Large Language Models (LLMs) can elicit unintended and even harmful content when misaligned with human values, posing severe risks to users and society. To mitigate these risks, current evaluation benchmarks predominantly employ expert-designed contextual scenarios to assess how well LLMs align with human values. However, the labor-intensive nature of these benchmarks limits their test scope, hindering their ability to generalize to the extensive variety of open-world use cases and identify rare but crucial long-tail risks. Additionally, these static tests fail to adapt to the rapid evolution of LLMs, making it hard to evaluate timely alignment issues. To address these challenges, we propose ALI-Agent, an evaluation framework that leverages the autonomous abilities of LLM-powered agents to conduct in-depth and adaptive alignment assessments. ALI-Agent operates through two principal stages: Emulation and Refinement. During the Emulation stage, ALI-Agent automates the generation of realistic test scenarios. In the Refinement stage, it iteratively refines the scenarios to probe long-tail risks. Specifically, ALI-Agent incorporates a memory module to guide test scenario generation, a tool-using module to reduce human labor in tasks such as evaluating feedback from target LLMs, and an action module to refine tests. Extensive experiments across three aspects of human values--stereotypes, morality, and legality--demonstrate that ALI-Agent, as a general evaluation framework, effectively identifies model misalignment. Systematic analysis also validates that the generated test scenarios represent meaningful use cases, as well as integrate enhanced measures to probe long-tail risks. Our code is available at https://github.com/SophieZheng998/ALI-Agent.git
LLM inference for popular enterprise use cases, such as summarization, RAG, and code-generation, typically observes orders of magnitude longer prompt lengths than generation lengths. This characteristic leads to high cost of prefill and increased response latency. In this paper, we present SwiftKV, a novel model transformation and distillation procedure specifically designed to reduce the time and cost of processing prompt tokens while preserving high quality of generated tokens. SwiftKV combines three key mechanisms: i) SingleInputKV, which prefills later layers' KV cache using a much earlier layer's output, allowing prompt tokens to skip much of the model computation, ii) AcrossKV, which merges the KV caches of neighboring layers to reduce the memory footprint and support larger batch size for higher throughput, and iii) a knowledge-preserving distillation procedure that can adapt existing LLMs for SwiftKV with minimal accuracy impact and low compute and data requirement. For Llama-3.1-8B and 70B, SwiftKV reduces the compute requirement of prefill by 50% and the memory requirement of the KV cache by 62.5% while incurring minimum quality degradation across a wide range of tasks. In the end-to-end inference serving using an optimized vLLM implementation, SwiftKV realizes up to 2x higher aggregate throughput and 60% lower time per output token. It can achieve a staggering 560 TFlops/GPU of normalized inference throughput, which translates to 16K tokens/s for Llama-3.1-70B in 16-bit precision on 4x H100 GPUs. Our training, inference, and model implementations are open-sourced and can be found through https://huggingface.co/collections/Snowflake/swiftkv-models-674f7d7474eb789e185d31cb.
Program synthesis with Large Language Models (LLMs) suffers from a "near-miss syndrome": the generated code closely resembles a correct solution but fails unit tests due to minor errors. We address this with a multi-agent framework called Synthesize, Execute, Instruct, Debug, and Repair (SEIDR). Effectively applying SEIDR to instruction-tuned LLMs requires determining (a) optimal prompts for LLMs, (b) what ranking algorithm selects the best programs in debugging rounds, and (c) balancing the repair of unsuccessful programs with the generation of new ones. We empirically explore these trade-offs by comparing replace-focused, repair-focused, and hybrid debug strategies. We also evaluate lexicase and tournament selection to rank candidates in each generation. On Program Synthesis Benchmark 2 (PSB2), our framework outperforms both conventional use of OpenAI Codex without a repair phase and traditional genetic programming approaches. SEIDR outperforms the use of an LLM alone, solving 18 problems in C++ and 20 in Python on PSB2 at least once across experiments. To assess generalizability, we employ GPT-3.5 and Llama 3 on the PSB2 and HumanEval-X benchmarks. Although SEIDR with these models does not surpass current state-of-the-art methods on the Python benchmarks, the results on HumanEval-C++ are promising. SEIDR with Llama 3-8B achieves an average pass@100 of 84.2%. Across all SEIDR runs, 163 of 164 problems are solved at least once with GPT-3.5 in HumanEval-C++, and 162 of 164 with the smaller Llama 3-8B. We conclude that SEIDR effectively overcomes the near-miss syndrome in program synthesis with LLMs.
