HatePrototypes: Interpretable representations for hate speech detection covering implicit and explicit hate. Addresses content moderation with transferable embeddings.
UnfoldLDM combines deep unfolding networks with latent diffusion models for blind image restoration. Model-based interpretable approach to image processing.
Probabilistic certification framework improving SmoothLLM defense against LLM jailbreaking attacks. Addresses robustness guarantees with realistic assumptions.
Yo'City: Agentic framework using self-critic expansion for personalized, boundless 3D city generation. Demonstrates AI agent reasoning in creative generation tasks.
Automated pipeline for generating multi-turn conversational jailbreak attacks against LLMs using psychological principles like FITD without manual dataset creation.
Contrastive learning approach for adapting foundation models to domain-specific tasks in Earth observation without full retraining.
AltNet addresses plasticity loss in RL-trained neural networks via parameter reset strategies. Research on continual learning for RL agents.
arXiv paper on evaluating agentic systems via process-centric analysis of trajectories and reasoning patterns rather than outcomes alone. Foundational agent analysis framework.
SALVE framework for neural network interpretability and control using sparse autoencoders. Mechanistic interpretability research not focused on LLMs or agents.
LaMer: Meta-RL framework enabling LLM agents to actively explore and learn from trial-and-error in multi-turn tasks. Research on agent training methodology.
arXiv paper analyzing cost trade-offs between reasoning and non-reasoning LLMs for Text-to-SQL tasks on cloud platforms. Empirical efficiency comparison.
DrivingGen benchmark for generative video world models in autonomous driving. Research on agent simulation and synthetic data generation.
NC-Bench: arXiv benchmark evaluating LLM conversational competence on form/structure vs content. Research paper on LLM evaluation methodology.
Audit of LAION-Aesthetics Predictor studying whose aesthetic values are embedded in visual generative AI training datasets.
Benchmark with 1,800 code completion instances across 6 languages derived from real developer telemetry; avoids contamination, enables detailed diagnostics.
Training-free caching framework for Flow Matching inference using average-velocity perspective and Jacobian-vector products for acceleration.
Open-source cybersecurity LLM trained on 11.8B tokens of curated domain data; supports diverse security workflows while protecting sensitive data.
Data Shapley attribution method for adaptive optimizers like Adam, extending in-run attribution beyond SGD's linear structure.
Multi-label classification study using Schwartz value hierarchies for sentence-level human value detection on sparse, imbalanced datasets.
Reward shaping method for LLM reasoning via reinforcement learning, addressing entropy collapse and exploration challenges in verification-based training.
Study of recurring vulnerabilities in LLM-generated code; introduces FSTab for black-box attacks predicting backend security issues from frontend patterns.
Semantic search system over 9M mathematical theorems using embeddings to retrieve specific results for mathematicians and theorem-proving agents.
LLM-driven recommendation system using multimodal motivation modeling to improve content preference prediction by incorporating review text and heterogeneous data.
Diffusion-guided pretraining for brain graph foundation models, using learnable augmentation for connectome data instead of random dropping.
CoCoA decoder mitigates LLM hallucinations by detecting representational instability across layers, requiring no training.
SToRM token reduction technique optimizes multimodal LLMs for end-to-end autonomous driving with natural language interaction.
Uses agent guidance from learned policies to accelerate robotic RL, reducing sample inefficiency without 1:1 human supervision.
TrasMuon optimizer improves Muon-style methods by adding trust-region adaptive scaling for robust gradient updates.
Variational flow-matching framework for simulation-based inference with structured domain constraints on posteriors.
Token-level noise filtering method for LLM fine-tuning datasets, addressing mismatch between sentence-level annotation and token optimization.
LongAudio-RAG hybrid framework for QA over multi-hour audio with temporal grounding and minimal hallucination.
CogitoRAG framework simulates human cognitive memory for retrieval-augmented generation, using semantic diffusion to preserve integrity.
Condition-gated reasoning system for biomedical QA that handles patient-specific conditional logic in clinical decision-making.
Analyzes deployment tradeoffs in conformal predictors beyond coverage, examining commit vs defer vs error exposure operational metrics.
CrystaL enables latent chain-of-thought reasoning in multimodal LLMs without predefined supervision, improving vision-language integration.
ModernBERT-based multilingual encoder family (150M-300M params) pretrained on 35 languages with domain and dimensional adaptation.
Continual multi-task training framework for universal audio representation across speech, environmental sounds, and music.
CeRA improves low-rank adaptation for LLM fine-tuning by adding manifold expansion via gating and dropout, addressing linear limitations.
Combines behavioral and textual relevance signals using LLMs to improve app store search ranking at scale.
First systematic 4-bit quantization-aware training study for attention mechanisms enabling end-to-end FP4 computation on emerging GPUs.
PEPA: embodied AI agent framework with personality-driven persistent autonomy enabling self-sustaining goals without external task specification.
Conformal prediction framework providing finite-sample coverage guarantees for LLM-based medical entity extraction across clinical domains.
Architectural model for trustworthy AI-assisted software via human-certified module repositories ensuring reliability of AI-assembled systems.
iGVLM: framework enabling dynamic instruction-guided vision encoding in LVLMs for task-specific visual understanding.
MASS: meta-learning framework enabling LLMs to self-adapt at test time by generating synthetic training data for improved downstream performance.
Neuro-symbolic approach combining LLMs with deterministic fact ledgers and hallucination detection for financial reasoning without arithmetic errors.
Research on merging task-specific models into consolidated ones, analyzing parameter competition and domain generalization effects.
vLLM Hook v0 plugin enabling programmable access to LLM model internals for test-time alignment and inference optimization in vLLM serving engine.
Interpretability study on attention sinks in LLMs, explaining why models allocate disproportionate attention to specific tokens including first token bias.
FuzzingRL approach using reinforcement learning for fuzz testing Vision Language Models to automatically generate failure-inducing queries.