Factored trees let agentic teams reuse past solution paths across problems
GRAFT-ATHENA embeds decision sequences as metric fingerprints so new physics tasks draw on accumulated experience and invent new solvers.
cs.MA
Multiagent Systems
Covers multiagent systems, distributed artificial intelligence, intelligent agents, coordinated interactions. and practical applications. Roughly covers ACM Subject Class I.2.11.
sort pith recommended most recent
CPU scheduling cuts agentic AI latency up to 3.9x
Analysis of CPU orchestration in autonomous agents yields two methods that raise hardware overlap and balance mixed requests on hybrid CPU-G
· “Towards Understanding, Analyzing, and Optimizing Agentic AI Execution: A CPU-Centric Perspective”
Shared memory enables consistent graph building in RAG
A multi-agent system with global context resolves conflicts and fragmentation to improve retrieval accuracy on complex tasks.
· “MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation”
Local rerollouts fix unfair credit assignment in memory LLM agents
By comparing memory operations from identical states, LoGo-GRPO supplies precise signals while retaining end-to-end trajectory rewards.
· “Memory-R2: Fair Credit Assignment for Long-Horizon Memory-Augmented LLM Agents”
AI multi-agent pipeline automates 97% of invoices
Production run shows 70% labor reduction and major drops in emissions and water use for document workflows.
· “MADP: A Multi-Agent Pipeline for Sustainable Document Processing with Human-in-the-Loop”
Data-augmented starts cut exploitability in hard-to-explore games
Sampling intermediate states from offline demonstrations lets regularized gradients find lower-exploitability equilibria under fixed compute
· “Data-Augmented Game Starts for Accelerating Self-Play Exploration in Imperfect Information Games”
Chip-design agents: 20% on multi-objective PPA
The benchmark's 145 tasks show the gap is trade-off reasoning and geometry planning, not rule recall.
LLM decision maker raises swarm foraging yield across setups
Inserting an unmodified model at three points in CPFA lets robots collect more resources with less variance when team or arena size changes.
· “LLM-Foraging: Large Language Models for Decentralized Swarm Robot Foraging”
Users and AI co-edit knowledge graphs for better organization
An interactive visual system let participants create slide decks with higher expert-rated coverage and structure while lowering cognitive负荷.
· “MindTrellis: Co-Creating Knowledge Structures with AI through Interactive Visual Exploration”
Embedding summaries route queries to the right agents
Iterative aggregation and refinement let the system assemble answers from distributed specialist agents without broadcasting every query.
· “Talk to Right Specialists: Iterative Routing in Multi-agent Systems for Question Answering”
AI buyer agents leak willingness to pay from dialogue
Seller inference recovers budgets nearly one-for-one from natural-language profiles, and confidentiality instructions do not stop it.
· “When Agents Shop for You: Role Coherence in AI-Mediated Markets”
Multi-agent AI lifts tool execution success 87.5 points in machining
By routing tasks to specialized agents and verifying against physics models, the system generates auditable plans that simulations indicate
One frozen LLM matches multi-agent accuracy at 20x fewer tokens
MoRe composes learned role vectors per query, giving one model adaptive multi-perspective reasoning in a single pass.
· “One Model, Many Minds: Unlocking Multi-Agent Synergy in a Single Agent via Mixture of Roles”
Lineage-first agent clears 80% MLE-bench tasks at 1/12 cost
Across 75 tasks it scores 60 medals, 49 gold, versus 55 and 34 for Claude Code, with auditable lineages.
· “Praxist: From Experimental Artifacts to Solution Lineages”
AI agent tunes 3D-printer settings to hit factory specs in a few runs
An LLM with memory and tool access sets laser, temperature, and overlap from past builds and ASTM test results.
· “AI Agentic Selective Laser Sintering Process Optimization”
Stage-specific skills lift AI hypothesis generation and testing
HypoForge learns from adversarial critique and empirical outcomes, improving both without fine-tuning.
Specialist-agent relay solves 62% of CodeContests tasks
A prompt-only relay with public-test checks at every step closes most of the gap to heavier systems at one-third the cost.
· “MARS: Multi-Specialist LLM Relay System for Competitive Programming”
Agent labor market beats central planners at routing LLMs
Bidding, reputation, and subcontracting outperform every baseline across math, code, quizzes, and agent tasks.