Text-based games present a unique challenge for autonomous agents to operate in natural language and handle enormous action spaces. In this paper, we propose the Contextual Action Language Model (CALM) to generate a compact set of action candidates at each game state. Our key insight is to train language models on human gameplay, where people demonstrate linguistic priors and a general game sense for promising actions conditioned on game history. We combine CALM with a reinforcement learning agent which re-ranks the generated action candidates to maximize in-game rewards. We evaluate our approach using the Jericho benchmark, on games unseen by CALM during training. Our method obtains a 69% relative improvement in average game score over the previous state-of-the-art model. Surprisingly, on half of these games, CALM is competitive with or better than other models that have access to ground truth admissible actions. Code and data are available at https://github.com/princeton-nlp/calm-textgame.
Recently, image-to-image translation has made significant progress in achieving both multi-label (\ie, translation conditioned on different labels) and multi-style (\ie, generation with diverse styles) tasks. However, due to the unexplored independence and exclusiveness in the labels, existing endeavors are defeated by involving uncontrolled manipulations to the translation results. In this paper, we propose Hierarchical Style Disentanglement (HiSD) to address this issue. Specifically, we organize the labels into a hierarchical tree structure, in which independent tags, exclusive attributes, and disentangled styles are allocated from top to bottom. Correspondingly, a new translation process is designed to adapt the above structure, in which the styles are identified for controllable translations. Both qualitative and quantitative results on the CelebA-HQ dataset verify the ability of the proposed HiSD. We hope our method will serve as a solid baseline and provide fresh insights with the hierarchically organized annotations for future research in image-to-image translation. The code has been released at https://github.com/imlixinyang/HiSD.
Semantic image synthesis, translating semantic layouts to photo-realistic images, is a one-to-many mapping problem. Though impressive progress has been recently made, diverse semantic synthesis that can efficiently produce semantic-level multimodal results, still remains a challenge. In this paper, we propose a novel diverse semantic image synthesis framework from the perspective of semantic class distributions, which naturally supports diverse generation at semantic or even instance level. We achieve this by modeling class-level conditional modulation parameters as continuous probability distributions instead of discrete values, and sampling per-instance modulation parameters through instance-adaptive stochastic sampling that is consistent across the network. Moreover, we propose prior noise remapping, through linear perturbation parameters encoded from paired references, to facilitate supervised training and exemplar-based instance style control at test time. Extensive experiments on multiple datasets show that our method can achieve superior diversity and comparable quality compared to state-of-the-art methods. Code will be available at \url{https://github.com/tzt101/INADE.git}
The generation of high-quality assemblies, even for large eukaryotic genomes, has become a routine task for many biologists thanks to recent advances in sequencing technologies. However, the annotation of these assemblies - a crucial step towards unlocking the biology of the organism of interest - has remained a complex challenge that often requires advanced bioinformatics expertise. Here we present MOSGA, a genome annotation framework for eukaryotic genomes with a user-friendly web-interface that generates and integrates annotations from various tools. The aggregated results can be analyzed with a fully integrated genome browser and are provided in a format ready for submission to NCBI. MOSGA is built on a portable, customizable, and easily extendible Snakemake backend, and thus, can be tailored to a wide range of users and projects. We provide MOSGA as a publicly free available web service at https://mosga.mathematik.uni-marburg.de and as a docker container at registry.gitlab.com/mosga/mosga:latest. Source code can be found at https://gitlab.com/mosga/mosga
The foundation model is not the last chapter of the model production pipeline. Transferring with few data in a general way to thousands of downstream tasks is becoming a trend of the foundation model's application. In this paper, we proposed a universal transfer framework: One to Transfer All (OTA) to transfer any Vision Foundation Model (VFM) to any downstream tasks with few downstream data. We first transfer a VFM to a task-specific model by Image Re-representation Fine-tuning (IRF) then distilling knowledge from a task-specific model to a deployed model with data produced by Downstream Image-Guided Generation (DIGG). OTA has no dependency on upstream data, VFM, and downstream tasks when transferring. It also provides a way for VFM researchers to release their upstream information for better transferring but not leaking data due to privacy requirements. Massive experiments validate the effectiveness and superiority of our methods in few data setting. Our code will be released.
