ICML 2024 Papers

Large language models (LLMs) demonstrate emergent in-context learning capabilities, where they adapt to new tasks based on example demonstrations. However, in-context learning has seen limited effectiveness in many settings, is difficult to quantitatively control and takes up context window space. To overcome these limitations, we propose an alternative approach that recasts in-context learning as in-context vectors (ICV). Using ICV has two steps. We first use a forward pass on demonstration examples to create the in-context vector from the latent embedding of the LLM. This vector captures essential information about the intended task. On a new query, instead of adding demonstrations to the prompt, we shift the latent states of the LLM using the ICV. The ICV approach has several benefits: 1) it enables the LLM to more effectively follow the demonstration examples; 2) it's easy to control by adjusting the magnitude of the ICV; 3) it reduces the length of the prompt by removing the in-context demonstrations; 4) ICV is computationally much more efficient than fine-tuning. We demonstrate that ICV achieves better performance compared to standard in-context learning and fine-tuning on diverse tasks including safety, style transfer, role-playing and formatting. Moreover, we show that we can flexibly teach LLM to simultaneously follow different types of instructions by simple vector arithmetics on the corresponding ICVs.

The fine-tuning paradigm in addressing long-tail learning tasks has sparked significant interest since the emergence of foundation models. Nonetheless, how fine-tuning impacts performance in long-tail learning was not explicitly quantified. In this paper, we disclose that heavy fine-tuning may even lead to non-negligible performance deterioration on tail classes, and lightweight fine-tuning is more effective. The reason is attributed to inconsistent class conditions caused by heavy fine-tuning. With the observation above, we develop a low-complexity and accurate long-tail learning algorithms LIFT with the goal of facilitating fast prediction and compact models by adaptive lightweight fine-tuning. Experiments clearly verify that both the training time and the learned parameters are significantly reduced with more accurate predictive performance compared with state-of-the-art approaches. The implementation code is available at https://github.com/shijxcs/LIFT.

In real-world applications, time series data frequently exhibit non-stationarity, with statistics changing over time. This variability undermines the forecasting accuracy of deep learning models that are trained on historical data but deployed for future prediction. A common approach to mitigate this issue involves normalizing the data to counteract statistical drift, followed by denormalization on the prediction. However, existing methods often employ heuristic normalization techniques that do not fully account for the unique characteristics of the series. Our paper addresses the critical question in this context: which statistics should be removed and restored? We argue that the statistics selected for normalization should exhibit both local invariance and global variability to ensure their correctness and helpfulness. To this end, we propose the Selective and Interpretable Normalization methodology, dubbed SIN. This approach maximizes the covariance between a given look-back window and its subsequent future values, thereby identifying key statistics for normalization and simultaneously learning the corresponding normalization transformations. The interpretable framework can be used to explain the success and limitations of some popular normalization methods. By integrating SIN, we demonstrate improvements in the performance of several prevalent forecasting models, thereby validating the utility of our approach.

In-context learning (ICL) has emerged as a particularly remarkable characteristic of Large Language Models (LLM): given a pretrained LLM and an observed dataset, LLMs can make predictions for new data points from the same distribution without fine-tuning. Numerous works have postulated ICL as approximately Bayesian inference, rendering this a natural hypothesis. In this work, we analyse this hypothesis from a new angle through the *martingale property*, a fundamental requirement of a Bayesian learning system for exchangeable data. We show that the martingale property is a necessary condition for unambiguous predictions in such scenarios, and enables a principled, decomposed notion of uncertainty vital in trustworthy, safety-critical systems. We derive actionable checks with corresponding theory and test statistics which must hold if the martingale property is satisfied. We also examine if uncertainty in LLMs decreases as expected in Bayesian learning when more data is observed. In three experiments, we provide evidence for violations of the martingale property, and deviations from a Bayesian scaling behaviour of uncertainty, falsifying the hypothesis that ICL is Bayesian.

Many neural network architectures are known to be Turing Complete, and can thus, in principle implement arbitrary algorithms. However, Transformers are unique in that they can implement gradient-based learning algorithms *under simple parameter configurations*. This paper provides theoretical and empirical evidence that (non-linear) Transformers naturally learn to implement gradient descent *in function space*, which in turn enable them to learn non-linear functions in context. Our results apply to a broad class of combinations of non-linear architectures and non-linear in-context learning tasks. Additionally, we show that the optimal choice of non-linear activation depends in a natural way on the class of functions that need to be learned.

