Measuring Progress in Dictionary Learning for Language Model Interpretability with Board Game Models
What latent features are encoded in language model (LM) representations? Recent work on training sparse autoencoders (SAEs) to disentangle interpretable features in LM representations has shown significant promise. However, evaluating the quality of these SAEs is difficult because we lack a ground-truth collection of interpretable features that we expect good SAEs to recover. We thus propose to measure progress in interpretable dictionary learning by working in the setting of LMs trained on chess and Othello transcripts. These settings carry natural collections of interpretable features -- for example, "there is a knight on F3" -- which we leverage into $\textit{supervised}$ metrics for SAE quality. To guide progress in interpretable dictionary learning, we introduce a new SAE training technique, $\textit{p-annealing}$, which improves performance on prior unsupervised metrics as well as our new metrics.
相關內容
ACM/IEEE第23屆模型驅動工程語言和系統國際會議，是模型驅動軟件和系統工程的首要會議系列，由ACM-SIGSOFT和IEEE-TCSE支持組織。自1998年以來，模型涵蓋了建模的各個方面，從語言和方法到工具和應用程序。模特的參加者來自不同的背景，包括研究人員、學者、工程師和工業專業人士。MODELS 2019是一個論壇，參與者可以圍繞建模和模型驅動的軟件和系統交流前沿研究成果和創新實踐經驗。今年的版本將為建模社區提供進一步推進建模基礎的機會，并在網絡物理系統、嵌入式系統、社會技術系統、云計算、大數據、機器學習、安全、開源等新興領域提出建模的創新應用以及可持續性。
官網鏈接： ·
欠估計
·
語言模型化
·
可辨認的
·
大語言模型
·
Large Language Models are trained on extensive datasets that often contain sensitive, human-generated information, raising significant concerns about privacy breaches. While certified unlearning approaches offer strong privacy guarantees, they rely on restrictive model assumptions that are not applicable to LLMs. As a result, various unlearning heuristics have been proposed, with the associated privacy risks assessed only empirically. The standard evaluation pipelines typically randomly select data for removal from the training set, apply unlearning techniques, and use membership inference attacks to compare the unlearned models against models retrained without the to-be-unlearned data. However, since every data point is subject to the right to be forgotten, unlearning should be considered in the worst-case scenario from the privacy perspective. Prior work shows that data outliers may exhibit higher memorization effects. Intuitively, they are harder to be unlearn and thus the privacy risk of unlearning them is underestimated in the current evaluation. In this paper, we leverage minority data to identify such a critical flaw in previously widely adopted evaluations. We substantiate this claim through carefully designed experiments, including unlearning canaries related to minority groups, inspired by privacy auditing literature. Using personally identifiable information as a representative minority identifier, we demonstrate that minority groups experience at least 20% more privacy leakage in most cases across six unlearning approaches, three MIAs, three benchmark datasets, and two LLMs of different scales. Given that the right to be forgotten should be upheld for every individual, we advocate for a more rigorous evaluation of LLM unlearning methods. Our minority-aware evaluation framework represents an initial step toward ensuring more equitable assessments of LLM unlearning efficacy.
Self-supervised learning through masked autoencoders (MAEs) has recently attracted great attention for remote sensing (RS) image representation learning, and thus embodies a significant potential for content-based image retrieval (CBIR) from ever-growing RS image archives. However, the existing MAE based CBIR studies in RS assume that the considered RS images are acquired by a single image sensor, and thus are only suitable for uni-modal CBIR problems. The effectiveness of MAEs for cross-sensor CBIR, which aims to search semantically similar images across different image modalities, has not been explored yet. In this paper, we take the first step to explore the effectiveness of MAEs for sensor-agnostic CBIR in RS. To this end, we present a systematic overview on the possible adaptations of the vanilla MAE to exploit masked image modeling on multi-sensor RS image archives (denoted as cross-sensor masked autoencoders [CSMAEs]) in the context of CBIR. Based on different adjustments applied to the vanilla MAE, we introduce different CSMAE models. We also provide an extensive experimental analysis of these CSMAE models. We finally derive a guideline to exploit masked image modeling for uni-modal and cross-modal CBIR problems in RS. The code of this work is publicly available at //github.com/jakhac/CSMAE.