· “Markets, Not Planners: Decentralized Orchestration of LLM Agents with Private Information”
Open-world AI agents set five new math records
The discoveries include a 604-point kissing configuration, a new Kakeya family, and an improved Erdős bound.
· “Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment”
Full-output exchange erases multi-agent LLM diversity in one round
New tests show agents converge after reading each other's solutions; independent generation preserves the gain.
· “The Interaction Tax: When Communication Erases Diversity in Multi-Agent Teams”
Middle-size agent splits lead
Two- and three-worker splits hit 0.830 on VAT cases vs 0.720 and 0.770 at the extremes; one bad record degrades every setup.
· “Right-Sizing LLM-Agent Decomposition in VAT Determination: A Pilot Controlled Sweep”
Multi-agent LLM hits 97.7% F1 on live million-link fiber nets
Ten-week production run also cut fault incidents by more than 60 percent, the paper reports.
Automatic harness tuning lifts agent pass rates by up to 10 points
Patches prompts, tools, and control logic from failure traces, beating hand-tuned harnesses with far fewer rollouts.
· “AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces”
One audit predicts which social mechanisms survive scaling up
Gossip's failure population equals message reach times lifetime; pre-filed predictions held on third-party code.
· “Predicting the scale limits of social mechanisms in agent societies”
Two tiny simulators expose the full LLM/VLM agent loop
Run locally on macOS, Windows, and Linux, logging every perception and decision for study.
· “Minimal Local Simulation Foundations for LLM- and VLM-Driven Agents in 2D and 3D Environments”
A runtime gate, not prompt text, stops unsafe robot actions
Informed LLM planners still dispatch 23–29% of faulty steps; a deterministic orchestrator blocks all of them.
· “Physical Agentic AI: An Architecture for Orchestrating a Robot Crew with LLMs”
Align peaks, don't denoise: the real X-ray gap is structural
Correcting a small rigid peak drift more than doubles retrieval; recalibration on measured spectra restores coverage.
Opinion-guided strategies coordinate with any rational agent
Two identical agents settle on one admissible joint behavior from their interaction alone.
· “Opinion-Guided Layered Strategies for Decentralized Coordination”
Simulation-coupled LLM agents beat language-only reasoning
In crystallization tests, simulator experiments cut vague wording fourfold and lifted expert ratings above 4.
· “LLM Agents Perform Controlled Experiments Using Simulation Models”
One spine plus parallel branches raises task success 6–16.5%
Keeping live state in one VM lineage and discarding branch VMs cuts per-task cost by 34–70% on 200 tasks.
One identity record binds agent capability
TessIndex anchors compact proofs on-chain and keeps rich metadata off-chain, so agents can be verified before payment flows.
· “TessIndex: Capability Verified Identity System for the Agent Economy”
Evolved agent teams beat hand-crafted ones on 4 benchmarks
A single task-agnostic optimizer treats agent trajectories as loss signals and improves every backbone it runs on.
Recursive FractionalVeto ranks all candidates at optimal distortion 3
Proving that building a full ranking need not cost extra metric distortion over picking a single winner.
Reused architecture knowledge wins 9 of 12 EHR modeling tasks
One shared-weight supernet plus a cross-hospital prior finds better transformer designs with less compute.
Adaptive dispatch cuts worst-case robot task waits by 10 percent
In hospital tests, the forecast-aware rollout serves all requests and shortens the longest waits.
Count wrong, respond right: LLMs miss at the wrong depth
New games make each reasoning depth yield a distinct action, exposing where models fail.
· “Level-k Distinguishable Mechanisms for Evaluating Bounded Rationality in LLMs”
Goal-driven AI can show self-preservation without any survival instinct
Adversarial tests show shutdown resistance is an incentive problem, so safer agents need better goals, tools, and supervision.
0.9991 monitor coverage is the key to fail-operational drone swarms
Catching 99.91 percent of flight-critical faults meets the one-in-ten-million per hour crash target.
· “A Safety-Driven Architectural Framework for Fail-Operational Drone Swarms in Critical Missions”
Agent fan-out creates bursts Poisson traffic models miss
One task can hit a shared LLM backend as a calm sequence, a burst, or a mesh—reasoning gaps fit a log-normal curve.
