EmbedLLM: Learning Compact Representations of Large Language Models

With hundreds of thousands of language models available on Huggingface today, efficiently evaluating and utilizing these models across various downstream, tasks has become increasingly critical. Many existing methods repeatedly learn task-specific representations of Large Language Models (LLMs), which leads to inefficiencies in both time and computational resources. To address this, we propose EmbedLLM, a framework designed to learn compact vector representations, of LLMs that facilitate downstream applications involving many models, such as model routing. We introduce an encoder-decoder approach for learning such embeddings, along with a systematic framework to evaluate their effectiveness. Empirical results show that EmbedLLM outperforms prior methods in model routing both in accuracy and latency. Additionally, we demonstrate that our method can forecast a model's performance on multiple benchmarks, without incurring additional inference cost. Extensive probing experiments validate that the learned embeddings capture key model characteristics, e.g. whether the model is specialized for coding tasks, even without being explicitly trained on them. We open source our dataset, code and embedder to facilitate further research and application.

Transformers on Markov Data: Constant Depth Suffices

Attention-based transformers have been remarkably successful at modeling generative processes across various domains and modalities. In this paper, we study the behavior of transformers on data drawn from \kth Markov processes, where the conditional distribution of the next symbol in a sequence depends on the previous $k$ symbols observed. We observe a surprising phenomenon empirically which contradicts previous findings: when trained for sufficiently long, a transformer with a fixed depth and $1$ head per layer is able to achieve low test loss on sequences drawn from \kth Markov sources, even as $k$ grows. Furthermore, this low test loss is achieved by the transformer's ability to represent and learn the in-context conditional empirical distribution. On the theoretical side, our main result is that a transformer with a single head and three layers can represent the in-context conditional empirical distribution for \kth Markov sources, concurring with our empirical observations. Along the way, we prove that \textit{attention-only} transformers with $O(\log_2(k))$ layers can represent the in-context conditional empirical distribution by composing induction heads to track the previous $k$ symbols in the sequence. These results provide more insight into our current understanding of the mechanisms by which transformers learn to capture context, by understanding their behavior on Markov sources.

Toward a Theory of Tokenization in LLMs

While there has been a large body of research attempting to circumvent tokenization for language modeling (Clark et al., 2022; Xue et al., 2022), the current consensus is that it is a necessary initial step for designing state-of-the-art performant language models. In this paper, we investigate tokenization from a theoretical point of view by studying the behavior of transformers on simple data generating processes. When trained on data drawn from certain simple $k^{\text{th}}$-order Markov processes for $k > 1$, transformers exhibit a surprising phenomenon - in the absence of tokenization, they empirically fail to learn the right distribution and predict characters according to a unigram model (Makkuva et al., 2024). With the addition of tokenization, however, we empirically observe that transformers break through this barrier and are able to model the probabilities of sequences drawn from the source near-optimally, achieving small cross-entropy loss. With this observation as starting point, we study the end-to-end cross-entropy loss achieved by transformers with and without tokenization. With the appropriate tokenization, we show that even the simplest unigram models (over tokens) learnt by transformers are able to model the probability of sequences drawn from $k^{\text{th}}$-order Markov sources near optimally. Our analysis provides a justification for the use of tokenization in practice through studying the behavior of transformers on Markovian data.

