Nexus-Agents Research Index

Generated: 2026-09-24 (ET) Total Papers: 180 | Techniques: 43 | Topics: 12

Quick Stats

Note: Paper-level status tracking deprecated. Technique status is source of truth.

Topics

Priority 1 (P1) Techniques

These techniques are high-impact and align well with the current architecture.

Priority 2 (P2) Techniques

Medium-impact or requiring moderate changes.

Recently Reviewed Papers

Papers by Topic

Consensus (12 papers)

• Multi-Agent Collaboration Mechanisms: A Survey of LLMs - Taxonomy of collaboration types (cooperation, competition, coopetition)
• Voting or Consensus? Decision-Making in Multi-Agent Debate - Empirical comparison of 7 decision protocols for multi-agent systems.
• Aegean: Formal Consensus Protocol for Stochastic Reasoning - Formal consensus protocol for stochastic reasoning with
• CP-WBFT: Confidence Probe-based Weighted Byzantine Fault Tolerant - Confidence Probe-based Weighted Byzantine Fault Tolerant consensus.
• Free-MAD: Score-Based Decision with Anti-Conformity - Score-based decision with anti-conformity to prevent majority
• MAR: Multi-Agent Reflexion Improves Reasoning Abilities - Multiple agents reflect and critique each other’s outputs.
• Higher-Order Voting: Optimal Weighting and Information Score Prior - Bayesian-optimal aggregation methods (Optimal Weighting and Information
• Auditing Multi-Agent LLM Reasoning Trees Outperforms Majority Vote and LLM-as-Judge -
• ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs -
• EMS: Multi-Agent Voting via Efficient Majority-then-Stopping - Adaptive Incremental Voting (AIV) — sequentially invokes agents and terminate…
• Representational Collapse in Multi-Agent LLM Committees: Measurement and Diversity-Aware Consensus -
• Heterogeneous Consensus-Progressive Reasoning for Efficient Multi-Agent Debate -

Routing (33 papers)

• A Unified Approach to Routing and Cascading for LLMs - Derives optimal strategy for cascading, proves optimality of existing
• RouteLLM: Cost-Quality Routing for LLM Inference - Quality-constrained routing to cheapest model meeting
• RouteLLM: Learning to Route LLMs with Preference Data - Train router on human preference data for dynamic selection
• OptiRoute - kNN + hierarchical filtering with cost/ethics tradeoffs.
• Capability Instruction Tuning - Achieves 80% GPT-4o coverage with smaller model zoo.
• MoMA: Towards Generalized Routing - Unified LLM + agent routing with TOPSIS algorithm for
• Cross-Attention Routing - Single-head cross-attention for query-model matching.
• SATER: Dual-Mode Routing with Confidence-Aware Rejection - Dual-mode routing with shortest-response preference optimization
• IPR: Intelligent Prompt Routing - Quality-constrained routing with user-controlled tolerance parameter
• PILOT: Preference-Prior Routing with Budget Constraints - Contextual bandit (LinUCB) with preference-prior routing and
• STRMAC: State-Aware Routing - State-aware routing with separate encoding of history and
• Edge Multi-LLM: Hybrid Routing with Cascade/ABC Patterns - Hybrid routing with cascade/ABC (Agreement-Based Cascading) patterns.
• DAAO: Difficulty-Aware Agent Orchestration via VAE - VAE-based difficulty estimation for intelligent task routing. Maps tasks
• AdaptOrch: Task-Adaptive Multi-Agent Orchestration in the Era of LLM Performance Convergence -
• Towards Efficient Multi-LLM Inference: Characterization and Analysis of LLM Routing and Hierarchical Techniques -
• ODAR: Principled Adaptive Routing for LLM Reasoning via Active Inference -
• Learning to Route LLMs from Bandit Feedback: One Policy, Many Trade-offs -
• LLM Routing with Dueling Feedback -
• HieraMAS: Optimizing Intra-Node LLM Mixtures and Inter-Node Topology for Multi-Agent Systems -
• Dynamic Tool Dependency Retrieval for Efficient Function Calling -
• Routing, Cascades, and User Choice for LLMs -
• ORCH: many analyses, one merge-a deterministic multi-agent orchestrator for discrete-choice reasoning with EMA-guided routing -
• Learning to Route and Schedule LLMs from User Retrials via Contextual Queueing Bandits -
• Symbolic Mixture-of-Experts: Adaptive Skill-based Routing for Heterogeneous Reasoning -
• HierRouter: Coordinated Routing of Specialized Large Language Models via Reinforcement Learning -
• Router-R1: Teaching LLMs Multi-Round Routing and Aggregation via Reinforcement Learning -
• RLAR: An Agentic Reward System for Multi-task Reinforcement Learning on Large Language Models -
• FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance -
• Sustainable LLM Inference using Context-Aware Model Switching -
• From Myopic Selection to Long-Horizon Awareness: Sequential LLM Routing for Multi-Turn Dialogue - MCTS over dialogue branches induced by different LLM selections; distills a l…
• Triage: Routing Software Engineering Tasks to Cost-Effective LLM Tiers via Code Quality Signals - Proposal paper (not yet evaluated). Routes SE tasks to light/standard/heavy t…
• RouteGoT: Node-Adaptive Routing for Cost-Efficient Graph of Thoughts Reasoning -
• Rethinking Predictive Modeling for LLM Routing: When Simple kNN Beats Complex Learned Routers - Demonstrates that a well-tuned k-Nearest Neighbors (kNN) router matches

