The Archive

ChunkAct reduces LLM reasoning overhead in robot control by chunking primitives into routines, improving efficiency without retraining.

Cepheid decouples scale from structure in cross-family motion completion via dimensionless priors to handle sparse-to-dense mesh recovery.

Benchmark study compares weighting, filtering, and conditioning strategies for demonstration quality in VLA pretraining across 24 tasks.

Tev-Gym introduces first long-horizon agent dataset with verifiable dense rewards to enable training on multi-hour autonomous tasks.

Diffusion models trained on self-generated samples using target score identity for lattice field theory sampling without paired data.

DoubtLess uses contrastive decoding with smaller steering model to shorten reasoning chains without training or hidden state steering.

AesCode generates code for visual artifacts using decoupled cross-modal rewards from image generators to improve aesthetic quality.

Hybrid Reasoning Agents dynamically switch between latent and text reasoning via HybridLink for efficient internal computation with external tool access.

Topology-aware graph generation enforcing exact persistence diagram matching via local and global constraint-guided approaches.

Two-clock framework analyzes why larger rollout batches in RLVR slow learning but preserve collapse timing under square-root scaling.

LLM-based method for multi-turn molecular optimization using iterative feedback to improve drug candidates across multiple objectives.

Coding agent autonomously designs and improves Earth observation models via recursive self-improvement without human tuning.

LIBERO-RIG benchmark evaluates vision-language-action robot policies across 11 hardware perturbation axes including camera calibration and gripper wear.

Physical Recursive Self-Improvement framework enables robots to autonomously update capabilities and meta-level strategies across tasks despite noisy observations.

Theoretical analysis proves KL-regularized Q-learning achieves O(1/ε) sample complexity for offline finite-horizon MDPs.

Method scales minibatch optimal transport for flow matching to reduce training stochasticity in generative models.

RoboSimStudio interactive toolkit enables humans and agents to iteratively refine physics simulations for robot manipulation via real2sim.

OmniCompass benchmark systematically evaluates multimodal LLM video understanding across vision-language and audio-visual grounding tasks.

Neural inference pipeline with Set Attention learns summary statistics for Bayesian inference over chemical reaction networks.

Scenario-Actor framework adaptively selects LLM decoding temperature based on inferred situational context from Transformer embeddings.

Back-translation outperforms continued pretraining for low-resource machine translation with LLMs when monolingual data is available.

Framework for self-evolving agents that revise prompts, memory, workflows, and code while keeping model weights fixed via feedback-driven optimization.

Analytical study of grokking phenomenon identifying sharp conditions and asymptotic expressions for delayed generalization in neural networks.

Depth Memory framework unifies information flow across network layers as storage-retrieval problem, improving over standard residual connections.

Adaptive Diversity Control framework optimizes behavioral diversity in multi-agent RL via extremum-seeking control for improved coordination.

SkillScaffold enables LLM-based agents to learn skills via agentic RL with policy-adaptive credit assignment and selective skill injection.

3D foundation model approach to whole-body PET/CT image registration using learned representations instead of raw intensities.

SEAR framework demonstrates language-model agents autonomously improving robot control via code-as-policy and skill persistence across sim-to-real transfer.

Byte-sequence transformer for malware detection operates on raw executable bytes to capture long-range dependencies without handcrafted features.

Analysis decomposing LLM ensemble gains into candidate value, diversity, and aggregation effects across four experimental conditions.