It is a notable trend to use Large Language Models (LLMs) to tackle complex tasks, e.g., tasks that require a sequence of actions and dynamic interaction with tools and external environments. In this paper, we propose StateFlow, a novel LLM-based task-solving paradigm that conceptualizes complex task-solving processes as state machines. In StateFlow, we distinguish between "process grounding" (via state and state transitions) and "sub-task solving" (through actions within a state), enhancing control and interpretability of the task-solving procedure. A state represents the status of a running process. The transitions between states are controlled by heuristic rules or decisions made by the LLM, allowing for a dynamic and adaptive progression. Upon entering a state, a series of actions is executed, involving not only calling LLMs guided by different prompts, but also the utilization of external tools as needed. Our results show that StateFlow significantly enhances LLMs' efficiency. For instance, StateFlow achieves 13% and 28% higher success rates compared to ReAct in InterCode SQL and ALFWorld benchmark, with 5x and 3x less cost respectively. We also show that StateFlow can be combined with iterative refining methods like Reflexion to further improve performance.

Some have criticised Generative AI Systems for replicating the familiar pathologies of already widely-deployed AI systems. Other critics highlight how they foreshadow vastly more powerful future systems, which might threaten humanity's survival. The first group says there is nothing new here; the other looks through the present to a perhaps distant horizon. In this paper, I instead pay attention to what makes these particular systems distinctive: both their remarkable scientific achievement, and the most likely and consequential ways in which they will change society over the next five to ten years. In particular, I explore the potential societal impacts and normative questions raised by the looming prospect of 'Generative Agents', in which multimodal large language models (LLMs) form the executive centre of complex, tool-using AI systems that can take unsupervised sequences of actions towards some goal.

Recently, the large language model (LLM) community has shown increasing interest in enhancing LLMs' capability to handle extremely long documents. As various long-text techniques and model architectures emerge, the precise and detailed evaluation of models' long-text capabilities has become increasingly important. Existing long-text evaluation benchmarks, such as L-Eval and LongBench, construct long-text test sets based on open-source datasets, focusing mainly on QA and summarization tasks. These datasets include test samples of varying lengths (from 2k to 32k+) entangled together, making it challenging to assess model capabilities across different length ranges. Moreover, they do not cover the ultralong settings (100k+ tokens) that the latest LLMs claim to achieve. In this paper, we introduce Ada-LEval, a length-adaptable benchmark for evaluating the long-context understanding of LLMs. Ada-LEval includes two challenging subsets, TSort and BestAnswer, which enable a more reliable evaluation of LLMs' long context capabilities. These benchmarks support intricate manipulation of the length of test cases, and can easily produce text samples up to 128k tokens. We evaluate 4 state-of-the-art closed-source API models and 6 open-source models with Ada-LEval. The evaluation results demonstrate the limitations of current LLMs, especially in ultra-long-context settings. Our code is available at https://github.com/open-compass/Ada-LEval.

There is a growing interest in applying large language models (LLMs) in robotic tasks, due to their remarkable reasoning ability and extensive knowledge learned from vast training corpora. Grounding LLMs in the physical world remains an open challenge as they can only process textual input. Recent advancements in large vision-language models (LVLMs) have enabled a more comprehensive understanding of the physical world by incorporating visual input, which provides richer contextual information than language alone. In this work, we proposed a novel paradigm that leveraged GPT-4V(ision), the state-of-the-art LVLM by OpenAI, to enable embodied agents to perceive liquid objects via image-based environmental feedback. Specifically, we exploited the physical understanding of GPT-4V to interpret the visual representation (e.g., time-series plot) of non-visual feedback (e.g., F/T sensor data), indirectly enabling multimodal perception beyond vision and language using images as proxies. We evaluated our method using 10 common household liquids with containers of various geometry and material. Without any training or fine-tuning, we demonstrated that our method can enable the robot to indirectly perceive the physical response of liquids and estimate their viscosity. We also showed that by jointly reasoning over the visual and physical attributes learned through interactions, our method could recognize liquid objects in the absence of strong visual cues (e.g., container labels with legible text or symbols), increasing the accuracy from 69.0% -- achieved by the best-performing vision-only variant -- to 86.0%.

Text simplification seeks to improve readability while retaining the original content and meaning. Our study investigates whether pre-trained classifiers also maintain such coherence by comparing their predictions on both original and simplified inputs. We conduct experiments using 11 pre-trained models, including BERT and OpenAI's GPT 3.5, across six datasets spanning three languages. Additionally, we conduct a detailed analysis of the correlation between prediction change rates and simplification types/strengths. Our findings reveal alarming inconsistencies across all languages and models. If not promptly addressed, simplified inputs can be easily exploited to craft zero-iteration model-agnostic adversarial attacks with success rates of up to 50%

Large language models (LLMs) such as Llama 2 perform very well on tasks that involve both natural language and source code, particularly code summarization and code generation. We show that for the task of code summarization, the performance of these models on individual examples often depends on the amount of (subword) token overlap between the code and the corresponding reference natural language descriptions in the dataset. This token overlap arises because the reference descriptions in standard datasets (corresponding to docstrings in large code bases) are often highly similar to the names of the functions they describe. We also show that this token overlap occurs largely in the function names of the code and compare the relative performance of these models after removing function names versus removing code structure. We also show that using multiple evaluation metrics like BLEU and BERTScore gives us very little additional insight since these metrics are highly correlated with each other.