Memory (23 papers)

• Acon: Optimizing Context Compression - Task-specific context compression techniques.
• CCF: Context Compression Framework - Learned compression modules for context management.
• Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory - Scalable memory architecture achieving 91% latency reduction
• Context Engineering Survey - Taxonomy of context management strategies.
• xKV: Cross-Layer SVD for KV-Cache Compression - Cross-layer SVD for KV-cache compression exploiting singular
• TreeKV: Tree-Structured Cache Compression - Tree-structured cache compression with smooth context transitions.
• BET: Behavior-Equivalent Token - Single-token compression of system prompts via reconstruction
• A-MEM: Agentic Memory for LLM Agents - Zettelkasten-inspired agentic memory system where memories are
• MIRIX: Six-Type Memory System - Six-type memory system with multi-agent management architecture.
• MobiMem: Post-Deployment Evolution via Memory Modules - Post-deployment evolution via Profile, Experience, and Action
• MemGPT: Towards LLMs as Operating Systems - Operating system-inspired memory management for LLMs with virtual
• ICAL: Continual Learning of Multimodal Agents - Build memory of multimodal experience from suboptimal
• Lifelong Learning of Large Language Model based Agents: A Roadmap - Three core modules: Perception (multimodal input), Memory
• ARIA: Human-in-the-Loop Test-Time Learning - Agents identify knowledge gaps through self-dialogue, request
• Hindsight Belief Memory for Reasoning Agents - Belief Memory layer for reasoning agents that maintains and updates
• Scaling LLM Multi-turn RL with End-to-end Summarization-based Context Management -
• TeleMem: Building Long-Term and Multimodal Memory for Agentic AI -
• LEGOMem: Modular Procedural Memory for Multi-agent LLM Systems for Workflow Automation -
• Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection -
• Multi-Layered Memory Architectures for LLM Agents: An Experimental Evaluation of Long-Term Context Retention -
• Escaping the Context Bottleneck: Active Context Curation for LLM Agents via Reinforcement Learning -
• GAM: Hierarchical Graph-based Agentic Memory for LLM Agents -
• Springdrift: An Auditable Persistent Runtime for LLM Agents with Case-Based Memory, Normative Safety, and Ambient Self-Perception -

Code Generation (24 papers)

