SemanticOpt uses LLMs to optimize expensive black-box problems by incorporating domain knowledge, expert heuristics, and scientific papers beyond traditional Bayesian optimization.
Regularization technique improving pruned neural network robustness at high sparsity by concentrating weights during training.
Reinforcement learning with evolving rubrics for training deep research agents on long-form attributed answers without verifiable rewards.
LLM-enhanced cross-domain sequential recommendation system addressing domain imbalance and transition issues via dual-phase training.
Evaluation framework for measuring human-AI collaboration in coding, beyond traditional benchmarks to capture collaborative problem-solving.
Theoretical framework for composing learned policies from free energy minimization principles, modeling context-dependent gating mechanisms.
Graph-regularized sparse autoencoders for steering LLM safety behaviors by learning distributed structure in activation space.
Theoretical analysis of how equivariance constraints affect neural network expressiveness in 2-layer ReLU networks.
Language model regression approach for predicting resource consumption and runtime of cloud workflow jobs from semi-structured configurations.
Two-sided guardrail system for enterprise LLMs preventing sensitive data leakage and unethical outputs with human-in-the-loop feedback.
Self-supervised evaluation framework for LLMs in web systems, decoupling structural fidelity from content quality for API invocation tasks.
Foundation model for tabular data that estimates feature importance without model access, enabling zero-shot feature attribution.
Method for optimizing data mixtures in LLM pre-training by decoupling search from training via model merging, avoiding expensive exploration.
Fingerprinting technique to detect when student LLMs are trained on teacher model outputs, balancing detection strength with generation quality.
Method for improving LLMs through continual learning from user interaction logs and deployment feedback instead of just scaling data/parameters.
Unified framework for detecting forged content across text, images, and videos using vision-language models and reinforcement learning.
Layer gradient analysis method for identifying optimal layers to edit in LLMs, improving knowledge editing performance.
Flowette framework for generative graph modeling with recurring subgraph motifs using flow matching and graph transformers.
Systematic assessment of LLM safety benchmark code quality, runnability, and adoption factors across research community.
FormulaCode benchmark evaluates LLM coding agents on repository-level optimization tasks with multi-objective evaluation.
First systematic safety evaluation of LLMs across 12 Indic languages using 6,000 culturally grounded prompts.
Framework using LLMs as adaptive controllers for topology optimization, replacing fixed continuation schedules with state-conditioned decisions.
DMax paradigm enables aggressive parallel decoding in diffusion language models through progressive self-refinement from mask to token embeddings.
Technique for offloading key-value cache to reduce memory footprint and inference latency in long-context LLM applications.
Characterizes hardware utilization and inference performance of YOLO object detection models on NVIDIA Jetson under fault injection.
Corpus2Skill system distills document corpora into hierarchical skill directories for LLM agents to navigate during retrieval-augmented generation.
Benchmark evaluating LLM reasoning capabilities for designing and verifying differential privacy algorithms.
Method for selective knowledge unlearning in LLMs using controllable alignment prompting without modifying model weights.
Pipeline converting imperative programs into typed, attributed graphs for verification artifact reuse across C, Java, and Dafny.
Proposes trust schema and verification framework for agent skills as first-class deployment artifacts, addressing runtime validation and correctness verification.
Demonstrates Trojan Hippo attack enabling data exfiltration from LLM agents via dormant payloads planted in long-term memory through single untrusted tool call.
Mechanistic analysis of VLM hallucinations traces failures to geometric over-alignment in decoder layers; proposes geometric debiasing techniques.
Proposes AIPO method enabling LLMs to learn reasoning through active interaction with environments, extending exploration beyond policy model capability boundaries.
Shows extremely quantized LLMs suffer smoothness degradation beyond numerical precision loss; proposes smoothness-aware quantization method for improved performance.
Presents ShadowMerge poisoning attack on graph-based agent memory via relation-channel conflicts, demonstrating new attack surface for LLM agents.
Studies how FFN architecture choices reshape attention computation in transformers, showing sparse MoE routing shifts computational behavior across model layers.
Proposes Key-Value Means attention block enabling transformers with fixed-size or expandable recurrent compressed memory, achieving O(N) complexity on long-context tasks.
Introduces portable specification for multi-agent coordination that allows agents to codify and systematically improve collaboration protocols beyond framework-specific implementations.
Identifies confound in chain-of-thought faithfulness evaluation: corruption studies measure answer placement rather than intermediate computation importance.
Proposes reasoning-prefix masking technique to distill VLM thinking capabilities into compact models, reducing computational cost while preserving reasoning performance.
Introduces visual latent reasoning method for MLLMs that generates intermediate visual evidence as continuous tokens without external tools, addressing feature-space mismatch.
Proposes four-stage post-training workflow for LLM reasoning combining sparse-reward RL, forward-KL warmup, and on-policy distillation. Evaluated on math reasoning tasks.
ODRPO method for robust LLM alignment using ordinal decomposition of discrete rewards to handle stochastic multi-tier rating schemes.
Protocol-Driven Development framework using machine-executable protocols with invariants and continuous evidence to govern AI-generated software.
Reframing pairwise ranking prompting reranking as active learning problem to improve top-K ranking under noisy and intransitive LLM judgments.
Concept-level machine unlearning for vision-language models using interpretable concept decomposition to remove target knowledge precisely.
Method reframing GUI agent critique from binary classification to continuous semantic alignment for improved action ranking in test-time scaling.
MARS framework for memory-augmented LLM agents in recommendation, using hierarchical belief-state memory instead of flat representations.
Context-Driven Decomposition probe diagnosing when RAG systems prioritize retrieved context over parametric knowledge under conflicting information.
Study evaluating whether coding agents understand least-privilege authorization principles, introducing permission-boundary inference task.