Neural networks learn moments of increasing order

The distributional simplicity bias (DSB) posits that neural networks learn low-order moments of the data distribution first, before moving on to higher-order correlations. In this work, we present compelling new evidence for the DSB by showing that networks automatically learn to perform well on maximum-entropy distributions whose low-order statistics match those of the training set early in training, then lose this ability later. We also extend the DSB to discrete domains by proving an equivalence between token n-gram frequencies and the moments of embedding vectors, and by finding empirical evidence for the bias in LLMs. Finally we use optimal transport methods to surgically edit the low-order statistics of one class to match those of another, and show that early-training networks treat the edited samples as if they were drawn from the target class. Code is available at this https URL.

Sparse Autoencoders Find Highly Interpretable Features in Language Models

One of the roadblocks to a better understanding of neural networks' internals is \textit{polysemanticity}, where neurons appear to activate in multiple, semantically distinct contexts. Polysemanticity prevents us from identifying concise, human-understandable explanations for what neural networks are doing internally. One hypothesised cause of polysemanticity is superposition, where neural networks represent more features than they have neurons by assigning features to an overcomplete set of directions in activation space, rather than to individual neurons. Here, we attempt to identify those directions, using sparse autoencoders to reconstruct the internal activations of a language model. These autoencoders learn sets of sparsely activating features that are more interpretable and monosemantic than directions identified by alternative approaches, where interpretability is measured by automated methods. Moreover, we show that with our learned set of features, we can pinpoint the features that are causally responsible for counterfactual behaviour on the indirect object identification task (Wang et al., 2022) to a finer degree than previous decompositions. This work indicates that it is possible to resolve superposition in language models using a scalable, unsupervised method. Our method may serve as a foundation for future mechanistic interpretability work, which we hope will enable greater model transparency and steerability.

Eliciting Language Model Behaviors using Reverse Language Models

Despite advances in fine-tuning methods, language models (LMs) continue to output toxic and harmful responses on worst-case inputs, including adversarial attacks and jailbreaks. We train an LM on tokens in reverse order---a reverse LM---as a tool for identifying such worst-case inputs. By prompting a reverse LM with a problematic string, we can sample prefixes that are likely to precede the problematic suffix. We test our reverse LM by using it to guide beam search for prefixes that have high probability of generating toxic statements when input to a forwards LM. Our 160m parameter reverse LM outperforms the existing state-of-the-art adversarial attack method, GCG, when measuring the probability of toxic continuations from the Pythia-160m LM. We also find that the prefixes generated by our reverse LM for the Pythia model are more likely to transfer to other models, eliciting toxic responses also from Llama 2 when compared to GCG-generated attacks.

Eliciting Language Model Behaviors using Reverse Language Models

Despite advances in fine-tuning methods, language models (LMs) continue to output toxic and harmful responses on worst-case inputs, including adversarial attacks and jailbreaks. We train an LM on tokens in reverse order---a reverse LM---as a tool for identifying such worst-case inputs. By prompting a reverse LM with a problematic string, we can sample prefixes that are likely to precede the problematic suffix. We test our reverse LM by using it to guide beam search for prefixes that have high probability of generating toxic statements when input to a forwards LM. Our 160m parameter reverse LM outperforms the existing state-of-the-art adversarial attack method, GCG, when measuring the probability of toxic continuations from the Pythia-160m LM. We also find that the prefixes generated by our reverse LM for the Pythia model are more likely to transfer to other models, eliciting toxic responses also from Llama 2 when compared to GCG-generated attacks.

Emergent and Predictable Memorization in Large Language Models

Memorization, or the tendency of large language models (LLMs) to output entire sequences from their training data verbatim, is a key concern for safely deploying language models. In particular, it is vital to minimize a model's memorization of sensitive datapoints such as those containing personal identifiable information (PII). The prevalence of such undesirable memorization can pose issues for model trainers, and may even require discarding an otherwise functional model. We therefore seek to predict which sequences will be memorized before a large model's full train-time by extrapolating the memorization behavior of lower-compute trial runs. We measure memorization of the Pythia model suite, and find that intermediate checkpoints are better predictors of a model's memorization behavior than smaller fully-trained models. We additionally provide further novel discoveries on the distribution of memorization scores across models and data.

LEACE: Perfect linear concept erasure in closed form

Concept erasure aims to remove specified features from a neural representation. It can be used to improve fairness (e.g. preventing a model from using gender or race) and interpretability (e.g. removing a concept to observe changes in model behavior). In this paper, we introduce LEAst-squares Concept Erasure (LEACE), a fast closed-form method which provably prevents all linear classifiers from detecting a concept while inflicting the least possible damage to the representation. We apply LEACE to large language models with a novel procedure called “concept scrubbing,” which erases information about the target concept from every hidden layer in the network. We demonstrate the usefulness of our method on two tasks: measuring the extent to which language models rely on part-of-speech information, and reducing gender bias in BERT embeddings.

