Reward models trained on human preference data have been proven to effectively align Large Language Models (LLMs) with human intent within the framework of reinforcement learning from human feedback (RLHF). However, current reward models have limited generalization capabilities to unseen prompts and responses, which can lead to an unexpected phenomenon known as reward over-optimization, resulting in a decline in actual performance due to excessive optimization of rewards. While previous research has advocated for constraining policy optimization, our study introduces a novel approach to enhance the reward model's generalization ability against distribution shifts by regularizing the hidden states. Specifically, we retain the base model's language model head and incorporate a suite of text-generation losses to preserve the hidden states' text-generation capabilities, while concurrently learning a reward head behind the same hidden states. Our experimental results demonstrate that the introduced regularization technique markedly improves the accuracy of learned reward models across a variety of out-of-distribution (OOD) tasks and effectively alleviates the over-optimization issue in RLHF, offering a more reliable and robust preference learning paradigm.

While many contemporary large language models (LLMs) can process lengthy input, they still struggle to fully utilize information within the long context, known as the challenge.We hypothesize that it stems from insufficient explicit supervision during the long-context training, which fails to emphasize that any position in a long context can hold crucial information.Based on this intuition, our study presents information-intensive (IN2) training

Web-crawled pretraining datasets underlie the impressive zero-shot evaluation performance of multimodal models, such as CLIP for classification and Stable-Diffusion for image generation. However, it is unclear how meaningful the notion of zero-shot generalization is for such multimodal models, as it is not known to what extent their pretraining datasets encompass the downstream concepts targeted for during zero-shot evaluation. In this work, we ask: How is the performance of multimodal models on downstream concepts influenced by the frequency of these concepts in their pretraining datasets?We comprehensively investigate this question across 34 models and 5 standard pretraining datasets (CC-3M, CC-12M, YFCC-15M, LAION-400M, LAION-Aesthetics), generating over 300GB of data artifacts. We consistently find that, far from exhibiting zero-shot generalization, multimodal models require exponentially more data to achieve linear improvements in downstream zero-shot performance, following a sample inefficient log-linear scaling trend. This trend persists even when controlling for sample-level similarity between pretraining and downstream datasets, and testing on purely synthetic data distributions. Furthermore, upon benchmarking models on long-tailed data sampled based on our analysis, we demonstrate that multimodal models across the board perform poorly. We contribute this long-tail test set as the Let it Wag! benchmark to further research in this direction. Taken together, our study reveals an exponential need for training data which implies that the key to zero-shot generalization capabilities under large-scale training data and compute paradigms remains to be found.

Large language models (LLMs) have made fundamental contributions over the last a few years. To train an LLM, one needs to alternatingly run `forward' computations and backward computations. The forward computation can be viewed as attention function evaluation, and the backward computation can be viewed as a gradient computation. In previous work by [Alman and Song, NeurIPS 2023], it was proved that the forward step can be performed in almost-linear time in certain parameter regimes, but that there is no truly sub-quadratic time algorithm in the remaining parameter regimes unless the popular hypothesis $\mathsf{SETH}$ is false. In this work, we show nearly identical results for the harder-seeming problem of computing the gradient of loss function of one layer attention network, and thus for the entire process of LLM training. This completely characterizes the fine-grained complexity of every step of LLM training.

When prompting a language model (LM), users often expect the model to adhere to a set of behavioral principles across diverse tasks, such as producing insightful content while avoiding harmful or biased language. Instilling such principles (i.e., a constitution) into a model is resource-intensive, technically challenging, and generally requires human preference labels or examples. We introduce SAMI, an iterative algorithm that finetunes a pretrained language model (without requiring preference labels or demonstrations) to increase the conditional mutual information between constitutions and self-generated responses given queries from a dataset. On single-turn dialogue and summarization, a SAMI-trained mistral-7b outperforms the initial pretrained model, with win rates between 66% and 77%.

