Math Word Problems (MWPs) are crucial for evaluating the capability of Large Language Models (LLMs), with current research primarily focusing on questions with concise contexts. However, as real-world math problems often involve complex circumstances, LLMs' ability to solve long MWPs is vital for their applications in these scenarios, yet remains under-explored. This study pioneers the exploration of Context Length Generalizability (CoLeG), the ability of LLMs to solve long MWPs. We introduce Extended Grade-School Math (E-GSM), a collection of MWPs with lengthy narratives. Two novel metrics are proposed to assess the efficacy and resilience of LLMs in solving these problems. Our examination of existing zero-shot prompting techniques and both proprietary and open-source LLMs reveals a general deficiency in CoLeG. To alleviate these challenges, we propose distinct approaches for different categories of LLMs. For proprietary LLMs, a new instructional prompt is proposed to mitigate the influence of long context. For open-source LLMs, a new data augmentation task is developed to improve CoLeG. Our comprehensive results demonstrate the effectiveness of our proposed methods, showing not only improved performance on E-GSM but also generalizability across several other MWP benchmarks. Our findings pave the way for future research in employing LLMs for complex, real-world applications, offering practical solutions to current limitations and opening avenues for further exploration of model generalizability and training methodologies.

Language models as intelligent agents push the boundaries of sequential decision-making agents but struggle with limited knowledge of environmental dynamics and exponentially huge action space. Recent efforts like GLAM and TWOSOME manually constrain the action space to a restricted subset and employ reinforcement learning to align agents' knowledge with specific environments. However, they overlook fine-grained credit assignments for intra-action tokens, which is essential for efficient language agent optimization, and rely on human's prior knowledge to restrict action space. This paper proposes decomposing language agent optimization from the action level to the token level, offering finer supervision for each intra-action token and manageable optimization complexity in environments with unrestricted action spaces. Beginning with the simplification of flattening all actions, we theoretically explore the discrepancies between action-level optimization and this naive token-level optimization. We then derive the Bellman backup with Action Decomposition (BAD) to integrate credit assignments for both intra-action and inter-action tokens, effectively eliminating the discrepancies. Implementing BAD within the PPO algorithm, we introduce Policy Optimization with Action Decomposition (POAD). POAD benefits from a finer-grained credit assignment process and lower optimization complexity, leading to enhanced learning efficiency and generalization abilities in aligning language agents with interactive environments. We validate POAD across diverse testbeds, with results affirming the advantages of our approach and the correctness of our theoretical analysis.

Recent work has proposed the linear representation hypothesis: that language models perform computation by manipulating one-dimensional representations of concepts ("features") in activation space. In contrast, we explore whether some language model representations may be inherently multi-dimensional. We begin by developing a rigorous definition of irreducible multi-dimensional features based on whether they can be decomposed into either independent or non-co-occurring lower-dimensional features. Motivated by these definitions, we design a scalable method that uses sparse autoencoders to automatically find multi-dimensional features in GPT-2 and Mistral 7B. These auto-discovered features include strikingly interpretable examples, e.g. circular features representing days of the week and months of the year. We identify tasks where these exact circles are used to solve computational problems involving modular arithmetic in days of the week and months of the year. Finally, we provide evidence that these circular features are indeed the fundamental unit of computation in these tasks with intervention experiments on Mistral 7B and Llama 3 8B, and we find further circular representations by breaking down the hidden states for these tasks into interpretable components.

Understanding the connections between artificial and biological intelligent systems can reveal fundamental principles underlying general intelligence. While many artificial intelligence (AI) models have a neuroscience counterpart, such connections are largely missing in Transformer models and the self-attention mechanism. Here, we examine the relationship between attention heads and human episodic memory. We focus on the induction heads, which contribute to the in-context learning capabilities of Transformer-based large language models (LLMs). We demonstrate that induction heads are behaviorally, functionally, and mechanistically similar to the contextual maintenance and retrieval (CMR) model of human episodic memory. Our analyses of LLMs pre-trained on extensive text data show that CMR-like heads often emerge in the intermediate model layers and that their behavior qualitatively mirrors the memory biases seen in humans. Our findings uncover a parallel between the computational mechanisms of LLMs and human memory, offering valuable insights into both research fields.

Large Language Models (LLMs) have garnered considerable attention owing to their remarkable capabilities, leading to an increasing number of companies offering LLMs as services. Different LLMs achieve different performance at different costs. A challenge for users lies in choosing the LLMs that best fit their needs, balancing cost and performance. In this paper, we propose a framework for addressing the cost-effective query allocation problem for LLMs. Given a set of input queries and candidate LLMs, our framework, named OptLLM, provides users with a range of optimal solutions to choose from, aligning with their budget constraints and performance preferences, including options for maximizing accuracy and minimizing cost. OptLLM predicts the performance of candidate LLMs on each query using a multi-label classification model with uncertainty estimation and then iteratively generates a set of non-dominated solutions by destructing and reconstructing the current solution. To evaluate the effectiveness of OptLLM, we conduct extensive experiments on various types of tasks, including text classification, question answering, sentiment analysis, reasoning, and log parsing. Our experimental results demonstrate that OptLLM substantially reduces costs by 2.40% to 49.18% while achieving the same accuracy as the best LLM. Compared to other multi-objective optimization algorithms, OptLLM improves accuracy by 2.94% to 69.05% at the same cost or saves costs by 8.79% and 95.87% while maintaining the highest attainable accuracy.