Learning to Understand: Identifying Interactions via the Mobius Transform

One of the most fundamental problems in machine learning is finding interpretable representations of the functions we learn. The Mobius transform is a useful tool for this because its coefficients correspond to unique importance scores on sets of input variables. The Mobius Transform is strongly related (and in some cases equivalent) to the concept of Shapley value, which is a widely used game-theoretic notion of importance. This work focuses on the (typical) regime where the fraction of non-zero Mobius coefficients (and thus interactions between inputs) is small compared to the set of all $2^n$ possible interactions between $n$ inputs. When there are $K = O(2^{n \delta})$ with $\delta \leq \frac{1}{3}$ non-zero coefficients chosen uniformly at random, our algorithm exactly recovers the Mobius transform in $O(Kn)$ samples and $O(Kn^2)$ time with vanishing error as $K \rightarrow \infty$, the first non-adaptive algorithm to do so. We also uncover a surprising connection between group testing and the Mobius transform. In the case where all interactions are between at most $t = \Theta(n^{\alpha})$ inputs, for $\alpha < 0.409$, we are able to leverage results from group testing to provide the first algorithm that computes the Mobius transform in $O(Kt\log n)$ sample complexity and $O(K\mathrm{poly}(n))$ time with vanishing error as $K \rightarrow \infty$. Finally, we present a robust version of this algorithm that achieves the same sample and time complexity under some assumptions, but with a factor depending on noise variance. Our work is deeply interdisciplinary, drawing from tools spanning across signal processing, algebra, information theory, learning theory and group testing to address this important problem at the forefront of machine learning.

Pairwise Proximal Policy Optimization: Harnessing Relative Feedback for LLM Alignment

Large Language Models (LLMs) can acquire extensive world knowledge through pre-training on large corpora. However, due to exposure to low-quality data, LLMs may exhibit harmful behavior without aligning with human values. The dominant approach for steering LLMs towards beneficial behavior involves Reinforcement Learning with Human Feedback (RLHF), with Proximal Policy Optimization (PPO) serving as the default RL optimizer. Despite its effectiveness, PPO has limitations when optimizing rewards trained from comparison-based loss. Primarily, PPO is not invariant to equivalent reward functions containing identical preference information due to the need to calibrate the reward scale. Additionally, PPO's necessity for token-wise updates introduces complexity in both function approximation and algorithm design compared to trajectory-wise optimization. This paper proposes a new framework, reinforcement learning with relative feedback, and a novel trajectory-wise policy gradient algorithm, Pairwise Proximal Policy Optimization (P3O) that operates directly on comparative rewards. We show theoretically that P3O is invariant to equivalent rewards and avoids the complexity of PPO. Empirical evaluations demonstrate that P3O outperforms PPO in the KL-Reward trade-off and can align with human preferences as well as or better than prior methods. In summary, this work introduces a simpler yet effective approach for aligning LLMs to human preferences through relative feedback.

Conditional Score-Based Reconstructions for Multi-contrast MRI

Magnetic resonance imaging (MRI) exam protocols consist of multiple contrast-weighted images of the same anatomy to emphasize different tissue properties. Due to the long acquisition times required to collect fully sampled k-space measurements, it is common to only collect a fraction of k-space for some, or all, of the scans and subsequently solve an inverse problem for each contrast to recover the desired image from sub-sampled measurements. Recently, there has been a push to further accelerate MRI exams using data-driven priors, and generative models in particular, to regularize the ill-posed inverse problem of image reconstruction. These methods have shown promising improvements over classical methods. However, many of the approaches neglect the multi-contrast nature of clinical MRI exams and treat each scan as an independent reconstruction. In this work we show that by learning a joint Bayesian prior over multi-contrast data with a score-based generative model we are able to leverage the underlying structure between multi-contrast images and thus improve image reconstruction fidelity over generative models that only reconstruct images of a single contrast.

Greedy Pruning with Group Lasso Provably Generalizes for Matrix Sensing and Neural Networks with Quadratic Activations

