Tag
This paper proposes a training-adaptive convolutional sparse coding framework that leverages information bottleneck principles for robust visual representation, achieving improved performance on CIFAR and ImageNet under input perturbations.
RoleBreak is an open benchmark for evaluating long-horizon role-playing robustness in spoken dialogue systems, revealing gaps in current models' ability to maintain role consistency and vocal emotion over extended interactions.
This paper studies how reinforcement learning can lead LLM agents to learn spurious tool-use policies based on superficial cues rather than task requirements, and introduces a dense reward method to mitigate this issue.
This paper demonstrates that ordinary typos in user inputs can significantly reduce the effectiveness of activation probes designed to detect prompt injection in LLMs, but introduces a KV-cache fork method that recovers most of the performance loss.
This paper proposes Fed-Equilibrium, a federated learning framework that balances robustness and fairness in clinical networks using topological Pareto control to ensure minority nodes achieve convergence comparable to dominant hubs.
DART proposes a distributional adversarial training framework for recurrent reasoning models, improving their robustness and performance on structured problems by using a local target distribution instead of single-point supervision.
This paper evaluates the robustness of LLM-generated SystemVerilog assertions under semantics-preserving RTL transformations, finding that point accuracy can hide substantial instability and advocating for robustness-aware evaluation in AI-assisted hardware verification.
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 evaluates the robustness of PII detection systems under distribution shifts, identifies failure modes in different architectures, and proposes a hybrid detection pipeline with a QA-driven feedback loop for improved privacy protection.
This paper investigates the structural sensitivity of multilingual large language models to semantics-preserving perturbations in Hindi and Malayalam, showing significant degradation in mathematical reasoning performance and introducing the IndicReStruct benchmark for evaluation.
This paper conducts a multi-level analysis of how input perturbations propagate through decoder-only language models, assessing robustness via output behavior, hidden-state geometry, and attention-head function across models like GPT-2 and Qwen2.5.
Topological Steering is a new framework for controlling large language model behavior using topological data analysis to capture global structures in activation spaces, enabling more robust behavioral control.
This research investigates whether near-tied rankings of large language models remain robust to changes in benchmark item composition using psychometric methods, finding that while overall rankings are stable, individual close orderings can reverse based on item selection.
RL-FAT is a reinforcement learning framework for fair adversarial training that improves robustness while reducing class-wise disparities. Experiments demonstrate competitive accuracy and better fairness compared to standard methods.
Flow-JEPA introduces a conditional flow matching approach to JEPA world models, improving robustness and accuracy in predicting future latent states under noisy conditions.
SnapBench introduces a paired corruption benchmark for robust snap-and-ask multimodal retrieval on mobile, revealing image noise as a key degradation factor and proposing an adaptive fusion method for modality reliability.
This paper proposes a novel adversarial training method that eliminates input gradients by using low-rank Householder expansions, reducing computational cost while achieving robustness comparable to standard techniques for small perturbations.
The paper audits large language models on their refusal and fabrication behavior in clinical pain speech transcripts, finding that authority-framed prompts lead to confident fabrication in models like Gemini 2.5 Flash and Llama 3.1 8B, while cooperative prompting shows robust abstention.
The article discusses how agentic coding and AI could transform open-source software by making coding less scarce, using biological analogies to contrast short-term efficiency with long-term robustness.
This paper evaluates self-supervised pretraining for tabular models under label scarcity and missing data, finding mixed efficacy but reliable improvements under test-time missingness.