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.
Metamemory agent framework for data-free code generation in LLMs, enabling reference example generation without curated training sets.
Controlled benchmark comparing seven lightweight CNNs on image classification tasks under unified training protocol, measuring accuracy and efficiency trade-offs.
MetaTT: Tensor Train adapter framework for parameter-efficient fine-tuning of pre-trained transformers using shared factorization across layers and matrix types.
Domain-adaptive continuous pre-training method for resource-efficient specialization of LLMs in cybersecurity with minimal tokens.
RedCoder: Automated multi-turn red teaming system for identifying vulnerabilities in code generation LLMs without extensive human effort.
MedRepBench: Benchmark for medical report understanding with 1,925 Chinese medical reports for evaluating VLM and LLM performance.
PsySET benchmark for evaluating LLM steering effectiveness across emotion and personality domains for human-centered interactions.
Model merging technique to navigate alignment-calibration trade-off in LLMs, improving both task accuracy and model calibration through weight interpolation.
CreativityPrism: Cross-domain evaluation framework for measuring LLM creativity across diverse scenarios without heavy reliance on human evaluation.
UniSE: unified decoder-only LM framework for speech enhancement tasks including restoration, speaker extraction, and speech separation.
Research on using LLMs to synthesize and summarize cloud access control policies, reducing manual policy writing complexity and errors.
SEPS framework for fine-grained cross-modal alignment in vision-language models addressing patch redundancy in multimodal LLMs.
Hireca pathology foundation model for interpretable biomarker assessment from histology images using pretrained vision-language architecture.
Theoretical framework combining game theory and process reward modeling to attribute system-level evaluation to individual agents and messages in multi-LLM systems.
Method for continual concept removal from diffusion models addressing stability issues in sequential unlearning applications.
Benchmark with 1,300+ human-curated PowerPoint edits for evaluating agents' ability to modify slides from natural language instructions.
Adaptive parallel reasoning framework for LLMs that reduces inference latency by parallelizing generation while maintaining sequential reasoning quality.
Benchmark for evaluating LLM-powered web agents' robustness to prompt injection attacks hidden in interface elements that redirect task completion.
Theoretical analysis comparing predictive inverse dynamics models to behavior cloning for imitation learning with limited expert demonstrations.
Adaptive batch size selection using gradient noise scales for optimizers based on generalized norms, improving hardware utilization without manual tuning.
Unsupervised time series anomaly detection using multi-view fusion of tokens across time, frequency, and mixed domains.
Research on efficient model upscaling by initializing larger neural networks from trained smaller ones, with principled hyperparameter transfer to reduce tuning costs.
Train-free attention recalibration method to restore linguistic grounding in vision-language-action models for robotic manipulation.
Framework for consolidating scientific knowledge in AI-driven computational physics, addressing knowledge accumulation beyond routine execution.
Adaptive contracts framework for cost-effective AI delegation in text generation tasks with selective evaluation methods.
DriveVLM-RL combines vision-language models with reinforcement learning for autonomous driving, addressing safety and latency challenges.
CaP-X framework for benchmarking Code-as-Policy agents in robot manipulation tasks using executable code and vision-language-action methods.
Lightweight uncertainty quantification method for neural networks using gradient norms and isotropy assumptions.
RAG reranker optimization using LLM feedback and reinforcement learning to align with downstream generation tasks.
Modular framework deploying VLM-guided reinforcement learning for autonomous driving via sim-to-real transfer.
Evaluation of multilingual prompt localization in agent-as-judge frameworks across 55 developer tasks and six backbones.
Autonomous LLM agents reproduce and critique published computational physics papers using first-principles ground truth.
Contract-based interface model for modular embodied agent capabilities with versioning and governance support.
Analysis of Claude Code architecture and design principles for agentic coding tools compared with open-source alternatives.
LLM-based agentic framework for XANES simulation and curation in computational chemistry with workflow automation.
Study of misaligned behaviors in frontier AI models including peer-preservation behavior across eight models.
Communication-efficient federated fine-tuning of LLMs on edge devices using Fisher-guided token quantization.
Claim verification system for semantic aggregates generated by LLMs from relational databases.
Regression test selection framework for ML systems with modular capability components.
Semantic code search system using concept-to-code alignment for explainable and generalizable retrieval.
LLM-based agent framework using contrastive learning to guide automated vulnerability repair in software systems.