· “Towards Traffic Modelling of Multi-Agent Systems: The Role of Coordination Topology”
LLM reward feedback cuts lane error in hybrid autonomous driving
An orchestrator fuses RL, PID, and common-sense rules while the LLM works offline to refine rewards.
Reward-guided tuning cuts multi-agent token use 20.5%
Accuracy matches ARG-Designer on six reasoning and code benchmarks while token bills shrink by about a fifth.
Restricted evidence beats full visibility in 9 of 10 LM societies
When each module reads only its own slot, held-out composition succeeds; fully visible twins mostly memorize.
A like-ranked feed makes LLM agents echo each other's wording
In 448 trials, ranked peer posts raise wording similarity; four sources show no reliable stance edge.
LLM agents earn their cost only when demand surges open headroom
A controlled test shows the LLM layer gains a small but significant edge only when a mid-run surge creates headroom
Three tiers defend V2X cars inside a 100-millisecond window
A proposed architecture splits each safety-message verdict among the car, roadside edge, and cloud, biasing ambiguity toward escalation.
· “Autonomous Cyber Defense in Connected Vehicles: A Multi-Agent Approach to V2X Security”
93.2% on MedQA via agent memory and reflection
Role-specific memory and reflection lift medical exam accuracy several points above GPT-4 and prior multi-agent baselines.
· “Adaptive Memory and Reflection Multi-Agent System for Medical Question Answering”
Dental AI that routes evidence beats senior specialists by 17.3 points
Five specialist agents feed a shared blackboard so every diagnosis traces back to specific findings and measurements.
· “DentAgent: Evidence-Centric Multi-Agent Coordination for Multimodal Dental Reasoning”
Persistent memory wins all six chip-flow tuning benchmarks
A state-carrying LLM tuner beats five baselines everywhere; removing its memory cuts frontier quality by 58.5%.
· “StateTune: Transforming LLM-Assisted EDA Flow Tuning into a Stateful, Closed-Loop Process”
Best AI player is not best AI coach in five of seven work tasks
Automation rankings diverge from assistant rankings; leaderboards miss a separate capability.
· “CentaurBench: Benchmarking LLM Capabilities on Augmenting vs. Automating Real-World Work Tasks”
Localized fixes beat full-trajectory training for tool agents
Training only the recovery steps after a failure point lets a small model overtake its teacher and cut tool calls.
Contradiction-gated replanning cuts replans 10x with same reward
Bayesian partner tracking turns belief into action, closing the belief-action gap from 0.41 to 0.20.
· “Bayesian Partner Modelling enables Adaptive Replanning for LLM Coordination”
LLM agents route methane monitoring from hours to minutes
Locally deployed agents coordinate plume analysis and reporting with 92% routing, 85% emission-estimate success.
Hierarchical MPC plus supervisory control keeps UAV fleets conflict-free
A two-layer scheme lets a central UTM ban unsafe airspace events while each drone optimizes only inside the safe set.
· “Model Predictive Supervisory Control for Hierarchical and Distributed UAS Traffic Management”
One contract share can control which AI model an agent picks
A threshold reward share triggers the switch from cheap to capable models; bandit learners match theory closely.
· “Contracting for LLM Delegation: Moral Hazard in Technology and Effort Choice”
Team relay beats private views
A formal model shows iterated cross-agent steps can reach goals no member plans alone, yet goals hidden to all can never be certified.
Parallel group planning preserves near-optimality of lifelong MAPF
A new bound shows spatial grouping and time horizon trade off evenly, cutting plan cost by 25x.
· “A Theoretical Framework for Parallel Lifelong MAPF Using Group Decentralized Planning”
Two-stage AI research cuts unsupported claims to 5 issues
Independent review catches bad experiments while generated ideas still improve image-retrieval recall by 1.85 points.
Physics-aware world model cuts offline state error 77%
True-state MAE drops from 0.41 to 0.10 as physics supervision improves offline CAV control.
10-number code keeps robot formations converging when Wi-Fi fails
In overloaded Wi-Fi tests, uncompressed DMPC received 0.00% of messages; the LSTM code received 98% and converged.
· “Communication Reduction via Semantic-Based Encoding in DMPC Using LSTMs”
Interactive agent nails 3D edit targets: 8x better than one-shot
Plan–perceive–act loop clarifies vague requests and anchors edits to 3D geometry for consistent novel views.
