Timezone: »
We have recently witnessed a number of impressive results on hard mathematical reasoning problems with large language models (LLMs). At the same time, the robustness of these models has also been called into question.Building on the idea of behavioral testing, we propose a novel framework, which pins down the causal effect of each factor in the input, e.g., the surface form of the problem text, the operands, and math operators, on the output. By grounding the behavioral analysis in a causal graph describing an intuitive reasoning process, we study the behavior of LLMs in terms of robustness and sensitivity to direct interventions in the input space. We apply our framework on a test bed of bivariate math word problems.Our analysis shows that robustness does not appear to continuously improve as a function of scale, but that the recent LLM, GPT-3-Instruct (175B), achieves a dramatic improvement in both robustness and sensitivity, compared to all other GPT variants.
Author Information
Alessandro Stolfo (ETH Zürich)
Zhijing Jin (ETH Zürich)
Kumar Shridhar (ETH Zurich)
Bernhard Schölkopf (MPI for Intelligent Systems, Tübingen)
Mrinmaya Sachan (ETH Zurich)
More from the Same Authors
-
2021 Spotlight: Iterative Teaching by Label Synthesis »
Weiyang Liu · Zhen Liu · Hanchen Wang · Liam Paull · Bernhard Schölkopf · Adrian Weller -
2021 Spotlight: DiBS: Differentiable Bayesian Structure Learning »
Lars Lorch · Jonas Rothfuss · Bernhard Schölkopf · Andreas Krause -
2022 : Automatic Generation of Socratic Questions for Learning to Solve Math Word Problems »
Kumar Shridhar · Jakub Macina · Menna El-Assady · tanmay sinha · Mrinmaya Sachan -
2022 Poster: When to Make Exceptions: Exploring Language Models as Accounts of Human Moral Judgment »
Zhijing Jin · Sydney Levine · Fernando Gonzalez Adauto · Ojasv Kamal · Maarten Sap · Mrinmaya Sachan · Rada Mihalcea · Josh Tenenbaum · Bernhard Schölkopf -
2021 : Natural Language Processing meets Educational Data Science »
Mrinmaya Sachan -
2021 : Boxhead: A Dataset for Learning Hierarchical Representations »
Yukun Chen · Andrea Dittadi · Frederik Träuble · Stefan Bauer · Bernhard Schölkopf -
2021 Poster: Dynamic Inference with Neural Interpreters »
Nasim Rahaman · Muhammad Waleed Gondal · Shruti Joshi · Peter Gehler · Yoshua Bengio · Francesco Locatello · Bernhard Schölkopf -
2021 Poster: Causal Influence Detection for Improving Efficiency in Reinforcement Learning »
Maximilian Seitzer · Bernhard Schölkopf · Georg Martius -
2021 Poster: Independent mechanism analysis, a new concept? »
Luigi Gresele · Julius von Kügelgen · Vincent Stimper · Bernhard Schölkopf · Michel Besserve -
2021 Poster: Iterative Teaching by Label Synthesis »
Weiyang Liu · Zhen Liu · Hanchen Wang · Liam Paull · Bernhard Schölkopf · Adrian Weller -
2021 Poster: The Inductive Bias of Quantum Kernels »
Jonas Kübler · Simon Buchholz · Bernhard Schölkopf -
2021 Poster: Backward-Compatible Prediction Updates: A Probabilistic Approach »
Frederik Träuble · Julius von Kügelgen · Matthäus Kleindessner · Francesco Locatello · Bernhard Schölkopf · Peter Gehler -
2021 Poster: Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style »
Julius von Kügelgen · Yash Sharma · Luigi Gresele · Wieland Brendel · Bernhard Schölkopf · Michel Besserve · Francesco Locatello -
2021 Poster: DiBS: Differentiable Bayesian Structure Learning »
Lars Lorch · Jonas Rothfuss · Bernhard Schölkopf · Andreas Krause -
2021 Poster: Regret Bounds for Gaussian-Process Optimization in Large Domains »
Manuel Wuethrich · Bernhard Schölkopf · Andreas Krause -
2019 : Bernhard Schölkopf »
Bernhard Schölkopf -
2018 : Learning Independent Mechanisms »
Bernhard Schölkopf