Recently, extensive studies in Reinforcement Learning have been carried out on the ability of transformers to adapt in-context to various environments and tasks. Current in-context RL methods are limited by their strict requirements for data, which needs to be generated by RL agents or labeled with actions from an optimal policy. In order to address this prevalent problem, we propose AD$^\varepsilon$, a new data acquisition approach that enables in-context Reinforcement Learning from noise-induced curriculum. We show that it is viable to construct a synthetic noise injection curriculum which helps to obtain learning histories. Moreover, we experimentally demonstrate that it is possible to alleviate the need for generation using optimal policies, with in-context RL still able to outperform the best suboptimal policy in a learning dataset by a 2x margin.

The rise of foundation models holds immense promise for advancing AI, but this progress may amplify existing risks and inequalities, leaving marginalized communities behind. In this position paper, we discuss that disparities towards marginalized communities – performance, representation, privacy, robustness, interpretability and safety – are not isolated concerns but rather interconnected elements of a cascading disparity phenomenon. We contrast foundation models with traditional models and highlight the potential for exacerbated disparity against marginalized communities. Moreover, we emphasize the unique threat of cascading impacts in foundation models, where interconnected disparities can trigger long-lasting negative consequences, specifically to the people on the margin. We define marginalized communities within the machine learning context and explore the multifaceted nature of disparities. We analyze the sources of these disparities, tracing them from data creation, training and deployment procedures to highlight the complex technical and socio-technical landscape. To mitigate the pressing crisis, we conclude with a set of calls to action to mitigate disparity at its source.

The emergence of In-Context Learning (ICL) in LLMs remains a remarkable phenomenon that is partially understood. To explain ICL, recent studies have created theoretical connections to Gradient Descent (GD). We ask, do such connections hold up in actual pre-trained language models? We highlight the limiting assumptions in prior works that make their setup considerably different from the practical setup in which language models are trained. For example, their experimental verification uses *ICL objective* (training models explicitly for ICL), which differs from the emergent ICL in the wild. Furthermore, the theoretical hand-constructed weights used in these studies have properties that don't match those of real LLMs. We also look for evidence in real models. We observe that ICL and GD have different sensitivity to the order in which they observe demonstrations. Finally, we probe and compare the ICL vs. GD hypothesis in a natural setting. We conduct comprehensive empirical analyses on language models pre-trained on natural data (LLaMa-7B). Our comparisons of three performance metrics highlight the inconsistent behavior of ICL and GD as a function of various factors such as datasets, models, and the number of demonstrations. We observe that ICL and GD modify the output distribution of language models differently. These results indicate that *the equivalence between ICL and GD remains an open hypothesis* and calls for further studies.

Large language models (LLM) have emerged as a powerful tool for AI, with the key ability of in-context learning (ICL), where they can perform well on unseen tasks based on a brief series of task examples without necessitating any adjustments to the model parameters. One recent interesting mysterious observation is that models of different scales may have different ICL behaviors: larger models tend to be more sensitive to noise in the test context. This work studies this observation theoretically aiming to improve the understanding of LLM and ICL. We analyze two stylized settings: (1) linear regression with one-layer single-head linear transformers and (2) parity classification with two-layer multiple attention heads transformers (non-linear data and non-linear model). In both settings, we give closed-form optimal solutions and find that smaller models emphasize important hidden features while larger ones cover more hidden features; thus, smaller models are more robust to noise while larger ones are more easily distracted, leading to different ICL behaviors. This sheds light on where transformers pay attention to and how that affects ICL. Preliminary experimental results on large base and chat models provide positive support for our analysis.

NLP has recently made exciting progress toward training language models (LMs) with strong scientific problem-solving skills. However, model development has not focused on real-life use-cases of LMs for science, including applications in education that require processing long scientific documents. To address this, we introduce TutorEval and TutorChat. TutorEval is a diverse question-answering benchmark consisting of questions about long chapters from STEM textbooks, written by experts. TutorEval helps measure real-life usability of LMs as scientific assistants, and it is the first benchmark combining long contexts, free-form generation, and multi-disciplinary scientific knowledge. Moreover, we show that fine-tuning base models with existing dialogue datasets leads to poor performance on TutorEval. Therefore, we create TutorChat, a dataset of 80,000 long synthetic dialogues about textbooks. We use TutorChat to fine-tune Llemma models with 7B and 34B parameters. These LM tutors specialized in math have a 32K-token context window, and they excel at TutorEval while performing strongly on GSM8K and MATH. Our datasets build on open-source materials, and we release our models, data, and evaluations publicly.

