Obeying precise constraints on top of multiple external attributes is a common computational problem underlying seemingly different domains, from controlled text generation to protein engineering. Existing language model (LM) controllability methods for multi-attribute constraint satisfaction often rely on specialized architectures or gradient-based classifiers, limiting their flexibility to work with arbitrary black-box evaluators and pretrained models. Current general-purpose large language models, while capable, cannot achieve fine-grained multi-attribute control over external attributes. Thus, we create Multi-Attribute Constraint Satisfaction (MACS), a generalized method capable of finetuning language models on any sequential domain to satisfy user-specified constraints on multiple external real-value attributes. Our method trains LMs as editors by sampling diverse multi-attribute edit pairs from an initial set of paraphrased outputs. During inference, LM iteratively improves upon its previous solution to satisfy constraints for all attributes by leveraging our designed constraint satisfaction reward. We additionally experiment with reward-weighted behavior cloning to further improve the constraint satisfaction rate of LMs. To evaluate our approach, we present a new Fine-grained Constraint Satisfaction (FineCS) benchmark, featuring two challenging tasks: (1) Text Style Transfer, where the goal is to simultaneously modify the sentiment and complexity of reviews, and (2) Protein Design, focusing on modulating fluorescence and stability of Green Fluorescent Proteins (GFP). Our empirical results show that MACS achieves the highest threshold satisfaction in both FineCS tasks, outperforming strong domain-specific baselines. Our work opens new avenues for generalized and real-value multi-attribute control, with implications for diverse applications spanning NLP and bioinformatics.
Long Context Language Models (LCLMs) have emerged as a new paradigm to perform Information Retrieval (IR), which enables the direct ingestion and retrieval of information by processing an entire corpus in their single context, showcasing the potential to surpass traditional sparse and dense retrieval methods. However, processing a large number of passages within in-context for retrieval is computationally expensive, and handling their representations during inference further exacerbates the processing time; thus, we aim to make LCLM retrieval more efficient and potentially more effective with passage compression. Specifically, we propose a new compression approach tailored for LCLM retrieval, which is trained to maximize the retrieval performance while minimizing the length of the compressed passages. To accomplish this, we generate the synthetic data, where compressed passages are automatically created and labeled as chosen or rejected according to their retrieval success for a given query, and we train the proposed Compression model for Long context Retrieval (CoLoR) with this data via preference optimization while adding the length regularization loss on top of it to enforce brevity. Through extensive experiments on 9 datasets, we show that CoLoR improves the retrieval performance by 6% while compressing the in-context size by a factor of 1.91.
Emerging topics in biomedical research are continuously expanding, providing a wealth of information about genes and their function. This rapid proliferation of knowledge presents unprecedented opportunities for scientific discovery and formidable challenges for researchers striving to keep abreast of the latest advancements. One significant challenge is navigating the vast corpus of literature to extract vital gene-related information, a time-consuming and cumbersome task. To enhance the efficiency of this process, it is crucial to address several key challenges: (1) the overwhelming volume of literature, (2) the complexity of gene functions, and (3) the automated integration and generation. In response, we propose GeneSUM, a two-stage automated gene summary extractor utilizing a large language model (LLM). Our approach retrieves and eliminates redundancy of target gene literature and then fine-tunes the LLM to refine and streamline the summarization process. We conducted extensive experiments to validate the efficacy of our proposed framework. The results demonstrate that LLM significantly enhances the integration of gene-specific information, allowing more efficient decision-making in ongoing research.
Forecasting is a task that is difficult to evaluate: the ground truth can only be known in the future. Recent work showing LLM forecasters rapidly approaching human-level performance begs the question: how can we benchmark and evaluate these forecasters instantaneously? Following the consistency check framework, we measure the performance of forecasters in terms of the consistency of their predictions on different logically-related questions. We propose a new, general consistency metric based on arbitrage: for example, if a forecasting AI illogically predicts that both the Democratic and Republican parties have 60% probability of winning the 2024 US presidential election, an arbitrageur can trade against the forecaster's predictions and make a profit. We build an automated evaluation system that generates a set of base questions, instantiates consistency checks from these questions, elicits the predictions of the forecaster, and measures the consistency of the predictions. We then build a standard, proper-scoring-rule forecasting benchmark, and show that our (instantaneous) consistency metrics correlate with LLM forecasters' ground truth Brier scores (which are only known in the future). We also release a consistency benchmark that resolves in 2028, providing a long-term evaluation tool for forecasting.
