Tag
MuteBench is a benchmark for evaluating multimodal fusion models under modality missing and within-modality missing conditions across clinical datasets. It provides insights into architecture robustness and suggests that diffusion-based imputation can help.
This paper empirically studies LLMs' legal reasoning in tax law, showing that data contamination inflates performance and that neuro-symbolic hybrid systems offer more reliable and robust generalization than monolithic LLMs.
This paper introduces an Information Bottleneck Adapter (IB-Adapter) for Vision-Language-Action (VLA) models to improve robustness against unseen visual disturbances without requiring extra data, achieving up to 30% improvement with minimal parameter overhead.
This paper investigates the resilience of AI-generated text detection methods (fine-tuned RoBERTa, Binoculars, text feature analysis, and ensembles) against paraphrasing attacks, finding that Binoculars-inclusive ensembles are most effective but also most vulnerable to attacks, highlighting a dichotomy between performance and resilience.
This paper proposes a novel transformer verification approach that uses ReLU to represent precise but non-linear bounds for dot products, enabling precise and efficient verification. The method outperforms state-of-the-art baselines on sentiment analysis models.
This paper evaluates six open-weight LLMs on biomedical QA under conflicting evidence conditions, revealing accuracy drops and prediction flips, and proposes a conflict-aware abstention score that improves selective accuracy.
This paper proposes a method for improving reasoning in large language models by recoding data to explicitly represent relationships, enabling efficient principled reasoning with polynomial-time learnability for relational rules, which addresses hallucinations and supports sound reasoning across multiple calls.
SAE-FT introduces a novel fine-tuning method for CLIP models that uses sparse autoencoder constraints to regularize visual representations, improving robustness against distribution shifts while maintaining performance and enabling interpretability.
This paper shows that mixing post-training data into pretraining (early exposure) improves how robustly a model retains capabilities after subsequent fine-tuning, challenging the notion that immediate post-training performance predicts retention. Controlled experiments on 135M and 1B models demonstrate that early exposure consistently improves the trade-off between upstream retention and downstream performance.
This paper proposes FedMPO, a robust federated multimodal graph learning method that addresses modality heterogeneity and missing modalities through topology-aware cross-modal generation, missing-aware expert routing, and reliability-aware aggregation, achieving performance gains on multiple datasets.
Introduces Persona Policies (PPol), a plug-and-play control layer that uses LLM-driven evolutionary program search to generate diverse, human-like user personas for evaluating LLM agents. Achieves 33–62% fitness gains over baseline, with human-likeness rated at 80.4%, and improves agent robustness with +17% task success.
Proposes VeGAS, a test-time framework for MLLM-based embodied agents that samples multiple candidate actions and uses a generative verifier to select the most reliable, achieving up to 36% relative improvement over CoT baselines on challenging tasks.
This paper introduces SAGE, a framework for scalable automated robustness augmentation of LLM knowledge evaluation benchmarks. It uses fine-tuned smaller models with reinforcement learning to generate and verify question variants at a lower cost than existing methods.
This paper presents a robust evaluation framework and training strategies for biomedical publication type and study design classification, using knowledge-guided perturbations to mitigate reliance on spurious features.
EVA-Bench introduces a comprehensive end-to-end framework for evaluating voice agents, simulating realistic multi-turn conversations and measuring performance across voice-specific failure modes with novel accuracy (EVA-A) and experience (EVA-X) metrics. The benchmark includes 213 scenarios across enterprise domains and a perturbation suite for accent and noise robustness, revealing substantial gaps in current systems.
This paper introduces a context optimization method that uses active information seeking via Wikipedia search and browser tools, combined with a search-based training procedure, to achieve robust performance improvements across diverse domains without updating model weights.
This paper introduces a controlled-invariance methodology and two oracle tests (Force and Remove) to determine if LLM hallucination detectors rely on reasoning traces or final answer artifacts. It proposes TRACT, a lightweight scorer using lexical features, which demonstrates robust performance independent of answer-level cues.
This paper demonstrates the robustness of refugee matching impact evaluations using off-policy methods like IPW and AIPW, confirming previous findings on algorithmic refugee assignment.
This paper introduces ROMA, an RL fine-tuning framework that enhances the robustness of multimodal large language models against visual degradations like blur and compression artifacts. It achieves this through a dual-forward-pass strategy and specialized regularization techniques, improving performance on reasoning benchmarks without sacrificing accuracy on clean inputs.
This paper introduces NATD-GSSL, a framework evaluating the robustness of Graph Self-Supervised Learning on noisy, text-driven biomedical graphs. It demonstrates that certain GNN architectures and pretext tasks maintain performance despite real-world noise, offering practical guidance for unsupervised learning in imperfect datasets.