We use CLEVER to evaluate several state-of-the-art LLMs prompted in a few-shot manner and show that they can only solve up to end-to-end verified code generation 1/161 problem, establishing CLEVER as a challenging frontier benchmark for program synthesis and formal reasoning. In summary, our contributions include: 1.
Our analysis yields a novel robustness metric called CLEVER, which is short for Cross Lipschitz Extreme Value for nEtwork Robustness. The proposed CLEVER score is attack-agnostic and is computationally feasible for large neural networks.
This survey on spurious correlations uses the Clever Hans metaphor to motivate the problem, formalizes a group-based setup g=(y,a) with core metrics (worst-group, average-group, bias-conflicting), and explains why models latch onto shortcuts (simplicity bias, training dynamics).
579 In this paper, we have proposed a novel counter- factual framework CLEVER for debiasing fact- checking models. Unlike existing works, CLEVER is augmentation-free and mitigates biases on infer- ence stage. In CLEVER, the claim-evidence fusion model and the claim-only model are independently trained to capture the corresponding information.
While, as we mentioned earlier, there can be thorny “clever hans” issues about humans prompting LLMs, an automated verifier mechanically backprompting the LLM doesn’t suffer from these. We tested this setup on a subset of the failed instances in the one-shot natural language prompt configuration using GPT-4, given its larger context window.
One common approach is training models to refuse unsafe queries, but this strategy can be vulnerable to clever prompts, often referred to as jailbreak attacks, which can trick the AI into providing harmful responses. Our method, STAIR (SafeTy Alignment with Introspective Reasoning), guides models to think more carefully before responding.