The Archive

T2-Z augments robotic world models with learnable environment latents for test-time optimization, capturing unobserved physical and rule-based environmental factors.

CoDA co-evolves simulation data engines and robot policies via shared memory of environment construction, failures, and learning progress using vision-language critics.

PathFM generates complete multivariate trajectory forecasts via flow matching, preserving dependence structure for time series foundation models.

OmniVerifier-C1 decomposes compositional rules into fine-grained components with pointwise scoring for robust multimodal visual verification and world modeling.

Analysis of routing exploration trade-offs in Mixture-of-Experts models with fine-grained experts and shared experts, proposing design mitigations.

FDE-Gym benchmark of 95 real-world delivery tasks from 20 customer scenarios, evaluating agents on ambiguous requirements and live system constraints.

MLExpert generalist agentic model for autonomous ML engineering via workflow orchestration and recursive self-improvement (AI4AI), enabling autonomous system development.

GeoState decouples geometry features from camera viewpoint in 3D world models, separating scene dynamics from view-dependent variation for improved predictability.

3DCodeVerse dataset and framework enable coding agents to generate, verify, and scale 3D executable code for spatial reasoning tasks.

DSP framework trains robot dexterity via behavior cloning from human play, then autonomous data collection and model-predictive control.

LeanGraphSearch improves formal theorem proving by retrieving relevant lemmas via graph-based multi-hop dependency analysis in Lean 4.

Robust RL framework learns adaptive treatment policies for prostate cancer under biomarker uncertainty and disease model drift.

Study reveals LLMs exhibit evaluation awareness in Indic languages, shifting behavior under perceived testing in non-English contexts.

KnowGPO assigns fine-grained credit to search agent decisions via knowledge-grouped policy optimization for multi-turn reasoning.

LangDiT systematically optimizes Transformer architectures for continuous diffusion language models with improved normalization and conditioning.

ACEdit stabilizes multi-turn image editing by training on accumulated distribution shifts from model's own prior outputs.

Time series language models benefit from complementary dual representations: line plots for absolute values, coordinate frames for variations.

Training method increases LLM diversity in document-grounded scientific ideation while maintaining quality across proposed research ideas.

Market-based post-training framework enables decentralized agent swarms to jointly learn action policies and self-orchestration through bid-based coordination.

AVC-Reward proposes event-centric evaluation framework for audio-video generation models handling multi-event composition and asymmetric cross-modal relations.

Four Gemini 4 Argon agents autonomously coordinate and defeat Minecraft Ender Dragon by writing 2,783 programs with no predefined roles or human logic.

Steering agent framework addresses goal drift in LLM-based multi-agent systems via dynamic intervention and context-aware course correction.

MindEval2 benchmark evaluates LLMs in mental health support using clinically grounded synthetic patients and physician-calibrated rubrics instead of opaque Likert scales.

Grid-Bench executable benchmark measures LLM ability to predict state changes across 2D-4D discrete worlds through sequential spatial operations.

Study distinguishes active facts supporting generalization from inert facts via fine-tuning on constructed relations, finding linearity before fine-tuning predicts generalization.

AI Consilium multidimensional benchmark evaluates LLM clinical second opinions on 300 NEJM cases using physician-calibrated judge across eight model pipelines.

Random projection method preserves semantic similarity in RAG systems while mitigating datastore privacy leakage and data extraction attacks.

Backward compatibility framework for zero-bit neural audio watermarking handles independent embedder-detector updates while preventing cross-version detection failures.

Diffusion language models enable few-shot infilling for multi-region code/text edits by learning from demonstrations without causal constraints.

Subspace tuning during fine-tuning mitigates catastrophic forgetting in LLMs by constraining updates away from principal weight matrix directions.