This work presents AstroPT, an autoregressive pretrained transformer developed with astronomical use-cases in mind. The AstroPT models presented here have been pretrained on 8.6 million $512 \times 512$ pixel $grz$-band galaxy postage stamp observations from the DESI Legacy Survey DR8. We train a selection of foundation models of increasing size from 1 million to 2.1 billion parameters, and find that AstroPT follows a similar saturating log-log scaling law to textual models. We also find that the models' performances on downstream tasks as measured by linear probing improves with model size up to the model parameter saturation point. We believe that collaborative community development paves the best route towards realising an open source `Large Observation Model' -- a model trained on data taken from the observational sciences at the scale seen in natural language processing. To this end, we release the source code, weights, and dataset for AstroPT under the MIT license, and invite potential collaborators to join us in collectively building and researching these models.

We introduce Bitune, a method that improves instruction-tuning of pretrained decoder-only large language models, leading to consistent gains on downstream tasks. Bitune applies both causal and bidirectional attention to the prompt, to obtain a better representation of the query or instruction. We realize this by introducing two sets of parameters, for which we apply parameter-efficient finetuning techniques. These causal and bidirectional features are then combined into a weighted average with trainable coefficients, which is subsequently used to generate new tokens. We demonstrate significant improvements in zero-shot performance on commonsense reasoning, arithmetic, and language understanding tasks, while extensive ablation studies validate the role of each component and demonstrate the method's agnosticism to different PEFT techniques.

Large language models (LLMs) improve their performance in downstream tasks when they generate Chain of Thought reasoning text before producing an answer. Our research investigates how LLMs recover from errors in Chain of Thought, reaching the correct final answer despite mistakes in the reasoning text. Through analysis of these error recovery behaviors, we find evidence for unfaithfulness in Chain of Thought, but we also identify many clear examples of faithful error recovery behaviors. We identify factors that shift LLM recovery behavior: LLMs recover more frequently from obvious errors and in contexts that provide more evidence for the correct answer. However, unfaithful recoveries show the opposite behavior, occurring more frequently for more difficult error positions. Our results indicate that there are distinct mechanisms driving faithful and unfaithful error recoveries. Our results challenge the view that LLM reasoning is a uniform, coherent process.

Implicit Personalization (IP) is a phenomenon of language models inferring a user's background from the implicit cues in the input prompts and tailoring the response based on this inference. While previous work has touched upon various instances of this problem, there lacks a unified framework to study this behavior. This work systematically studies IP through a rigorous mathematical formulation, a multi-perspective moral reasoning framework, and a set of case studies. Our theoretical foundation for IP relies on a structural causal model and introduces a novel method, indirect intervention, to estimate the causal effect of a mediator variable that cannot be directly intervened upon. Beyond the technical approach, we also introduce a set of moral reasoning principles based on three schools of moral philosophy to study when IP may or may not be ethically appropriate. Equipped with both mathematical and ethical insights, we present three diverse case studies illustrating the varied nature of the IP problem and offer recommendations for future research. Our code and data are at https://github.com/jiarui-liu/IP.

Cultural bias is pervasive in many large language models (LLMs), largely due to the deficiency of data representative of different cultures. Typically, cultural datasets and benchmarks are constructed either by extracting subsets of existing datasets or by aggregating from platforms such as Wikipedia and social media. However, these approaches are highly dependent on real-world data and human annotations, making them costly and difficult to scale. Inspired by cognitive theories on social communication, this paper introduces CulturePark, an LLM-powered multi-agent communication framework for cultural data collection. CulturePark simulates cross-cultural human communication with LLM-based agents playing roles in different cultures. It generates high-quality cross-cultural dialogues encapsulating human beliefs, norms, and customs. Using CulturePark, we generated 41,000 cultural samples to fine-tune eight culture-specific LLMs. We evaluated these models across three downstream tasks: content moderation, cultural alignment, and cultural education. Results show that for content moderation, our GPT-3.5-based models either match or outperform GPT-4 on datasets. Regarding cultural alignment, our models surpass GPT-4 on Hofstede's VSM 13 framework. Furthermore, for cultural education of human participants, our models demonstrate superior outcomes in both learning efficacy and user experience compared to GPT-4. CulturePark proves an important step in addressing cultural bias and advancing the democratization of AI, highlighting the critical role of culturally inclusive data in model training.