NeurIPS 2024 Papers

Pre-trained language encoders---functions that represent text as vectors---are an integral component of many NLP tasks. We tackle a natural question in language encoder analysis: What does it mean for two encoders to be similar? We contend that a faithful measure of similarity needs to be \emph{intrinsic}, that is, task-independent, yet still be informative of \emph{extrinsic} similarity---the performance on downstream tasks. It is common to consider two encoders similar if they are \emph{homotopic}, i.e., if they can be aligned through some transformation. In this spirit, we study the properties of \emph{affine} alignment of language encoders and its implications on extrinsic similarity. We find that while affine alignment is fundamentally an asymmetric notion of similarity, it is still informative of extrinsic similarity. We confirm this on datasets of natural language representations. Beyond providing useful bounds on extrinsic similarity, affine intrinsic similarity also allows us to begin uncovering the structure of the space of pre-trained encoders by defining an order over them.

This paper introduces AutoSurvey, a speedy and well-organized methodology for automating the creation of comprehensive literature surveys in rapidly evolving fields like artificial intelligence. Traditional survey paper creation faces challenges due to the vast volume and complexity of information, prompting the need for efficient survey methods. While large language models (LLMs) offer promise in automating this process, challenges such as context window limitations, parametric knowledge constraints, and the lack of evaluation benchmarks remain. AutoSurvey addresses these challenges through a systematic approach that involves initial retrieval and outline generation, subsection drafting by specialized LLMs, integration and refinement, and rigorous evaluation and iteration. Our contributions include a comprehensive solution to the survey problem, a reliable evaluation method, and experimental validation demonstrating AutoSurvey's effectiveness.

Watermarking is a technique that involves embedding nearly unnoticeable statistical signals within generated content to help trace its source. This work focuses on a scenario where an untrusted third-party user sends prompts to a trusted language model (LLM) provider, who then generates a text from their LLM with a watermark. This setup makes it possible for a detector to later identify the source of the text if the user publishes it. The user can modify the generated text by substitutions, insertions, or deletions. Our objective is to develop a statistical method to detect if a published text is LLM-generated from the perspective of a detector. We further propose a methodology to segment the published text into watermarked and non-watermarked sub-strings. The proposed approach is built upon randomization tests and change point detection techniques. We demonstrate that our method ensures Type I and Type II error control and can accurately identify watermarked sub-strings by finding the corresponding change point locations. To validate our technique, we apply it to texts generated by several language models with prompts extracted from Google's C4 dataset and obtain encouraging numerical results. We release all code publicly at https://github.com/doccstat/llm-watermark-cpd.

Position embedding is a core component of current Large Language Models (LLMs). Rotary position embedding (RoPE), a technique that encodes the position information with a rotation matrix, has been the de facto choice for position embedding in many LLMs, such as the Llama series. RoPE has been further utilized to extend long context capability, which is roughly based on adjusting the \textit{base} parameter of RoPE to mitigate out-of-distribution (OOD) problems in position embedding. However, in this paper, we find that LLMs may obtain a superficial long-context ability based on the OOD theory. We revisit the role of RoPE in LLMs and propose a novel property of long-term decay, we derive that the \textit{base of RoPE bounds context length}: there is an absolute lower bound for the base value to obtain certain context length capability. Our work reveals the relationship between context length and RoPE base both theoretically and empirically, which may shed light on future long context training.

We propose **Quan**tum-informed **T**ensor **A**daptation (**QuanTA**), a novel, easy-to-implement, fine-tuning method with no inference overhead for large-scale pre-trained language models. By leveraging quantum-inspired methods derived from quantum circuit structures, QuanTA enables efficient *high-rank* fine-tuning, surpassing the limitations of Low-Rank Adaptation (LoRA)---low-rank approximation may fail for complicated downstream tasks. Our approach is theoretically supported by the universality theorem and the rank representation theorem to achieve efficient high-rank adaptations. Experiments demonstrate that QuanTA significantly enhances commonsense reasoning, arithmetic reasoning, and scalability compared to traditional methods. Furthermore, QuanTA shows superior performance with fewer trainable parameters compared to other approaches and can be designed to integrate with existing fine-tuning algorithms for further improvement, providing a scalable and efficient solution for fine-tuning large language models and advancing state-of-the-art in natural language processing.

