Evaluating Language Model Context Windows: A "Working Memory" Test and Inference-time Correction

Large language models are prominently used in real-world applications, often tasked with reasoning over large volumes of documents. An exciting development in this space is models boasting extended context capabilities, with some accommodating over 2 million tokens. Such long context model capabilities remain uncertain in production systems, motivating the need to benchmark their performance on real world use cases. We address this challenge by proposing SWiM, an evaluation framework that addresses the limitations of standard tests. Testing the framework on eight long context models, we find that even strong models such as GPT-4 and Claude 3 Opus degrade in performance when information is present in the middle of the context window (lost-in-the-middle effect). Next, in addition to our benchmark, we propose medoid voting, a simple, but effective training-free approach that helps alleviate this effect, by generating responses a few times, each time randomly permuting documents in the context, and selecting the medoid answer. We evaluate medoid voting on single document QA tasks, achieving up to a 24% lift in accuracy.

Is Free Self-Alignment Possible?

Aligning pretrained language models (LMs) is a complex and resource-intensive process, often requiring access to large amounts of ground-truth preference data and substantial compute. Are these costs necessary? That is, it is possible to align using only inherent model knowledge and without additional training? We tackle this challenge with AlignEZ, a novel approach that uses (1) self-generated preference data and (2) representation editing to provide nearly cost-free alignment. During inference, AlignEZ modifies LM representations to reduce undesirable and boost desirable components using subspaces identified via self-generated preference pairs. Our experiments reveal that this nearly cost-free procedure significantly narrows the gap between base pretrained and tuned models by an average of 31.6%, observed across six datasets and three model architectures. Additionally, we explore the potential of using AlignEZ as a means of expediting more expensive alignment procedures. Our experiments show that AlignEZ improves DPO models tuned only using a small subset of ground-truth preference data. Lastly, we study the conditions under which improvement using AlignEZ is feasible, providing valuable insights into its effectiveness.

OTTER: Improving Zero-Shot Classification via Optimal Transport

Popular zero-shot models suffer due to artifacts inherited from pretraining. A particularly detrimental artifact, caused by unbalanced web-scale pretraining data, is mismatched label distribution. Existing approaches that seek to repair the label distribution are not suitable in zero-shot settings, as they have incompatible requirements such as access to labeled downstream task data or knowledge of the true label balance in the pretraining distribution. We sidestep these challenges and introduce a simple and lightweight approach to adjust pretrained model predictions via optimal transport. Our technique requires only an estimate of the label distribution of a downstream task. Theoretically, we characterize the improvement produced by our procedure under certain mild conditions and provide bounds on the error caused by misspecification. Empirically, we validate our method in a wide array of zero-shot image and text classification tasks, improving accuracy by 4.8% and 15.9% on average, and beating baselines like Prior Matching -- often by significant margins -- in 17 out of 21 datasets.

Zero-Shot Robustification of Zero-Shot Models With Foundation Models

Zero-shot inference is a powerful paradigm that enables the use of large pretrained models for downstream classification tasks without further training. However, these models are vulnerable to inherited biases that can impact their performance. The traditional solution is fine-tuning, but this undermines the key advantage of pretrained models, which is their ability to be used out-of-the-box. We propose RoboShot, a method that improves the robustness of pretrained model embeddings in a fully zero-shot fashion. First, we use zero-shot language models (LMs) to obtain useful insights from task descriptions. These insights are embedded and used to remove harmful and boost useful components in embeddings -- without any supervision. Theoretically, we provide a simple and tractable model for biases in zero-shot embeddings and give a result characterizing under what conditions our approach can boost performance. Empirically, we evaluate RoboShot on nine image and NLP classification tasks and show an average improvement of 15.98% over several zero-shot baselines. Additionally, we demonstrate that RoboShot is compatible with a variety of pretrained and language models.

