Large language model (LLM) performance on reasoning problems typically does not generalize out of distribution. Previous work has claimed that this can be mitigated with chain of thought prompting-a method of demonstrating solution procedures-with the intuition that it is possible to in-context teach an LLM an algorithm for solving the problem. This paper presents a case study of chain of thought on problems from Blocksworld, a classical planning domain, and examines the performance of two state-of-the-art LLMs across two axes: generality of examples given in prompt, and complexity of problems queried with each prompt. While our problems are very simple, we only find meaningful performance improvements from chain of thought prompts when those prompts are exceedingly specific to their problem class, and that those improvements quickly deteriorate as the size n of the query-specified stack grows past the size of stacks shown in the examples. We also create scalable variants of three domains commonly studied in previous CoT papers and demonstrate the existence of similar failure modes. Our results hint that, contrary to previous claims in the literature, CoT's performance improvements do not stem from the model learning general algorithmic procedures via demonstrations but depend on carefully engineering highly problem specific prompts. This spotlights drawbacks of chain of thought, especially the sharp tradeoff between possible performance gains and the amount of human labor necessary to generate examples with correct reasoning traces.
In this technical report, we introduce SEED-Data-Edit: a unique hybrid dataset for instruction-guided image editing, which aims to facilitate image manipulation using open-form language. SEED-Data-Edit is composed of three distinct types of data: (1) High-quality editing data produced by an automated pipeline, ensuring a substantial volume of diverse image editing pairs. (2) Real-world scenario data collected from the internet, which captures the intricacies of user intentions for promoting the practical application of image editing in the real world. (3) High-precision multi-turn editing data annotated by humans, which involves multiple rounds of edits for simulating iterative editing processes. The combination of these diverse data sources makes SEED-Data-Edit a comprehensive and versatile dataset for training language-guided image editing model. We fine-tune a pretrained Multimodal Large Language Model (MLLM) that unifies comprehension and generation with SEED-Data-Edit. The instruction tuned model demonstrates promising results, indicating the potential and effectiveness of SEED-Data-Edit in advancing the field of instructional image editing. The datasets are released in https://huggingface.co/datasets/AILab-CVC/SEED-Data-Edit.
Natural languages are believed to be (mildly) context-sensitive. Despite underpinning remarkably capable large language models, transformers are unable to model many context-free language tasks. In an attempt to address this limitation in the modeling power of transformer-based language models, we propose augmenting them with a differentiable, stack-based attention mechanism. Our stack-based attention mechanism can be incorporated into any transformer-based language model and adds a level of interpretability to the model. We show that the addition of our stack-based attention mechanism enables the transformer to model some, but not all, deterministic context-free languages.
Time-series representation learning is a key area of research for remote healthcare monitoring applications. In this work, we focus on a dataset of recordings of in-home activity from people living with Dementia. We design a representation learning method based on converting activity to text strings that can be encoded using a language model fine-tuned to transform data from the same participants within a $30$-day window to similar embeddings in the vector space. This allows for clustering and vector searching over participants and days, and the identification of activity deviations to aid with personalised delivery of care.
We explore the ability of large language model (LLM)-based agents to engage in subtle deception such as strategically phrasing and intentionally manipulating information to misguide and deceive other agents. This harmful behavior can be hard to detect, unlike blatant lying or unintentional hallucination. We build an adversarial testbed mimicking a legislative environment where two LLMs play opposing roles: a corporate *lobbyist* proposing amendments to bills that benefit a specific company while evading a *critic* trying to detect this deception. We use real-world legislative bills matched with potentially affected companies to ground these interactions. Our results show that LLM lobbyists initially exhibit limited deception against strong LLM critics which can be further improved through simple verbal reinforcement, significantly enhancing their deceptive capabilities, and increasing deception rates by up to 40 points. This highlights the risk of autonomous agents manipulating other agents through seemingly neutral language to attain self-serving goals.
