Large language models (LLMs) have shown amazing performance on tasks that require planning and reasoning. Motivated by this, we investigate the internal mechanisms that underpin a network's ability to perform complex logical reasoning. We first construct a synthetic propositional logic problem that serves as a concrete test-bed for network training and evaluation. Crucially, this problem demands nontrivial planning to solve. We perform our study on two fronts. First, we pursue an understanding of precisely how a three-layer transformer, trained from scratch and attains perfect test accuracy, solves this problem. We are able to identify certain "planning" and "reasoning" mechanisms in the network that necessitate cooperation between the attention blocks to implement the desired logic. Second, we study how pretrained LLMs, namely Mistral-7B and Gemma-2-9B, solve this problem. We characterize their reasoning circuits through causal intervention experiments, providing necessity and sufficiency evidence for the circuits. We find evidence suggesting that the two models' latent reasoning strategies are surprisingly similar, and human-like. Overall, our work systemically uncovers novel aspects of small and large transformers, and continues the study of how they plan and reason.

Augmenting LLMs with context leads to improved performance across many applications. Despite much research on Retrieval Augmented Generation (RAG) systems, an open question is whether errors arise because LLMs fail to utilize the context from retrieval or the context itself is insufficient to answer the query. To shed light on this, we develop a new notion of sufficient context, along with a way to classify instances that have enough information to answer the query. We then use sufficient context to analyze several models and datasets. By stratifying errors based on context sufficiency, we find that proprietary LLMs (Gemini, GPT, Claude) excel at answering queries when the context is sufficient, but often output incorrect answers instead of abstaining when the context is not. We further categorize cases when the context is useful, and improves accuracy, even though it does not fully answer the query and the model errs without the context. Building on our findings, we explore ways to reduce hallucinations in RAG systems, including a new selective generation method that leverages sufficient context information for guided abstention. Our method improves the fraction of correct answers among times where the model responds by 2-10% for Gemini, GPT, and Gemma. Providing Large Language Models (LLMs) with additional context, such as in Retrieval Augmented Generation (RAG) systems, has led to major improvements in LLM factuality and verifiability when adapting to new domains (Lewis et al., 2020). In the case of open-domain question answering, a retrieval model provides context at inference time in the form of snippets or long-form text (Zhu et al., 2021). Then, the model synthesizes the query along with this added context to generate the answer. The ideal outcome is for the LLM to output the correct answer if the provided context contains enough information to answer the question when combined with the model's parametric knowledge. Otherwise, the model should abstain from answering and/or ask for more information. One core challenge in achieving this ideal outcome is building models that can use the provided context only when it helps answer the question correctly. Several works have investigated this issue by evaluating models in the presence of irrelevant information in the context (discussed in Section 2). However, "relevant information" can range from directly containing the answer to simply being topically related Work done during an internship at Google. Work done during an internship at Google. Question: Who is Lya L. married to?

Large language models (LLMs) have demonstrated remarkable capabilities, but their outputs can sometimes be unreliable or factually incorrect. To address this, we introduce Self Logits Evolution Decoding (SLED), a novel decoding framework that enhances the truthfulness of LLMs without relying on external knowledge bases or requiring further fine-tuning. From an optimization perspective, our SLED framework leverages the latent knowledge embedded within the LLM by contrasting the output logits from the final layer with those from early layers. It then utilizes an approximate gradient approach to enable latent knowledge to guide the self-refinement of outputs, thereby effectively improving factual accuracy. Extensive experiments have been conducted on established benchmarks across a diverse range of model families (LLaMA 2, LLaMA 3, Gemma) and scales (from 2B to 70B), including more advanced architectural configurations such as the mixture of experts (MoE). Our evaluation spans a wide variety of tasks, including multi-choice, open-generation, and adaptations to chain-of-thought reasoning tasks. The results demonstrate that SLED consistently improves factual accuracy by up to 20\% compared to existing decoding methods while maintaining natural language fluency and negligible latency overhead. Furthermore, it can be flexibly combined with other decoding methods to further enhance their performance.

Automated content filtering and moderation is an important tool that allows online platforms to build striving user communities that facilitate cooperation and prevent abuse. Unfortunately, resourceful actors try to bypass automated filters in a bid to post content that violate platform policies and codes of conduct. To reach this goal, these malicious actors may obfuscate policy violating images (e.g. overlay harmful images by carefully selected benign images or visual patterns) to prevent machine learning models from reaching the correct decision. In this paper, we invite researchers to tackle this specific issue and present a new image benchmark. This benchmark, based on ImageNet, simulates the type of obfuscations created by malicious actors. It goes beyond ImageNet-$\textrm{C}$ and ImageNet-$\bar{\textrm{C}}$ by proposing general, drastic, adversarial modifications that preserve the original content intent. It aims to tackle a more common adversarial threat than the one considered by $\ell_p$-norm bounded adversaries. We evaluate 33 pretrained models on the benchmark and train models with different augmentations, architectures and training methods on subsets of the obfuscations to measure generalization. We hope this benchmark will encourage researchers to test their models and methods and try to find new approaches that are more robust to these obfuscations.

