Aaron J. Li

BenchEvolver addresses benchmark saturation in large language models by introducing an automated framework that transforms existing coding problems into more difficult variants. Rather than creating problems from scratch, the approach evolves reference solutions through structured modifications and derives corresponding problem statements and test cases from these evolved solutions. This grounding in executable semantics enables scalable generation of challenging, diverse, and valid tasks. Applied to LiveCodeBench and SciCode, the resulting tasks substantially increase difficulty while preserving correctness and diversity. The curated LiveCodeBench-Plus benchmark shows model Pass@1 scores ranging from 27.5% to 62.6% across frontier models, restoring discrimination among strong systems. Additionally, reinforcement learning on evolved tasks yields measurable improvements on held-out benchmarks.

Large language models (LLMs) are increasingly deployed, yet their outputs can be highly sensitive to routine, non-adversarial variation in how users phrase queries, a gap not well addressed by existing red-teaming efforts. We introduce Green Shielding and propose the CUE criteria — benchmarks with authentic Context, reference standards capturing true Utility, and realistic Elicitation perturbations. We instantiate this through HealthCareMagic-Diagnosis (HCM-Dx), a medical diagnosis benchmark built from patient-authored queries with clinically grounded evaluation metrics. Testing across frontier LLMs reveals that prompt-level factors shift model behavior along clinically meaningful dimensions, and that neutralization — which removes common user-level factors while preserving clinical content — increases plausibility and yields more concise, clinician-like differentials. Our findings show that deployment choices systematically influence model output properties, supporting safer implementation in high-stakes domains.

Sparse autoencoders (SAEs) are commonly used to interpret the internal activations of large language models (LLMs) by mapping them to human-interpretable concept representations. While existing evaluations of SAEs focus on metrics such as the reconstruction-sparsity tradeoff, human (auto-)interpretability, and feature disentanglement, they overlook a critical aspect: the robustness of concept representations to input perturbations. We argue that robustness must be a fundamental consideration for concept representations, reflecting the fidelity of concept labeling. To this end, we formulate robustness quantification as input-space optimization problems and develop a comprehensive evaluation framework featuring realistic scenarios in which adversarial perturbations are crafted to manipulate SAE representations. Empirically, we find that tiny adversarial input perturbations can effectively manipulate concept-based interpretations in most scenarios without notably affecting the outputs of the base LLMs themselves. Overall, our results suggest that SAE concept representations are fragile and may be ill-suited for applications in model monitoring and oversight.

The trustworthiness of Large Language Models (LLMs) refers to the extent to which their outputs are reliable, safe, and ethically aligned, and it has become a crucial consideration alongside their cognitive performance. In practice, Reinforcement Learning From Human Feedback (RLHF) has been widely used to align LLMs with labeled human preferences, but its assumed effect on model trustworthiness hasn’t been rigorously evaluated. To bridge this knowledge gap, this study investigates how models aligned with general-purpose preference data perform across five trustworthiness verticals: toxicity, stereotypical bias, machine ethics, truthfulness, and privacy. Our results demonstrate that RLHF on human preferences doesn’t automatically guarantee trustworthiness, and reverse effects are often observed. Furthermore, we propose to adapt efficient influence function based data attribution methods to the RLHF setting to better understand the influence of fine-tuning data on individual trustworthiness benchmarks, and show its feasibility by providing our estimated attribution scores. Together, our results underscore the need for more nuanced approaches for model alignment from both the data and framework perspectives, and we hope this research will guide the community towards developing language models that are increasingly capable without sacrificing trustworthiness.

In recent years, work has gone into developing deep interpretable methods for image classification that clearly attributes a model’s output to specific features of the data. One such of these methods is the Prototypical Part Network (ProtoPNet), which attempts to classify images based on meaningful parts of the input. While this architecture is able to produce visually interpretable classifications, it often learns to classify based on parts of the image that are not semantically meaningful. To address this problem, we propose the Reward Reweighing, Reselecting, and Retraining (R3) post-processing framework, which performs three additional corrective updates to a pretrained ProtoPNet in an offline and efficient manner. The first two steps involve learning a reward model based on collected human feedback and then aligning the prototypes with human preferences. The final step is retraining, which realigns the base features and the classifier layer of the original model with the updated prototypes. We find that our R3 framework consistently improves both the interpretability and the predictive accuracy of ProtoPNet and its variants.