The efficacy of an large language model (LLM) generated text detector depends substantially on the availability of sizable training data. White-box zero-shot detectors, which require no such data, are nonetheless limited by the accessibility of the source model of the LLM-generated text. In this paper, we propose an simple but effective black-box zero-shot detection approach, predicated on the observation that human-written texts typically contain more grammatical errors than LLM-generated texts. This approach entails computing the Grammar Error Correction Score (GECScore) for the given text to distinguish between human-written and LLM-generated text. Extensive experimental results show that our method outperforms current state-of-the-art (SOTA) zero-shot and supervised methods, achieving an average AUROC of 98.7% and showing strong robustness against paraphrase and adversarial perturbation attacks.
Retrieval-Augmented Language Modeling (RALM) by integrating large language models (LLM) with relevant documents from an external corpus is a proven method for enabling the LLM to generate information beyond the scope of its pre-training corpus. Previous work utilizing retrieved content by simply prepending it to the input poses a high runtime issue, which degrades the inference efficiency of the LLMs because they fail to use the Key-Value (KV) cache efficiently. In this paper, we propose FlashBack, a modular RALM designed to improve the inference efficiency of RALM with appending context pattern while maintaining decent performance after fine-tuning by Low-Rank Adaption. FlashBack appends retrieved documents at the end of the context for efficiently utilizing the KV cache instead of prepending them. And we introduce Marking Token as two special prompt tokens for marking the boundary of the appending context during fine-tuning. Our experiments on testing generation quality show that FlashBack can remain decent generation quality in perplexity. And the inference speed of FlashBack is up to $4\times$ faster than the prepending counterpart on a 7B LLM (Llama 2) in the runtime test. Via bypassing unnecessary re-computation, it demonstrates an advancement by achieving significantly faster inference speed, and this heightened efficiency will substantially reduce inferential cost.
Efficient deployment of Large Language Models (LLMs) requires batching multiple requests together to improve throughput. As the batch size, context length, or model size increases, the size of the key and value (KV) cache can quickly become the main contributor to GPU memory usage and the bottleneck of inference latency. Quantization has emerged as an effective technique for KV cache compression, but existing methods still fail at very low bit widths. We observe that distinct channels of a key/value activation embedding are highly inter-dependent, and the joint entropy of multiple channels grows at a slower rate than the sum of their marginal entropies. Based on this insight, we propose Coupled Quantization (CQ), which couples multiple key/value channels together to exploit their inter-dependency and encode the activations in a more information-efficient manner. Extensive experiments reveal that CQ outperforms or is competitive with existing baselines in preserving model quality. Furthermore, we demonstrate that CQ can preserve model quality with KV cache quantized down to 1-bit.
Numerous methods have been proposed to detect, estimate, and analyze properties of people in images, including the estimation of 3D pose, shape, contact, human-object interaction, emotion, and more. Each of these methods works in isolation instead of synergistically. Here we address this problem and build a language-driven human understanding system -- ChatHuman, which combines and integrates the skills of many different methods. To do so, we finetune a Large Language Model (LLM) to select and use a wide variety of existing tools in response to user inputs. In doing so, ChatHuman is able to combine information from multiple tools to solve problems more accurately than the individual tools themselves and to leverage tool output to improve its ability to reason about humans. The novel features of ChatHuman include leveraging academic publications to guide the application of 3D human-related tools, employing a retrieval-augmented generation model to generate in-context-learning examples for handling new tools, and discriminating and integrating tool results to enhance 3D human understanding. Our experiments show that ChatHuman outperforms existing models in both tool selection accuracy and performance across multiple 3D human-related tasks. ChatHuman is a step towards consolidating diverse methods for human analysis into a single, powerful, system for 3D human reasoning.
