Leading language model (LM) providers like OpenAI and Google offer fine-tuning APIs that allow customers to adapt LMs for specific use cases. To prevent misuse, these LM providers implement filtering mechanisms to block harmful fine-tuning data. Consequently, adversaries seeking to produce unsafe LMs via these APIs must craft adversarial training data that are not identifiably harmful. We make three contributions in this context: 1. We show that many existing attacks that use harmless data to create unsafe LMs rely on eliminating model refusals in the first few tokens of their responses. 2. We show that such prior attacks can be blocked by a simple defense that pre-fills the first few tokens from an aligned model before letting the fine-tuned model fill in the rest. 3. We describe a new data-poisoning attack, ``No, Of course I Can Execute'' (NOICE), which exploits an LM's formulaic refusal mechanism to elicit harmful responses. By training an LM to refuse benign requests on the basis of safety before fulfilling those requests regardless, we are able to jailbreak several open-source models and a closed-source model (GPT-4o). We show an attack success rate (ASR) of 57% against GPT-4o; our attack earned a Bug Bounty from OpenAI. Against open-source models protected by simple defenses, we improve ASRs by an average of 3.25 times compared to the best performing previous attacks that use only harmless data. NOICE demonstrates the exploitability of repetitive refusal mechanisms and broadens understanding of the threats closed-source models face from harmless data.

The proliferation of AI-generated content online has fueled concerns over \emph{model collapse}, a degradation in future generative models' performance when trained on synthetic data generated by earlier models. Industry leaders, premier research journals and popular science publications alike have prophesied catastrophic societal consequences stemming from model collapse. In this position piece, we contend this widespread narrative fundamentally misunderstands the scientific evidence. We highlight that research on model collapse actually encompasses eight distinct and at times conflicting definitions of model collapse, and argue that inconsistent terminology within and between papers has hindered building a comprehensive understanding of model collapse. To assess how significantly different interpretations of model collapse threaten future generative models, we posit what we believe are realistic conditions for studying model collapse and then conduct a rigorous assessment of the literature's methodologies through this lens. While we leave room for reasonable disagreement, our analysis of research studies, weighted by how faithfully each study matches real-world conditions, leads us to conclude that certain predicted claims of model collapse rely on assumptions and conditions that poorly match real-world conditions, and in fact several prominent collapse scenarios are readily avoidable. Altogether, this position paper argues that model collapse has been warped from a nuanced multifaceted consideration into an oversimplified threat, and that the evidence suggests specific harms more likely under society's current trajectory have received disproportionately less attention.

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.

The explosion of high-performing conversational language models (LMs) has spurred a shift from classic natural language processing (NLP) benchmarks to expensive, time-consuming and noisy human evaluations - yet the relationship between these two evaluation strategies remains hazy. In this paper, we conduct a large-scale study of four Chat Llama 2 models, comparing their performance on 160 standard NLP benchmarks (e.g., MMLU, ARC, BIG-Bench Hard) against extensive human preferences on more than 11k single-turn and 2k multi-turn dialogues from over 2k human annotators. Our findings are striking: most NLP benchmarks strongly correlate with human evaluations, suggesting that cheaper, automated metrics can serve as surprisingly reliable predictors of human preferences. Three human evaluations, such as adversarial dishonesty and safety, are anticorrelated with NLP benchmarks, while two are uncorrelated. Moreover, through overparameterized linear regressions, we show that NLP scores can accurately predict human evaluations across different model scales, offering a path to reduce costly human annotation without sacrificing rigor. Overall, our results affirm the continued value of classic benchmarks and illuminate how to harness them to anticipate real-world user satisfaction - pointing to how NLP benchmarks can be leveraged to meet evaluation needs of our new era of conversational AI.

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. 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.

The increasing presence of AI-generated content on the internet raises a critical question: What happens when generative machine learning models are pretrained on web-scale datasets containing data created by earlier models? Some authors prophesy \textit{model collapse} under a `{\it replace}' scenario: a sequence of models, the first trained with real data and each later one trained {\it only on} synthetic data from its preceding model. In this scenario, models successively degrade. Others see collapse as avoidable; in an `{\it accumulate}' scenario, a sequence of models is trained, but each training uses all real and synthetic data generated so far. In this work, we deepen and extend the study of these contrasting scenarios. First, collapse versus avoidance of collapse is studied by comparing the replace and accumulate scenarios on each of three prominent generative modeling settings; we find the same contrast emerges in all three settings. Second, we study a compromise scenario; the available data remains the same as in the {\it accumulate} scenario -- but unlike {\it accumulate} and like {\it replace}, each model is trained using a fixed compute budget; we demonstrate that model test loss on real data is larger than in the {\it accumulate} scenario, but apparently plateaus, unlike the divergence seen with {\it replace}. Third, we study the relative importance of cardinality and proportion of real data for avoiding model collapse. Surprisingly, we find a non-trivial interaction between real and synthetic data, where the value of synthetic data for reducing test loss depends on the absolute quantity of real data. Our insights are particularly important when forecasting whether future frontier generative models will collapse or thrive, and our results open avenues for empirically and mathematically studying the context-dependent value of synthetic data.

