Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

Humans are capable of strategically deceptive behavior: behaving helpfully in most situations, but then behaving very differently in order to pursue alternative objectives when given the opportunity. If an AI system learned such a deceptive strategy, could we detect it and remove it using current state-of-the-art safety training techniques? To study this question, we construct proof-of-concept examples of deceptive behavior in large language models (LLMs). For example, we train models that write secure code when the prompt states that the year is 2023, but insert exploitable code when the stated year is 2024. We find that such backdoor behavior can be made persistent, so that it is not removed by standard safety training techniques, including supervised fine-tuning, reinforcement learning, and adversarial training (eliciting unsafe behavior and then training to remove it). The backdoor behavior is most persistent in the largest models and in models trained to produce chain-of-thought reasoning about deceiving the training process, with the persistence remaining even when the chain-of-thought is distilled away. Furthermore, rather than removing backdoors, we find that adversarial training can teach models to better recognize their backdoor triggers, effectively hiding the unsafe behavior. Our results suggest that, once a model exhibits deceptive behavior, standard techniques could fail to remove such deception and create a false impression of safety.

Characterizing tradeoffs between teaching via language and demonstrations in multi-agent systems

Humans teach others about the world through language and demonstration. When might one of these modalities be more effective than the other? In this work, we study the factors that modulate the effectiveness of language vs. demonstration using multi-agent systems to model human communication. Specifically, we train neural network agents to teach via language or demonstration in a grounded communication task, manipulating 1) the inherent difficulty of the task and 2) the competence of the teacher. We find that teaching by demonstration is more effective in the simplest settings, but language is more effective as task difficulty increases, due to its ability to generalize more effectively to unseen scenarios. Overall, these results provide converging evidence for a tradeoff between language and demonstration as teaching modalities in humans, and make the novel predictions that demonstration may be optimal for easy tasks, while language enables generalization in more challenging settings.

Learning to Compress Prompts with Gist Tokens

Prompting is now the primary way to utilize the multitask capabilities of language models (LMs), but prompts occupy valuable space in the input context window, and re-encoding the same prompt is computationally inefficient. Finetuning and distillation methods allow for specialization of LMs without prompting, but require retraining the model for each task. To avoid this trade-off entirely, we present gisting, which trains an LM to compress prompts into smaller sets of "gist" tokens which can be reused for compute efficiency. Gist models can be easily trained as part of instruction finetuning via a restricted attention mask that encourages prompt compression. On decoder (LLaMA-7B) and encoder-decoder (FLAN-T5-XXL) LMs, gisting enables up to 26x compression of prompts, resulting in up to 40% FLOPs reductions, 4.2% wall time speedups, storage savings, and minimal loss in output quality.

Improving Policy Learning via Language Dynamics Distillation

Recent work has shown that augmenting environments with language descriptions improves policy learning. However, for environments with complex language abstractions, learning how to ground language to observations is difficult due to sparse, delayed rewards. We propose Language Dynamics Distillation (LDD), which pretrains a model to predict environment dynamics given demonstrations with language descriptions, and then fine-tunes these language-aware pretrained representations via reinforcement learning (RL). In this way, the model is trained to both maximize expected reward and retain knowledge about how language relates to environment dynamics. On SILG, a benchmark of five tasks with language descriptions that evaluate distinct generalization challenges on unseen environments (NetHack, ALFWorld, RTFM, Messenger, and Touchdown), LDD outperforms tabula-rasa RL, VAE pretraining, and methods that learn from unlabeled demonstrations in inverse RL and reward shaping with pretrained experts. In our analyses, we show that language descriptions in demonstrations improve sample-efficiency and generalization across environments, and that dynamics modelling with expert demonstrations is more effective than with non-experts.