The pretraining data of today's strongest language models remains opaque, even when their parameters are open-sourced. In particular, little is known about the proportions of different domains, languages, or code represented in the data. While a long line of membership inference attacks aim to identify training examples on an instance level, they do not extend easily to *global* statistics about the corpus. In this work, we tackle a task which we call *data mixture inference*, which aims to uncover the distributional make-up of the pretraining data. We introduce a novel attack based on a previously overlooked source of information — byte-pair encoding (BPE) tokenizers, used by the vast majority of modern language models. Our key insight is that the ordered vocabulary learned by a BPE tokenizer naturally reveals information about the token frequencies in its training data: the first token is the most common byte pair, the second is the most common pair after merging the first token, and so on. Given a tokenizer's merge list along with data samples for each category of interest (e.g., different natural languages), we formulate a linear program that solves for the relative proportion of each category in the tokenizer's training set. Importantly, to the extent to which tokenizer training data is representative of the pretraining data, we indirectly learn about the pretraining data. In controlled experiments, we show that our attack can recover mixture ratios with high precision for tokenizers trained on known mixtures of natural languages, programming languages, and data sources. We then apply our approach to off-the-shelf tokenizers released alongside recent LMs. We confirm much publicly disclosed information about these models, and also make several new inferences: GPT-4o is much more multilingual than its predecessors, training on 10x more non-English data than GPT-3.5, Llama 3 and Claude are trained on predominantly code, and many recent models are trained on 7-16% books. We hope our work sheds light on current design practices for pretraining data, and inspires continued research into data mixture inference for LMs.

We study interactive learning of language agents based on user edits made to the agent's output. In a typical setting such as writing assistants, the user interacts with a language agent to generate a response given a context, and may optionally edit the agent response to personalize it based on their latent preference, in addition to improving the correctness. The edit feedback is naturally generated, making it a suitable candidate for improving the agent's alignment with the user's preference, and for reducing the cost of user edits over time. We propose a learning framework, PRELUDE that infers a description of the user's latent preference based on historic edit data and using it to define a prompt policy that drives future response generation. This avoids fine-tuning the agent, which is costly, challenging to scale with the number of users, and may even degrade its performance on other tasks. Furthermore, learning descriptive preference improves interpretability, allowing the user to view and modify the learned preference. However, user preference can be complex and vary based on context, making it challenging to learn. To address this, we propose a simple yet effective algorithm named CIPHER that leverages a large language model (LLM) to infer the user preference for a given context based on user edits. In the future, CIPHER retrieves inferred preferences from the k-closest contexts in the history, and forms an aggregate preference for response generation. We introduce two interactive environments -- summarization and email writing, for evaluation using a GPT-4 simulated user. We compare with algorithms that directly retrieve user edits but do not learn descriptive preference, and algorithms that learn context-agnostic preference. On both tasks, CIPHER outperforms baselines by achieving the lowest edit distance cost. Meanwhile, CIPHER has a lower computational expense, as using learned preference results in a shorter prompt than directly using user edits. Our further analysis reports that the user preference learned by CIPHER shows significant similarity to the ground truth latent preference.

A central piece in enabling intelligent agentic behavior in foundation models is to make them capable of introspecting upon their behavior, reasoning, and correcting their mistakes as more computation or interaction is available. Even the strongest proprietary large language models (LLMs) do not quite exhibit the ability of continually improving their responses sequentially. In this paper, we develop $\textbf{RISE:}$ $\textbf{R}$ecursive $\textbf{I}$ntro$\textbf{S}$p$\textbf{E}$ction, an approach for fine-tuning LLMs to introduce this capability, despite prior work hypothesizing that this capability may not be possible to attain. Our approach prescribes an iterative fine-tuning procedure, which attempts to teach the model how to alter its response after having executed previously unsuccessful attempts to solve a hard test-time problem, with optionally additional environment feedback. RISE poses fine-tuning for a single-turn prompt as solving a multi-turn Markov decision process (MDP), where the initial state is the prompt. Inspired by principles in online imitation and offline reinforcement learning, we propose strategies for multi-turn data collection and training so as to imbue an LLM with the capability to recursively detect and correct its previous mistakes in subsequent iterations. Our experiments show that RISE enables Llama2, Llama3, and Mistral models to improve themselves with more turns on reasoning tasks, outperforming several single-turn strategies given an equal amount of inference-time computation. We also find that RISE scales well, often attaining larger benefits with more capable models, without disrupting one-turn abilities as a result of expressing more complex distributions.