• MAR: Multi-Agent Reflexion Improves Reasoning Abilities - Multiple agents reflect and critique each other’s outputs.
• Self-Refine: Iterative Refinement with Self-Feedback - A single LLM acts as generator, feedback provider, and refiner
• Teaching Large Language Models to Self-Debug - Self-Debugging teaches LLMs to debug their own code via few-shot
• Reflexion: Language Agents with Verbal Reinforcement Learning - Agents maintain an episodic memory of verbal reflections that
• LATS: Language Agent Tree Search - Combines Monte Carlo Tree Search (MCTS) with LLM-based value
• Godel Agent: A Self-Referential Agent Framework - Inspired by Godel machines, agents can modify their own logic
• SICA: A Self-Improving Coding Agent - A unified agent that performs tasks AND improves its own
• Self-Improving AI Agents through Self-Play - Formalizes self-improvement as a Generator-Verifier-Updater (GVU)
• Voyager: An Open-Ended Embodied Agent with Large Language Models - Lifelong learning agent that builds an ever-growing library of
• CycleQD: Quality-Diversity for Agent Skill Acquisition - Uses Quality-Diversity framework with cyclic task focus, model
• EXIF: Automated Skill Discovery for Language Agents - Exploration-first strategy using two agents (Alice explores,
• RLAIF vs. RLHF: Scaling Reinforcement Learning with AI Feedback - RLAIF achieves comparable performance to RLHF at 10x lower cost.
• Constitutional AI: Harmlessness from AI Feedback - Two-phase approach: supervised phase (model critiques and revises)
• Agent Q: Advanced Reasoning and Learning for Autonomous AI Agents - Combines guided MCTS with self-critique and DPO for learning
• On the Role of Feedback in Test-Time Scaling of Agentic AI Workflows - Feedback mechanisms that enable agents to improve during
• A Survey of Self-Evolving Agents - Survey of self-evolving agent architectures and techniques.
• Confucius Code Agent: Scalable Agent Scaffolding for Real-World Codebases - Production-scale code agent with Confucius SDK. Achieves 54.3% Resolve@1
• SWE-Refactor: A Repository-Level Benchmark for Real-World LLM-Based Code Refactoring -
• CL4SE: A Context Learning Benchmark For Software Engineering Tasks -
• Retrieval-Augmented Code Generation: A Survey with Focus on Repository-Level Approaches -
• In Line with Context: Repository-Level Code Generation via Context Inlining -
• Are LLMs Reliable Code Reviewers? Systematic Overcorrection in Requirement Conformance Judgement -
• TAM-Eval: Evaluating LLMs for Automated Unit Test Maintenance -
• MARS$^2$: Scaling Multi-Agent Tree Search via Reinforcement Learning for Code Generation -

Orchestration (32 papers)