Federated learning (FL) is a promising paradigm in distributed learning while preserving the privacy of users. However, the increasing size of recent models makes it unaffordable for a few users to encompass the model. It leads the users to adopt heterogeneous models based on their diverse computing capabilities and network bandwidth. Correspondingly, FL with heterogeneous models should be addressed, given that FL typically involves training a single global model. In this paper, we propose Generative Model-Aided Federated Learning (GeFL), incorporating a generative model that aggregates global knowledge across users of heterogeneous models. Our experiments on various classification tasks demonstrate notable performance improvements of GeFL compared to baselines, as well as limitations in terms of privacy and scalability. To tackle these concerns, we introduce a novel framework, GeFL-F. It trains target networks aided by feature-generative models. We empirically demonstrate the consistent performance gains of GeFL-F, while demonstrating better privacy preservation and robustness to a large number of clients. Codes are available at [1].
Effective pre-training of large language models (LLMs) has been challenging due to the immense resource demands and the complexity of the technical processes involved. This paper presents a detailed technical report on YuLan-Mini, a highly capable base model with 2.42B parameters that achieves top-tier performance among models of similar parameter scale. Our pre-training approach focuses on enhancing training efficacy through three key technical contributions: an elaborate data pipeline combines data cleaning with data schedule strategies, a robust optimization method to mitigate training instability, and an effective annealing approach that incorporates targeted data selection and long context training. Remarkably, YuLan-Mini, trained on 1.08T tokens, achieves performance comparable to industry-leading models that require significantly more data. To facilitate reproduction, we release the full details of the data composition for each training phase. Project details can be accessed at the following link: https://github.com/RUC-GSAI/YuLan-Mini.
Given the ease of creating synthetic data from machine learning models, new models can be potentially trained on synthetic data generated by previous models. This recursive training process raises concerns about the long-term impact on model quality. As models are recursively trained on generated data from previous rounds, their ability to capture the nuances of the original human-generated data may degrade. This is often referred to as \emph{model collapse}. In this work, we ask how fast model collapse occurs for some well-studied distribution families under maximum likelihood (ML or near ML) estimation during recursive training. Surprisingly, even for fundamental distributions such as discrete and Gaussian distributions, the exact rate of model collapse is unknown. In this work, we theoretically characterize the rate of collapse in these fundamental settings and complement it with experimental evaluations. Our results show that for discrete distributions, the time to forget a word is approximately linearly dependent on the number of times it occurred in the original corpus, and for Gaussian models, the standard deviation reduces to zero roughly at $n$ iterations, where $n$ is the number of samples at each iteration. Both of these findings imply that model forgetting, at least in these simple distributions under near ML estimation with many samples, takes a long time.
Low-light image enhancement (LLIE) is a fundamental task in computational photography, aiming to improve illumination, reduce noise, and enhance the image quality of low-light images. While recent advancements primarily focus on customizing complex neural network models, we have observed significant redundancy in these models, limiting further performance improvement. In this paper, we investigate and rethink the model redundancy for LLIE, identifying parameter harmfulness and parameter uselessness. Inspired by the rethinking, we propose two innovative techniques to mitigate model redundancy while improving the LLIE performance: Attention Dynamic Reallocation (ADR) and Parameter Orthogonal Generation (POG). ADR dynamically reallocates appropriate attention based on original attention, thereby mitigating parameter harmfulness. POG learns orthogonal basis embeddings of parameters and prevents degradation to static parameters, thereby mitigating parameter uselessness. Experiments validate the effectiveness of our techniques. We will release the code to the public.