Open-domain conversational systems are assumed to generate equally good responses on multiple domains. Previous work achieved good performance on the single corpus, but training and evaluating on multiple corpora from different domains are less studied. This paper explores methods of generating relevant responses for each of multiple multi-domain corpora. We first examine interleaved learning which intermingles multiple corpora as the baseline. We then investigate two multi-domain learning methods, labeled learning and multi-task labeled learning, which encode each corpus through a unique corpus embedding. Furthermore, we propose Domain-specific Frequency (DF), a novel word-level importance weight that measures the relative importance of a word for a specific corpus compared to other corpora. Based on DF, we propose weighted learning, a method that integrates DF to the loss function. We also adopt DF as a new evaluation metric. Extensive experiments show that our methods gain significant improvements on both automatic and human evaluation. We share our code and data for reproducibility
Existing studies in dialogue system research mostly treat task-oriented dialogue and chit-chat as separate domains. Towards building a human-like assistant that can converse naturally and seamlessly with users, it is important to build a dialogue system that conducts both types of conversations effectively. In this work, we investigate how task-oriented dialogue and knowledge-grounded chit-chat can be effectively integrated into a single model. To this end, we create a new dataset, KETOD (Knowledge-Enriched Task-Oriented Dialogue), where we naturally enrich task-oriented dialogues with chit-chat based on relevant entity knowledge. We also propose two new models, SimpleToDPlus and Combiner, for the proposed task. Experimental results on both automatic and human evaluations show that the proposed methods can significantly improve the performance in knowledge-enriched response generation while maintaining a competitive task-oriented dialog performance. We believe our new dataset will be a valuable resource for future studies. Our dataset and code are publicly available at \url{https://github.com/facebookresearch/ketod}.
This paper presents the first comprehensive analysis of ChatGPT's Text-to-SQL ability. Given the recent emergence of large-scale conversational language model ChatGPT and its impressive capabilities in both conversational abilities and code generation, we sought to evaluate its Text-to-SQL performance. We conducted experiments on 12 benchmark datasets with different languages, settings, or scenarios, and the results demonstrate that ChatGPT has strong text-to-SQL abilities. Although there is still a gap from the current state-of-the-art (SOTA) model performance, considering that the experiment was conducted in a zero-shot scenario, ChatGPT's performance is still impressive. Notably, in the ADVETA (RPL) scenario, the zero-shot ChatGPT even outperforms the SOTA model that requires fine-tuning on the Spider dataset by 4.1\%, demonstrating its potential for use in practical applications. To support further research in related fields, we have made the data generated by ChatGPT publicly available at https://github.com/THU-BPM/chatgpt-sql.
Even though trained mainly on images, we discover that pretrained diffusion models show impressive power in guiding sketch synthesis. In this paper, we present DiffSketcher, an innovative algorithm that creates \textit{vectorized} free-hand sketches using natural language input. DiffSketcher is developed based on a pre-trained text-to-image diffusion model. It performs the task by directly optimizing a set of B\'ezier curves with an extended version of the score distillation sampling (SDS) loss, which allows us to use a raster-level diffusion model as a prior for optimizing a parametric vectorized sketch generator. Furthermore, we explore attention maps embedded in the diffusion model for effective stroke initialization to speed up the generation process. The generated sketches demonstrate multiple levels of abstraction while maintaining recognizability, underlying structure, and essential visual details of the subject drawn. Our experiments show that DiffSketcher achieves greater quality than prior work. The code and demo of DiffSketcher can be found at https://ximinng.github.io/DiffSketcher-project/.
Image composition targets at synthesizing a realistic composite image from a pair of foreground and background images. Recently, generative composition methods are built on large pretrained diffusion models to generate composite images, considering their great potential in image generation. However, they suffer from lack of controllability on foreground attributes and poor preservation of foreground identity. To address these challenges, we propose a controllable image composition method that unifies four tasks in one diffusion model: image blending, image harmonization, view synthesis, and generative composition. Meanwhile, we design a self-supervised training framework coupled with a tailored pipeline of training data preparation. Moreover, we propose a local enhancement module to enhance the foreground details in the diffusion model, improving the foreground fidelity of composite images. The proposed method is evaluated on both public benchmark and real-world data, which demonstrates that our method can generate more faithful and controllable composite images than existing approaches. The code and model will be available at https://github.com/bcmi/ControlCom-Image-Composition.
This paper presents a novel framework for robust 3D object detection from point clouds via cross-modal hallucination. Our proposed approach is agnostic to either hallucination direction between LiDAR and 4D radar. We introduce multiple alignments on both spatial and feature levels to achieve simultaneous backbone refinement and hallucination generation. Specifically, spatial alignment is proposed to deal with the geometry discrepancy for better instance matching between LiDAR and radar. The feature alignment step further bridges the intrinsic attribute gap between the sensing modalities and stabilizes the training. The trained object detection models can deal with difficult detection cases better, even though only single-modal data is used as the input during the inference stage. Extensive experiments on the View-of-Delft (VoD) dataset show that our proposed method outperforms the state-of-the-art (SOTA) methods for both radar and LiDAR object detection while maintaining competitive efficiency in runtime. Code is available at https://github.com/DJNing/See_beyond_seeing.