In-Context Learning (ICL) is the ability of Large Language Models (LLMs) to perform new tasks when conditioned on prompts comprising a few task examples. However, ICL performance can be critically sensitive to the choice of examples. To dynamically select the best examples for every test input, we propose Example Gisting, a novel approach for training example encoders through supervised finetuning with an attention bottleneck between the inputs and outputs. These gist models form the basis for GistScore, a novel metric for scoring and selecting informative examples. Further, we experiment with two variations: (1) finetuning gist models for each dataset and (2) multi-task training a single model on a large collection of datasets. The latter can be used for new tasks out-of-the-box, enabling a training-free ICL pipeline. Evaluations with 21 datasets spanning 9 tasks and 8 diverse LLMs show that our fine-tuned models get state-of-the-art ICL performance with over 20% absolute gain over off-the-shelf retrievers and 5% over the best prior methods. Further, our multi-task model generalizes well to new tasks, datasets, and prompt templates. Selection using this model matches or outperforms prior methods while being three orders of magnitude faster than the strongest training-free baseline.

We study 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 *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 *quantity* and *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 *domain balance* and *length upsampling*. Concretely, naïvely upsampling longer data on certain domains like books, a common practice of existing work, gives suboptimal performance; a balanced domain mixture is equally 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.

To mitigate the computational complexity in the self-attention mechanism on long sequences, linear attention utilizes computation tricks to achieve linear complexity, while state space models (SSMs) popularize a favourable practice of using non-data-dependent memory pattern, *i.e.,* emphasize the near and neglect the distant, to processing sequences. Recent studies have shown the priorities by combining them as one. However, the efficiency of linear attention remains only at the theoretical level in a causal setting, and SSMs require various designed constraints to operate effectively on specific data. Therefore, in order to unveil the true power of the hybrid design, the following two issues need to be addressed: (1) hardware-efficient implementation for linear attention and (2) stabilization of SSMs. To achieve this, we leverage the thought of tiling and hierarchy to propose CHELA (short-long Convolutions with Hardware-Efficient Linear Attention), which replaces SSMs with short-long convolutions and implements linear attention in a divide-and-conquer manner. This approach enjoys global abstraction and data-dependent selection from stable SSM and linear attention while maintaining real linear complexity. Our comprehensive experiments on the Long Range Arena benchmark and language modeling tasks demonstrate the effectiveness of the proposed method.

In-context learning (ICL, also known as few-shot prompting) has been the standard method of adapting LLMs to downstream tasks, by learning from a few input-output examples. Nonetheless, all ICL-based approaches only learn from correct input-output pairs. In this paper, we revisit this paradigm, by learning more from the few given input-output examples. We introduce Learning Principles (LEAP): First, we intentionally induce the model to make mistakes on these few examples; then we reflect on these mistakes, and learn explicit task-specific “principles” from them, which help solve similar problems and avoid common mistakes; finally, we prompt the model to answer unseen test questions using the original few-shot examples and these learned general principles. We evaluate LEAP on a wide range of benchmarks, including multi-hop question answering (Hotpot QA), textual QA (DROP), Big-Bench Hard reasoning, and math problems (GSM8K and MATH); in all these benchmarks, LEAP improves the strongest available LLMs such as GPT-3.5-turbo, GPT-4, GPT-4-turbo and Claude-2.1. For example, LEAP improves over the standard few-shot prompting using GPT-4 by 7.5% in DROP, and by 3.3% in HotpotQA. Importantly, LEAP does not require any more input or examples than the standard few-shot prompting settings.

Current Large Language Models (LLMs) are not only limited to some maximum context length, but also are not able to robustly consume long inputs. To address these limitations, we propose ReadAgent, an LLM agent system that increases effective context length up to 20x in our experiments. Inspired by how humans interactively read long documents, we implement ReadAgent as a simple prompting system that uses the advanced language capabilities of LLMs to (1) decide what content to store together in a memory episode, (2) compress those memory episodes into short episodic memories called *gist memories*, and (3) take actions to look up passages in the original text if ReadAgent needs to remind itself of relevant details to complete a task. We evaluate ReadAgent against baselines using retrieval methods, using the original long contexts, and using the gist memories. These evaluations are performed on three long-document reading comprehension tasks: QuALITY, NarrativeQA, and QMSum. ReadAgent outperforms the baselines on all three tasks while extending the effective context window by 3.5-20x.