Multimodal large language models (MLLMs) have significantly advanced tasks like caption generation and visual question answering by integrating visual and textual data. However, they sometimes produce misleading or hallucinate content due to discrepancies between their pre-training data and real user prompts. Existing approaches using Direct Preference Optimization (DPO) in vision-language tasks often rely on strong models like GPT-4 or CLIP to determine positive and negative responses. Here, we propose a new framework in generating synthetic data using a reward model as a proxy of human preference for effective multimodal alignment with DPO training. The resulting DPO dataset ranges from 2K to 9K image-text pairs, was evaluated on LLaVA-v1.5-7B, where our approach demonstrated substantial improvements in both the trustworthiness and reasoning capabilities of the base model across multiple hallucination and vision-language benchmark. The experiment results indicate that integrating selected synthetic data, such as from generative and rewards models can effectively reduce reliance on human-annotated data while enhancing MLLMs' alignment capability, offering a scalable solution for safer deployment.
Model merging combines knowledge from task-specific models into a unified multi-task model to avoid joint training on all task data. However, current methods face challenges due to representation bias, which can interfere with tasks performance. As a remedy, we propose IntervMerge, a novel approach to multi-task model merging that effectively mitigates representation bias across the model using taskspecific interventions. To further enhance its efficiency, we introduce mini-interventions, which modify only part of the representation, thereby reducing the additional parameters without compromising performance. Experimental results demonstrate that IntervMerge consistently outperforms the state-of-the-art approaches using fewer parameters.
Learning-based solutions for long-tailed recognition face difficulties in generalizing on balanced test datasets. Due to imbalanced data prior, the learned \textit{a posteriori} distribution is biased toward the most frequent (head) classes, leading to an inferior performance on the least frequent (tail) classes. In general, the performance can be improved by removing such a bias by eliminating the effect of imbalanced prior modeled using the number of class samples (frequencies). We first observe that the \textit{effective prior} on the classes, learned by the model at the end of the training, can differ from the empirical prior obtained using class frequencies. Thus, we propose a novel approach to accurately model the effective prior of a trained model using \textit{a posteriori} probabilities. We propose to correct the imbalanced prior by adjusting the predicted \textit{a posteriori} probabilities (Prior2Posterior: P2P) using the calculated prior in a post-hoc manner after the training, and show that it can result in improved model performance. We present theoretical analysis showing the optimality of our approach for models trained with naive cross-entropy loss as well as logit adjusted loss. Our experiments show that the proposed approach achieves new state-of-the-art (SOTA) on several benchmark datasets from the long-tail literature in the category of logit adjustment methods. Further, the proposed approach can be used to inspect any existing method to capture the \textit{effective prior} and remove any residual bias to improve its performance, post-hoc, without model retraining. We also show that by using the proposed post-hoc approach, the performance of many existing methods can be improved further.
Recent advancements in AI alignment techniques have significantly improved the alignment of large language models (LLMs) with static human preferences. However, the dynamic nature of human preferences can render some prior training data outdated or even erroneous, ultimately causing LLMs to deviate from contemporary human preferences and societal norms. Existing methodologies, whether they involve the curation of new data for continual alignment or the manual correction of outdated data for re-alignment, demand costly human resources. To address this challenge, we propose a novel approach, Large Language Model Behavior Correction with Influence Function Recall and Post-Training (LANCET), which requires no human involvement. LANCET consists of two phases: (1) using influence functions to identify the training data that significantly impact undesirable model outputs, and (2) applying an Influence function-driven Bregman Optimization (IBO) technique to adjust the model's behavior based on these influence distributions. Our experiments demonstrate that LANCET effectively and efficiently correct inappropriate behaviors of LLMs. Furthermore, LANCET can outperform methods that rely on collecting human preferences, and it enhances the interpretability of learning human preferences within LLMs.