\emph{Metacognitive knowledge} refers to humans' intuitive knowledge of their own thinking and reasoning processes. Today's best LLMs clearly possess some reasoning processes. The paper gives evidence that they also have metacognitive knowledge, including ability to name skills and procedures to apply given a task. We explore this primarily in context of math reasoning, developing a prompt-guided interaction procedure to get a powerful LLM to assign sensible skill labels to math questions, followed by having it perform semantic clustering to obtain coarser families of skill labels. These coarse skill labels look interpretable to humans. To validate that these skill labels are meaningful and relevant to the LLM's reasoning processes we perform the following experiments. (a) We ask GPT-4 to assign skill labels to training questions in math datasets GSM8K and MATH. (b) When using an LLM to solve the test questions, we present it with the full list of skill labels and ask it to identify the skill needed. Then it is presented with randomly selected exemplar solved questions associated with that skill label. This improves accuracy on GSM8k and MATH for several strong LLMs, including code-assisted models. The methodology presented is domain-agnostic, even though this article applies it to math problems.

Temporal Knowledge Graph Reasoning (TKGR) is the process of utilizing temporal information to capture complex relations within a Temporal Knowledge Graph (TKG) to infer new knowledge. Conventional methods in TKGR typically depend on deep learning algorithms or temporal logical rules. However, deep learning-based TKGRs often lack interpretability, whereas rule-based TKGRs struggle to effectively learn temporal rules that capture temporal patterns. Recently, Large Language Models (LLMs) have demonstrated extensive knowledge and remarkable proficiency in temporal reasoning. Consequently, the employment of LLMs for Temporal Knowledge Graph Reasoning (TKGR) has sparked increasing interest among researchers. Nonetheless, LLMs are known to function as black boxes, making it challenging to comprehend their reasoning process. Additionally, due to the resource-intensive nature of fine-tuning, promptly updating LLMs to integrate evolving knowledge within TKGs for reasoning is impractical. To address these challenges, in this paper, we propose a Large Language Models-guided Dynamic Adaptation (LLM-DA) method for reasoning on TKGs. Specifically, LLM-DA harnesses the capabilities of LLMs to analyze historical data and extract temporal logical rules. These rules unveil temporal patterns and facilitate interpretable reasoning. To account for the evolving nature of TKGs, a dynamic adaptation strategy is proposed to update the LLM-generated rules with the latest events. This ensures that the extracted rules always incorporate the most recent knowledge and better generalize to the predictions on future events. Experimental results show that without the need of fine-tuning, LLM-DA significantly improves the accuracy of reasoning over several common datasets, providing a robust framework for TKGR tasks.

In the face of uncertainty, the ability to *seek information* is of fundamental importance. In many practical applications, such as medical diagnosis and troubleshooting, the information needed to solve the task is not initially given, and has to be actively sought by asking follow-up questions (for example, a doctor asking a patient for more details about their symptoms). In this work, we introduce **Uncertainty of Thoughts (UoT)**, an algorithm to augment large language models with the ability to actively seek information by asking effective questions. UoT combines: 1. An *uncertainty-aware simulation approach* which enables the model to simulate possible future scenarios and how likely they are to occur, 2. *Uncertainty-based rewards* motivated by information gain which incentivizes the model to seek information, and 3. A *reward propagation scheme* to select the optimal question to ask in a way that maximizes the expected reward. In experiments on medical diagnosis, troubleshooting and the `20 Questions' game, UoT achieves an average performance improvement of 38.1% in the rate of successful task completion across multiple LLMs compared with direct prompting, and also improves efficiency (i.e., the number of questions needed to complete the task).