Despite their wide-spread success, Text-to-Image models (T2I) still struggle to produce images that are both aesthetically pleasing and faithful to the user's input text. We introduce DreamSync, a model-agnostic training algorithm by design that improves T2I models to be faithful to the text input. DreamSync builds off a recent insight from TIFA's evaluation framework -- that large vision-language models (VLMs) can effectively identify the fine-grained discrepancies between generated images and the text inputs. DreamSync uses this insight to train T2I models without any labeled data; it improves T2I models using its own generations. First, it prompts the model to generate several candidate images for a given input text. Then, it uses two VLMs to select the best generation: a Visual Question Answering model that measures the alignment of generated images to the text, and another that measures the generation's aesthetic quality. After selection, we use LoRA to iteratively finetune the T2I model to guide its generation towards the selected best generations. DreamSync does not need any additional human annotation. model architecture changes, or reinforcement learning. Despite its simplicity, DreamSync improves both the semantic alignment and aesthetic appeal of two diffusion-based T2I models, evidenced by multiple benchmarks (+1.7% on TIFA, +2.9% on DSG1K, +3.4% on VILA aesthetic) and human evaluation.

Probes are small networks that predict properties of underlying data from embeddings, and they provide a targeted, effective way to illuminate the information contained in embeddings. While analysis through the use of probes has become standard in NLP, there has been much less exploration in vision. Image foundation models have primarily been evaluated for semantic content. Better understanding the non-semantic information in popular embeddings (e.g., MAE, SimCLR, or CLIP) will shed new light both on the training algorithms and on the uses for these foundation models. We design a systematic transformation prediction task and measure the visual content of embeddings along many axes, including image style, quality, and a range of natural and artificial transformations. Surprisingly, six embeddings (including SimCLR) encode enough non-semantic information to identify dozens of transformations. We also consider a generalization task, where we group similar transformations and hold out several for testing. We find that image-text models (CLIP and ALIGN) are better at recognizing new examples of style transfer than masking-based models (CAN and MAE). Overall, our results suggest that the choice of pre-training algorithm impacts the types of information in the embedding, and certain models are better than others for non-semantic downstream tasks.

Out-of-distribution generalization is a core challenge in machine learning. We introduce and propose a solution to a new type of out-of-distribution evaluation, which we call close category generalization. This task specifies how a classifier should extrapolate to unseen classes by considering a bi-criteria objective: (i) on in-distribution examples, output the correct label, and (ii) on out-of-distribution examples, output the label of the nearest neighbor in the training set. In addition to formalizing this problem, we present a new training algorithm to improve the close category generalization of neural networks. We compare to many baselines, including robust algorithms and out-of-distribution detection methods, and we show that our method has better or comparable close category generalization. Then, we investigate a related representation learning task, and we find that performing well on close category generalization correlates with learning a good representation of an unseen class and with finding a good initialization for few-shot learning. Code available at https://github.com/yangarbiter/close-category-generalization

Hierarchical clustering aims to find an iterative grouping of increasingly large subsets of data, terminating when one cluster contains all of the data. This results in a tree structure, where each level determines a partition of the data. Organizing the data in such a way has many applications, including automated taxonomy construction and phylogeny reconstruction [6, 26, 54, 69, 87]. Motivated by these applications, we address the simultaneous goals of recovering an underlying clustering and building a hierarchical tree of the entire dataset. We imagine that we have access to many samples from each individual cluster (e.g., images of plants and animals), while we lack labels or any direct information about the hierarchical relationships. In this scenario, our objective is to correctly cluster the samples in each group and also build a dendrogram within and on top of the clusters that matches the natural supergroups. Despite significant effort, hierarchical clustering remains a challenging task, theoretically and empirically. Many objectives are NP-Hard [13, 14, 18, 19, 35, 59], and many theoretical algorithms may not be practical because they require solving a hard subproblem (e.g., sparsest cut) at each step [2, 4, 13, 14, 19, 51].

Despite the popularity of explainable AI, there is limited work on effective methods for unsupervised learning. We study algorithms for $k$-means clustering, focusing on a trade-off between explainability and accuracy. Following prior work, we use a small decision tree to partition a dataset into $k$ clusters. This enables us to explain each cluster assignment by a short sequence of single-feature thresholds. While larger trees produce more accurate clusterings, they also require more complex explanations. To allow flexibility, we develop a new explainable $k$-means clustering algorithm, ExKMC, that takes an additional parameter $k' \geq k$ and outputs a decision tree with $k'$ leaves. We use a new surrogate cost to efficiently expand the tree and to label the leaves with one of $k$ clusters. We prove that as $k'$ increases, the surrogate cost is non-increasing, and hence, we trade explainability for accuracy. Empirically, we validate that ExKMC produces a low cost clustering, outperforming both standard decision tree methods and other algorithms for explainable clustering. Implementation of ExKMC available at https://github.com/navefr/ExKMC.

Clustering is a popular form of unsupervised learning for geometric data. Unfortunately, many clustering algorithms lead to cluster assignments that are hard to explain, partially because they depend on all the features of the data in a complicated way. To improve interpretability, we consider using a small decision tree to partition a data set into clusters, so that clusters can be characterized in a straightforward manner. We study this problem from a theoretical viewpoint, measuring cluster quality by the $k$-means and $k$-medians objectives: Must there exist a tree-induced clustering whose cost is comparable to that of the best unconstrained clustering, and if so, how can it be found? In terms of negative results, we show, first, that popular top-down decision tree algorithms may lead to clusterings with arbitrarily large cost, and second, that any tree-induced clustering must in general incur an $\Omega(\log k)$ approximation factor compared to the optimal clustering. On the positive side, we design an efficient algorithm that produces explainable clusters using a tree with $k$ leaves. For two means/medians, we show that a single threshold cut suffices to achieve a constant factor approximation, and we give nearly-matching lower bounds. For general $k \geq 2$, our algorithm is an $O(k)$ approximation to the optimal $k$-medians and an $O(k^2)$ approximation to the optimal $k$-means. Prior to our work, no algorithms were known with provable guarantees independent of dimension and input size.