• Multi-Agent Collaboration Mechanisms: A Survey of LLMs - Taxonomy of collaboration types (cooperation, competition, coopetition)
• Multi-Agent Collaboration via Evolving Orchestration - Puppeteer-style paradigm with RL-trained orchestrator for dynamic
• AgentsNet - Benchmark for 100-agent coordination scenarios.
• TRINITY: Evolved LLM Coordinator - Evolved LLM coordinator with Thinker/Worker/Verifier roles
• LATTS: Locally Adaptive Test-Time Scaling - Locally adaptive test-time scaling with verifier-based
• Hybrid Architectures for LLMs - Transformer + SSM hybrid analysis.
• Multi-LLM Orchestration Engine - Temporal graph + vector DB integration for multi-LLM orchestration.
• Pick and Spin: Efficient Multi-Model Orchestration - Unified Helm-based deployment with adaptive scale-to-zero
• LatentMAS: Direct Latent Space Collaboration - Direct latent space collaboration through hidden state sharing.
• LATS: Language Agent Tree Search - Combines Monte Carlo Tree Search (MCTS) with LLM-based value
• Multi-AI Agent System for Autonomous Optimization - Five specialized agents (Refinement, Execution, Evaluation,
• Meta-Thinking in LLMs via Multi-Agent Reinforcement Learning - Develop meta-thinking (self-reflection, assessment, control of
• A Survey on LLM-based Multi-Agent System - Comprehensive survey on LLM-based multi-agent systems.
• Confucius Code Agent: Scalable Agent Scaffolding for Real-World Codebases - Production-scale code agent with Confucius SDK. Achieves 54.3% Resolve@1
• AFlow: Automatic Workflow Discovery via MCTS - Monte Carlo Tree Search (MCTS) for automatic workflow discovery and
• SEW: Self-Evolving Workflows for Agent Systems - Self-evolving workflow patterns that adapt and improve through execution
• Forest-of-Thought: Multi-Tree Reasoning with Sparse Activation - Multi-tree reasoning architecture with sparse activation for parallel
• Scaling Agent Systems: Coordination Prediction and Resource Optimization - Prediction of coordination requirements for scaling multi-agent systems.
• AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration - Framework for dynamic sub-agent creation via unified agent abstraction.
• Rethinking the Role of Entropy in Optimizing Tool-Use Behaviors for Large Language Model Agents - Entropy reduction correlates with higher-quality tool invocations. Dense proc…
• Towards Reducible Uncertainty Modeling for Reliable Large Language Model Agents - Framework for uncertainty quantification in LLM agents using conditional unce…
• Codified Context: Infrastructure for AI Agents in a Complex Codebase -
• Agentic Lybic: Multi-Agent Execution System with Tiered Reasoning and Orchestration -
• Multi-Agent LLM Committees for Autonomous Software Beta Testing -
• Multi-Agent Teams Hold Experts Back -
• CommCP: Efficient Multi-Agent Coordination via LLM-Based Communication with Conformal Prediction -
• Towards Adaptive, Scalable, and Robust Coordination of LLM Agents: A Dynamic Ad-Hoc Networking Perspective -
• ScaleCall – Agentic Tool Calling at Scale for Fintech: Challenges, Methods, and Deployment Insights -
• MASPOB: Bandit-Based Prompt Optimization for Multi-Agent Systems with Graph Neural Networks -
• CAMEL: Communicative Agents for “Mind” Exploration of Large Language Model Society -
• Enhanced Mycelium of Thought (EMoT): A Bio-Inspired Hierarchical Reasoning Architecture with Strategic Dormancy and Mnemonic Encoding -
• Reasoning Topology Matters: Network-of-Thought for Complex Reasoning Tasks -

Security (16 papers)

• STPA MCP Framework: System-Theoretic Process Analysis for MCP Safety - Formal STPA safety analysis methodology for MCP tools. Systematically
• Agent-SafetyBench: Comprehensive Safety Evaluation for AI Agents - Comprehensive safety evaluation suite for agent behaviors across multiple
• Beyond Max Tokens: Stealthy Resource Amplification via Tool Calling Chains in LLM Agents -
• Zero-Shot Embedding Drift Detection: A Lightweight Defense Against Prompt Injections in LLMs -
• ICON: Indirect Prompt Injection Defense for Agents based on Inference-Time Correction -
• Silent Egress: When Implicit Prompt Injection Makes LLM Agents Leak Without a Trace -
• AdapTools: Adaptive Tool-based Indirect Prompt Injection Attacks on Agentic LLMs -
• From Secure Agentic AI to Secure Agentic Web: Challenges, Threats, and Future Directions -
• Security Risks of AI Agents Hiring Humans: An Empirical Marketplace Study -
• A Safety and Security Framework for Real-World Agentic Systems -
• MCP Security Bench (MSB): Benchmarking Attacks Against Model Context Protocol in LLM Agents -
• AgentSentry: Mitigating Indirect Prompt Injection in LLM Agents via Temporal Causal Diagnostics and Context Purification -
• Red-Teaming LLM Multi-Agent Systems via Communication Attacks -
• ClawGuard: A Runtime Security Framework for Tool-Augmented LLM Agents Against Indirect Prompt Injection - Deterministic interceptor at tool-call boundary. LLM-based rule induction der…
• Architecting Secure AI Agents: Perspectives on System-Level Defenses Against Indirect Prompt Injection Attacks -
• ShieldNet: Network-Level Guardrails against Emerging Supply-Chain Injections in Agentic Systems -

