perturbation

This paper proposes TASCO, a framework for test-time adaptation in LLMs that incorporates local stability into confidence-based optimization to improve reasoning accuracy and token efficiency without updating model parameters.

The paper proposes RMS-RSP, a perturbation-sensitive method for selecting medical questions to receive rationale supervision, improving robust accuracy and semantic consistency in QA systems.

The article explores the philosophical problem of Plato's Cave in the context of LLMs, proposing an experiment to compare how different conversational regimes—reconstructive versus perturbation-sensitive—might yield measurable differences in interaction behavior.

This paper introduces PRISM, a perturbation-based method for spatially resolved interpretability of large language models, adapting neuroimaging subtraction analysis to transformers and applying it in parallel to post-stroke aphasia patients to recover shared phonemic-favoring dissociations.

Introduces DECAF, a method that decomposes perturbation responses into evidence, contradiction, and fragility components, improving interpretability over raw response magnitude and achieving strong results across vision benchmarks.

Introduces AndroidReality, a perturbation-based framework for evaluating and improving the robustness of mobile agents, with a taxonomy of real-world interface perturbations and a training-free Test-Time Introspective Recovery (TTIR) mechanism.

This paper systematically studies perturbation-based continued pre-training (CPT) for improving zero-shot dialect robustness in multilingual LLMs, comparing six training conditions across German, Italian, and Arabic. It finds that character-noised CPT is the most effective general strategy and reveals that different perturbation methods induce distinct robustness mechanisms.

This paper proposes a method to identify spuriously correlated samples after model convergence by measuring prediction fragility under input perturbation, requiring no group labels or early-stopping epochs. Rebalancing training with detected samples improves worst-group accuracy on Waterbirds from 57.3% to 80.8%.

This paper proposes using generative inpainting to create photorealistic perturbations for LIME, improving the quality of explanations by avoiding out-of-distribution artifacts common in traditional occlusion methods.

Proposes JaiLIP, a method that jailbreaks vision-language models by generating imperceptible adversarial images using loss-guided perturbation, achieving high toxicity and outperforming existing methods.

Arc Institute's PerturbSpace enables high-throughput single-cell profiling of transcriptome, location, CRISPR guides, clonal relationships, and surface proteins from many samples in one day, using standard single-cell sequencing.

This paper identifies Footprint Bias in document layout analysis robustness evaluation and proposes a structure-aware auditing framework that decouples probe construction and pathway attribution, showing that small structurally targeted probes cause comparable downstream degradation to larger perturbations.

This paper introduces Shesha, a geometric stability metric that quantifies directional coherence of single-cell CRISPR perturbation responses using mean cosine similarity, revealing regulatory architecture and predicting cellular stress across 2,200+ perturbations in five CRISPR datasets.