While Large Language Models (LLMs) display versatile functionality, they continue to generate harmful, biased, and toxic content, as demonstrated by the prevalence of human-designed . In this work, we present (TAP), an automated method for generating jailbreaks that only requires black-box access to the target LLM. TAP utilizes an attacker LLM to iteratively refine candidate (attack) prompts until one of the refined prompts jailbreaks the target. In addition, before sending prompts to the target, TAP assesses them and prunes the ones unlikely to result in jailbreaks, reducing the number of queries sent to the target LLM. In empirical evaluations, we observe that TAP generates prompts that jailbreak state-of-the-art LLMs (including GPT4-Turbo and GPT4o) for more than 80% of the prompts. This significantly improves upon the previous state-of-the-art black-box methods for generating jailbreaks while using a smaller number of queries than them. Furthermore, TAP is also capable of jailbreaking LLMs protected by state-of-the-art, e.g., LlamaGuard.

The increasing size of language models necessitates a thorough analysis across multiple dimensions to assess trade-offs among crucial hardware metrics such as latency, energy consumption, GPU memory usage, and performance. Identifying optimal model configurations under specific hardware constraints is becoming essential but remains challenging due to the computational load of exhaustive training and evaluation on multiple devices. To address this, we introduce HW-GPT-Bench, a hardware-aware benchmark that utilizes surrogate predictions to approximate various hardware metrics across 13 devices of architectures in the GPT-2 family, with architectures containing up to 1.55B parameters. Our surrogates, via calibrated predictions and reliable uncertainty estimates, faithfully model the heteroscedastic noise inherent in the energy and latency measurements. To estimate perplexity, we employ weight-sharing techniques from Neural Architecture Search (NAS), inheriting pretrained weights from the largest GPT-2 model. Finally, we demonstrate the utility of HW-GPT-Bench by simulating optimization trajectories of various multi-objective optimization algorithms in just a few seconds.

This paper aims to overcome the ``lost-in-the-middle'' challenge of large language models (LLMs). While recent advancements have successfully enabled LLMs to perform stable language modeling with up to 4 million tokens, the persistent difficulty faced by most LLMs in identifying relevant information situated in the middle of the context has not been adequately tackled. To address this problem, this paper introduces Multi-scale Positional Encoding (Ms-PoE) which is a simple yet effective plug-and-play approach to enhance the capacity of LLMs to handle the relevant information located in the middle of the context, without fine-tuning or introducing any additional overhead. Ms-PoE leverages the position indice rescaling to relieve the long-term decay effect introduced by RoPE, while meticulously assigning distinct scaling ratios to different attention heads to preserve essential knowledge learned during the pre-training step, forming a multi-scale context fusion from short to long distance. Extensive experiments with a wide range of LLMs demonstrate the efficacy of our approach. Notably, Ms-PoE achieves an average accuracy gain of up to 3.8 on the Zero-SCROLLS benchmark over the original LLMs. Code will be made public upon acceptence.

Despite the recent observation that large language models (LLMs) can store substantial factual knowledge, there is a limited understanding of the mechanisms of how they acquire factual knowledge through pretraining. This work addresses this gap by studying how LLMs acquire factual knowledge during pretraining. The findings reveal several important insights into the dynamics of factual knowledge acquisition during pretraining. First, counterintuitively, we observe that pretraining on more data shows no significant improvement in the model's capability to acquire and maintain factual knowledge. Next, LLMs undergo forgetting of memorization and generalization of factual knowledge, and LLMs trained with duplicated training data exhibit faster forgetting. Third, training LLMs with larger batch sizes can enhance the models' robustness to forgetting. Overall, our observations suggest that factual knowledge acquisition in LLM pretraining occurs by progressively increasing the probability of factual knowledge presented in the pretraining data at each step. However, this increase is diluted by subsequent forgetting. Based on this interpretation, we demonstrate that we can provide plausible explanations on recently observed behaviors of LLMs, such as the poor performance of LLMs on long-tail knowledge and the benefits of deduplicating the pretraining corpus.

This paper reveals the phenomenon of parameter heterogeneity in large language models (LLMs). We find that a small subset of ``cherry'' parameters exhibit a disproportionately large influence on model performance, while the vast majority of parameters have minimal impact. This heterogeneity is found to be prevalent across different model families, scales, and types. Motivated by this observation, we propose CherryQ, a novel quantization method that unifies the optimization of mixed-precision parameters. CherryQ identifies and preserves the critical cherry parameters in high precision while aggressively quantizing the remaining parameters to low precision. Extensive experiments demonstrate the effectiveness of CherryQ. CherryQ outperforms existing quantization approaches in terms of perplexity and downstream task performance. Notably, our 3-bit quantized Vicuna-1.5 exhibits competitive performance compared to their 16-bit counterparts.