# Multi-agent system

Automated design of multi-agent systems has mostly worked at the task level, optimizing one fixed workflow per benchmark and applying it to every query. OOPMAS moves design to the individual query: a frozen LLM as meta-agent first generates object-style agents with roles, tools and persistent state, then an executable main-function workflow, while a dynamic skill library accumulates structured lessons from execution feedback without gradient updates. On a benchmark mixing code, math and QA, OOPMAS reaches 89.6% overall accuracy, 18.1 percentage points above the strongest baseline, ADAS.

Automated design of multi-agent systems has mostly worked at the task level, optimizing one fixed workflow per benchmark and applying it to every query. OOPMAS moves design to the individual query: a frozen LLM as meta-agent first generates object-style agents with roles, tools and persistent state, then an executable main-function workflow, while a dynamic skill library accumulates structured lessons from execution feedback without gradient updates. On a benchmark mixing code, math and QA, OOPMAS reaches 89.6% overall accuracy, 18.1 percentage points above the strongest baseline, ADAS.

Adapting a protein foundation model to one laboratory's assay is an expensive, failure-prone step, and existing agents treat fine-tuning as a reliable tool call whose feasibility is estimated and whose infeasibility surfaces only after compute is spent. CAPO reframes adaptation as a gated decision: it measures feasibility on the target accelerator, admits or rejects a configuration against a budget, runs a 200-step canary inside the real training loop, and may decline. One unchanged system chose partial fine-tuning of ESM-2 8M, refused to fine-tune Boltz-2 in favor of a calibrated head over frozen representations, and applied rank-4 low-rank adaptation to ESM-2 150M across three protein tasks.

Adapting a protein foundation model to one laboratory's assay is an expensive, failure-prone step, and existing agents treat fine-tuning as a reliable tool call whose feasibility is estimated and whose infeasibility surfaces only after compute is spent. CAPO reframes adaptation as a gated decision: it measures feasibility on the target accelerator, admits or rejects a configuration against a budget, runs a 200-step canary inside the real training loop, and may decline. One unchanged system chose partial fine-tuning of ESM-2 8M, refused to fine-tune Boltz-2 in favor of a calibrated head over frozen representations, and applied rank-4 low-rank adaptation to ESM-2 150M across three protein tasks.

Edge AI development has relied on developers hand-building fixed agent workflows that struggle with heterogeneous data modalities and diverse task pipelines. SE2 distills development into four atomic operations, retrieve, reason, code and test, and lets a coordinator agent pick the next one from the current context, while mobile runtime feedback and reinforcement fine-tuning let the agents keep evolving. On the IoTBench benchmark and three multimodal edge AI applications, its synthesized programs reach up to 26% higher accuracy and 28% lower resource consumption than eight state-of-the-art baselines.

Edge AI development has relied on developers hand-building fixed agent workflows that struggle with heterogeneous data modalities and diverse task pipelines. SE2 distills development into four atomic operations, retrieve, reason, code and test, and lets a coordinator agent pick the next one from the current context, while mobile runtime feedback and reinforcement fine-tuning let the agents keep evolving. On the IoTBench benchmark and three multimodal edge AI applications, its synthesized programs reach up to 26% higher accuracy and 28% lower resource consumption than eight state-of-the-art baselines.

A laboratory of large-language-model agents, each holding a separate research role, asked how many calibration measurements are needed to predict a diamond nitrogen-vacancy spin ensemble's response to arbitrary phase-controlled microwave pulse sequences. The principal-investigator agent tightened a proposed bound of 91 calibrations to an exact minimum of 10; the first acquisition met all ten precommitted accuracy criteria but failed phase and reference controls and was ruled inconclusive, and after the agents revised the protocol without relaxing the criteria, the final measurement supported the prediction while two interpolation methods predicted zero.

A laboratory of large-language-model agents, each holding a separate research role, asked how many calibration measurements are needed to predict a diamond nitrogen-vacancy spin ensemble's response to arbitrary phase-controlled microwave pulse sequences. The principal-investigator agent tightened a proposed bound of 91 calibrations to an exact minimum of 10; the first acquisition met all ten precommitted accuracy criteria but failed phase and reference controls and was ruled inconclusive, and after the agents revised the protocol without relaxing the criteria, the final measurement supported the prediction while two interpolation methods predicted zero.