Evaluation (8 papers)

• Predictive Coding and Information Bottleneck for Hallucination Detection in Large Language Models -
• Duel-Evolve: Reward-Free Test-Time Scaling via LLM Self-Preferences -
• Are We on the Right Way to Assessing LLM-as-a-Judge? -
• HackSynth: LLM Agent and Evaluation Framework for Autonomous Penetration Testing -
• MobileWorld: Benchmarking Autonomous Mobile Agents in Agent-User Interactive and MCP-Augmented Environments -
• Style Over Substance: Evaluation Biases for Large Language Models -
• Benchmarks for Trajectory Safety Evaluation and Diagnosis in OpenClaw and Codex: ATBench-Claw and ATBench-CodeX - CLI-specific trajectory safety benchmarks (ATBench-Claw for OpenClaw, ATBench…
• Does RL Expand the Capability Boundary of LLM Agents? A PASS@(k,T) Analysis -

Tool Use (3 papers)

• MCP-Zero: Active Tool Discovery for Autonomous LLM Agents -
• Tool-Augmented Policy Optimization: Synergizing Reasoning and Adaptive Tool Use with Reinforcement Learning -
• Towards Reliable Benchmarking: A Contamination Free, Controllable Evaluation Framework for Multi-step LLM Function Calling -

GitHub Issues

Search Tags

#adaptive #agent-specialization #agentic-memory #agreement #anti-conformity #attribute-extraction #automatic-discovery #bayesian #belief-state #benchmarking #budget-constraint #byzantine #cascade #code-repair #comprehensive #confidence-aware #constitutional #context-injection #context-management #contextual-bandit #coordination #coordinator #correlation-aware #cost-optimization #cross-critique #debate #deduplication #difficulty-estimation #dynamic #dynamic-linking #dynamic-selection #embedding #ensemble #episodic-memory #evolution #evolutionary #executable #execution-feedback #failure-recovery #fault-tolerance #feedback-learning #formal-safety #formal-verification #graph #hazard-analysis #hindsight #inter-agent #iterative #latent-space #lifelong-learning #lightweight #linucb #long-term #mcp #mcts #memory-efficiency #memory-evolution #multi-criteria #multi-tree #multimodal #observability #optimal-weighting #parallel-execution #pareto #policy #post-deployment #prediction #preference-data #principles #priority #production #protocol-selection #pruning #puppeteer #quality-constrained #quorum #reasoning #recency-decay #reflection #reflexion #reinforcement-learning #relational #relevance #risk-assessment #role-based #routing #rubber-duck #safety-evaluation #scalable #scaling #scoring #self-critique #self-debug #self-evolving #self-feedback #self-improvement #semantic #semantic-similarity #shortest-response #skill-library #sparse-activation #stpa #streaming #structured-memory #task-classification #task-routing #test-time #three-module #token-efficiency #tolerance #training-free #transfer-learning #universal-difficulty #vae #verbal-rl #verifier #versioning #wave-coordination #weighted #worker-dispatch #workflow-generation #workflow-optimization #zettelkasten

Registry Files

• papers.yaml - All 180 papers with metadata
• techniques.yaml - All 43 techniques with status
• sources.yaml - Product docs and other sources

How to Contribute

See CONTRIBUTING.md for guidelines on adding new research.

Generated from YAML registries. Last updated: 2026-09-24 (ET)