Searching through chemical space is an exceptionally challenging problem because the number of possible molecules grows combinatorially with the number of atoms. Large, autoregressive models trained on databases of chemical compounds have yielded powerful generators, but we still lack robust strategies for generating molecules with desired properties. This molecular search problem closely resembles the "alignment" problem for large language models, though for many chemical tasks we have a specific and easily evaluable reward function. Here, we introduce an algorithm called energy rank alignment (ERA) that leverages an explicit reward function to produce a gradient-based objective that we use to optimize autoregressive policies. We show theoretically that this algorithm is closely related to proximal policy optimization (PPO) and direct preference optimization (DPO), but has a minimizer that converges to an ideal Gibbs-Boltzmann distribution with the reward playing the role of an energy function. Furthermore, this algorithm is highly scalable, does not require reinforcement learning, and performs well relative to DPO when the number of preference observations per pairing is small. We deploy this approach to align molecular transformers to generate molecules with externally specified properties and find that it does so robustly, searching through diverse parts of chemical space. While our focus here is on chemical search, we also obtain excellent results on an AI supervised task for LLM alignment, showing that the method is scalable and general.

Large Language Models (LLMs) are frequently discussed in academia and the general public as support tools for virtually any use case that relies on the production of text, including software engineering. Currently there is much debate, but little empirical evidence, regarding the practical usefulness of LLM-based tools such as ChatGPT for engineers in industry. We conduct an observational study of 24 professional software engineers who have been using ChatGPT over a period of one week in their jobs, and qualitatively analyse their dialogues with the chatbot as well as their overall experience (as captured by an exit survey). We find that, rather than expecting ChatGPT to generate ready-to-use software artifacts (e.g., code), practitioners more often use ChatGPT to receive guidance on how to solve their tasks or learn about a topic in more abstract terms. We also propose a theoretical framework for how (i) purpose of the interaction, (ii) internal factors (e.g., the user's personality), and (iii) external factors (e.g., company policy) together shape the experience (in terms of perceived usefulness and trust). We envision that our framework can be used by future research to further the academic discussion on LLM usage by software engineering practitioners, and to serve as a reference point for the design of future empirical LLM research in this domain.

Extrapolation in Large language models (LLMs) for open-ended inquiry encounters two pivotal issues: (1) hallucination and (2) expensive training costs. These issues present challenges for LLMs in specialized domains and personalized data, requiring truthful responses and low fine-tuning costs. Existing works attempt to tackle the problem by augmenting the input of a smaller language model with information from a knowledge graph (KG). However, they have two limitations: (1) failing to extract relevant information from a large one-hop neighborhood in KG and (2) applying the same augmentation strategy for KGs with different characteristics that may result in low performance. Moreover, open-ended inquiry typically yields multiple responses, further complicating extrapolation. We propose a new task, the extreme multi-label KG link prediction task, to enable a model to perform extrapolation with multiple responses using structured real-world knowledge. Our retriever identifies relevant one-hop neighbors by considering entity, relation, and textual data together. Our experiments demonstrate that (1) KGs with different characteristics require different augmenting strategies, and (2) augmenting the language model's input with textual data improves task performance significantly. By incorporating the retrieval-augmented framework with KG, our framework, with a small parameter size, is able to extrapolate based on a given KG. The code can be obtained on GitHub: https://github.com/exiled1143/Retrieval-Augmented-Language-Model-for-Multi-Label-Knowledge-Graph-Link-Prediction.git

The strategic significance of Large Language Models (LLMs) in economic expansion, innovation, societal development, and national security has been increasingly recognized since the advent of ChatGPT. This study provides a comprehensive comparative evaluation of American and Chinese LLMs in both English and Chinese contexts. We proposed a comprehensive evaluation framework that encompasses natural language proficiency, disciplinary expertise, and safety and responsibility, and systematically assessed 16 prominent models from the US and China under various operational tasks and scenarios. Our key findings show that GPT 4-Turbo is at the forefront in English contexts, whereas Ernie-Bot 4 stands out in Chinese contexts. The study also highlights disparities in LLM performance across languages and tasks, stressing the necessity for linguistically and culturally nuanced model development. The complementary strengths of American and Chinese LLMs point to the value of Sino-US collaboration in advancing LLM technology. The research presents the current LLM competition landscape and offers valuable insights for policymakers and businesses regarding strategic LLM investments and development. Future work will expand on this framework to include emerging LLM multimodal capabilities and business application assessments.

Large language models (LLMs) are commonly trained on datasets consisting of fixed-length token sequences. These datasets are created by randomly concatenating documents of various lengths and then chunking them into sequences of a predetermined target length. However, this method of concatenation can lead to cross-document attention within a sequence, which is neither a desirable learning signal nor computationally efficient. Additionally, training on long sequences becomes computationally prohibitive due to the quadratic cost of attention. In this study, we introduce dataset decomposition, a novel variable sequence length training technique, to tackle these challenges. We decompose a dataset into a union of buckets, each containing sequences of the same size extracted from a unique document. During training, we use variable sequence length and batch size, sampling simultaneously from all buckets with a curriculum. In contrast to the concat-and-chunk baseline, which incurs a fixed attention cost at every step of training, our proposed method incurs a penalty proportional to the actual document lengths at each step, resulting in significant savings in training time. We train an 8k context-length 1B model at the same cost as a 2k context-length model trained with the baseline approach. Experiments on a web-scale corpus demonstrate that our approach significantly enhances performance on standard language evaluations and long-context benchmarks, reaching target accuracy 3x faster compared to the baseline. Our method not only enables efficient pretraining on long sequences but also scales effectively with dataset size. Lastly, we shed light on a critical yet less studied aspect of training large language models: the distribution and curriculum of sequence lengths, which results in a non-negligible difference in performance.

