efficient, the proposed framework achieves state-of-the-art results on real-world causal graphs of varying sizes. The results demonstrate the effectiveness and efficiency of the proposed method in discovering causal relationships, showcasing its potential for broad applicability in causal graph discovery tasks across different domains.

Recent advancements on Large Language Models (LLMs) enable AI Agents to automatically generate and execute multi-step plans to solve complex tasks. However, since LLM's content generation process is hardly controllable, current LLM-based agents frequently generate invalid or non-executable plans, which jeopardizes the performance of the generated plans and corrupts users' trust in LLM-based agents. In response, this paper proposes a novel ``Formal-LLM'' framework for LLM-based agents by integrating the expressiveness of natural language and the precision of formal language. Specifically, the framework allows human users to express their requirements or constraints for the planning process as an automaton. A stack-based LLM plan generation process is then conducted under the supervision of the automaton to ensure that the generated plan satisfies the constraints, making the planning process controllable. We conduct experiments on both benchmark tasks and practical real-life tasks, and our framework achieves over 50% overall performance increase, which validates the feasibility and effectiveness of employing Formal-LLM to guide the plan generation of agents, preventing the agents from generating invalid and unsuccessful plans. Further, more controllable LLM-based agents can facilitate the broader utilization of LLM in application scenarios where high validity of planning is essential. The work is open-sourced at https://github.com/agiresearch/Formal-LLM.

The burgeoning capabilities of large language models (LLMs) have raised growing concerns about abuse. DetectGPT, a zero-shot metric-based unsupervised machine-generated text detector, first introduces perturbation and shows great performance improvement. However, DetectGPT's random perturbation strategy might introduce noise, limiting the distinguishability and further performance improvements. Moreover, its logit regression module relies on setting the threshold, which harms the generalizability and applicability of individual or small-batch inputs. Hence, we propose a novel detector, Pecola, which uses selective strategy perturbation to relieve the information loss caused by random masking, and multi-pair contrastive learning to capture the implicit pattern information during perturbation, facilitating few-shot performance. The experiments show that Pecola outperforms the SOTA method by 1.20% in accuracy on average on four public datasets. We further analyze the effectiveness, robustness, and generalization of our perturbation method.

Autoregressive decoding makes the inference of Large Language Models (LLMs) time-consuming. In this paper, we reconsider speculative sampling and derive two key observations. Firstly, autoregression at the feature (second-to-top-layer) level is more straightforward than at the token level. Secondly, the inherent uncertainty in feature (second-to-top-layer) level autoregression constrains its performance. Based on these insights, we introduce EAGLE (Extrapolation Algorithm for Greater Language-model Efficiency), a simple yet highly efficient speculative sampling framework. By incorporating a token sequence advanced by one time step, EAGLE effectively resolves the uncertainty, enabling precise second-to-top-layer feature prediction with minimal overhead. We conducted comprehensive evaluations of EAGLE, including all models from the Vicuna and LLaMA2-Chat series, the MoE model Mixtral 8x7B Instruct, and tasks in dialogue, code generation, mathematical reasoning, and instruction following. For LLaMA2-Chat 70B, EAGLE achieved a latency speedup ratio of 2.7x-3.5x, doubled throughput, while maintaining the distribution of the generated text.

We explore the impact of extending the context length, releasing models with context lengths of 4096 and 32768 tokens, and further refining performance with a specialized 16384 context length instruction-tuned model, we called it Malaysian Mistral. Our experiments demonstrate the efficacy of continue pretraining and the influence of extended context lengths on Mistral 7B's language understanding capabilities. Additionally, we release a model specifically tuned with a 16384 context length instruction, showcasing its potential for capturing nuanced language intricacies. Furthermore, our research contributes to the benchmarking of Malaysian Mistral against prominent language models, including ChatGPT3.5 and Claude 2. We present compelling results indicating Malaysian Mistral's superior performance on Tatabahasa (Malay grammar) test set, particularly when fine-tuned with instructions. All models released at HuggingFace Mesolitica Malaysian Mistral 7B Collection.

