Observational study of 90 agentic code generation runs showing reasoning effort matters more than tool access for first-try reliability.
Neuron-aware data selection approach for annotation-free LLM self-distillation in specialized domains without human supervision.
Data-agnostic quantization method for diffusion transformers enabling efficient post-training quantization without calibration data recalibration.
Task-agnostic pretraining approach for Vision-Language-Action models separates physical competence learning from semantic alignment to reduce expert demonstration requirements.
Executable benchmark for evaluating test generation agents on code-test co-evolution with real semantic verification of test-code relationships.
Entropy-aware token pruning method for vision-language models to reduce redundant visual tokens while preserving critical information under dense instructions.
Systematic comparison of matrix-structured optimizers (Muon, SOAP) versus Adam for training machine learning interatomic potentials with improved efficiency.
DemoPSD method improves LLM reasoning via self-distillation with disagreement modulation to reduce overfitting and improve cross-domain generalization.
Benchmark dataset (DramaSR-532K) and reasoning LLM approach for speaker recognition in TV dramas using dialogue attribution tasks.
Program-as-Weights paradigm compiles natural-language specifications into compact, locally-executable neural artifacts for fuzzy functions.
LACUNA testbed evaluates localization precision in LLM unlearning methods that remove sensitive training data and PII.
Model-agnostic global interpretability method for understanding perturbation robustness in image models via spectral analysis.
Black-box method for computing image classifier explanations using formal causal theory and actual causality definitions.
Attention-based diffusion model for completing missing modalities in multimodal emotion and intent recognition tasks.
Psychological network analysis comparing mental imagery structure between humans and LLMs across populations and languages.
Aria agent system for theorem formalization in Lean using retrieval and iterative auto-formalization to improve LLM accuracy in mathematics.
BuilderBench benchmark for developing AI agents that learn through interaction and exploration rather than mimicry alone.
Ophiuchus tool-augmented framework enabling medical MLLMs to dynamically focus on fine-grained visual regions for clinical reasoning tasks.
HAL framework for aligning language models to human-likeness through interpretable, data-driven alignment methods.
Attention-driven neural backbone for solving mixed-integer linear programming using graph neural networks with improved representation power.
Memora harmonic memory representation system for agent memory balancing abstraction and specificity for efficient context-aware retrieval.
BRIDGE psychometric framework predicting human task completion time from model performance without direct human annotations.
PreScience dataset and benchmark for forecasting scientific advances using 98K AI papers with citations and author histories.
OmniGAIA benchmark evaluating omni-modal AI agents with vision, audio, and language integration for complex reasoning and tool usage.
Reinforcement learning approach for multi-agent Formula 1 race strategy optimization, modeling energy, tire degradation, and competitor behavior.
Conformal policy control method using safe reference policies to regulate untested agent policies, balancing exploration and safety constraints.
Dual-helix governance framework stabilizing agentic AI for WebGIS by using knowledge graphs and protocol enforcement to address context and instruction failures.
Formal verification framework for LLM agent protocols, comparing Schema-Guided Dialogue and Model Context Protocol for agent-tool integration.
Category-theoretic framework for defining and comparing AGI systems, addressing lack of formal AGI definitions and benchmarking approaches.
Activation steering methods to prevent LLM misalignment at runtime by manipulating linear structures in activation space.
Framework for interpreting temporal evolution of concepts in LLM agents using conformal inference, improving transparency of sequential behavior.
Hamiltonian-based approach to generative world modeling combining video synthesis, 3D scene reconstruction, and latent predictive models.
LGMT framework uses first-order logic for oracle-free evaluation of LLM reasoning robustness under logically equivalent transformations.
Unified evaluation framework for LLM agentic capabilities that separates model capability from benchmark implementation choices for fair cross-benchmark comparison.
SkillDAG framework models inter-skill relationships as typed directed graphs for LLM agent skill selection at scale, improving over similarity-matching approaches.
Analysis of benchmark contamination detection methods for LLMs, showing limitations of statistical tools in realistic auditing scenarios with distribution shift.
Study of stochasticity sources in AI agents, examining how foundation models and orchestration loops produce variability in planning, tool calls, and outputs.
Research on model collapse from recursive training on synthetic data and how sample selection bias affects model verification in low-resource regimes.
HarnessX: foundry for composable, adaptive agent harnesses combining prompts, tools, memory, and control flow with systematic evolution from execution traces.
Power Systems Agent Benchmark: executable evaluation framework for tool-using AI agents applied to power engineering tasks with concrete outcome verification.
GroundEval: deterministic alternative to LLM judges for agent evaluation, verifying agent search, retrieval, and citation behavior through execution traces.
Grounded Iterative Language Planning: parameterized world models for LLM agents reducing hallucination propagation through measurable transition prediction.
Dynamic representation editing framework steering LLM reasoning trajectories toward truth by analyzing geometry of correctness in reasoning chains.
SAGA: scene-aware multi-agent system for long-horizon strategy planning in CivRealm addressing scene blindness, context overflow, and cross-game learning.
AgentBound: behavioral governance framework for autonomous AI agents controlling consequential actions (transactions, communications) based on operational context.
Framework for autoformalization: automatic translation of natural language mathematics to Lean 4 verifiable code using LLM agents beyond standard libraries.
ClawArena-Team: benchmark for evaluating LLM agents managing subagents through dynamic workflows with parallel asynchronous orchestration.
Framework for spatial reasoning via switching between language and symbolic representations (layouts, grids) to improve multi-hop reasoning in LLMs.
ProtoPilot: self-evolving multi-agent system for automated generation and execution of biological lab protocols with alignment between design and physical execution.
Introduction to Transformer architecture covering basic concepts, model refinements, and NLP applications.