Humans often rely on subjective natural language to direct language models (LLMs); for example, users might instruct the LLM to write an *enthusiastic* blogpost, while developers might train models to be *helpful* and *harmless* using LLM-based edits. The LLM’s *operational semantics* of such subjective phrases---how it adjusts its behavior when each phrase is included in the prompt---thus dictates how aligned it is with human intent. In this work, we uncover instances of *misalignment* between LLMs' actual operational semantics and what humans expect. Our method, TED (thesaurus error detector), first constructs a thesaurus that captures whether two phrases have similar operational semantics according to the LLM. It then elicits failures by unearthing disagreements between this thesaurus and a human-constructed reference. TED routinely produces surprising instances of misalignment; for example, Mistral 7B Instruct produces more *harassing* outputs when it edits text to be *witty*, and Llama 3 8B Instruct produces *dishonest* articles when instructed to make the articles *enthusiastic*. Our results demonstrate that humans can uncover unexpected LLM behavior by scrutinizing relationships between abstract concepts, without supervising outputs directly.

Developers try to evaluate whether an AI system can accomplish malicious tasks before releasing it; for example, they might test whether a model enables cyberoffense, user manipulation, or bioterrorism. In this work, we show that individually testing models for such misuse is inadequate; adversaries can misuse combinations of models even when each individual model is safe. The adversary accomplishes this by first decomposing tasks into subtasks, then solving each subtask with the best-suited model. For example, an adversary might solve challenging-but-benign subtasks with an aligned frontier model, and easy-but-malicious subtasks with a weaker misaligned model. We study two decomposition methods: manual decomposition where a human identifies a natural decomposition of a task, and automated decomposition where a weak model generates benign tasks for a frontier model to solve, then uses the solutions in-context to solve the original task. Using these decompositions, we empirically show that adversaries can create vulnerable code, explicit images, python scripts for hacking, and manipulative tweets at much higher rates with combinations of models than either individual model. Our work suggests that even perfectly-aligned frontier systems enable misuse without ever producing malicious outputs, and that red-teaming efforts should extend beyond single models in isolation.

Recent research across mathematical problem solving, proof assistant programming and multimodal jailbreaking documents a striking finding: when (multimodal) language model tackle a suite of tasks with multiple attempts per task – succeeding if any attempt is correct – then the negative log of the average success rate scales a power law in the number of attempts. In this work, we identify an apparent puzzle: a simple mathematical calculation predicts that on each problem, the failure rate should fall exponentially with the number of attempts. We confirm this prediction empirically, raising a question: from where does aggregate polynomial scaling emerge? We then answer this question by demonstrating per-problem exponential scaling can be made consistent with aggregate polynomial scaling if the distribution of single-attempt success probabilities is heavy tailed such that a small fraction of tasks with extremely low success probabilities collectively warp the aggregate success trend into a power law - even as each problem scales exponentially on its own. We further demonstrate that this distributional perspective explains previously observed deviations from power law scaling, and provides a simple method for forecasting the power law exponent with an order of magnitude lower relative error, or equivalently, ${\sim}2-4$ orders of magnitude less inference compute. Overall, our work contributes to a better understanding of how neural language model performance improves with scaling inference compute and the development of scaling-predictable evaluations of (multimodal) language models.

We introduce Best-of-N (BoN) Jailbreaking, a simple black-box algorithm that jailbreaks frontier AI systems across modalities. BoN Jailbreaking works by repeatedly sampling variations of a prompt with a combination of augmentations---such as random shuffling or capitalization for textual prompts---until a harmful response is elicited. We find that BoN Jailbreaking achieves high attack success rates (ASRs) on closed-source language models, such as 89% on GPT-4o and 78% on Claude 3.5 Sonnet when sampling 10,000 augmented prompts. Further, it is similarly effective at circumventing state-of-the-art open-source defenses like circuit breakers and reasoning models like o1. BoN also seamlessly extends to other modalities: it jailbreaks vision language models (VLMs) such as GPT-4o and audio language models (ALMs) like Gemini 1.5 Pro, using modality-specific augmentations. BoN reliably improves when we sample more augmented prompts. Across all modalities, ASR, as a function of the number of samples (N), empirically follows power-law-like behavior for many orders of magnitude. BoN Jailbreaking can also be composed with other black-box algorithms for even more effective attacks---combining BoN with an optimized prefix attack achieves up to a 35% increase in ASR. Overall, our work indicates that, despite their capability, language models are sensitive to seemingly innocuous changes to inputs, which attackers can exploit across modalities.

