UniMamba combines state-space models and attention mechanisms for efficient multivariate time series forecasting with lower computational cost than Transformers.
Framework for evaluating adaptive personalization of educational readings using theory-grounded simulated learners with ontology-based content labeling.
RoboWM-Bench evaluates world models for robotic manipulation, testing both visual realism and physical consistency of generated interactions for robot learning.
Shows image generators develop zero-shot visual understanding capabilities similar to LLM emergent abilities, with evidence of strong generalist vision learning.
Quotient-space diffusion models exploit symmetries in generative tasks like 3D molecular structure generation for improved learning efficiency.
Quantifies and mitigates Self-Preference Bias in LLM-based evaluation systems, addressing systematic favoritism toward models' own outputs in automated assessment.
SIEVES method for selective prediction in multimodal LLMs using visual evidence scoring to improve out-of-distribution reliability while maintaining coverage.
ClawGym framework for building and training AI agents that interact with file systems, tools, and persistent workspaces. Includes systematic data synthesis and diagnostic evaluation.
Motion-aware caching strategy for autoregressive video generation that skips redundant denoising steps through fine-grained pixel-level optimization.
Proposes Predict-then-Diffuse method to dynamically adjust response length in diffusion-based LLMs, addressing fixed-size constraints while maintaining parallel generation efficiency.
PIQL framework integrates privileged information to accelerate training and improve generalization in tabular foundation models, reducing computational burden.
arXiv paper optimizing vision-language-action models by reducing visual bandwidth in world modules through single token-per-frame representation.
Research paper on Flow-OPD: on-policy distillation technique for multi-task text-to-image flow matching models addressing reward sparsity and gradient interference.
Online auditing system for early failure prediction in multi-agent LLM systems, enabling intervention before trajectory-level failures cascade.
Decentralized framework organizing coding agents into co-evolving system for algorithmic discovery, evolving agent behaviors through cumulative guidance.
Security research on poisoning attacks targeting graph-based memory in LLM agents, introducing ShadowMerge attack via relation-channel conflicts.
Study of FP4 quantization in transformer training, progressively enabling low-precision arithmetic across forward/backward passes for Llama 3.1 pretraining.
Novel technique for improving LLM activation steering by addressing KV-cache contamination issues in multi-turn dialogue via attention-level interventions.
Research on applying FP8 low-precision arithmetic to large recommendation models, addressing numerical sensitivity challenges in GPU training.
Introduces successor-representation diagnostic for multi-agent LLM systems to predict communication topology robustness without post-hoc evaluation.
Proposes training-free method to correct pace and path predictions in Vision-Language-Action models blind to temporal dynamics.
GEAR enables granularity-adaptive credit assignment for LLM agents via self-distillation, improving fine-grained supervision beyond outcome-level rewards.
Audits omni-modal benchmarks for visual bias and studies post-training with visually debiased evaluation to improve audio-visual-language integration.
Introduces TokenRatio for token-level preference optimization in LLMs via ratio matching, improving on sequence-level DPO methods.
Argues sparse-to-dense reward principle is more efficient than GRPO alone for LLM post-training with scarce labeled data.
Proposes dual-learning approach combining parameter updates and in-context learning to enable LLMs to adapt continually while retaining plasticity.
Introduces WriteSAE, sparse autoencoder decomposing matrix cache writes in state-space and recurrent language models like Mamba-2 and RWKV-7.
Equips LLMs with active information-seeking during context training to adapt to tasks requiring newly produced or niche domain knowledge without weight updates.
Demonstrates few-shot spoken word classification scaling to 1000 classes using generative meta-continual learning with only 5 shots per class.
Applies generative meta-learning algorithm to few-shot multilingual spoken word classification.
Argues watermarking in generative models should be treated as monitoring primitive considering internal detection inevitability with per-entity attribution keys.
Addresses over-squashing in Graph Neural Networks for multi-label graphs using information bottleneck approach.
LeanSearch v2 retrieves relevant library lemmas globally for Lean 4 theorem proving, addressing gap between semantic search and step-by-step premise selection.
Demonstrates vulnerability in large reasoning models where incomplete inputs cause excessive reasoning traces, enabling denial-of-service attacks via genetic algorithms.
Proposes test-time self-training method for LLMs that enables parameter updates during inference to adapt to specific queries without correcting model misconceptions.
Studies whether LLMs understand semantic meaning of High-Level Message Sequence Charts used in software architecture design.
Applies sparse autoencoders to interpret internal computations of EEG foundation models for clinical applications.
Addresses out-of-distribution generalization for molecular property prediction in drug discovery using improved data selection and domain adaptation.
TraFL trajectory-balance post-training for diffusion language models addresses trajectory locking and improves solution coverage.
Cross-domain benchmark for federated fine-tuning of LLMs on private data across healthcare, finance and other regulated sectors.
EvolveMem self-evolving memory system for LLM agents that co-evolves stored knowledge and retrieval mechanisms across sessions.
EMA framework for efficient model adaptation in learning-based systems operating in heterogeneous dynamic environments.
Unified geometric framework interprets weighted contrastive learning as Distance Geometry Problems with target geometry specification.
Collider-Bench benchmark evaluates LLM agents on reproducing particle physics analyses from LHC using public papers and software.
End-to-end energy accounting for LLM distillation pipelines analyzing full resource costs including teacher-side workloads.
TabPFN-3 foundation model for tabular data scaled to 1M rows with improved training/inference speed and synthetic pretraining.
HodgeCover uses higher-order topological coverage to compress sparse Mixture-of-Experts layers without retraining.
Analysis of discrete diffusion models showing hierarchy between support and frequency information in denoising objectives.
Dywave dynamic tokenization framework for heterogeneous IoT sensing signals addressing non-stationary multi-scale data.
R2R2 regularization method for reinforcement learning reduces overfitting in self-predictive learning under high update-to-data regimes.