LangFIR uses sparse autoencoders on monolingual data to discover language-specific features for steering LLM output language without parallel corpora.
AgenticFlict dataset of merge conflicts from AI coding agent pull requests on GitHub, studying integration challenges in collaborative AI-assisted development.
Video diffusion framework (CRAFT) for generating synthetic bimanual robot manipulation demonstrations with temporal coherence.
Phase-aware suppression method to reduce hallucinations in Vision-Language Models without iterative optimization overhead.
SecPI framework for secure code generation using reasoning LLMs through security reasoning internalization, addressing inference-time vulnerability mitigation.
Actor-critic reinforcement learning approach for multi-robot task allocation with asymmetric arrivals and switching delays.
Neural method for black-box global optimization using iterative refinement from noisy samples, addressing multi-modal function optimization.
LLM-based approach for multi-file repository code generation with executable validation, addressing dependency resolution and integration challenges.
LiveCoder framework for repository-level code generation preserving and reusing task-specific state across multiple LLM attempts.
Generative foundation model for multimodal histopathology that imputes missing modalities from incomplete medical data.
Reinforcement learning approach for environments with delayed feedback using homomorphic state representation.
Method for stable unsupervised self-evolution of multimodal LLMs using continuous softened retracing resampling for feedback quality.
Adaptive Relational Transformer for pedestrian trajectory prediction using temporal-aware relations in robotics.
Microservice system using NLP and deep learning to automate classification of citizen appeals in government services.
Unlocks prompt infilling in masked diffusion language models by applying full-sequence masking during supervised finetuning.
LightThinker++ enables LLMs to dynamically compress intermediate reasoning thoughts into compact representations for efficiency.
Uses LLMs to capture semantic relationships for tail-item sequential recommendation, addressing sparse interaction problem.
RDEx-CMOP is a differential evolution algorithm variant for constrained multiobjective optimization under budget constraints.
Graph learning approach for melanoma detection in dermoscopic images using graph signal processing.
Scientometric analysis of 15 years of augmented human research, examining conference evolution and core themes.
CREBench evaluates LLMs on cryptographic binary reverse engineering, assessing capabilities for vulnerability discovery and malware analysis.
Research identifying limitations in universality of linear truth directions in LLM activation spaces across different settings.
Study measuring human ability to distinguish LLM-generated news from human-written content across six LLM models.
AutoReSpec uses LLMs to generate formal specifications for programs, addressing syntax and logic errors through techniques for complex control flow.
Neuro-symbolic framework for robot manipulation using vision-language models and autonomous domain construction.
Method for discovering repeated attention patterns in large language models at scale for mechanistic interpretability.
Compares vision-language models and CNNs for spectrum management in satellite-terrestrial networks.
CountsDiff extends diffusion models to discrete ordinal data on natural numbers for generation and imputation tasks.
Automated framework for research-level mathematical problem solving combining LLMs with formal verification to reliably resolve conjectures and verify proofs.
Representational collapse in multi-agent LLM committees: measurement of similarity showing agents produce redundant rationales despite different role prompts, with diversity-aware consensus.
InCaRPose: Transformer-based model for relative camera pose estimation in automotive in-cabin monitoring with distorted imaging environments.
k-Maximum Inner Product Attention for efficient graph transformers, reducing quadratic complexity while maintaining expressiveness for large-scale graphs.
Analysis of analogical reasoning in LLMs comparing probed representations with prompted performance, revealing limitations in latent abstraction and generalization.
Field experiment on LLM agent providing iterative personalized behavioral nudges for electricity and hot-water conservation across intervention rounds.
Regime-calibrated demand priors for ride-hailing dispatch using historical segmentation and multi-metric similarity ensemble for fleet repositioning.
Lorentz-Invariant Auction mechanism for bandwidth allocation across heterogeneous-delay networks including LEO satellites and deep-space relays.
I-CALM: prompt-only intervention reducing LLM hallucinations by incentivizing confidence-aware abstention through reward scheme announcements and humility principles.
DC-Ada: reward-only decentralized adaptation for heterogeneous multi-robot teams, adapting frozen policies to mismatched sensor configurations.
Secure-by-design GenAI framework for cloud security and forensics using LLMs with defenses against prompt injection and forensic rigor requirements.
Spatio-temporal sparse autoencoders for interpretable video representation learning, using contrastive objectives and hierarchical grouping to preserve temporal coherence.
Multi-turn decision making framework for goal-oriented conversational systems balancing information acquisition and target commitment under user intent uncertainty.
AdaptFuse: training-free framework for LLMs to perform Bayesian belief updating across multi-turn interactions without fine-tuning on user data.
Regime-calibrated approach for ride-hailing demand prediction using historical trip segmentation and similarity ensemble matching across temporal patterns.
Low-bit mixed-precision attention kernel using MXFP format for efficient LLM inference, reducing memory bandwidth and computational costs of transformer attention mechanisms.
Symbolic-Vector Attention Fusion (SVAF): mechanism for multi-agent communication enabling agents to evaluate which signal dimensions to use in collective intelligence systems.
VLA-Forget: unlearning framework for vision-language-action embodied models in robotic manipulation, removing unsafe behaviors while preserving perception and language grounding.
TraceGuard: structured multi-dimensional monitoring protocol for detecting attacks on untrusted AI agents, addressing collusion risks through five-dimensional evaluation of agent reasoning and actions.
Gram-anchored prompt learning method for Vision-Language Models using second-order statistics for parameter-efficient adaptation.
Analysis of noisy label robustness in Reinforcement Learning with Verifiable Rewards for training LLM reasoning models.
Causality laundering: security vulnerability in tool-calling LLM agents where adversaries exfiltrate information through denial-feedback patterns.