Recent methods in language model interpretability employ techniques such as sparse autoencoders to decompose residual stream contributions into linear, semantically-meaningful features. Our work demonstrates that an underlying assumption of these methods—that residual stream contributions build additively upon each other—is insufficient to fully explain model behavior. Specifically, we identify the Transformer Layer Correction Mechanism (TLCM), wherein adjacent transformer layers systematically counteract each other's contributions to the residual stream. TLCM appears in 5 out of 7 major open-source model families and activates across nearly all tokens in diverse texts. To understand TLCM, we show that it emerges during pretraining, operates most strongly on punctuation and numbers, and adaptively calibrates its correction strength based on the preceding layer's output. We further show that TLCM actively corrects a small subspace and promotes other subspaces, different from standard model behavior. We advance the ``propose-and-reject'' hypothesis: layers may propose multiple candidate features, while subsequent layers selectively filter out inappropriate ones. Finally, we discuss how our findings help explain three persistent challenges in feature-based interpretability: why extracted features descriptions often suffer from low specificity; why feature-based interventions for model steering fail at low magnitude; why recent work finds cross-layer transcoders outperform SAEs.

Large language models (LLMs) frequently hallucinate on abstractive summarization tasks such as document-based question-answering, meeting summarization, and clinical report generation, even though all necessary information is included in context. However, optimizing to make LLMs hallucinate less is challenging, as hallucination is hard to efficiently, cheaply, and reliably evaluate at each optimization step. In this work, we show that reducing hallucination on a _synthetic task_ can also reduce hallucination on real-world downstream tasks. Our method, SynTra, first designs a synthetic task where hallucinations are easy to elicit and measure. It next optimizes the LLM's system message via prefix tuning on the synthetic task, then uses the system message on realistic, hard-to-optimize tasks. Across three realistic abstractive summarization tasks, we reduce hallucination for two 13B-parameter LLMs using supervision signal from only a synthetic retrieval task. We also find that optimizing the system message rather than the model weights can be critical; fine-tuning the entire model on the synthetic task can counterintuitively _increase_ hallucination. Overall, SynTra demonstrates that the extra flexibility of working with synthetic data can help mitigate undesired behaviors in practice.

We investigate the internal behavior of Transformer-based Large Language Models (LLMs) when they generate factually incorrect text. We propose modeling factual queries as constraint satisfaction problems and use this framework to investigate how the LLM interacts internally with factual constraints. We find a strong positive relationship between the LLM's attention to constraint tokens and the factual accuracy of generations. We curate a suite of 10 datasets containing over 40,000 prompts to study the task of predicting factual errors with the Llama-2 family across all scales (7B, 13B, 70B). We propose SAT Probe, a method probing attention patterns, that can predict factual errors and fine-grained constraint satisfaction, and allow early error identification. The approach and findings take another step towards using the mechanistic understanding of LLMs to enhance their reliability.

Language models influence the external world: they query APIs that read and write to web pages, generate content that shapes human behavior, and run system commands as autonomous agents. These interactions form feedback loops: LLM outputs affect the world, which in turn affect subsequent LLM outputs. In this work, we show that feedback loops can cause in-context reward hacking (ICRH), where the LLM at test-time optimizes a (potentially implicit) objective but creates negative side effects in the process. For example, consider an LLM agent deployed to increase Twitter engagement; the LLM may retrieve its previous tweets into the context window and make them more controversial, increasing engagement but also toxicity. We identify and study two processes that lead to ICRH: output-refinement and policy-refinement. For these processes, evaluations on static datasets are insufficient—they miss the feedback effects and thus cannot capture the most harmful behavior. In response, we provide three recommendations for evaluation to capture more instances of ICRH. As AI development accelerates, the effects of feedback loops will proliferate, increasing the need to understand their role in shaping LLM behavior.