Teams of coding agents generally beat a single agent at rebuilding programs, but at four to seven times the cost. A controlled study across 200 ProgramBench programs finds that putting cheaper models only on exploration and planning comes within 2.7 points of an all-strong-model team that delegates every stage, at roughly a fifth of that team's cost, because discovering the original program's behavior is the real bottleneck.

Teams of coding agents generally beat a single agent at rebuilding programs, but at four to seven times the cost. A controlled study across 200 ProgramBench programs finds that putting cheaper models only on exploration and planning comes within 2.7 points of an all-strong-model team that delegates every stage, at roughly a fifth of that team's cost, because discovering the original program's behavior is the real bottleneck.

Reinforcement learning from verifiable rewards (RLVR) can stall when a jump in task difficulty makes every sampled trajectory fail, leaving zero reward and no learning signal. MENTOR keeps those failed trajectories, has a set of agents diagnose the error in real time and write a targeted instructional hint, then prepends the hint to the original problem for a new rollout. On math and code generation it beats GRPO, DAPO and static hint libraries, with average relative math gains up to 78.3% on Llama-3.2-3B-Instruct.

Reinforcement learning from verifiable rewards (RLVR) can stall when a jump in task difficulty makes every sampled trajectory fail, leaving zero reward and no learning signal. MENTOR keeps those failed trajectories, has a set of agents diagnose the error in real time and write a targeted instructional hint, then prepends the hint to the original problem for a new rollout. On math and code generation it beats GRPO, DAPO and static hint libraries, with average relative math gains up to 78.3% on Llama-3.2-3B-Instruct.

The work presents MEA, a multi-agent framework in which a Proposer agent selects and configures explanation tools based on the question and modality while an Actor agent is optimized end-to-end against faithfulness to turn outputs into natural-language explanations; the authors also introduce question types spanning feature attribution, counterfactual reasoning, and spurious feature detection, each paired with a perturbation-based faithfulness metric, find that frontier LLMs systematically produce unfaithful explanations, and report that MEA outperforms post hoc explainers, agentic, and closed-source baselines across six datasets, with faithfulness gains of +28% (tabular), +21% (text), and +34% (vision) over the untrained backbone.

The work presents MEA, a multi-agent framework in which a Proposer agent selects and configures explanation tools based on the question and modality while an Actor agent is optimized end-to-end against faithfulness to turn outputs into natural-language explanations; the authors also introduce question types spanning feature attribution, counterfactual reasoning, and spurious feature detection, each paired with a perturbation-based faithfulness metric, find that frontier LLMs systematically produce unfaithful explanations, and report that MEA outperforms post hoc explainers, agentic, and closed-source baselines across six datasets, with faithfulness gains of +28% (tabular), +21% (text), and +34% (vision) over the untrained backbone.

The authors present LATHE, a geometric toolkit that expresses seven classes of crystallographic properties, namely bond length, bond angle, dihedral angle, coordination environment, lattice parameters, cell volume, and space group, as differentiable objectives so that a given crystal structure can be edited directly; on single-property benchmarks all seven property types reach near-perfect constraint satisfaction while optimized structures stay close to local energy minima, the LATHE-equipped agent translates over 87.5% of natural-language prompts into valid executable configurations, and in a band-gap inverse-design case study nine of ten independent runs reach the target tolerance window at a typical cost of thirteen hypothesis-evaluation cycles.

The authors present LATHE, a geometric toolkit that expresses seven classes of crystallographic properties, namely bond length, bond angle, dihedral angle, coordination environment, lattice parameters, cell volume, and space group, as differentiable objectives so that a given crystal structure can be edited directly; on single-property benchmarks all seven property types reach near-perfect constraint satisfaction while optimized structures stay close to local energy minima, the LATHE-equipped agent translates over 87.5% of natural-language prompts into valid executable configurations, and in a band-gap inverse-design case study nine of ten independent runs reach the target tolerance window at a typical cost of thirteen hypothesis-evaluation cycles.