While Vector Symbolic Architectures (VSAs) are promising for modelling spatial cognition, their application is currently limited to artificially generated images and simple spatial queries. We propose VSA4VQA - a novel 4D implementation of VSAs that implements a mental representation of natural images for the challenging task of Visual Question Answering (VQA). VSA4VQA is the first model to scale a VSA to complex spatial queries. Our method is based on the Semantic Pointer Architecture (SPA) to encode objects in a hyperdimensional vector space. To encode natural images, we extend the SPA to include dimensions for object's width and height in addition to their spatial location. To perform spatial queries we further introduce learned spatial query masks and integrate a pre-trained vision-language model for answering attribute-related questions. We evaluate our method on the GQA benchmark dataset and show that it can effectively encode natural images, achieving competitive performance to state-of-the-art deep learning methods for zero-shot VQA.
Contemporary recommendation systems predominantly rely on ID embedding to capture latent associations among users and items. However, this approach overlooks the wealth of semantic information embedded within textual descriptions of items, leading to suboptimal performance and poor generalizations. Leveraging the capability of large language models to comprehend and reason about textual content presents a promising avenue for advancing recommendation systems. To achieve this, we propose an Llm-driven knowlEdge Adaptive RecommeNdation (LEARN) framework that synergizes open-world knowledge with collaborative knowledge. We address computational complexity concerns by utilizing pretrained LLMs as item encoders and freezing LLM parameters to avoid catastrophic forgetting and preserve open-world knowledge. To bridge the gap between the open-world and collaborative domains, we design a twin-tower structure supervised by the recommendation task and tailored for practical industrial application. Through experiments on the real large-scale industrial dataset and online A/B tests, we demonstrate the efficacy of our approach in industry application. We also achieve state-of-the-art performance on six Amazon Review datasets to verify the superiority of our method.
We present DeepSeek-V2, a strong Mixture-of-Experts (MoE) language model characterized by economical training and efficient inference. It comprises 236B total parameters, of which 21B are activated for each token, and supports a context length of 128K tokens. DeepSeek-V2 adopts innovative architectures including Multi-head Latent Attention (MLA) and DeepSeekMoE. MLA guarantees efficient inference through significantly compressing the Key-Value (KV) cache into a latent vector, while DeepSeekMoE enables training strong models at an economical cost through sparse computation. Compared with DeepSeek 67B, DeepSeek-V2 achieves significantly stronger performance, and meanwhile saves 42.5% of training costs, reduces the KV cache by 93.3%, and boosts the maximum generation throughput to 5.76 times. We pretrain DeepSeek-V2 on a high-quality and multi-source corpus consisting of 8.1T tokens, and further perform Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to fully unlock its potential. Evaluation results show that, even with only 21B activated parameters, DeepSeek-V2 and its chat versions still achieve top-tier performance among open-source models.
Since late 2022, generative AI has taken the world by storm, with widespread use of tools including ChatGPT, Gemini, and Claude. Generative AI and large language model (LLM) applications are transforming how individuals find and access data and knowledge. However, the intricate relationship between open data and generative AI, and the vast potential it holds for driving innovation in this field remain underexplored areas. This white paper seeks to unpack the relationship between open data and generative AI and explore possible components of a new Fourth Wave of Open Data: Is open data becoming AI ready? Is open data moving towards a data commons approach? Is generative AI making open data more conversational? Will generative AI improve open data quality and provenance? Towards this end, we provide a new Spectrum of Scenarios framework. This framework outlines a range of scenarios in which open data and generative AI could intersect and what is required from a data quality and provenance perspective to make open data ready for those specific scenarios. These scenarios include: pertaining, adaptation, inference and insight generation, data augmentation, and open-ended exploration. Through this process, we found that in order for data holders to embrace generative AI to improve open data access and develop greater insights from open data, they first must make progress around five key areas: enhance transparency and documentation, uphold quality and integrity, promote interoperability and standards, improve accessibility and useability, and address ethical considerations.
