As fine-tuning large language models (LLMs) becomes increasingly prevalent, users often rely on third-party services with limited visibility into their fine-tuning processes. This lack of transparency raises the question: how do consumers verify that fine-tuning services are performed correctly? For instance, a service provider could claim to fine-tune a model for each user, yet simply send all users back the same base model. To address this issue, we propose vTune, a simple method that uses a small number of backdoor data points added to the training data to provide a statistical test for verifying that a provider fine-tuned a custom model on a particular user's dataset. Unlike existing works, vTune is able to scale to verification of fine-tuning on state-of-the-art LLMs, and can be used both with open-source and closed-source models. We test our approach across several model families and sizes as well as across multiple instruction-tuning datasets, and find that the statistical test is satisfied with p-values on the order of $\sim 10^{-40}$, with no negative impact on downstream task performance. Further, we explore several attacks that attempt to subvert vTune and demonstrate the method's robustness to these attacks.

Test set contamination, wherein test data from a benchmark ends up in a newer model's training set, is a well-documented obstacle for fair LLM evaluation and can quickly render benchmarks obsolete. To mitigate this, many recent benchmarks crowdsource new prompts and evaluations from human or LLM judges; however, these can introduce significant biases, and break down when scoring hard questions. In this work, we introduce a new benchmark for LLMs designed to be immune to both test set contamination and the pitfalls of LLM judging and human crowdsourcing. We release LiveBench, the first benchmark that (1) contains frequently-updated questions from recent information sources, (2) scores answers automatically according to objective ground-truth values, and (3) contains a wide variety of challenging tasks, spanning math, coding, reasoning, language, instruction following, and data analysis. To achieve this, LiveBench contains questions that are based on recently-released math competitions, arXiv papers, news articles, and datasets, and it contains harder, contamination-free versions of tasks from previous benchmarks such as Big-Bench Hard, AMPS, and IFEval. We evaluate many prominent closed-source models, as well as dozens of open-source models ranging from 0.5B to 110B in size. LiveBench is difficult, with top models achieving below 65% accuracy. We release all questions, code, and model answers. Questions will be added and updated on a monthly basis, and we will release new tasks and harder versions of tasks over time so that LiveBench can distinguish between the capabilities of LLMs as they improve in the future. We welcome community engagement and collaboration for expanding the benchmark tasks and models.

When using large language models (LLMs) in high-stakes applications, we need to know when we can trust their predictions. Some works argue that prompting high-performance LLMs is sufficient to produce calibrated uncertainties, while others introduce sampling methods that can be prohibitively expensive. In this work, we first argue that prompting on its own is insufficient to achieve good calibration and then show that fine-tuning on a small dataset of correct and incorrect answers can create an uncertainty estimate with good generalization and small computational overhead. We show that a thousand graded examples are sufficient to outperform baseline methods and that training through the features of a model is necessary for good performance and tractable for large open-source models when using LoRA. We also investigate the mechanisms that enable reliable LLM uncertainty estimation, finding that many models can be used as general-purpose uncertainty estimators, applicable not just to their own uncertainties but also the uncertainty of other models. Lastly, we show that uncertainty estimates inform human use of LLMs in human-AI collaborative settings through a user study.

Modern large language models (LLMs) that rely on attention mechanisms are typically trained with fixed context lengths which enforce upper limits on the length of input sequences that they can handle at evaluation time. To use these models on sequences longer than the train-time context length, one might employ techniques from the growing family of context length extrapolation methods -- most of which focus on modifying the system of positional encodings used in the attention mechanism to indicate where tokens or activations are located in the input sequence. We conduct a wide survey of existing methods of context length extrapolation on a base LLaMA or LLaMA 2 model, and introduce some of our own design as well -- in particular, a new truncation strategy for modifying the basis for the position encoding. We test these methods using three new evaluation tasks (FreeFormQA, AlteredNumericQA, and LongChat-Lines) as well as perplexity, which we find to be less fine-grained as a measure of long context performance of LLMs. We release the three tasks publicly as datasets on HuggingFace. We discover that linear scaling is the best method for extending context length, and show that further gains can be achieved by using longer scales at evaluation time. We also discover promising extrapolation capabilities in the truncated basis. To support further research in this area, we release three new 13B parameter long-context models which we call Giraffe: 4k and 16k context models trained from base LLaMA-13B, and a 32k context model trained from base LLaMA2-13B. We also release the code to replicate our results.

We present new intuitions and theoretical assessments of the emergence of disentangled representation in variational autoencoders. Taking a rate-distortion theory perspective, we show the circumstances under which representations aligned with the underlying generative factors of variation of data emerge when optimising the modified ELBO bound in $\beta$-VAE, as training progresses. From these insights, we propose a modification to the training regime of $\beta$-VAE, that progressively increases the information capacity of the latent code during training. This modification facilitates the robust learning of disentangled representations in $\beta$-VAE, without the previous trade-off in reconstruction accuracy.

Domain adaptation is an important open problem in deep reinforcement learning (RL). In many scenarios of interest data is hard to obtain, so agents may learn a source policy in a setting where data is readily available, with the hope that it generalises well to the target domain. We propose a new multi-stage RL agent, DARLA (DisentAngled Representation Learning Agent), which learns to see before learning to act. DARLA's vision is based on learning a disentangled representation of the observed environment. Once DARLA can see, it is able to acquire source policies that are robust to many domain shifts - even with no access to the target domain. DARLA significantly outperforms conventional baselines in zero-shot domain adaptation scenarios, an effect that holds across a variety of RL environments (Jaco arm, DeepMind Lab) and base RL algorithms (DQN, A3C and EC).

The natural world is infinitely diverse, yet this diversity arises from a relatively small set of coherent properties and rules, such as the laws of physics or chemistry. We conjecture that biological intelligent systems are able to survive within their diverse environments by discovering the regularities that arise from these rules primarily through unsupervised experiences, and representing this knowledge as abstract concepts. Such representations possess useful properties of compositionality and hierarchical organisation, which allow intelligent agents to recombine a finite set of conceptual building blocks into an exponentially large set of useful new concepts. This paper describes SCAN (Symbol-Concept Association Network), a new framework for learning such concepts in the visual domain. We first use the previously published beta-VAE (Higgins et al., 2017a) architecture to learn a disentangled representation of the latent structure of the visual world, before training SCAN to extract abstract concepts grounded in such disentangled visual primitives through fast symbol association. Our approach requires very few pairings between symbols and images and makes no assumptions about the choice of symbol representations. Once trained, SCAN is capable of multimodal bi-directional inference, generating a diverse set of image samples from symbolic descriptions and vice versa. It also allows for traversal and manipulation of the implicit hierarchy of compositional visual concepts through symbolic instructions and learnt logical recombination operations. Such manipulations enable SCAN to invent and learn novel visual concepts through recombination of the few learnt concepts.

Automated discovery of early visual concepts from raw image data is a major open challenge in AI research. Addressing this problem, we propose an unsupervised approach for learning disentangled representations of the underlying factors of variation. We draw inspiration from neuroscience, and show how this can be achieved in an unsupervised generative model by applying the same learning pressures as have been suggested to act in the ventral visual stream in the brain. By enforcing redundancy reduction, encouraging statistical independence, and exposure to data with transform continuities analogous to those to which human infants are exposed, we obtain a variational autoencoder (VAE) framework capable of learning disentangled factors. Our approach makes few assumptions and works well across a wide variety of datasets. Furthermore, our solution has useful emergent properties, such as zero-shot inference and an intuitive understanding of "objectness".