We study Online Convex Optimization (OCO) with adversarial constraints, where an online algorithm must make repeated decisions to minimize both convex loss functions and cumulative constraint violations. We focus on a setting where the algorithm has access to predictions of the loss and constraint functions. Our results show that we can improve the current best bounds of $ O(\sqrt{T}) $ regret and $ \tilde{O}(\sqrt{T}) $ cumulative constraint violations to $ O(\sqrt{E_T(f)}) $ and $ \tilde{O}(\sqrt{E_T(g)}) $, respectively, where $ E_T(f) $ and $ E_T(g) $ represent the cumulative prediction errors of the loss and constraint functions. In the worst case, where $ E_T(f) = O(T) $ and $ E_T(g) = O(T) $ (assuming bounded loss and constraint functions), our rates match the prior $ O(\sqrt{T}) $ results. However, when the loss and constraint predictions are accurate, our approach yields significantly smaller regret and cumulative constraint violations. Notably, if the constraint function remains constant over time, we achieve $ \tilde{O}(1) $ cumulative constraint violation, aligning with prior results.
Training data attribution (TDA) methods aim to attribute model outputs back to specific training examples, and the application of these methods to large language model (LLM) outputs could significantly advance model transparency and data curation. However, it has been challenging to date to apply these methods to the full scale of LLM pretraining. In this paper, we refine existing gradient-based methods to work effectively at scale, allowing us to retrieve influential examples for an 8B-parameter language model from a pretraining corpus of over 160B tokens with no need for subsampling or pre-filtering. Our method combines several techniques, including optimizer state correction, a task-specific Hessian approximation, and normalized encodings, which we find to be critical for performance at scale. In quantitative evaluations on a fact tracing task, our method performs best at identifying examples that influence model predictions, but classical, model-agnostic retrieval methods such as BM25 still perform better at finding passages which explicitly contain relevant facts. These results demonstrate a misalignment between factual *attribution* and causal *influence*. With increasing model size and training tokens, we find that influence more closely aligns with factual attribution. Finally, we examine different types of examples identified as influential by our method, finding that while many directly entail a particular fact, others support the same output by reinforcing priors on relation types, common entities, and names. We release our prompt set and model outputs, along with a web-based visualization tool to explore influential examples for factual predictions, commonsense reasoning, arithmetic, and open-ended generation for an 8B-parameter LLM.
Counterfactual explanations (CFEs) guide users on how to adjust inputs to machine learning models to achieve desired outputs. While existing research primarily addresses static scenarios, real-world applications often involve data or model changes, potentially invalidating previously generated CFEs and rendering user-induced input changes ineffective. Current methods addressing this issue often support only specific models or change types, require extensive hyperparameter tuning, or fail to provide probabilistic guarantees on CFE robustness to model changes. This paper proposes a novel approach for generating CFEs that provides probabilistic guarantees for any model and change type, while offering interpretable and easy-to-select hyperparameters. We establish a theoretical framework for probabilistically defining robustness to model change and demonstrate how our BetaRCE method directly stems from it. BetaRCE is a post-hoc method applied alongside a chosen base CFE generation method to enhance the quality of the explanation beyond robustness. It facilitates a transition from the base explanation to a more robust one with user-adjusted probability bounds. Through experimental comparisons with baselines, we show that BetaRCE yields robust, most plausible, and closest to baseline counterfactual explanations.
Large vision-language models (LVLMs) have demonstrated remarkable capabilities in multimodal understanding and generation tasks. However, these models occasionally generate hallucinatory texts, resulting in descriptions that seem reasonable but do not correspond to the image. This phenomenon can lead to wrong driving decisions of the autonomous driving system. To address this challenge, this paper proposes HCOENet, a plug-and-play chain-of-thought correction method designed to eliminate object hallucinations and generate enhanced descriptions for critical objects overlooked in the initial response. Specifically, HCOENet employs a cross-checking mechanism to filter entities and directly extracts critical objects from the given image, enriching the descriptive text. Experimental results on the POPE benchmark demonstrate that HCOENet improves the F1-score of the Mini-InternVL-4B and mPLUG-Owl3 models by 12.58% and 4.28%, respectively. Additionally, qualitative results using images collected in open campus scene further highlight the practical applicability of the proposed method. Compared with the GPT-4o model, HCOENet achieves comparable descriptive performance while significantly reducing costs. Finally, two novel semantic understanding datasets, CODA_desc and nuScenes_desc, are created for traffic scenarios to support future research. The codes and datasets are publicly available at //github.com/fjq-tongji/HCOENet.