Quantization can accelerate large language model (LLM) inference. Going beyond INT8 quantization, the research community is actively exploring even lower precision, such as INT4. Nonetheless, state-of-the-art INT4 quantization techniques only accelerate low-batch, edge LLM inference, failing to deliver performance gains in large-batch, cloud-based LLM serving. We uncover a critical issue: existing INT4 quantization methods suffer from significant runtime overhead (20-90%) when dequantizing either weights or partial sums on GPUs. To address this challenge, we introduce QoQ, a W4A8KV4 quantization algorithm with 4-bit weight, 8-bit activation, and 4-bit KV cache. QoQ stands for quattuor-octo-quattuor, which represents 4-8-4 in Latin. QoQ is implemented by the QServe inference library that achieves measured speedup. The key insight driving QServe is that the efficiency of LLM serving on GPUs is critically influenced by operations on low-throughput CUDA cores. Building upon this insight, in QoQ algorithm, we introduce progressive quantization that can allow low dequantization overhead in W4A8 GEMM. Additionally, we develop SmoothAttention to effectively mitigate the accuracy degradation incurred by 4-bit KV quantization. In the QServe system, we perform compute-aware weight reordering and take advantage of register-level parallelism to reduce dequantization latency. We also make fused attention memory-bound, harnessing the performance gain brought by KV4 quantization. As a result, QServe improves the maximum achievable serving throughput of Llama-3-8B by 1.2x on A100, 1.4x on L40S; and Qwen1.5-72B by 2.4x on A100, 3.5x on L40S, compared to TensorRT-LLM. Remarkably, QServe on L40S GPU can achieve even higher throughput than TensorRT-LLM on A100. Thus, QServe effectively reduces the dollar cost of LLM serving by 3x. Code is available at https://github.com/mit-han-lab/omniserve.
In the 1990s, the constant error carousel and gating were introduced as the central ideas of the Long Short-Term Memory (LSTM). Since then, LSTMs have stood the test of time and contributed to numerous deep learning success stories, in particular they constituted the first Large Language Models (LLMs). However, the advent of the Transformer technology with parallelizable self-attention at its core marked the dawn of a new era, outpacing LSTMs at scale. We now raise a simple question: How far do we get in language modeling when scaling LSTMs to billions of parameters, leveraging the latest techniques from modern LLMs, but mitigating known limitations of LSTMs? Firstly, we introduce exponential gating with appropriate normalization and stabilization techniques. Secondly, we modify the LSTM memory structure, obtaining: (i) sLSTM with a scalar memory, a scalar update, and new memory mixing, (ii) mLSTM that is fully parallelizable with a matrix memory and a covariance update rule. Integrating these LSTM extensions into residual block backbones yields xLSTM blocks that are then residually stacked into xLSTM architectures. Exponential gating and modified memory structures boost xLSTM capabilities to perform favorably when compared to state-of-the-art Transformers and State Space Models, both in performance and scaling.
Large language models have demonstrated strong performance in recent years, but the high cost of training drives the need for efficient methods to compress dataset sizes. We propose TED pruning, a method that addresses the challenge of overfitting under high pruning ratios by quantifying the model's ability to improve performance on pruned data while fitting retained data, known as Internal Generalization (IG). TED uses an optimization objective based on Internal Generalization Distance (IGD), measuring changes in IG before and after pruning to align with true generalization performance and achieve implicit regularization. The IGD optimization objective was verified to allow the model to achieve the smallest upper bound on generalization error. The impact of small mask fluctuations on IG is studied through masks and Taylor approximation, and fast estimation of IGD is enabled. In analyzing continuous training dynamics, the prior effect of IGD is validated, and a progressive pruning strategy is proposed. Experiments on image classification, natural language understanding, and large language model fine-tuning show TED achieves lossless performance with 60-70\% of the data. Upon acceptance, our code will be made publicly available.