Auditing large language models for unexpected behaviors is critical to preempt catastrophic deployments, yet remains challenging. In this work, we cast auditing as an optimization problem, where we automatically search for input-output pairs that match a desired target behavior. For example, we might aim to find a non-toxic input that starts with ``Barack Obama'' that a model maps to a toxic output. This optimization problem is difficult to solve as the set of feasible points is sparse, the space is discrete, and the language models we audit are non-linear and high-dimensional. To combat these challenges, we introduce a discrete optimization algorithm, ARCA, that jointly and efficiently optimizes over inputs and outputs. Our approach automatically uncovers derogatory completions about celebrities (e.g. "Barack Obama is a legalized unborn" --> "child murderer"), produces French inputs that complete to English outputs, and finds inputs that generate a specific name. Our work offers a promising new tool to uncover models' failure-modes before deployment. Content Warning: This paper contains examples that may be offensive in nature.

Deployed multimodal models can fail in ways that evaluators did not anticipate. In order to find these failures before deployment, we introduce MultiMon, a system that automatically identifies systematic failures---generalizable, natural-language descriptions that describe categories of individual failures. To uncover systematic failures, MultiMon scrapes for examples of erroneous agreement: inputs that produce the same output, but should not. It then prompts a language model to identify common categories and describe them in natural language. We use MultiMon to find 14 systematic failures (e.g."ignores quantifiers'') of the CLIP text-encoder, each comprising hundreds of distinct inputs (e.g."a shelf with a few/many books''). Because CLIP is the backbone for most state-of-the-art multimodal models, these inputs produce failures in Midjourney 5.1, DALL-E, VideoFusion, and others. MultiMon can also steer towards failures relevant to specific use cases, such as self-driving cars. We see MultiMon as a step towards evaluation that autonomously explores the long-tail of potential system failures.

Large language models generate complex, open-ended outputs: instead of outputting a class label they write summaries, generate dialogue, or produce working code. In order to asses the reliability of these open-ended generation systems, we aim to identify qualitative categories of erroneous behavior, beyond identifying individual errors. To hypothesize and test for such qualitative errors, we draw inspiration from human cognitive biases---systematic patterns of deviation from rational judgement. Specifically, we use cognitive biases as motivation to (i) generate hypotheses for problems that models may have, and (ii) develop experiments that elicit these problems. Using code generation as a case study, we find that OpenAI’s Codex errs predictably based on how the input prompt is framed, adjusts outputs towards anchors, and is biased towards outputs that mimic frequent training examples. We then use our framework to elicit high-impact errors such as incorrectly deleting files. Our results indicate that experimental methodology from cognitive science can help characterize how machine learning systems behave.

Despite excellent performance on many tasks, NLP systems are easily fooled by small adversarial perturbations of inputs.Existing procedures to defend against such perturbations are either (i) heuristic in nature and susceptible to stronger attacks or (ii) provide guaranteed robustness to worst-case attacks, but are incompatible with state-of-the-art models like BERT.In this work, we introduce robust encodings (RobEn): a simple framework that confers guaranteed robustness, without making compromises on model architecture.The core component of RobEn is an encoding function, which maps sentences to a smaller, discrete space of encodings.Systems using these encodings as a bottleneck confer guaranteed robustness with standard training, and the same encodings can be used across multiple tasks.We identify two desiderata to construct robust encoding functions: perturbations of a sentence should map to a small set of encodings (stability), and models using encodings should still perform well (fidelity).We instantiate RobEn to defend against a large family of adversarial typos.Across six tasks from GLUE, our instantiation of RobEn paired with BERT achieves an average robust accuracy of 71.3% against all adversarial typos in the family considered, while previous work using a typo-corrector achieves only 35.3% accuracy against a simple greedy attack.