Safety of Large Language Models (LLMs) has become a central issue given their rapid progress and wide applications. Greedy Coordinate Gradient (GCG) is shown to be effective in constructing prompts containing adversarial suffixes to break the presumingly safe LLMs, but the optimization of GCG is time-consuming and limits its practicality. To reduce the time cost of GCG and enable more comprehensive studies of LLM safety, in this work, we study a new algorithm called $\texttt{Probe sampling}$ to accelerate the GCG algorithm. At the core of the algorithm is a mechanism that dynamically determines how similar a smaller draft model's predictions are to the target model's predictions for prompt candidates. When the target model is similar to the draft model, we rely heavily on the draft model to filter out a large number of potential prompt candidates to reduce the computation time. Probe sampling achieves up to $5.6$ times speedup using Llama2-7b-chat and leads to equal or improved attack success rate (ASR) on the AdvBench. Furthermore, probe sampling is also able to accelerate other prompt optimization techniques and adversarial attack methods, leading to acceleration of $1.8\times$ for AutoPrompt, $2.4\times$ for APE and $2.4\times$ for AutoDAN.

Personalization has emerged as a critical research area in modern intelligent systems, focusing on mining users' behavioral history and adapting to their preferences for delivering tailored experiences. Despite the remarkable few-shot capabilities exhibited by black-box large language models (LLMs), the inherent opacity of their model parameters presents significant challenges in aligning the generated output with individual expectations. Existing solutions have primarily focused on prompt design to incorporate user-specific profiles and behaviors; however, such approaches often struggle to generalize effectively due to their inability to capture shared knowledge among all users. To address these challenges, we propose HYDRA, a model factorization framework that captures both user-specific behavior patterns from historical data and shared general knowledge among all users to deliver personalized generation. In order to capture user-specific behavior patterns, we first train a reranker to prioritize the most useful information from top-retrieved relevant historical records. By combining the prioritized history with the corresponding query, we train an adapter to align the output with individual user-specific preferences, eliminating the reliance on access to inherent model parameters of black-box LLMs. Both the reranker and the adapter can be decomposed into a base model with multiple user-specific heads, resembling a hydra. The base model maintains shared knowledge across users, while the multiple personal heads capture user-specific preferences. Experimental results demonstrate that \method outperforms existing state-of-the-art prompt-based methods by an average relative improvement of 9.01% across five diverse personalization tasks in the LaMP benchmark.

A critical question about Large Language Models (LLMs) is whether their apparent deficiency in mathematical reasoning is inherent, or merely a result of insufficient exposure to high-quality mathematical data. To explore this, we developed an automated method for generating high-quality, supervised mathematical datasets. The method carefully mutates existing math problems, ensuring both diversity and validity of the newly generated problems. This is achieved by a neuro-symbolic data generation framework combining the intuitive informalization strengths of LLMs, and the precise symbolic reasoning of math solvers along with projected Markov chain Monte Carlo sampling in the highly-irregular symbolic space. Empirical experiments demonstrate the high quality of data generated by the proposed method, and that the LLMs, specifically LLaMA-2 and Mistral, when realigned with the generated data, surpass their state-of-the-art counterparts.

Large language models (LLMs) have exhibited impressive performance in language comprehension and various reasoning tasks. However, their abilities in spatial reasoning, a crucial aspect of human cognition, remain relatively unexplored. Human possess a remarkable ability to create mental images of unseen objects and actions through a process known as the Mind's Eye, enabling the imagination of the unseen world. Inspired by this cognitive capacity, we propose Visualization-of-Thought (VoT) prompting. VoT aims to elicit spatial reasoning of LLMs by visualizing their reasoning traces, thereby guiding subsequent reasoning steps. We employed VoT for multi-hop spatial reasoning tasks, including natural language navigation, visual navigation, and visual tiling in 2D grid worlds. Experimental results demonstrated that VoT significantly enhances the spatial reasoning abilities of LLMs. Notably, VoT outperformed existing multimodal large language models (MLLMs) in these tasks. While VoT works surprisingly well on LLMs, the ability to generate mental images to facilitate spatial reasoning resembles the mind's eye process, suggesting its potential viability in MLLMs. Please find the dataset and codes in our [project page](https://microsoft.github.io/visualization-of-thought).