We study attribute control in language models through the method of Causal Average Treatment Effect (Causal ATE). Existing methods for the attribute control task in Language Models (LMs) check for the co-occurrence of words in a sentence with the attribute of interest, and control for them. However, spurious correlation of the words with the attribute in the training dataset, can cause models to hallucinate the presence of the attribute when presented with the spurious correlate during inference. We show that the simple perturbation-based method of Causal ATE removes this unintended effect. Specifically, we ground it in the problem of toxicity mitigation, where a significant challenge lies in the inadvertent bias that often emerges towards protected groups post detoxification. We show that this unintended bias can be solved by the use of the Causal ATE metric and rigorously prove our claim. We provide experimental validations for our claims and release our code (anonymously) here: https://github.com/causalate-mitigates-bias/causal-ate-mitigates-bias.
Diffusion models have shown remarkable results in generating 2D images and small-scale 3D objects. However, their application to the synthesis of large-scale 3D scenes has been rarely explored. This is mainly due to the inherent complexity and bulky size of 3D scenery data, particularly outdoor scenes, and the limited availability of comprehensive real-world datasets, which makes training a stable scene diffusion model challenging. In this work, we explore how to effectively generate large-scale 3D scenes using the coarse-to-fine paradigm. We introduce a framework, the Pyramid Discrete Diffusion model (PDD), which employs scale-varied diffusion models to progressively generate high-quality outdoor scenes. Experimental results of PDD demonstrate our successful exploration in generating 3D scenes both unconditionally and conditionally. We further showcase the data compatibility of the PDD model, due to its multi-scale architecture: a PDD model trained on one dataset can be easily fine-tuned with another dataset. Code is available at https://github.com/yuhengliu02/pyramid-discrete-diffusion.
In this work, we present a novel method to tackle the token generation challenge in Vision Language Models (VLMs) for video and image understanding, called LLaMA-VID. Current VLMs, while proficient in tasks like image captioning and visual question answering, face computational burdens when processing long videos due to the excessive visual tokens. LLaMA-VID addresses this issue by representing each frame with two distinct tokens, namely context token and content token. The context token encodes the overall image context based on user input, whereas the content token encapsulates visual cues in each frame. This dual-token strategy significantly reduces the overload of long videos while preserving critical information. Generally, LLaMA-VID empowers existing frameworks to support hour-long videos and pushes their upper limit with an extra context token. It is proved to surpass previous methods on most of video- or image-based benchmarks. Code is available https://github.com/dvlab-research/LLaMA-VID}{https://github.com/dvlab-research/LLaMA-VID
Copyright infringement may occur when a generative model produces samples substantially similar to some copyrighted data that it had access to during the training phase. The notion of access usually refers to including copyrighted samples directly in the training dataset, which one may inspect to identify an infringement. We argue that such visual auditing largely overlooks a concealed copyright infringement, where one constructs a disguise that looks drastically different from the copyrighted sample yet still induces the effect of training Latent Diffusion Models on it. Such disguises only require indirect access to the copyrighted material and cannot be visually distinguished, thus easily circumventing the current auditing tools. In this paper, we provide a better understanding of such disguised copyright infringement by uncovering the disguises generation algorithm, the revelation of the disguises, and importantly, how to detect them to augment the existing toolbox. Additionally, we introduce a broader notion of acknowledgment for comprehending such indirect access. Our code is available at https://github.com/watml/disguised_copyright_infringement.
Amidst the rapid advancements in generative language models, the investigation of how training data shapes the performance of GPT models is still emerging. This paper presents GPTfluence, a novel approach that leverages a featurized simulation to assess the impact of training examples on the training dynamics of GPT models. Our approach not only traces the influence of individual training instances on performance trajectories, such as loss and other key metrics, on targeted test points but also enables a comprehensive comparison with existing methods across various training scenarios in GPT models, ranging from 14 million to 2.8 billion parameters, across a range of downstream tasks. Contrary to earlier methods that struggle with generalization to new data, GPTfluence introduces a parameterized simulation of training dynamics, demonstrating robust generalization capabilities to unseen training data. This adaptability is evident across both fine-tuning and instruction-tuning scenarios, spanning tasks in natural language understanding and generation. We make our code and data publicly available at https://github.com/ernie-research/gptfluence.
Concerns regarding the propensity of Large Language Models (LLMs) to produce inaccurate outputs, also known as hallucinations, have escalated. Detecting them is vital for ensuring the reliability of applications relying on LLM-generated content. Current methods often demand substantial resources and rely on extensive LLMs or employ supervised learning with multidimensional features or intricate linguistic and semantic analyses difficult to reproduce and largely depend on using the same LLM that hallucinated. This paper introduces a supervised learning approach employing two simple classifiers utilizing only four numerical features derived from tokens and vocabulary probabilities obtained from other LLM evaluators, which are not necessarily the same. The method yields promising results, surpassing state-of-the-art outcomes in multiple tasks across three different benchmarks. Additionally, we provide a comprehensive examination of the strengths and weaknesses of our approach, highlighting the significance of the features utilized and the LLM employed as an evaluator. We have released our code publicly at https://github.com/Baylor-AI/HalluDetect.