Interpretability analysis of world models in reinforcement learning using probing and causal interventions on IRIS and DIAMOND architectures.
Theoretical framework for hierarchical memory in language agents, unifying design choices for multi-level compression and retrieval under token budgets.
PRISM photonic accelerator achieving O(1) memory scaling for long-context LLM inference by selective KV cache block attention.
Multi-agent deep reinforcement learning with spatio-temporal attention for UAV coordination and relay communications optimization.
Framework for graph foundation models using Riemannian geometry for advancing graph learning beyond current graph neural networks.
mSFT algorithm for multi-task supervised fine-tuning addressing heterogeneous overfitting by dynamically adjusting dataset mixtures.
Bayesian framework for monitoring compliance in rule-governed domains by inferring latent state from observed data.
AgenticRec framework combining LLM agents with ranking feedback and tool integration for improved recommendation quality.
Training-free approach for detecting AI-generated images by capturing subtle discrepancies without model-specific training data.
Comprehensive review of security threats, defenses, and benchmarks for retrieval-augmented generation systems with LLMs.
ComicJailbreak benchmark revealing safety vulnerabilities in multimodal LLMs when exposed to visually-grounded harmful narratives.
Method for efficient token reduction in vision-language models for multi-turn QA, addressing practical inference cost challenges.
Multi-armed bandit algorithm for selecting among generative models under diversity-aware metrics without requiring classical UCB approaches.
SemEval-2026 shared task on abductive event reasoning for LLMs to infer direct causes from evidence in real-world scenarios.
FISFormer replaces transformer self-attention with fuzzy inference system for improved uncertainty modeling in time series forecasting.
Study on cognitive automation bias and epistemic risks from AI interfaces, proposing friction-based design to preserve user agency.
Framework for multimodal chain-of-thought reasoning with dynamic visual thought positioning for improved LLM efficiency.
Ctrl-A automated data augmentation algorithm using control theory for dynamic adjustment of augmentation strength during training.
SteelDefectX vision-language dataset with 7,778 images for steel surface defect detection with textual annotations.
CoRA lightweight adapter improves time series foundation models for multivariate forecasting by capturing inter-channel correlations.
IsalSR framework addresses structural redundancy in symbolic regression by encoding expression DAGs to reduce search space redundancy.
Active Testing framework reduces annotation costs in NLP by selecting informative test samples instead of annotating entire test sets.
Analyzes computational complexity of conditional independence tests in constraint-based causal discovery, proposing improvements to PC algorithm efficiency.
Novel sim-to-real method for humanoid locomotion using state-dependent joint torque perturbations to simulate broader reality gaps than standard domain randomization.
Adaptive video distillation addresses artifacts in few-step video generation by mitigating oversaturation and temporal collapse in diffusion models.
Manifold-aware exploration for GRPO in video generation, mitigating noise injection from ODE-to-SDE conversion to stabilize post-training alignment.
P²O jointly optimizes policy and prompts for reinforcement learning with verifiable rewards, improving exploration on hard samples in LLMs.
SmaAT-QMix-UNet combines parameter-efficient techniques and vector quantization for precipitation nowcasting with reduced computational overhead.
Adaptive LoRA ranks for personalized image generation, dynamically selecting per-layer rank based on subject complexity rather than using fixed values.
SHAPE addresses unsupervised domain adaptation for medical image segmentation using structure-aware feature alignment and plausibility constraints.
Examines two interpretations of temporal difference errors in deep RL literature and their implications for critic loss formulation.
ChronoCon applies contrastive learning to longitudinal medical imaging data for disease severity assessment without explicit annotations.
Suiren-1.0 introduces three molecular foundation model variants for modeling organic systems using DFT data and 3D conformational geometry.
Comparative study of parameter-efficient fine-tuning methods (LoRA, Prompt Tuning, Full Fine-Tuning) for medical text summarization with Flan-T5.
SecureBreak dataset enables evaluation of LLM safety and security alignment, addressing harmful generation risks in deployed models.
TREX provides trajectory-based explanations for multi-objective reinforcement learning agents handling conflicting objectives.
SegMaFormer combines state-space and transformer architectures for efficient 3D medical image segmentation with reduced computational complexity.
ROM mitigates overthinking in Large Reasoning Models by detecting and stopping redundant reasoning steps after reaching correct answers, reducing latency and compute cost.
Proposes hyperbolic geometry for vision-language models to better capture hierarchical and part-to-whole semantic relationships.
Studies directional abliteration for removing refusal behavior from instruction-tuned LLMs using contrast baselines.
Analyzes direction of parameter updates in reinforcement learning with verifiable rewards for improving LLM reasoning capabilities.
Mamba-VMR uses generated video augmentations for improved temporal video moment retrieval from text queries.
Applies locally-deployable LLMs for multimodal survival analysis integrating clinical text, tabular data, and genomic profiles.
Improves document-level machine translation using LLMs with filtered synthetic corpora and two-stage adaptation for coherence.
Uses multimodal LLMs with visual feedback loops for iterative refinement of text-to-layout generation for improved aesthetics.
SPA uses prompt engineering to generate synthetic data at scale for knowledge injection into LLMs in specialized domains.
Evaluates reliability and fidelity of using LLMs as judges to assess quality of other LLM outputs with engineered prompts.
Dyadic is a web-based platform for studying human-human and human-AI conversations in text or video format for research.
SpatialReward provides a verifiable reward model for fine-grained spatial consistency in text-to-image generation via reinforcement learning.
GEM-Rec framework integrates commercial relevance and monetization into generative recommendation systems using semantic IDs.