The Archive

Detection Transformer approach for fiber orientation and microstructure quantification from multi-shell diffusion MRI without Monte-Carlo inversion.

MEND detects, localizes, and corrects latent hallucinations in self-supervised world models at inference time without ground-truth labels.

Whitening preprocessing reduces reliance on spurious correlations in linear probes by equalizing covariance eigenvalues.

SVM theory in high-dimension low-sample-size settings under spiked covariance models using random matrix convergence.

Hybrid quantum-classical self-supervised learning for fingerprint recognition tested across SimCLR, MoCo v2, and BYOL frameworks.

Token-level reward signals using visual dependency and entropy metrics improve multimodal reasoning in LLMs via reinforcement learning.

Predictive 4D belief framework enables navigation agents to infer and revise moving target locations in dynamic environments.

Score-based variational inference via quantum maximally mixed states addresses high-dimensional degeneracy in eigenvalue problems.

Training-free affinity fusion integrates neural diarization partitions with embedding-based speaker diarization systems.

DAGent uses directed acyclic graphs with evaluate-then-grow planning to dynamically scaffold deep research tasks for multi-agent systems.

Diagonal linear recurrence with optional rank-k nonlinearity distills DreamerV3 world-model policies for robot air hockey control.

Rep2Skill enables LLM agents to evolve textual skills by leveraging internal representations as feedback for skill optimization.

Study evaluates whether self-evolving skills trained on task sets generalize to held-out test tasks across five methods and six benchmarks.

MADBench benchmark systematically evaluates multi-agent debate robustness to adversarial attacks, testing whether debate mitigates or amplifies manipulation.

Analysis of π0.5 and GR00T VLAs reveals distinct physical failure modes under camera faults; freezing causes extreme joint behavior, blackout causes object drops.

Theoretical proof of sharp Gaussian approximation for constant-stepsize SGD with bounded noise under strong convexity and Lipschitz Hessian assumptions.

RefCon method combines iterative refinement and contrastive learning for context-evolving agents to extract higher-quality memories without gold labels.

Schema agent harness uses interactive program induction to discover unknown environment rules, expressing learned dynamics as executable programs.

BELIEFRAG closes-loop controller maintains persistent evidence state across multi-step retrieval, addressing evidence-state fragmentation in adaptive RAG.

ID Balancing applies PID control to stabilize training of extremely sparse MoE models, unifying DeepSeek's loss-free method with proportional-integral-derivative control.

Empirical characterization of high-bandwidth flash memory for LLM serving, analyzing access costs and endurance trade-offs for KV cache retention.

CDMD trains a tabular diffusion model jointly across heterogeneous datasets with mixed-type features, enabling cross-dataset generalization.

E-value stopping rules for Gaussian process black-box optimization without pre-fixed hyperparameters.

On-policy self-distillation for reasoning LMs shows inconsistent gains; token-level analysis reveals collapse risk.

Row normalization in NorMuon optimizer introduces dimension-dependent convergence factors.

Bongard: open-weight System One model separating pattern recognition from judgment using T5Gemma 2 4B encoder-decoder.

T-Router uses parameter-efficient RL with thalamic routing to improve reasoning via reuse of learned computations.

MASCRDM: multi-agent system for real-time compliance and bias detection during LLM training, not just post-hoc filtering.

Reserve-C4B: baseline-relative contrast certificates for conservative bandit optimization under performance budgets.

Co-cheating in self-evolving agents: proposer-solver loop agreement masks errors; external validation prevents collapse.