Linear Representations of Sentiment in Large Language Models

Sentiment is a pervasive feature in natural language text, yet it is an open question how sentiment is represented within Large Language Models (LLMs). In this study, we reveal that across a range of models, sentiment is represented linearly: a single direction in activation space mostly captures the feature across a range of tasks with one extreme for positive and the other for negative. Through causal interventions, we isolate this direction and show it is causally relevant in both toy tasks and real world datasets such as Stanford Sentiment Treebank. Through this case study we model a thorough investigation of what a single direction means on a broad data distribution.

We further uncover the mechanisms that involve this direction, highlighting the roles of a small subset of attention heads and neurons. Finally, we discover a phenomenon which we term the summarization motif: sentiment is not solely represented on emotionally charged words, but is additionally summarized at intermediate positions without inherent sentiment, such as punctuation and names. We show that in Stanford Sentiment Treebank zero-shot classification, 76% of above-chance classification accuracy is lost when ablating the sentiment direction, nearly half of which (36%) is due to ablating the summarized sentiment direction exclusively at comma positions.

Representation Engineering: A Top-Down Approach to AI Transparency

In this paper, we identify and characterize the emerging area of representation engineering (RepE), an approach to enhancing the transparency of AI systems that draws on insights from cognitive neuroscience. RepE places population-level representations, rather than neurons or circuits, at the center of analysis, equipping us with novel methods for monitoring and manipulating high-level cognitive phenomena in deep neural networks (DNNs). We provide baselines and an initial analysis of RepE techniques, showing that they offer simple yet effective solutions for improving our understanding and control of large language models. We showcase how these methods can provide traction on a wide range of safety-relevant problems, including honesty, harmlessness, power-seeking, and more, demonstrating the promise of top-down transparency research. We hope that this work catalyzes further exploration of RepE and fosters advancements in the transparency and safety of AI systems.

Can Transformers Learn to Solve Problems Recursively?

Neural networks have in recent years shown promise for helping software engineers write programs and even formally verify them. While semantic information plays a crucial part in these processes, it remains unclear to what degree popular neural architectures like transformers are capable of modeling that information.

This paper examines the behavior of neural networks learning algorithms relevant to programs and formal verification proofs through the lens of mechanistic interpretability, focusing in particular on structural recursion. Structural recursion is at the heart of tasks on which symbolic tools currently outperform neural models, like inferring semantic relations between datatypes and emulating program behavior.

We evaluate the ability of transformer models to learn to emulate the behavior of structurally recursive functions from input-output examples. Our evaluation includes empirical and conceptual analyses of the limitations and capabilities of transformer models

in approximating these functions, as well as reconstructions of the “shortcut” algorithms the model learns. By reconstructing these algorithms, we are able to correctly predict 91% of failure cases for one of the approximated functions. Our work provides a new foundation for understanding the behavior of neural networks that fail to solve the very tasks they are trained for.

Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling

How do large language models (LLMs) develop and evolve over the course of training? How do these patterns change as models scale? To answer these questions, we introduce Pythia, a suite of 16 LLMs all trained on public data seen in the exact same order and ranging in size from 70M to 12B parameters. We provide public access to 154 checkpoints for each one of the 16 models, alongside tools to download and reconstruct their exact training dataloaders for further study. We intend Pythia to facilitate research in many areas, and we present several case studies including novel results in memorization, term frequency effects on few-shot arithmetic performance, and reducing gender bias. We demonstrate that this highly controlled setup can be used to yield novel insights toward LLMs and their training dynamics.

Eliciting Latent Predictions from Transformers with the Tuned Lens

We analyze transformers from the perspective of iterative inference, seeking to understand how model predictions are refined layer by layer. To do so, we train an affine probe for each block in a frozen pretrained model, making it possible to decode every hidden state into a distribution over the vocabulary. Our method, the tuned lens, is a refinement of the earlier ``logit lens'' technique, which yielded useful insights but is often brittle.

We test our method on various autoregressive language models with up to 20B parameters, showing it to be more predictive, reliable and unbiased than the logit lens. With causal experiments, we show the tuned lens uses similar features to the model itself. We also find the trajectory of latent predictions can be used to detect malicious inputs with high accuracy. All code needed to reproduce our results can be found at here.

OpenFold: Retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization

AlphaFold2 revolutionized structural biology with the ability to predict protein structures with exceptionally high accuracy. Its implementation, however, lacks the code and data required to train new models. These are necessary to (i) tackle new tasks, like protein-ligand complex structure prediction, (ii) investigate the process by which the model learns, which remains poorly understood, and (iii) assess the model’s generalization capacity to unseen regions of fold space. Here we report OpenFold, a fast, memory-efficient, and trainable implementation of AlphaFold2, and OpenProteinSet, the largest public database of protein multiple sequence alignments. We use OpenProteinSet to train OpenFold from scratch, fully matching the accuracy of AlphaFold2. Having established parity, we assess OpenFold’s capacity to generalize across fold space by retraining it using carefully designed datasets. We find that OpenFold is remarkably robust at generalizing despite extreme reductions in training set size and diversity, including near-complete elisions of classes of secondary structure elements. By analyzing intermediate structures produced by OpenFold during training, we also gain surprising insights into the manner in which the model learns to fold proteins, discovering that spatial dimensions are learned sequentially. Taken together, our studies demonstrate the power and utility of OpenFold, which we believe will prove to be a crucial new resource for the protein modeling community.