Survey of memory mechanisms in LLM agents. Structured account of how memory is designed, implemented, and evaluated to enable persistent learning across interactions.
Studies decision biases in LLMs via 20k bandit trials across decoding configurations. Shows positional order bias and rigid exploitation patterns independent of temperature/top-p.
Multi-agent architecture using LLM agents to reduce hallucinations in structural modeling. Agents operate analysis software (OpenSees) for automated model construction and analysis.
Evaluates LLM capabilities (Llama-3, ChatGPT) on discrete optimization problems using natural language datasets. Tests step-by-step reasoning across varied parameter scales.
Studies deception as engineered capability in LLM agents using multi-agent RPG testbed. Analyzes adversarial manipulation across 36 behavioral profiles with ethical grounding.
Research agent framework using synthetic plans to improve tool exploration. Addresses poor exploration in RL agents through better incentive design for dynamic web tool use.
Evaluates vision-language models on hotel/hospitality VQA tasks. Benchmark measuring VLM performance on decision-oriented visual understanding for consumer applications.
Foundation model for EHR data enabling zero-shot clinical predictions. Proposes task-conditioned pretraining to replace expensive autoregressive inference with direct conditioning on clinical questions.
Ares adaptive reasoning framework for efficient LLM agents, dynamically selecting reasoning effort levels to balance accuracy and inference cost.
Rel-MOSS addresses class imbalance in relational deep learning on databases using graph neural networks for heterogeneous entity structures.
Evolutionary stagewise algorithm design approach using LLMs to automate algorithm design, addressing hallucination through problem mechanism awareness.
Multi-agent LLM system with metacognitive policy optimization enabling continual learning and adaptive human collaboration beyond static pre-trained knowledge.
OSExpert introduces computer-use agents learning professional skills via exploration, with benchmark showing limitations in efficiency and transfer to unseen UIs.
CMMR-VLN framework for vision-language navigation using continual multimodal memory retrieval to improve performance on long-horizon unfamiliar scenarios.
PIRA-Bench transitioning GUI agents from reactive task execution to proactive intent recommendation from visual inputs without explicit user prompting.
CDRRM method for generating interpretable rubrics to improve reward modeling for LLM alignment, reducing expert annotation costs and mitigating evaluator biases.
S2S-FDD bridges industrial time series and natural language using LLMs for explainable zero-shot fault diagnosis with semantic gap bridging.
In-context reinforcement learning approach enabling LLMs to effectively use external tools like Python interpreters and search engines for complex task solving.
UIS-Digger research agent system addressing unindexed information seeking beyond search engines, including dynamic webpages and embedded content.
Condition Insight Agent for industrial maintenance using heterogeneous data sources to support evidence-driven decision-making on asset failures and actions.
Mechanistic analysis of continuation-triggered jailbreak vulnerabilities in LLMs, investigating root causes of safety alignment failures.
FinToolBench benchmark evaluating LLM agents on real-world financial tool use tasks, addressing compliance and data volatility challenges.
Method for efficient bias detection in financial language models through mutation analysis, reducing computational cost vs. exhaustive approaches.
Framework addressing perception, alignment, and reasoning challenges in multimodal mathematical reasoning for visual math problem-solving with LLMs.
CORE-Acu neuro-symbolic framework using structured chain-of-thought and knowledge graphs for interpretable acupuncture clinical decision support with LLMs.
Neurosymbolic AI agent combining LLMs with symbolic computation tools for mathematical discovery, producing new result in combinatorial design theory.
Multi-agentic approach to improve multimodal math reasoning by addressing visual perception bottlenecks in LLM-based mathematical problem solving.
HECG framework for autonomous agents using LLMs with multi-dimensional strategy alignment combining task quality, confidence, reward, and semantic reasoning metrics.
General AI assistant platform with orchestration core, MCP extensibility, tool execution, persistent memory, and task scheduling.
Study of world model-based anomaly detection thresholds under gradual observation drift in RL agents.
LLM agent training method using retrospective feedback and dual intrinsic motivation for continuous adaptation and experiential learning.
Code-based chain-of-thought approach for text-to-image generation with precise spatial and structural control.
Benchmark for evaluating AI agents on multi-document reasoning using 100 years of U.S. Treasury data.
Benchmark suite (AFIB) evaluating financial reasoning capabilities across LLM engines on five dimensions.
Training approach for LLM agents that uses contrastive learning to develop action quality awareness beyond imitation.
Evaluation of LLM-based agents using Model Context Protocol with SPARQL endpoints for federated knowledge graph question answering.
Research on isotonic layer framework for debiasing and calibrating large-scale recommendation systems.
Transformer-based system for ARC reasoning combining neural inference with structure-aware priors and online adaptation for few-shot generalization.
Shows LLM-as-Judge frameworks fail to reliably evaluate adversarial robustness in safety testing due to distribution shifts from red-teaming.
FuzzingRL uses reinforcement learning and fuzz testing to automatically generate adversarial queries that reveal VLM failures and vulnerabilities.
Benchmarking multi-agent DRL approaches for radio resource allocation in cellular V2X networks, addressing non-stationarity and coordination challenges.
Study shows scaling inference compute via consensus and pass@k fails to improve LLM truthfulness in domains without external verifiers, unlike math/code domains.
OptiRoulette is a stochastic meta-optimizer that dynamically selects optimizers during training, achieving 5.3x faster convergence and available as torch-compatible package.
RACER routes queries across multiple LLMs by formulating selection as α-VOR problem to minimize cost while controlling misrouting risk.
Evo unifies autoregressive and diffusion-based language generation through a latent trajectory model with evolving token embeddings.
NAT framework reduces computational cost of reinforcement learning in LLMs by selectively updating tokens instead of backpropagating through entire chain-of-thought trajectories.
GraphSkill applies hierarchical retrieval-augmented generation to LLM-based graph algorithm reasoning, using documentation and code debugging for complex graph problems.
Compares ARIMA, LSTM, BiLSTM, and Transformer models for short-term power load forecasting on energy consumption data.
Graph neural networks for imputing missing pavement condition data using collective learning across road network infrastructure.
Grouter decouples routing from representation in MoE training using preemptive routing to improve convergence and stability.