The rapid development and deployment of large language models (LLMs) have introduced a new frontier in artificial intelligence, marked by unprecedented capabilities in natural language understanding and generation. However, the increasing integration of these models into critical applications raises substantial safety concerns, necessitating a thorough examination of their potential risks and associated mitigation strategies. This survey provides a comprehensive overview of the current landscape of LLM safety, covering four major categories: value misalignment, robustness to adversarial attacks, misuse, and autonomous AI risks. In addition to the comprehensive review of the mitigation methodologies and evaluation resources on these four aspects, we further explore four topics related to LLM safety: the safety implications of LLM agents, the role of interpretability in enhancing LLM safety, the technology roadmaps proposed and abided by a list of AI companies and institutes for LLM safety, and AI governance aimed at LLM safety with discussions on international cooperation, policy proposals, and prospective regulatory directions. Our findings underscore the necessity for a proactive, multifaceted approach to LLM safety, emphasizing the integration of technical solutions, ethical considerations, and robust governance frameworks. This survey is intended to serve as a foundational resource for academy researchers, industry practitioners, and policymakers, offering insights into the challenges and opportunities associated with the safe integration of LLMs into society. Ultimately, it seeks to contribute to the safe and beneficial development of LLMs, aligning with the overarching goal of harnessing AI for societal advancement and well-being. A curated list of related papers has been publicly available at https://github.com/tjunlp-lab/Awesome-LLM-Safety-Papers.
Large Multimodal Models (LMMs) have demonstrated impressive performance across numerous academic benchmarks. However, fine-tuning still remains essential to achieve satisfactory performance on downstream tasks, while the task-specific tuning samples are usually not readily available or expensive and time-consuming to obtain. To address this, we propose an error-driven data-efficient tuning framework that aims to efficiently adapt generic LMMs to newly emerging tasks without requiring any task-specific training samples. In our approach, a generic LMM, acting as a student model, is first evaluated on a small validation set of the target task, and then a more powerful model, acting as a teacher model, identifies the erroneous steps within the student model's reasoning steps and analyzes its capability gaps from fully addressing the target task. Based on these gaps, targeted training samples are further retrieved from existing task-agnostic datasets to tune the student model and tailor it to the target task. We perform extensive experiments across three different training data scales and seven tasks, demonstrating that our training paradigm significantly and efficiently improves LMM's performance on downstream tasks, achieving an average performance boost of 7.01%.
Model immunization is an emerging direction that aims to mitigate the potential risk of misuse associated with open-sourced models and advancing adaptation methods. The idea is to make the released models' weights difficult to fine-tune on certain harmful applications, hence the name ``immunized''. Recent work on model immunization focuses on the single-concept setting. However, models need to be immunized against multiple concepts in real-world situations. To address this gap, we propose an immunization algorithm that, simultaneously, learns a single ``difficult initialization'' for adaptation methods over a set of concepts. We achieve this by incorporating a differentiable merging layer that combines a set of model weights adapted over multiple concepts. In our experiments, we demonstrate the effectiveness of multi-concept immunization by generalizing prior work's experiment setup of re-learning and personalization adaptation to multiple concepts.
Predicting human mobility is crucial for urban planning, traffic control, and emergency response. Mobility behaviors can be categorized into individual and collective, and these behaviors are recorded by diverse mobility data, such as individual trajectory and crowd flow. As different modalities of mobility data, individual trajectory and crowd flow have a close coupling relationship. Crowd flows originate from the bottom-up aggregation of individual trajectories, while the constraints imposed by crowd flows shape these individual trajectories. Existing mobility prediction methods are limited to single tasks due to modal gaps between individual trajectory and crowd flow. In this work, we aim to unify mobility prediction to break through the limitations of task-specific models. We propose a universal human mobility prediction model (named UniMob), which can be applied to both individual trajectory and crowd flow. UniMob leverages a multi-view mobility tokenizer that transforms both trajectory and flow data into spatiotemporal tokens, facilitating unified sequential modeling through a diffusion transformer architecture. To bridge the gap between the different characteristics of these two data modalities, we implement a novel bidirectional individual and collective alignment mechanism. This mechanism enables learning common spatiotemporal patterns from different mobility data, facilitating mutual enhancement of both trajectory and flow predictions. Extensive experiments on real-world datasets validate the superiority of our model over state-of-the-art baselines in trajectory and flow prediction. Especially in noisy and scarce data scenarios, our model achieves the highest performance improvement of more than 14% and 25% in MAPE and Accuracy@5.