Large language models (LLMs), trained on diverse data effectively acquire a breadth of information across various domains. However, their computational complexity, cost, and lack of transparency hinder their direct application for specialised tasks. In fields such as clinical research, acquiring expert annotations or prior knowledge about predictive models is often costly and time-consuming. This study proposes the use of LLMs to elicit expert prior distributions for predictive models. This approach also provides an alternative to in-context learning, where language models are tasked with making predictions directly. In this work, we compare LLM-elicited and uninformative priors, evaluate whether LLMs truthfully generate parameter distributions, and propose a model selection strategy for in-context learning and prior elicitation. Our findings show that LLM-elicited prior parameter distributions significantly reduce predictive error compared to uninformative priors in low-data settings. Applied to clinical problems, this translates to fewer required biological samples, lowering cost and resources. Prior elicitation also consistently outperforms and proves more reliable than in-context learning at a lower cost, making it a preferred alternative in our setting. We demonstrate the utility of this method across various use cases, including clinical applications. For infection prediction, using LLM-elicited priors reduced the number of required labels to achieve the same accuracy as an uninformative prior by 55%, 200 days earlier in the study.
Distributed training methods are crucial for large language models (LLMs). However, existing distributed training methods often suffer from communication bottlenecks, stragglers, and limited elasticity. Local SGD methods have been proposed to address these issues, but their effectiveness remains limited to small-scale training due to additional memory overhead and lack of concerns on efficiency and stability. To tackle these issues, we propose EDiT, an innovative Efficient Distributed Training method that combines a tailored Local SGD approach with model sharding techniques to enhance large-scale training efficiency. EDiT performs layer-wise parameter synchronization during forward pass, reducing communication and memory overhead and enabling the overlap of computation and communication. Besides, EDiT employs a pseudo gradient penalty strategy to suppress loss spikes, which ensures training stability and improve performance. Additionally, we introduce A-EDiT, a fully asynchronous variant of EDiT that accommodates heterogeneous clusters. Building on EDiT/A-EDiT, we conduct a series of experiments to validate large-scale asynchronous training for LLMs, accompanied by comprehensive analyses. Experimental results demonstrate the superior performance of EDiT/A-EDiT, establishing them as robust solutions for distributed LLM training in diverse computational ecosystems.
Despite significant advancements in vision-language models (VLMs), there lacks effective approaches to enhance response quality by scaling inference-time computation. This capability is known to be a core step towards the self-improving models in recent large language model studies. In this paper, we present Vision Value Model (VisVM) that can guide VLM inference-time search to generate responses with better visual comprehension. Specifically, VisVM not only evaluates the generated sentence quality in the current search step, but also anticipates the quality of subsequent sentences that may result from the current step, thus providing a long-term value. In this way, VisVM steers VLMs away from generating sentences prone to hallucinations or insufficient detail, thereby producing higher quality responses. Experimental results demonstrate that VisVM-guided search significantly enhances VLMs' ability to generate descriptive captions with richer visual details and fewer hallucinations, compared with greedy decoding and search methods with other visual reward signals. Furthermore, we find that self-training the model with the VisVM-guided captions improve VLM's performance across a wide range of multimodal benchmarks, indicating the potential for developing self-improving VLMs. Our value model and code are available at //github.com/si0wang/VisVM.
Translational distance-based knowledge graph embedding has shown progressive improvements on the link prediction task, from TransE to the latest state-of-the-art RotatE. However, N-1, 1-N and N-N predictions still remain challenging. In this work, we propose a novel translational distance-based approach for knowledge graph link prediction. The proposed method includes two-folds, first we extend the RotatE from 2D complex domain to high dimension space with orthogonal transforms to model relations for better modeling capacity. Second, the graph context is explicitly modeled via two directed context representations. These context representations are used as part of the distance scoring function to measure the plausibility of the triples during training and inference. The proposed approach effectively improves prediction accuracy on the difficult N-1, 1-N and N-N cases for knowledge graph link prediction task. The experimental results show that it achieves better performance on two benchmark data sets compared to the baseline RotatE, especially on data set (FB15k-237) with many high in-degree connection nodes.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191