Active Learning Helps Pretrained Models Learn the Intended Task

Models can fail in unpredictable ways during deployment due to task ambiguity, when multiple behaviors are consistent with the provided training data. An example is an object classifier trained on red squares and blue circles: when encountering blue squares, the intended behavior is undefined. We investigate whether pretrained models are better active learners, capable of disambiguating between the possible tasks a user may be trying to specify. Intriguingly, we find that better active learning is an emergent property of the pretraining process: pretrained models require up to 5 times fewer labels when using uncertainty-based active learning, while non-pretrained models see no or even negative benefit. We find these gains come from an ability to select examples with attributes that disambiguate the intended behavior, such as rare product categories or atypical backgrounds. These attributes are far more linearly separable in pretrained model's representation spaces vs non-pretrained models, suggesting a possible mechanism for this behavior.

Improving Intrinsic Exploration with Language Abstractions

Reinforcement learning (RL) agents are particularly hard to train when rewards are sparse. One common solution is to use intrinsic rewards to encourage agents to explore their environment. However, recent intrinsic exploration methods often use state-based novelty measures which reward low-level exploration and may not scale to domains requiring more abstract skills. Instead, we explore natural language as a general medium for highlighting relevant abstractions in an environment. Unlike previous work, we evaluate whether language can improve over existing exploration methods by directly extending (and comparing to) competitive intrinsic exploration baselines: AMIGo (Campero et al., 2021) and NovelD (Zhang et al., 2021). These language-based variants outperform their non-linguistic forms by 45-85% across 13 challenging tasks from the MiniGrid and MiniHack environment suites.

Calibrate your listeners! Robust communication-based training for pragmatic speakers

To be good conversational partners, natural language processing (NLP) systems should be trained to produce contextually useful utterances. Prior work has investigated training NLP systems with communication-based objectives, where a neural listener stands in as a communication partner. However, these systems commonly suffer from semantic drift where the learned language diverges radically from natural language. We propose a method that uses a population of neural listeners to regularize speaker training. We first show that language drift originates from the poor uncertainty calibration of a neural listener, which makes high-certainty predictions on novel sentences. We explore ensemble- and dropout-based populations of listeners and find that the former results in better uncertainty quantification. We evaluate both population-based objectives on reference games, and show that the ensemble method with better calibration enables the speaker to generate pragmatic utterances while scaling to a large vocabulary and generalizing to new games and listeners.

Emergent Communication of Generalizations

To build agents that can collaborate effectively with others, recent research has trained artificial agents to communicate with each other in Lewis-style referential games. However, this often leads to successful but uninterpretable communication. We argue that this is due to the game objective: communicating about a single object in a shared visual context is prone to overfitting and does not encourage language useful beyond concrete reference. In contrast, human language conveys a rich variety of abstract ideas. To promote such skills, we propose games that require communicating generalizations over sets of objects representing abstract visual concepts, optionally with separate contexts for each agent. We find that these games greatly improve systematicity and interpretability of the learned languages, according to several metrics in the literature. Finally, we propose a method for identifying logical operations embedded in the emergent languages by learning an approximate compositional reconstruction of the language.

Compositional Explanations of Neurons

We describe a procedure for explaining neurons in deep representations by identifying compositional logical concepts that closely approximate neuron behavior. Compared to prior work that uses atomic labels as explanations, analyzing neurons compositionally allows us to more precisely and expressively characterize their behavior. We use this procedure to answer several questions on interpretability in models for vision and natural language processing. First, we examine the kinds of abstractions learned by neurons. In image classification, we find that many neurons learn highly abstract but semantically coherent visual concepts, while other polysemantic neurons detect multiple unrelated features; in natural language inference (NLI), neurons learn shallow lexical heuristics from dataset biases. Second, we see whether compositional explanations give us insight into model performance: vision neurons that detect human-interpretable concepts are positively correlated with task performance, while NLI neurons that fire for shallow heuristics are negatively correlated with task performance. Finally, we show how compositional explanations provide an accessible way for end users to produce simple "copy-paste" adversarial examples that change model behavior in predictable ways.