Pruning schemes have been widely used in practice to reduce the complexity of trained models with a massive number of parameters. Several practical studies have shown that pruning an overparameterized model and fine-tuning generalizes well to new samples. Although the above pipeline, which we refer to as pruning + fine-tuning, has been extremely successful in lowering the complexity of trained models, there is very little known about the theory behind this success. In this paper we address this issue by investigating the pruning + fine-tuning framework on the overparameterized matrix sensing problem, with the ground truth denoted $U_\star \in \mathbb{R}^{d \times r}$ and the overparameterized model $U \in \mathbb{R}^{d \times k}$ with $k \gg r$. We study the approximate local minima of the empirical mean square error, augmented with a smooth version of a group Lasso regularizer, $\sum_{i=1}^k \| U e_i \|_2$ and show that pruning the low $\ell_2$-norm columns results in a solution $U_{\text{prune}}$ which has the minimum number of columns $r$, yet is close to the ground truth in training loss. Initializing the subsequent fine-tuning phase from $U_{\text{prune}}$, the resulting solution converges linearly to a generalization error of $O(\sqrt{rd/n})$ ignoring lower order terms, which is statistically optimal. While our analysis provides insights into the role of regularization in pruning, we also show that running gradient descent in the absence of regularization results in models which {are not suitable for greedy pruning}, i.e., many columns could have their $\ell_2$ norm comparable to that of the maximum. Lastly, we extend our results for the training and pruning of two-layer neural networks with quadratic activation functions. Our results provide the first rigorous insights on why greedy pruning + fine-tuning leads to smaller models which also generalize well.

Beyond UCB: Statistical Complexity and Optimal Algorithms for Non-linear Ridge Bandits

We consider the sequential decision-making problem where the mean outcome is a non-linear function of the chosen action. Compared with the linear model, two curious phenomena arise in non-linear models: first, in addition to the "learning phase" with a standard parametric rate for estimation or regret, there is an "burn-in period" with a fixed cost determined by the non-linear function; second, achieving the smallest burn-in cost requires new exploration algorithms. For a special family of non-linear functions named ridge functions in the literature, we derive upper and lower bounds on the optimal burn-in cost, and in addition, on the entire learning trajectory during the burn-in period via differential equations. In particular, a two-stage algorithm that first finds a good initial action and then treats the problem as locally linear is statistically optimal. In contrast, several classical algorithms, such as UCB and algorithms relying on regression oracles, are provably suboptimal.

The Fair Value of Data Under Heterogeneous Privacy Constraints

Modern data aggregation often takes the form of a platform collecting data from a network of users. More than ever, these users are now requesting that the data they provide is protected with a guarantee of privacy. This has led to the study of optimal data acquisition frameworks, where the optimality criterion is typically the maximization of utility for the agent trying to acquire the data. This involves determining how to allocate payments to users for the purchase of their data at various privacy levels. The main goal of this paper is to characterize a fair amount to pay users for their data at a given privacy level. We propose an axiomatic definition of fairness, analogous to the celebrated Shapley value. Two concepts for fairness are introduced. The first treats the platform and users as members of a common coalition and provides a complete description of how to divide the utility among the platform and users. In the second concept, fairness is defined only among users, leading to a potential fairness-constrained mechanism design problem for the platform. We consider explicit examples involving private heterogeneous data and show how these notions of fairness can be applied. To the best of our knowledge, these are the first fairness concepts for data that explicitly consider privacy constraints.

Efficiently Computing Sparse Fourier Transforms of $q$-ary Functions

Fourier transformations of pseudo-Boolean functions are popular tools for analyzing functions of binary sequences. Real-world functions often have structures that manifest in a sparse Fourier transform, and previous works have shown that under the assumption of sparsity the transform can be computed efficiently. But what if we want to compute the Fourier transform of functions defined over a $q$-ary alphabet? These types of functions arise naturally in many areas including biology. A typical workaround is to encode the $q$-ary sequence in binary, however, this approach is computationally inefficient and fundamentally incompatible with the existing sparse Fourier transform techniques. Herein, we develop a sparse Fourier transform algorithm specifically for $q$-ary functions of length $n$ sequences, dubbed $q$-SFT, which provably computes an $S$-sparse transform with vanishing error as $q^n \rightarrow \infty$ in $O(Sn)$ function evaluations and $O(S n^2 \log q)$ computations, where $S = q^{n\delta}$ for some $\delta < 1$. Under certain assumptions, we show that for fixed $q$, a robust version of $q$-SFT has a sample complexity of $O(Sn^2)$ and a computational complexity of $O(Sn^3)$ with the same asymptotic guarantees. We present numerical simulations on synthetic and real-world RNA data, demonstrating the scalability of $q$-SFT to massively high dimensional $q$-ary functions.