Research map / Multi-agent systems and games / LLM agents

AI-for-Science Low-code Platform with Bayesian Adversarial Multi-Agent Framework

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a Bayesian adversarial multi-agent low-code platform (TM challenger, SG solver, Evaluator) that co-optimizes code, test cases and prompts via non-LLM Bayesian updates, evaluated on SciCode, ScienceAgentBench, HumanEval/MBPP and APPS.

Problem

LLM code generation for scientific tasks is unreliable: hallucinated code and tests propagate through multi-agent workflows, and scientific correctness is hard to evaluate with standard unit tests. Non-expert users also write vague prompts.

Approach

A recursive adversarial loop in which a Task Manager agent generates and refines test cases that challenge a Solution Generator; test-case difficulty (S1), code scores (S2) and prompt scores (S3) are computed without an LLM judge, and prompt composition (test case and sample code pairs) is updated by a Bayesian rule. A Gaussian-process surrogate over AST/embedding features predicts untested code s

Key results

• On SciCode, LCP raises Qwen3-8B Sub-problem resolve rate from 13.2% to 24.7% (+87.1% relative) and Qwen3-14B with framework (30.6%) matches Qwen3-235B baseline (30.6%) in the without-knowledge setting.
• On ScienceAgentBench with GPT-4o, LCP reaches 90.2% VER without knowledge versus 83.3% for Self-Debug.
• With GPT-3.5-turbo, LCP reports 88.4% pass@1 on HumanEval and 91.1% on MBPP, versus 57.3% and 52.2% zero-shot baselines.

Surprising findings

• Qwen3-14B with the framework matches the 16x larger Qwen3-235B-A22B-Instruct baseline on SciCode sub-problems.
• The framework narrows the gap between basic and expert prompts, so a non-expert with the framework outperforms an expert with the baseline alone.

Limitations

• Performance depends on the quality of initial reference code and struggles to enforce implicit physical laws.
• Iterative adversarial refinement consumes substantially more tokens than one-shot prompting.
• Evaluating ML/DL models is resource-intensive and training variability challenges the update mechanism.

Experimental setup

Datasets: HumanEval HumanEval-ET MBPP MBPP-ET APPS SciCode ScienceAgentBench LGG MRI Segmentation (case study) beach profile dataset (case study)

Benchmarks: HumanEval HumanEval-ET MBPP MBPP-ET APPS SciCode ScienceAgentBench

Metrics: pass@1 Resolve Rate (Sub/Main) SR CBS VER

Baselines: Few-Shot CoT ReAct Reflexion Self-Debugging Self-Collaboration MetaGPT MapCoder AgentCoder CodeCoR OpenHands CodeAct LDB

Builds on: Reflexion Self-Debugging MapCoder AgentCoder CodeCoR

Closest papers by meaning

• SciCode: A Research Coding Benchmark Curated by Scientists (NeurIPS 2024 · similarity 0.88)
• From Reproduction to Replication: Evaluating Research Agents with Progressive Code Masking (ICLR 2026 · similarity 0.87)
• ScienceBoard: Evaluating Multimodal Autonomous Agents in Realistic Scientific Workflows (ICLR 2026 · similarity 0.86)
• Star-Agents: Automatic Data Optimization with LLM Agents for Instruction Tuning (NeurIPS 2024 · similarity 0.86)
• InnovatorBench: Evaluating Agents’ Ability to Conduct Innovative AI Research (ICLR 2026 · similarity 0.85)
• AI-Researcher: Autonomous Scientific Innovation (NeurIPS 2025 · similarity 0.85)