arXiv subjects

Michael Sklar

Publications and source records attributed to Michael Sklar.

11 recordsLinked to original sources

arXiv2025

HyperSteer: Activation Steering at Scale with Hypernetworks

Steering language models (LMs) by modifying internal activations is a popular approach for controlling text generation. Unsupervised dictionary learning methods, e.g., sparse autoencoders, can be scaled to produce many steering vectors, but lack guarantees on the individual efficacy of each vector and control over the coverage of relevant steering tasks. In contrast, supervised methods for constructing steering vectors are targeted and effective, but require more data collection and training for each additional steering vector produced. In this work, we introduce HyperSteer, a family of hypernetwork-based architectures which are trained end-to-end to generate steering vectors conditioned on the natural language steering prompts and the internals of the steered LM. In our evaluations, we show that scaling HyperSteer with thousands of steering prompts exceeds the performance of state-of-the-art activation steering methods, even on steering prompts never seen during training. Moreover, HyperSteer performs on par with steering-via-prompting.

cs.CL↗

arXiv2025

HyperDAS: Towards Automating Mechanistic Interpretability with Hypernetworks

Mechanistic interpretability has made great strides in identifying neural network features (e.g., directions in hidden activation space) that mediate concepts(e.g., the birth year of a person) and enable predictable manipulation. Distributed alignment search (DAS) leverages supervision from counterfactual data to learn concept features within hidden states, but DAS assumes we can afford to conduct a brute force search over potential feature locations. To address this, we present HyperDAS, a transformer-based hypernetwork architecture that (1) automatically locates the token-positions of the residual stream that a concept is realized in and (2) constructs features of those residual stream vectors for the concept. In experiments with Llama3-8B, HyperDAS achieves state-of-the-art performance on the RAVEL benchmark for disentangling concepts in hidden states. In addition, we review the design decisions we made to mitigate the concern that HyperDAS (like all powerful interpretabilty methods) might inject new information into the target model rather than faithfully interpreting it.

cs.CL↗

arXiv2024

Guarantees for Comprehensive Simulation Assessment of Statistical Methods

Simulation can evaluate a statistical method for properties such as Type I Error, FDR, or bias on a grid of hypothesized parameter values. But what about the gaps between the grid-points? Continuous Simulation Extension (CSE) is a proof-by-simulation framework which can supplement simulations with (1) confidence bands valid over regions of parameter space or (2) calibration of rejection thresholds to provide rigorous proof of strong Type I Error control. CSE extends simulation estimates at grid-points into bounds over nearby space using a model shift bound related to the Renyi divergence, which we analyze for models in exponential family or canonical GLM form. CSE can work with adaptive sampling, nuisance parameters, administrative censoring, multiple arms, multiple testing, Bayesian randomization, Bayesian decision-making, and inference algorithms of arbitrary complexity. As a case study, we calibrate for strong Type I Error control a Phase II/III Bayesian selection design with 4 unknown statistical parameters. Potential applications include calibration of new statistical procedures or streamlining regulatory review of adaptive trial designs. Our open-source software implementation imprint is available athttps://github.com/Confirm-Solutions/imprint

stat.ME↗

arXiv2024

FLRT: Fluent Student-Teacher Redteaming

Many publicly available language models have been safety tuned to reduce the likelihood of toxic or liability-inducing text. To redteam or jailbreak these models for compliance with toxic requests, users and security analysts have developed adversarial prompting techniques. One attack method is to apply discrete optimization techniques to the prompt. However, the resulting attack strings are often gibberish text, easily filtered by defenders due to high measured perplexity, and may fail for unseen tasks and/or well-tuned models. In this work, we improve existing algorithms (primarily GCG and BEAST) to develop powerful and fluent attacks on safety-tuned models like Llama-2 and Phi-3. Our technique centers around a new distillation-based approach that encourages the victim model to emulate a toxified finetune, either in terms of output probabilities or internal activations. To encourage human-fluent attacks, we add a multi-model perplexity penalty and a repetition penalty to the objective. We also enhance optimizer strength by allowing token insertions, token swaps, and token deletions and by using longer attack sequences. The resulting process is able to reliably jailbreak the most difficult target models with prompts that appear similar to human-written prompts. On Advbench we achieve attack success rates $>93$% for Llama-2-7B, Llama-3-8B, and Vicuna-7B, while maintaining model-measured perplexity $<33$; we achieve $95$% attack success for Phi-3, though with higher perplexity. We also find a universally-optimized single fluent prompt that induces $>88$% compliance on previously unseen tasks across Llama-2-7B, Phi-3-mini and Vicuna-7B and transfers to other black-box models.

cs.CL↗

arXiv2024

Principal-Agent Hypothesis Testing

Consider the relationship between a regulator (the principal) and an experimenter (the agent) such as a pharmaceutical company. The pharmaceutical company wishes to sell a drug for profit, whereas the regulator wishes to allow only efficacious drugs to be marketed. The efficacy of the drug is not known to the regulator, so the pharmaceutical company must run a costly trial to prove efficacy to the regulator. Critically, the statistical protocol used to establish efficacy affects the behavior of a strategic, self-interested agent; a lower standard of statistical evidence incentivizes the agent to run more trials that are less likely to be effective. The interaction between the statistical protocol and the incentives of the pharmaceutical company is crucial for understanding this system and designing protocols with high social utility. In this work, we discuss how the regulator can set up a protocol with payoffs based on statistical evidence. We show how to design protocols that are robust to an agent's strategic actions, and derive the optimal protocol in the presence of strategic entrants.

