An official website of the United States government
With the rapid growth of large language models, big data, and malicious online attacks, it has become increasingly important to have tools for anomaly detection that can distinguish machine from human, fair from unfair, and dangerous from safe. Prior work has shown that two-distribution (specified complexity) hypothesis tests are useful tools for such tasks, aiding in detecting bias in datasets and providing artificial agents with the ability to recognize artifacts that are likely to have been designed by humans and pose a threat. However, existing work on two-distribution hypothesis tests requires exact values for the specification function, which can often be costly or impossible to compute. In this work, we prove novel finite-sample bounds that allow for two-distribution hypothesis tests with only estimates of required quantities, such as specification function values. Significantly, the resulting bounds do not require knowledge of the true distribution, distinguishing them from traditional p-values. We apply our bounds to detect student cheating on multiple-choice tests, as an example where the exact specification function is unknown. We additionally apply our results to detect representational bias in machine-learning datasets and provide artificial agents with intention perception, showing that our results are consistent with prior work despite only requiring a finite sample of the space. Finally, we discuss additional applications and provide guidance for those applying these bounds to their own work.
more »
« less
This content will become publicly available on December 9, 2025
Hypothesis testing the circuit hypothesis in LLMs
Large language models (LLMs) demonstrate surprising capabilities, but we do not understand how they are implemented. One hypothesis suggests that these capabilities are primarily executed by small subnetworks within the LLM, known as circuits. Identifying these circuits is particularly useful in the context of building models that are robust to shortcut learning and distribution shifts. Identifying these shortcut encoding circuits allows us to "turn them off" by replacing their outputs with random values or zeros. Many papers have claimed to identify meaningful circuits in existing language models. In this paper, we focus on evaluating candidate circuits. Specifically, we formalize a set of criteria that a circuit is hypothesized to meet and develop a suite of hypothesis tests to evaluate how well circuits satisfy them. The criteria focus on the extent to which the LLM's behavior is preserved, the degree of localization of this behavior, and whether the circuit is minimal. We apply these tests to six circuits described in the research literature. We find that synthetic circuits -- circuits that are hard-coded in the model -- align with the idealized properties. Circuits discovered in Transformer models satisfy the criteria to varying degrees. To facilitate future empirical studies of circuits, we created the circuitry package, a wrapper around the TransformerLens library, which abstracts away lower-level manipulations of hooks and activations. The software is available at https://github.com/blei-lab/circuitry.
more »
« less
- PAR ID:
- 10569399
- Publisher / Repository:
- Neural Information Processing Systems
- Date Published:
- Journal Name:
- Advances in neural information processing systems
- ISSN:
- 1049-5258
- Format(s):
- Medium: X
- Location:
- Vancouver, Canada
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Most electronic devices have complex circuitry systems embedded in flat and thin structures called printed-circuit boards (PCBs) traditionally to learn at the university level. We systematically compared how 21 youth (ages 15-16) learned basic circuitry concepts and PCB layout design principles using two educational circuitry toolkits, paper circuits or the traditional solderless breadboard for prototyping. Statistical tests of pre- and post-assessments showed large effect sizes (Hedges’ g) of the score-gain differences favoring paper circuits.more » « less
-
This paper addresses how to fairly compare ROCs of ad hoc (or data driven) detectors with tests derived from statistical models of digital media. We argue that the ways ROCs are typically drawn for each detector type correspond to different hypothesis testing problems with different optimality criteria, making the ROCs incomparable. To understand the problem and why it occurs, we model a source of natural images as a mixture of scene oracles and derive optimal detectors for the task of image steganalysis. Our goal is to guarantee that, when the data follows the statistical model adopted for the hypothesis test, the ROC of the optimal detector bounds the ROC of the ad hoc detector. While the results are applicable beyond the field of image steganalysis, we use this setup to point out possi- ble inconsistencies when comparing both types of detectors and explain guidelines for their proper comparison. Experiments on an artificial cover source with a known model with real stegano- graphic algorithms and deep learning detectors are used to confirm our claims.more » « less
-
This paper addresses how to fairly compare ROCs of ad hoc (or data driven) detectors with tests derived from statistical models of digital media. We argue that the ways ROCs are typically drawn for each detector type correspond to different hypothesis testing problems with different optimality criteria, making the ROCs incomparable. To understand the problem and why it occurs, we model a source of natural images as a mixture of scene oracles and derive optimal detectors for the task of image steganalysis. Our goal is to guarantee that, when the data follows the statistical model adopted for the hypothesis test, the ROC of the optimal detector bounds the ROC of the ad hoc detector. While the results are applicable beyond the field of image steganalysis, we use this setup to point out possi- ble inconsistencies when comparing both types of detectors and explain guidelines for their proper comparison. Experiments on an artificial cover source with a known model with real stegano- graphic algorithms and deep learning detectors are used to confirm our claims.more » « less
-
Prior work applying semiparametric theory to causal inference has primarily focused on deriving estimators that exhibit statistical robustness under a prespecified causal model that permits identification of a desired causal parameter. However, a fundamental challenge is correct specification of such a model, which usually involves making untestable assumptions. Evidence factors is an approach to combining hypothesis tests of a common causal null hypothesis under two or more candidate causal models. Under certain conditions, this yields a test that is valid if at least one of the underlying models is correct, which is a form of causal robustness. We propose a method of combining semiparametric theory with evidence factors. We develop a causal null hypothesis test based on joint asymptotic normality of asymptotically linear semiparametric estimators, where each estimator is based on a distinct identifying functional derived from each of candidate causal models. We show that this test provides both statistical and causal robustness in the sense that it is valid if at least one of the proposed causal models is correct, while also allowing for slower than parametric rates of convergence in estimating nuisance functions. We demonstrate the effectiveness of our method via simulations and applications to the Framingham Heart Study and Wisconsin Longitudinal Study.more » « less