Combining multiple machine learning models has long been a technique for enhancing performance, particularly in distributed settings. Traditional approaches, such as model ensembles, work well, but are expensive in terms of memory and compute. Recently, methods based on averaging model parameters have achieved good results in some settings and have gained popularity. However, merging models initialized differently that do not share a part of their training trajectories can yield worse results than simply using the base models, even after aligning their neurons. In this paper, we introduce a novel approach, Non-uniform Parameter-wise Model Merging, or NP Merge, which merges models by learning the contribution of each parameter to the final model using gradient-based optimization. We empirically demonstrate the effectiveness of our method for merging models of various architectures in multiple settings, outperforming past methods. We also extend NP Merge to handle the merging of multiple models, showcasing its scalability and robustness.
Conflict scholars have used rule-based approaches to extract information about political violence from news reports and texts. Recent Natural Language Processing developments move beyond rigid rule-based approaches. We review our recent ConfliBERT language model (Hu et al. 2022) to process political and violence related texts. The model can be used to extract actor and action classifications from texts about political conflict. When fine-tuned, results show that ConfliBERT has superior performance in accuracy, precision and recall over other large language models (LLM) like Google's Gemma 2 (9B), Meta's Llama 3.1 (7B), and Alibaba's Qwen 2.5 (14B) within its relevant domains. It is also hundreds of times faster than these more generalist LLMs. These results are illustrated using texts from the BBC, re3d, and the Global Terrorism Dataset (GTD).
Sensor data collected by Internet of Things (IoT) devices carries detailed information about individuals in their vicinity. Sharing this data with a semi-trusted service provider may compromise the individuals' privacy, as sensitive information can be extracted by powerful machine learning models. Data obfuscation empowered by generative models is a promising approach to generate synthetic sensor data such that the useful information contained in the original data is preserved and the sensitive information is obscured. This newly generated data will then be shared with the service provider instead of the original sensor data. In this work, we propose PrivDiffuser, a novel data obfuscation technique based on a denoising diffusion model that attains a superior trade-off between data utility and privacy through effective guidance techniques. Specifically, we extract latent representations that contain information about public and private attributes from sensor data to guide the diffusion model, and impose mutual information-based regularization when learning the latent representations to alleviate the entanglement of public and private attributes, thereby increasing the effectiveness of guidance. Evaluation on three real-world datasets containing different sensing modalities reveals that PrivDiffuser yields a better privacy-utility trade-off than the state-of-the-art obfuscation model, decreasing the utility loss by up to $1.81\%$ and the privacy loss by up to $3.42\%$. Moreover, we showed that users with diverse privacy needs can use PrivDiffuser to protect their privacy without having to retrain the model.
Model misspecification analysis strategies, such as anomaly detection, model validation, and model comparison are a key component of scientific model development. Over the last few years, there has been a rapid rise in the use of simulation-based inference (SBI) techniques for Bayesian parameter estimation, applied to increasingly complex forward models. To move towards fully simulation-based analysis pipelines, however, there is an urgent need for a comprehensive simulation-based framework for model misspecification analysis. In this work, we provide a solid and flexible foundation for a wide range of model discrepancy analysis tasks, using distortion-driven model misspecification tests. From a theoretical perspective, we introduce the statistical framework built around performing many hypothesis tests for distortions of the simulation model. We also make explicit analytic connections to classical techniques: anomaly detection, model validation, and goodness-of-fit residual analysis. Furthermore, we introduce an efficient self-calibrating training algorithm that is useful for practitioners. We demonstrate the performance of the framework in multiple scenarios, making the connection to classical results where they are valid. Finally, we show how to conduct such a distortion-driven model misspecification test for real gravitational wave data, specifically on the event GW150914.