Mixture-of-Experts (MoE) has emerged as a prominent architecture for scaling model size while maintaining computational efficiency. In MoE, each token in the input sequence activates a different subset of experts determined by a routing mechanism. However, the unchosen experts in MoE models do not contribute to the output, potentially leading to underutilization of the model's capacity. In this work, we first conduct exploratory studies to demonstrate that increasing the number of activated experts does not necessarily improve and can even degrade the output quality. Then, we show that output distributions from an MoE model using different routing strategies substantially differ, indicating that different experts do not always act synergistically. Motivated by these findings, we propose **S**elf-**C**ontrast **M**ixture-**o**f-**E**xperts (SCMoE), a training-free strategy that utilizes unchosen experts in a self-contrast manner during inference. In SCMoE, the next-token probabilities are determined by contrasting the outputs from strong and weak activation using the same MoE model. Our method is conceptually simple and computationally lightweight, as it incurs minimal latency compared to greedy decoding. Experiments on several benchmarks (GSM8K, StrategyQA, MBPP and HumanEval) demonstrate that SCMoE can consistently enhance Mixtral 8x7B’s reasoning capability across various domains. For example, it improves the accuracy on GSM8K from 61.79 to 66.94. Moreover, combining SCMoE with self-consistency yields additional gains, increasing major@20 accuracy from 75.59 to 78.31.

With the widespread adoption of Large Language Models (LLMs), the prevalence of iterative interactions among these models is anticipated to increase. Notably, recent advancements in multi-round on-policy self-improving methods allow LLMs to generate new examples for training subsequent models. At the same time, multi-agent LLM systems, involving automated interactions among agents, are also increasing in prominence. Thus, in both short and long terms, LLMs may actively engage in an evolutionary process. We draw parallels between the behavior of LLMs and the evolution of human culture, as the latter has been extensively studied by cognitive scientists for decades. Our approach involves leveraging Iterated Learning (IL), a Bayesian framework that elucidates how subtle biases are magnified during human cultural evolution, to explain some behaviors of LLMs. This paper outlines key characteristics of agents' behavior in the Bayesian-IL framework, including predictions that are supported by experimental verification with various LLMs. This theoretical framework could help to more effectively predict and guide the evolution of LLMs in desired directions.

We introduce SELF-DISCOVER, a general framework for LLMs to self-discover the task-intrinsic reasoning structures to tackle complex reasoning problems that are challenging for typical prompting methods. Core to the framework is a self-discovery process where LLMs select multiple atomic reasoning modules such as critical thinking and step-by-step thinking, and compose them into an explicit reasoning structure for LLMs to follow during decoding. SELF-DISCOVER substantially improves GPT-4 and PaLM 2’s performance on challenging reasoning benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH, by as much as 32% compared to Chain of Thought (CoT). Furthermore, SELF-DISCOVER outperforms inference-intensive methods such as CoT-Self-Consistency by more than 20%, while requiring 10-40x fewer inference compute. Finally, we show that the self-discovered reasoning structures are universally applicable across model families: from PaLM 2-L to GPT-4, and from GPT-4 to Llama2, and share commonalities with human reasoning patterns.

Mixture of Experts (MoE) framework has become a popular architecture for large language models due to its superior performance compared to dense models. However, training MoEs from scratch in a large-scale regime is prohibitively expensive. Previous work addresses this challenge by independently training multiple dense expert models and using them to initialize an MoE. In particular, state-of-the-art approaches initialize MoE layers using experts' feed-forward parameters while merging all other parameters, limiting the advantages of the specialized dense models when upcycling them as MoEs. We propose BAM (Branch-Attend-Mix), a simple yet effective improvement to MoE training. BAM makes full use of specialized dense models by not only using their feed-forward network (FFN) to initialize the MoE layers but also leveraging experts' attention weights fully by leveraging them as mixture-of-attention (MoA) layers. We explore two methods for upcycling MoA layers: 1) initializing separate attention experts from dense models including key, value, and query matrices; and 2) initializing only Q projections while sharing key-value pairs across all experts to facilitate efficient inference. Our experiments using seed models ranging from 590 million to 2 billion parameters show that our approach outperforms state-of-the-art approaches under the same data and compute budget in both perplexity and downstream tasks evaluations, confirming the effectiveness of BAM.