Defending large language models against jailbreaks so that they never engage in a broadly-defined set of forbidden behaviors is an open problem. In this paper, we investigate the difficulty of jailbreak-defense when we only want to forbid a narrowly-defined set of behaviors. As a case study, we focus on preventing an LLM from helping a user make a bomb. We find that popular defenses such as safety training, adversarial training, and input/output classifiers are unable to fully solve this problem. In pursuit of a better solution, we develop a transcript-classifier defense which outperforms the baseline defenses we test. However, our classifier defense still fails in some circumstances, which highlights the difficulty of jailbreak-defense even in a narrow domain.

Data selection is crucial for optimizing language model (LM) performance on specific tasks, yet most existing methods fail to effectively consider the target task distribution. Current approaches either ignore task-specific requirements entirely or rely on approximations that fail to capture the nuanced patterns needed for tasks like Autoformalization or code generation. Methods that do consider the target distribution often rely on simplistic, sometimes noisy, representations, like hashed n-gram features, which can lead to collisions and introduce noise. We introduce ZIP-FIT, a data selection framework that uses gzip compression to directly measure alignment between potential training data and the target task distribution. Our key insight is that compression-based similarity captures both syntactic and structural patterns relevant to the target task, enabling more precise selection of truly task-relevant data. In extensive evaluations on Autoformalization and Python code generation, ZIP-FIT significantly outperforms leading baselines like DSIR and D4. Models trained on ZIP-FIT-selected data achieve their lowest cross-entropy loss up to 85.1% faster than baselines, demonstrating that better task alignment leads to more efficient learning. In addition, ZIP-FIT performs selection up to 65.8% faster than DSIR and two orders of magnitude faster than D4. Notably, ZIP-FIT shows that smaller, well-aligned datasets often outperform larger but less targeted ones, demonstrating that a small amount of higher quality data is superior to a large amount of lower quality data. Our results imply that task-aware data selection is crucial for efficient domain adaptation, and that compression offers a principled way to measure task alignment. By showing that targeted data selection can dramatically improve task-specific performance, our work provides new insights into the relationship between data quality, task alignment, and model learning efficiency.

Frontier AI systems are making transformative impacts across society, but such benefits are not without costs: models trained on web-scale datasets containing personal and private data raise profound concerns about data privacy and security. Language models are trained on extensive corpora including potentially sensitive or proprietary information, and the risk of data leakage -- where the model response reveals pieces of such information -- remains inadequately understood. Prior work has investigated what factors drive memorization and have identified that sequence complexity and the number of repetitions drive memorization. Here, we focus on the evolution of memorization over training. We begin by reproducing findings that the probability of memorizing a sequence scales logarithmically with the number of times it is present in the data. We next show that sequences which are apparently not memorized after the first encounter can be "uncovered" throughout the course of training even without subsequent encounters, a phenomenon we term "latent memorization". The presence of latent memorization presents a challenge for data privacy as memorized sequences may be hidden at the final checkpoint of the model but remain easily recoverable. To this end, we develop a diagnostic test relying on the cross entropy loss to uncover latent memorized sequences with high accuracy.

Evaluation benchmarks are the cornerstone of measuring capabilities of large language models (LLMs), as well as driving progress in said capabilities. Originally designed to make claims about capabilities (or lack thereof) in fully pretrained models, evaluation benchmarks are now also extensively used to decide between various training choices. Despite this widespread usage, we rarely quantify the variance in our evaluation benchmarks, which dictates whether differences in performance are meaningful. Here, we define and measure a range of metrics geared towards measuring variance in evaluation benchmarks, including seed variance across initialisations, and monotonicity during training. By studying a large number of models -- both openly available and pretrained from scratch -- we provide empirical estimates for a variety of variance metrics, with considerations and recommendations for practitioners. We also evaluate the utility and tradeoffs of continuous versus discrete performance measures and explore options for better understanding and reducing this variance. We find that simple changes, such as framing choice tasks (like MMLU) as completion tasks, can often reduce variance for smaller scale ($\sim$7B) models, while more involved methods inspired from human testing literature (such as item analysis and item response theory) struggle to meaningfully reduce variance. Overall, our work provides insights into variance in evaluation benchmarks, suggests LM-specific techniques to reduce variance, and more generally encourages practitioners to carefully factor in variance when comparing models.