Developing intelligent dialog systems has been one of the longest running goals in AI. In recent years, significant progress has been made in building dialog systems with the breakthrough of deep learning methods and the large amount of conversational data being made available for system development (Budzianowski et al., 2018; Ou et al., 2022a; Ouyang et al., 2022; Achiam et al., 2023). There are still full of challenges toward building future dialog systems. The first FutureDial challenge focused on building semi-supervised and reinforced task-oriented dialog systems (FutureDial-SereTOD) (Ou et al., 2022a;b), which was successfully held at EMNLP 2022 SereTOD workshop However, problems like hallucination and fabrication (Alkaissi & McFarlane, 2023) still hinder the usage of such systems in real-life applications like customer service systems, which requires pin-point accuracy. Retrieval augmented generation (RAG) (Lewis et al., 2020; Guu et al., 2020) has been introduced to enhance dialog systems with retrieved information from external knowledge bases and has attracted increasing interests.

Optimizing the deployment of Large language models (LLMs) is expensive today since it requires experimentally running an application workload against an LLM implementation while exploring large configuration space formed by system knobs such as parallelization strategies, batching techniques, and scheduling policies. To address this challenge, we present Vidur - a large-scale, high-fidelity, easily-extensible simulation framework for LLM inference performance. Vidur models the performance of LLM operators using a combination of experimental profiling and predictive modeling, and evaluates the end-to-end inference performance for different workloads by estimating several metrics of interest such as latency and throughput. We validate the fidelity of Vidur on several LLMs and show that it estimates inference latency with less than 9% error across the range. Further, we present Vidur-Search, a configuration search tool that helps optimize LLM deployment. Vidur-Search uses Vidur to automatically identify the most cost-effective deployment configuration that meets application performance constraints. For example, Vidur-Search finds the best deployment configuration for LLaMA2-70B in one hour on a CPU machine, in contrast to a deployment-based exploration which would require 42K GPU hours - costing ~218K dollars. Source code for Vidur is available at https://github.com/microsoft/vidur.

We introduce ReALLM, a novel approach for compression and memory-efficient adaptation of pre-trained language models that encompasses most of the post-training quantization and fine-tuning methods for a budget of <4 bits. Pre-trained matrices are decomposed into a high-precision low-rank component and a vector-quantized latent representation (using an autoencoder). During the fine-tuning step, only the low-rank components are updated. Our results show that pre-trained matrices exhibit different patterns. ReALLM adapts the shape of the encoder (small/large embedding, high/low bit VQ, etc.) to each matrix. ReALLM proposes to represent each matrix with a small embedding on $b$ bits and a neural decoder model $\mathcal{D}_\phi$ with its weights on $b_\phi$ bits. The decompression of a matrix requires only one embedding and a single forward pass with the decoder. Our weight-only quantization algorithm yields the best results on language generation tasks (C4 and WikiText-2) for a budget of $3$ bits without any training. With a budget of $2$ bits, ReALLM achieves state-of-the art performance after fine-tuning on a small calibration dataset.

Before deploying outputs from foundation models in high-stakes tasks, it is imperative to ensure that they align with human values. For instance, in radiology report generation, reports generated by a vision-language model must align with human evaluations before their use in medical decision-making. This paper presents Conformal Alignment, a general framework for identifying units whose outputs meet a user-specified alignment criterion. It is guaranteed that on average, a prescribed fraction of selected units indeed meet the alignment criterion, regardless of the foundation model or the data distribution. Given any pre-trained model and new units with model-generated outputs, Conformal Alignment leverages a set of reference data with ground-truth alignment status to train an alignment predictor. It then selects new units whose predicted alignment scores surpass a data-dependent threshold, certifying their corresponding outputs as trustworthy. Through applications to question answering and radiology report generation, we demonstrate that our method is able to accurately identify units with trustworthy outputs via lightweight training over a moderate amount of reference data. En route, we investigate the informativeness of various features in alignment prediction and combine them with standard models to construct the alignment predictor.

Last week OpenAI revealed a new conversational interface for ChatGPT with an expressive synthetic voice strikingly similar to that of the AI assistant played by Scarlett Johansson in the sci-fi movie Her--only to suddenly disable the new voice over the weekend. On Monday, Johansson issued a statement claiming to have forced that reversal, after her lawyers demanded OpenAI clarify how the new voice was created. Johansson's statement, relayed to WIRED by her publicist, claims that OpenAI CEO Sam Altman asked her last September to provide ChatGPT's new voice but that she declined. She describes being astounded to see the company demo a new voice for ChatGPT last week that sounded like her anyway. "When I heard the release demo I was shocked, angered, and in disbelief that Mr. Altman would pursue a voice that sounded so eerily similar to mine that my closest friends and news outlets could not tell the difference," the statement reads.