One critical challenge that has emerged is the presence of hallucinations in the output of large language models (LLMs) due to false or outdated knowledge. Since retraining LLMs with updated information is resource-intensive, there has been a growing interest in model editing. However, current model editing methods, while effective in improving editing performance in various scenarios, often overlook potential side effects on the general abilities of LLMs. In this paper, we raise concerns that model editing inherently improves the factuality of the model, but may come at the cost of a significant degradation of these general abilities. Systematically, we analyze side effects by evaluating four popular editing methods on three LLMs across eight representative task categories. Extensive empirical research reveals that current model editing methods are difficult to couple well with LLMs to simultaneously improve the factuality and maintain the general abilities such as reasoning, question answering, etc. Strikingly, the use of a specific method to edit LLaMA-1 (7B) resulted in a drastic performance degradation to nearly 0 on all selected tasks with just a single edit. Therefore, we advocate for more research efforts to minimize the loss of general abilities acquired during LLM pre-training and to ultimately preserve them during model editing.

Large language models (LLMs) have exhibited impressive competence in various tasks, but their opaque internal mechanisms hinder their use in mathematical problems. In this paper, we study a fundamental question: whether language models understand numbers, a basic element in math. Based on an assumption that LLMs should be capable of compressing numbers in their hidden states to solve mathematical problems, we construct a synthetic dataset comprising addition problems and utilize linear probes to read out input numbers from the hidden states. Experimental results support the existence of compressed numbers in LLMs. However, it is difficult to precisely reconstruct the original numbers, indicating that the compression process may not be lossless. Further experiments show that LLMs can utilize encoded numbers to perform arithmetic computations, and the computational ability scales up with the model size. Our preliminary research suggests that LLMs exhibit a partial understanding of numbers, offering insights for future investigations about the models' mathematical capability.

Cognitive diagnosis seeks to estimate the cognitive states of students by exploring their logged practice quiz data. It plays a pivotal role in personalized learning guidance within intelligent education systems. In this paper, we focus on an important, practical, yet often underexplored task: domain-level zero-shot cognitive diagnosis (DZCD), which arises due to the absence of student practice logs in newly launched domains. Recent cross-domain diagnostic models have been demonstrated to be a promising strategy for DZCD. These methods primarily focus on how to transfer student states across domains. However, they might inadvertently incorporate non-transferable information into student representations, thereby limiting the efficacy of knowledge transfer. To tackle this, we propose Zero-1-to-3, a domain-level zero-shot cognitive diagnosis framework via one batch of early-bird students towards three diagnostic objectives. Our approach initiates with pre-training a diagnosis model with dual regularizers, which decouples student states into domain-shared and domain-specific parts. The shared cognitive signals can be transferred to the target domain, enriching the cognitive priors for the new domain, which ensures the cognitive state propagation objective. Subsequently, we devise a strategy to generate simulated practice logs for cold-start students through analyzing the behavioral patterns from early-bird students, fulfilling the domain-adaption goal. Consequently, we refine the cognitive states of cold-start students as diagnostic outcomes via virtual data, aligning with the diagnosis-oriented goal. Finally, extensive experiments on six real-world datasets highlight the efficacy of our model for DZCD and its practical application in question recommendation. The code is publicly available at https://github.com/bigdata-ustc/Zero-1-to-3.

Autonomous driving systems are increasingly popular in today's technological landscape, where vehicles with partial automation have already been widely available on the market, and the full automation era with "driverless" capabilities is near the horizon. However, accurately understanding humans' commands, particularly for autonomous vehicles that have only passengers instead of drivers, and achieving a high level of personalization remain challenging tasks in the development of autonomous driving systems. In this paper, we introduce a Large Language Model (LLM)-based framework Talk-to-Drive (Talk2Drive) to process verbal commands from humans and make autonomous driving decisions with contextual information, satisfying their personalized preferences for safety, efficiency, and comfort. First, a speech recognition module is developed for Talk2Drive to interpret verbal inputs from humans to textual instructions, which are then sent to LLMs for reasoning. Then, appropriate commands for the Electrical Control Unit (ECU) are generated, achieving a 100% success rate in executing codes. Real-world experiments show that our framework can substantially reduce the takeover rate for a diverse range of drivers by up to 90.1%. To the best of our knowledge, Talk2Drive marks the first instance of employing an LLM-based system in a real-world autonomous driving environment.

We consider regret minimization in repeated games with a very large number of actions. Such games are inherent in the setting of AI safety via debate, and more generally games whose actions are language-based. Existing algorithms for online game playing require computation polynomial in the number of actions, which can be prohibitive for large games. We thus consider oracle-based algorithms, as oracles naturally model access to AI agents. With oracle access, we characterize when internal and external regret can be minimized efficiently. We give a novel efficient algorithm for internal regret minimization whose regret and computation complexity depend logarithmically on the number of actions. This implies efficient oracle-based computation of a correlated equilibrium in large games. We conclude with experiments in the setting of AI Safety via Debate that shows the benefit of insights from our algorithmic analysis.