· “DesignAgent3D: Interactive 3D Scene Editing via Designer-like Multimodal Reasoning”
Forget reversibly: suppress stale knowledge
A proposed memory controller moves knowledge through active, dormant, and retired states, enabling governance and later reactivation.
· “Towards Reversible Forgetting: Managing Obsolete Knowledge in Continual Enterprise AI Agents”
Shared multi-agent search doubles kernel speedup at same budget
On a paged-attention task, two agents sharing conclusions reach a 290.8x speedup versus 142.6x for one agent.
· “KernelArc: A Multi-Agent Framework for GPU Kernel Optimization”
A compact spin model predicts how AI agent groups decide
Fit to one-step opinion flips, the rule beats baselines and generalizes to network shapes it never saw.
· “Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents”
Schelling segregation has no critical point
A 12,500-run sweep finds no divergent susceptibility; subcritical cascades explain the sharp segregation transition.
· “Absence of critical scaling in the Schelling segregation model”
Skill version histories boost agent self-evolution by 3–5 points
Distilling public skill diffs into reusable experience lifts all five benchmarks and improves cross-model transfer.
· “VCE-Skill: Enhancing Skill Self-Evolution with Version-Change Experience”
LLM scientist tops protein fitness leaderboard
A hypothesis-driven agent using per-instance feedback outranks 96 models and writes a faster motif algorithm.
· “The Little Scientist: LLM Agent-Driven Discovery via the Scientific Method”
Ethics layer raises abstention from 40% to 62% on degraded inputs
A runtime ethics watchdog withholds more answers when imaging or labs are incomplete, and accuracy on answered cases rises.
· “ETHOS: Towards a Modular Ethics Framework for Clinical Multi-Agent Systems”
Self-reported signals decide when LLM ensembles think fast or slow
Each model scores its own confidence, familiarity, and conflict; hard queries get routed to a deeper pipeline.
Comm-dropout training alone is the strongest learned swarm strategy
Mission success rises from 35% to 62% as test-time communication denial increases.
AI insurance lifts simulated adoption to 84%
In a 300-firm agent simulation, pooling AI's tail losses keeps firms solvent and narrows adoption gaps.
· “Insurance as AI Risk Infrastructure: A Generative-Agent Simulation of AI Adoption”
Tools and state, not just models, set agent serving costs
Across ten apps and three production traces, non-LLM work dominates latency, memory, and cost; coordinated serving cuts latency 29-40%.
· “From LLM Inference to Agentic Workloads: Characterization and Implications for Serving Systems”
Attack and defense AI train each other in one closed loop
A range platform, an attack agent, and a detector turn exercise outcomes into mutual training signal.
· “SysEvolve: An AI-native, safe, autonomous adversarial attack-defense co-evolutionary system”
Conditional cooperation wins when strategies are public
Six of nine strategies are weakly dominated; the survivors are all conditional cooperators.
· “The Open-Strategy Dictator Game: Cooperation Under Mutual Transparency”
Distributed submodular task allocation gets a 1/3-optimal guarantee
Policy gradients match categorical execution; KL-mirror updates survive agents joining and leaving.
· “Submodular Policy Learning for Distributed Task Allocation in Open Multi-Agent Systems”
A tiny LLM can fix an agent's drift after it happens
One 2B model trained on five recovery roles restores most task goals lost to drift in AppWorld.
Zero-conflict offline slicing agents cut inference delay 88%
SHAP-guided reward relabeling lets slices trained on random data coordinate without messages.
· “XAI-Guided Conservative Decentralized Execution for Offline Multi-Agent Network Slicing”
4B model negotiates at frontier level after six-game training
Post-training on six negotiation and scheduling games lifts a 4-billion-parameter model to 0.627 average utility, on par with GPT-4.1 and…
Three conditions keep optimal control linear after agents share
Under them, joint communication-control design reduces to Riccati recursions plus a finite schedule search.
Heterogeneous AI agents sync beliefs without joint training
MEC-hosted latent translators cut belief error 68 percent at equal communication cost in a 6G case study.
· “Heterogeneity-Aware Belief Synchronization for Semantic Communication in AI-Native 6G Networks”
A unified benchmark scores AI research ideas on one 0–5 scale
It formalizes ideas, compares them pairwise against past papers, and Elo-aggregates scores to match expert judgment.
· “LigBench: A Unified and Human-Aligned Benchmark for LLM-based Research Idea Generation”