cs.GT↗

arXiv2024

Incentive-Theoretic Bayesian Inference for Collaborative Science

Contemporary scientific research is a distributed, collaborative endeavor, carried out by teams of researchers, regulatory institutions, funding agencies, commercial partners, and scientific bodies, all interacting with each other and facing different incentives. To maintain scientific rigor, statistical methods should acknowledge this state of affairs. To this end, we study hypothesis testing when there is an agent (e.g., a researcher or a pharmaceutical company) with a private prior about an unknown parameter and a principal (e.g., a policymaker or regulator) who wishes to make decisions based on the parameter value. The agent chooses whether to run a statistical trial based on their private prior and then the result of the trial is used by the principal to reach a decision. We show how the principal can conduct statistical inference that leverages the information that is revealed by an agent's strategic behavior -- their choice to run a trial or not. In particular, we show how the principal can design a policy to elucidate partial information about the agent's private prior beliefs and use this to control the posterior probability of the null. One implication is a simple guideline for the choice of significance threshold in clinical trials: the type-I error level should be set to be strictly less than the cost of the trial divided by the firm's profit if the trial is successful.

stat.ME↗

arXiv2024

Fluent dreaming for language models

Feature visualization, also known as "dreaming", offers insights into vision models by optimizing the inputs to maximize a neuron's activation or other internal component. However, dreaming has not been successfully applied to language models because the input space is discrete. We extend Greedy Coordinate Gradient, a method from the language model adversarial attack literature, to design the Evolutionary Prompt Optimization (EPO) algorithm. EPO optimizes the input prompt to simultaneously maximize the Pareto frontier between a chosen internal feature and prompt fluency, enabling fluent dreaming for language models. We demonstrate dreaming with neurons, output logits and arbitrary directions in activation space. We measure the fluency of the resulting prompts and compare language model dreaming with max-activating dataset examples. Critically, fluent dreaming allows automatically exploring the behavior of model internals in reaction to mildly out-of-distribution prompts. Code for running EPO is available at https://github.com/Confirm-Solutions/dreamy. A companion page demonstrating code usage is at https://confirmlabs.org/posts/dreamy.html

cs.CL↗

arXiv2022

Adaptive Experiments and a Rigorous Framework for Type I Error Verification and Computational Experiment Design

This PhD thesis covers breakthroughs in several areas of adaptive experiment design: (i) (Chapter 2) Novel clinical trial designs and statistical methods in the era of precision medicine. (ii) (Chapter 3) Multi-armed bandit theory, with applications to learning healthcare systems and clinical trials. (iii) (Chapter 4) Bandit and covariate processes, with finite and non-denumerable set of arms. (iv) (Chapter 5) A rigorous framework for simulation-based verification of adaptive design properties.

stat.ME↗

arXiv2021

Semi-Supervised linear regression

We study a regression problem where for some part of the data we observe both the label variable ($Y$) and the predictors (${\bf X}$), while for other part of the data only the predictors are given. Such a problem arises, for example, when observations of the label variable are costly and may require a skilled human agent. When the conditional expectation $E[Y | {\bf X}]$ is not exactly linear, one can consider the best linear approximation to the conditional expectation, which can be estimated consistently by the least squares estimates (LSE). The latter depends only on the labeled data. We suggest improved alternative estimates to the LSE that use also the unlabeled data. Our estimation method can be easily implemented and has simply described asymptotic properties.The new estimates asymptotically dominate the usual standard procedures under certain non-linearity condition of $E[Y | {\bf X}]$; otherwise, they are asymptotically equivalent.The performance of the new estimator for small sample size is investigated in an extensive simulation study. A real data example of inferring homeless population is used to illustrate the new methodology.

math.ST↗

arXiv2021

Optimal Rerandomization via a Criterion that Provides Insurance Against Failed Experiments

We present an optimized rerandomization design procedure for a non-sequential treatment-control experiment. Randomized experiments are the gold standard for finding causal effects in nature. But sometimes random assignments result in unequal partitions of the treatment and control group visibly seen as imbalance in observed covariates. There can additionally be imbalance on unobserved covariates. Imbalance in either observed or unobserved covariates increases treatment effect estimator error inflating the width of confidence regions and reducing experimental power. "Rerandomization" is a strategy that omits poor imbalance assignments by limiting imbalance in the observed covariates to a prespecified threshold. However, limiting this threshold too much can increase the risk of contracting error from unobserved covariates. We introduce a criterion that combines observed imbalance while factoring in the risk of inadvertently imbalancing unobserved covariates. We then use this criterion to locate the optimal rerandomization threshold based on the practitioner's level of desired insurance against high estimator error. We demonstrate the gains of our designs in simulation and in a dataset from a large randomized experiment in education. We provide an open source R package available on CRAN named OptimalRerandExpDesigns which generates designs according to our algorithm.

stat.ME↗

arXiv2020

Improving the Power of the Randomization Test

We consider the problem of evaluating designs for a two-arm randomized experiment with the criterion being the power of the randomization test for the one-sided null hypothesis. Our evaluation assumes a response that is linear in one observed covariate, an unobserved component and an additive treatment effect where the only randomness comes from the treatment allocations. It is well-known that the power depends on the allocations' imbalance in the observed covariate and this is the reason for the classic restricted designs such as rerandomization. We show that power is also affected by two other design choices: the number of allocations in the design and the degree of linear dependence among the allocations. We prove that the more allocations, the higher the power and the lower the variability in the power. Designs that feature greater independence of allocations are also shown to have higher performance. Our theoretical findings and extensive simulation studies imply that the designs with the highest power provide thousands of highly independent allocations that each provide nominal imbalance in the observed covariates. These high powered designs exhibit less randomization than complete randomization and more randomization than recently proposed designs based on numerical optimization. Model choices for a practicing experimenter are rerandomization and greedy pair switching, where both outperform complete randomization and numerical optimization. The tradeoff we find also provides a means to specify the imbalance threshold parameter when rerandomizing.

stat.ME↗