Mitigating Source Bias for Fairer Weak Supervision

Weak supervision overcomes the label bottleneck, enabling efficient development of training sets. Millions of models trained on such datasets have been deployed in the real world and interact with users on a daily basis. However, the techniques that make weak supervision attractive -- such as integrating any source of signal to estimate unknown labels -- also ensure that the pseudolabels it produces are highly biased. Surprisingly, given everyday use and the potential for increased bias, weak supervision has not been studied from the point of view of fairness. This work begins such a study. Our departure point is the observation that even when a fair model can be built from a dataset with access to ground-truth labels, the corresponding dataset labeled via weak supervision can be arbitrarily unfair. Fortunately, not all is lost: we propose and empirically validate a model for source unfairness in weak supervision, then introduce a simple counterfactual fairness-based technique that can mitigate these biases. Theoretically, we show that it is possible for our approach to simultaneously improve both accuracy and fairness metrics -- in contrast to standard fairness approaches that suffer from tradeoffs. Empirically, we show that our technique improves accuracy on weak supervision baselines by as much as 32% while reducing demographic parity gap by 82.5%.

Universalizing Weak Supervision

Weak supervision (WS) frameworks are a popular way to bypass hand-labeling large datasets for training data-hungry models. These approaches synthesize multiple noisy but cheaply-acquired estimates of labels into a set of high-quality pseudolabels for downstream training. However, the synthesis technique is specific to a particular kind of label, such as binary labels or sequences, and each new label type requires manually designing a new synthesis algorithm. Instead, we propose a universal technique that enables weak supervision over any label type while still offering desirable properties, including practical flexibility, computational efficiency, and theoretical guarantees. We apply this technique to important problems previously not tackled by WS frameworks including learning to rank, regression, and learning in hyperbolic manifolds. Theoretically, our synthesis approach produces a consistent estimator for learning a challenging but important generalization of the exponential family model. Experimentally, we validate our framework and show improvement over baselines in diverse settings including real-world learning-to-rank and regression problems along with learning on hyperbolic manifolds.

Subtask Gated Networks for Non-Intrusive Load Monitoring

Non-intrusive load monitoring (NILM), also known as energy disaggregation, is a blind source separation problem where a household's aggregate electricity consumption is broken down into electricity usages of individual appliances. In this way, the cost and trouble of installing many measurement devices over numerous household appliances can be avoided, and only one device needs to be installed. The problem has been well-known since Hart's seminal paper in 1992, and recently significant performance improvements have been achieved by adopting deep networks. In this work, we focus on the idea that appliances have on/off states, and develop a deep network for further performance improvements. Specifically, we propose a subtask gated network that combines the main regression network with an on/off classification subtask network. Unlike typical multitask learning algorithms where multiple tasks simply share the network parameters to take advantage of the relevance among tasks, the subtask gated network multiply the main network's regression output with the subtask's classification probability. When standby-power is additionally learned, the proposed solution surpasses the state-of-the-art performance for most of the benchmark cases. The subtask gated network can be very effective for any problem that inherently has on/off states.

On the Statistical and Information-theoretic Characteristics of Deep Network Representations

It has been common to argue or imply that a regularizer can be used to alter a statistical property of a hidden layer's representation and thus improve generalization or performance of deep networks. For instance, dropout has been known to improve performance by reducing co-adaptation, and representational sparsity has been argued as a good characteristic because many data-generation processes have a small number of factors that are independent. In this work, we analytically and empirically investigate the popular characteristics of learned representations, including correlation, sparsity, dead unit, rank, and mutual information, and disprove many of the \textit{conventional wisdom}. We first show that infinitely many Identical Output Networks (IONs) can be constructed for any deep network with a linear layer, where any invertible affine transformation can be applied to alter the layer's representation characteristics. The existence of ION proves that the correlation characteristics of representation is irrelevant to the performance. Extensions to ReLU layers are provided, too. Then, we consider sparsity, dead unit, and rank to show that only loose relationships exist among the three characteristics. It is shown that a higher sparsity or additional dead units do not imply a better or worse performance when the rank of representation is fixed. We also develop a rank regularizer and show that neither representation sparsity nor lower rank is helpful for improving performance even when the data-generation process has a small number of independent factors. Mutual information $I(\mathbf{z}_l;\mathbf{x})$ and $I(\mathbf{z}_l;\mathbf{y})$ are investigated, and we show that regularizers can affect $I(\mathbf{z}_l;\mathbf{x})$ and thus indirectly influence the performance. Finally, we explain how a rich set of regularizers can be used as a powerful tool for performance tuning.