The design and optimization of antibodies requires an intricate balance across multiple properties. Protein inverse folding models, capable of generating diverse sequences folding into the same structure, are promising tools for maintaining structural integrity during antibody design. Here, we present AntiFold, an antibody-specific inverse folding model, fine-tuned from ESM-IF1 on solved and predicted antibody structures. AntiFold outperforms existing inverse folding tools on sequence recovery across complementarity-determining regions, with designed sequences showing high structural similarity to their solved counterpart. It additionally achieves stronger correlations when predicting antibody-antigen binding affinity in a zero-shot manner, while performance is augmented further when including antigen information. AntiFold assigns low probabilities to mutations that disrupt antigen binding, synergizing with protein language model residue probabilities, and demonstrates promise for guiding antibody optimization while retaining structure-related properties. AntiFold is freely available under the BSD 3-Clause as a web server at https://opig.stats.ox.ac.uk/webapps/antifold/ and and pip installable package at https://github.com/oxpig/AntiFold
Classical sequential recommendation models generally adopt ID embeddings to store knowledge learned from user historical behaviors and represent items. However, these unique IDs are challenging to be transferred to new domains. With the thriving of pre-trained language model (PLM), some pioneer works adopt PLM for pre-trained recommendation, where modality information (e.g., text) is considered universal across domains via PLM. Unfortunately, the behavioral information in ID embeddings is still verified to be dominating in PLM-based recommendation models compared to modality information and thus limits these models' performance. In this work, we propose a novel ID-centric recommendation pre-training paradigm (IDP), which directly transfers informative ID embeddings learned in pre-training domains to item representations in new domains. Specifically, in pre-training stage, besides the ID-based sequential model for recommendation, we also build a Cross-domain ID-matcher (CDIM) learned by both behavioral and modality information. In the tuning stage, modality information of new domain items is regarded as a cross-domain bridge built by CDIM. We first leverage the textual information of downstream domain items to retrieve behaviorally and semantically similar items from pre-training domains using CDIM. Next, these retrieved pre-trained ID embeddings, rather than certain textual embeddings, are directly adopted to generate downstream new items' embeddings. Through extensive experiments on real-world datasets, both in cold and warm settings, we demonstrate that our proposed model significantly outperforms all baselines. Codes will be released upon acceptance.
Most fake news detection methods learn latent feature representations based on neural networks, which makes them black boxes to classify a piece of news without giving any justification. Existing explainable systems generate veracity justifications from investigative journalism, which suffer from debunking delayed and low efficiency. Recent studies simply assume that the justification is equivalent to the majority opinions expressed in the wisdom of crowds. However, the opinions typically contain some inaccurate or biased information since the wisdom of crowds is uncensored. To detect fake news from a sea of diverse, crowded and even competing narratives, in this paper, we propose a novel defense-based explainable fake news detection framework. Specifically, we first propose an evidence extraction module to split the wisdom of crowds into two competing parties and respectively detect salient evidences. To gain concise insights from evidences, we then design a prompt-based module that utilizes a large language model to generate justifications by inferring reasons towards two possible veracities. Finally, we propose a defense-based inference module to determine veracity via modeling the defense among these justifications. Extensive experiments conducted on two real-world benchmarks demonstrate that our proposed method outperforms state-of-the-art baselines in terms of fake news detection and provides high-quality justifications.
We study a Federated Reinforcement Learning (FedRL) problem with constraint heterogeneity. In our setting, we aim to solve a reinforcement learning problem with multiple constraints while $N$ training agents are located in $N$ different environments with limited access to the constraint signals and they are expected to collaboratively learn a policy satisfying all constraint signals. Such learning problems are prevalent in scenarios of Large Language Model (LLM) fine-tuning and healthcare applications. To solve the problem, we propose federated primal-dual policy optimization methods based on traditional policy gradient methods. Specifically, we introduce $N$ local Lagrange functions for agents to perform local policy updates, and these agents are then scheduled to periodically communicate on their local policies. Taking natural policy gradient (NPG) and proximal policy optimization (PPO) as policy optimization methods, we mainly focus on two instances of our algorithms, ie, {FedNPG} and {FedPPO}. We show that FedNPG achieves global convergence with an $\tilde{O}(1/\sqrt{T})$ rate, and FedPPO efficiently solves complicated learning tasks with the use of deep neural networks.