All Articles
16890 articles total
The Planar Case of Thomas Positive Circuits Conjecture
A mathematical study on the Planar Case of Thomas Positive Circuits Conjecture within dynamical systems.
Breaking Refusal in the First Half: A Mechanistic Study of the Prefill Jailbreak
A mechanistic study on how 'prefill' attacks can bypass refusal mechanisms in aligned language models.
"Trust Junk" Leads to Unjustified Support for Highly Discriminatory Predictive Models
Research showing that accurate but irrelevant data in XAI visualizations can lead users to trust discriminatory models.
Certified Domain Consistency for Multi-Domain Retrieval: Label-Free Per-Domain Contamination Control with Conformal Risk Guarantees
Introduction of C3R, a control layer for multi-domain retrieval that guarantees per-domain contamination control.
Towards Reliable AI-Assisted Analog Design: Template-Constrained LLM Agents for SAR ADC Generation
ATLAS, an LLM agentic framework for generating functional SAR ADC analog circuits through template-constrained generation.
Structured Feedback Improves Repair in an LLM Agent Loop
Research demonstrating that structured feedback (location, observed value, and alternatives) improves LLM agent repair capabilities.
Eta Given Delta: Defining LLM Tool Efficiency With Marginal Tool Utility
Introduces 'tool efficiency' and 'marginal tool utility' as new quantitative metrics to evaluate and optimize the use of tool calls in LLM agent trajectories.
Simplicity Paradox: Debunking myths about prompting and datasets for LLM evaluation
A study debunking the myth that complex prompting techniques always improve LLM performance, finding that baseline prompting often outperforms elaborate methods.
MAPS: Modeling Co-Existing Subjective Perspectives and Shared Meaning in Multi-Agent Cognitive Dialogue
Presents MAPS, a framework for multi-agent cognitive dialogue that allows agents to maintain individual perspectives while converging on shared meaning.
Introspection Fine-Tuning (IFT): Training Small LLMs to Introspect
Proposes Introspection Fine-Tuning (IFT) to train small LLMs to detect and report perturbations in their own internal activations, enhancing AI transparency.
Information-Theoretic Limits of Reliability and Scaling in Language Models
Derives a first-principles scaling law and an information-theoretic ceiling for LLM reliability, formalizing the limits of performance relative to data and capacity.
T5-CSBoost: Adversarial Perturbation Resistant LLM Fingerprinting
Introduces T5-CSBoost, a contrastive style-boosted classifier for robust LLM fingerprinting and AI-generated text detection resistant to adversarial perturbations.
CoEvoT: Co-Evolving Chain-of-Thought Prompting for Graph-LLM Reasoning
Proposes CoEvoT, a co-evolving Chain-of-Thought prompting framework that dynamically updates graph token evidence to improve Graph-LLM reasoning.
ReportMedSAM: Guiding Segmentation Through Radiology Reports
Presents ReportMedSAM, a framework that uses radiology reports and a learnable concept bank to guide medical image segmentation.
Heterogeneous Element-Aware Cross-Version Differencing of Scientific Documents via Layout-Aware Alignment and Structure-Aware Reasoning
Develops a layout-aware framework for cross-version differencing of scientific documents, handling complex elements like tables, formulas, and figures.
Position: Explainability Research Must Prioritize Foundations over Ad-hoc Methods
Argues that Explainable AI (XAI) research should shift from ad-hoc methods to foundational, human-centered, and action-oriented paradigms.
Camera Chase Vehicle
A discussion thread regarding camera chase vehicles used in cinematography.
Pretraining Data Can Be Poisoned through Computational Propaganda
Researchers demonstrate that pretraining data for language models can be poisoned via public discussion interfaces and introduce 'HalfLife' to estimate adversarial content inclusion.
All Polarized but Still Different: a Multi-factorial Metric to Discriminate between Polarization Behaviors on Social Media
The authors propose GRAIL, a multi-factorial metric for discriminating between different polarization behaviors on social media using an adaptable Generalized Additive Model.
Fast-Fading Channel and Power Optimization of the Magnetic Inductive Cellular Network
This paper models fast-fading channels in magnetic inductive cellular networks for underground communication and proposes a power control algorithm using multiagent Q-learning.