Large context window is a desirable feature in large language models (LLMs). However, due to high fine-tuning costs, scarcity of long texts, and catastrophic values introduced by new token positions, current extended context windows are limited to around 128k tokens. This paper introduces LongRoPE that, for the first time, extends the context window of pre-trained LLMs to an impressive 2048k tokens, with up to only 1k fine-tuning steps at within 256k training lengths, while maintaining performance at the original short context window. This is achieved by three key innovations: (i) we identify and exploit two forms of non-uniformities in positional interpolation through an efficient search, providing a better initialization for fine-tuning and enabling an 8x extension in non-fine-tuning scenarios; (ii) we introduce a progressive extension strategy that first fine-tunes a 256k length LLM and then conducts a second positional interpolation on the fine-tuned extended LLM to achieve a 2048k context window; (iii) we readjust LongRoPE on 8k length to recover the short context window performance. Extensive experiments on LLaMA2 and Mistral across various tasks demonstrate the effectiveness of our method. Models extended via LongRoPE retain the original architecture with minor modifications to the positional embedding, and can reuse most pre-existing optimizations. Code is available at https://github.com/microsoft/LongRoPE

Recent advancements in semi-supervised learning have focused on a more realistic yet challenging task: addressing imbalances in labeled data while the class distribution of unlabeled data remains both unknown and potentially mismatched. Current approaches in this sphere often presuppose rigid assumptions regarding the class distribution of unlabeled data, thereby limiting the adaptability of models to only certain distribution ranges. In this study, we propose a novel approach, introducing a highly adaptable framework, designated as **SimPro**, which does not rely on any predefined assumptions about the distribution of unlabeled data. Our framework, grounded in a probabilistic model, innovatively refines the expectation-maximization (EM) method by separating the modeling of conditional and marginal class distributions. This separation facilitates a closed-form solution for class distribution estimation during the maximization phase, leading to the formulation of a Bayes classifier. The Bayes classifier, in turn, enhances the quality of pseudo-labels in the expectation phase. Remarkably, the SimPro framework is not only straightforward to implement but also comes with theoretical guarantees. Moreover, we introduce two novel class distributions broadening the scope of the evaluation. Our method showcases consistent state-of-the-art performance across diverse benchmarks and data distribution scenarios. benchmarks and data distribution scenarios. Our code is available at https://github.com/LeapLabTHU/SimPro.

This paper tackles the memory hurdle of of processing long context sequences in Large Language Models (LLMs), by presenting a novel approach, Dropping In Convolutions for **Lo**ng **Co**ntext **Co**mpression (**LoCoCo**). LoCoCo employs only a fixed-size Key-Value (KV) cache, and can enhance efficiency in both inference and fine-tuning stages. Diverging from prior methods that selectively drop KV pairs based on heuristics, LoCoCo leverages a data-driven adaptive fusion technique, blending previous KV pairs with incoming tokens to minimize the loss of contextual information and ensure accurate attention modeling. This token integration is achieved through injecting one-dimensional convolutional kernels that dynamically calculate mixing weights for each KV cache slot. Designed for broad compatibility with existing LLM frameworks, LoCoCo allows for straightforward "drop-in" integration without needing architectural modifications, while incurring minimal tuning overhead. Experiments demonstrate that LoCoCo maintains consistently outstanding performance across various context lengths and can achieve a high context compression rate during both inference and fine-tuning phases. During inference, we successfully compressed up to $3482$ tokens into a $128$-size KV cache, while retaining comparable performance to the full sequence - an accuracy improvement of up to $0.2791$ compared to baselines at the same cache size. During post-training tuning, we also effectively extended the context length from 4K to 32K using a KV cache of fixed size 512, achieving performance similar to fine-tuning with entire sequences.

Previous theoretical results pertaining to meta-learning on sequences build on contrived and convoluted mixing time assumptions. We introduce new information-theoretic tools that lead to a concise yet general decomposition of error for a Bayes optimal predictor into two components: meta-learning error and intra-task error. These tools unify analyses across many meta-learning challenges. To illustrate, we apply them to establish new results about in-context learning with transformers and corroborate existing results a simple linear setting. Our theoretical results characterize how error decays in both the number of training sequences and sequence lengths. Our results are very general; for example, they avoid contrived mixing time assumptions made by all prior results that establish decay of error with sequence length.