Vision-language models (VLMs) are trained for thousands of GPU hours on carefully curated web datasets. In recent times, data curation has gained prominence with several works developing strategies to retain 'high-quality' subsets of 'raw' scraped data. For instance, the LAION public dataset retained only 10% of the total crawled data. However, these strategies are typically developed agnostic of the available compute for training. In this paper, we first demonstrate that making filtering decisions independent of training compute is often suboptimal: the limited high-quality data rapidly loses its utility when repeated, eventually requiring the inclusion of 'unseen' but 'lower-quality' data. To address this quality-quantity tradeoff ($\texttt{QQT}$), we introduce neural scaling laws that account for the non-homogeneous nature of web data, an angle ignored in existing literature. Our scaling laws (i) characterize the $\textit{differing}$ 'utility' of various quality subsets of web data; (ii) account for how utility diminishes for a data point at its 'nth' repetition; and (iii) formulate the mutual interaction of various data pools when combined, enabling the estimation of model performance on a combination of multiple data pools without ever jointly training on them. Our key message is that data curation $\textit{cannot}$ be agnostic of the total compute that a model will be trained for. Our scaling laws allow us to curate the best possible pool for achieving top performance on Datacomp at various compute budgets, carving out a pareto-frontier for data curation. Code is available at https://github.com/locuslab/scaling_laws_data_filtering.

In-context learning is a powerful emergent ability in transformer models. Prior work in mechanistic interpretability has identified a circuit element that may be critical for in-context learning -- the induction head (IH), which performs a match-and-copy operation. During training of large transformers on natural language data, IHs emerge around the same time as a notable phase change in the loss. Despite the robust evidence for IHs and this interesting coincidence with the phase change, relatively little is known about the diversity and emergence dynamics of IHs. Why is there more than one IH, and how are they dependent on each other? Why do IHs appear all of a sudden, and what are the subcircuits that enable them to emerge? We answer these questions by studying IH emergence dynamics in a controlled setting by training on synthetic data. In doing so, we develop and share a novel optogenetics-inspired causal framework for modifying activations throughout training. Using this framework, we delineate the diverse and additive nature of IHs. By clamping subsets of activations throughout training, we then identify three underlying subcircuits that interact to drive IH formation, yielding the phase change. Furthermore, these subcircuits shed light on data-dependent properties of formation, such as phase change timing, already showing the promise of this more in-depth understanding of subcircuits that need to "go right" for an induction head.

Recent advances in Automated Theorem Proving have shown the effectiveness of leveraging a (large) language model that generates tactics (i.e. proof steps) to search through proof states. The current model, while trained solely on successful proof paths, faces a discrepancy at the inference stage, as it must sample and try various tactics at each proof state until finding success, unlike its training which does not incorporate learning from failed attempts. Intuitively, a tactic that leads to a failed search path would indicate that similar tactics should receive less attention during the following trials. In this paper, we demonstrate the benefit of training models that additionally learn from failed search paths. Facing the lack of such trial-and-error data in existing open-source theorem-proving datasets, we curate a dataset on intuitionistic propositional logic theorems and formalize it in Lean, such that we can reliably check the correctness of proofs. We compare our model trained on relatively short trial-and-error information (TrialMaster) with models trained only on the correct paths and discover that the former solves more unseen theorems with lower trial searches.

Large language models (LLMs), while exhibiting exceptional performance, suffer from hallucinations, especially on knowledge-intensive tasks. Existing works propose to augment LLMs with individual text units retrieved from external knowledge corpora to alleviate the issue. However, in many domains, texts are interconnected (e.g., academic papers in a bibliographic graph are linked by citations and co-authorships) which form a (text-attributed) graph. The knowledge in such graphs is encoded not only in single texts/nodes but also in their associated connections. To facilitate the research of augmenting LLMs with graphs, we manually construct a Graph Reasoning Benchmark dataset called GRBench, containing 1,740 questions that can be answered with the knowledge from 10 domain graphs. Then, we propose a simple and effective framework called Graph Chain-of-thought (Graph-CoT) to augment LLMs with graphs by encouraging LLMs to reason on the graph iteratively. Each Graph-CoT iteration consists of three sub-steps: LLM reasoning, LLM-graph interaction, and graph execution. We conduct systematic experiments with three LLM backbones on GRBench, where Graph-CoT outperforms the baselines consistently. The code is available at https://github.com/PeterGriffinJin/Graph-CoT.