An Inspectable LLM Council for Multi-Model Research Answer Aggregation

Abstract:Long-form research questions often require a single answer that combines details, reconciles conflicting estimates, and retains qualifications spread across several large language model (LLM) outputs. We present an inspectable LLM Council that separates synthesis into two stages: an analyst converts three independent candidate answers into a structured state, and a writer composes the final response from that state. We evaluate the workflow under closed-book conditions on 100 tasks from the Deep Research Accuracy, Completeness, and Objectivity (DRACO) benchmark, retaining 600 final answers, 100 analyst states, and 720 whole-answer judgments across the main and supplementary experiments. Under the primary rubric judge, Council scores 70.98, exceeding GPT and Gemini by 6.18 and 8.34 points; its 0.47-point difference from Claude has a 95% paired interval of $[-0.69,1.61]$. In the full supplementary comparison, Council scores 1.87 points above direct fusion (95% paired interval $[0.33,3.48]$) and covers 51.1% of positive rubric criteria covered by exactly one candidate, compared with 43.7% for direct fusion. On an exploratory 68-task subset with matching returned model labels, direct fusion scores higher, so the full comparison cannot isolate the analyst state's effect. Traced cases show how estimates, provenance qualifications, and selection decisions pass through the saved state into final answers.

Computation and Language (cs.CL); Artificial Intelligence (cs.AI)

I.2.7; I.2.6; I.2.11

arXiv:2604.02923 [cs.CL]

(or arXiv:2604.02923v5 [cs.CL] for this version)

https://doi.org/10.48550/arXiv.2604.02923 Focus to learn more arXiv-issued DOI via DataCite

THIRD-PARTY-NOTICES.md

analysis/diagnostics_v3.py

analysis/reproduce_v3.py

analysis/restore_records.py

analysis/supplement.py

analysis/test_diagnostics_v3.py

analysis/test_reproduce_v3.py

analysis/test_supplement_analysis.py

data/task_manifest.csv

data_records.json

data_supplement/analysis/all_calls.csv

data_supplement/analysis/costs.csv

data_supplement/analysis/coverage.csv

data_supplement/analysis/coverage_comparison.pdf

data_supplement/analysis/coverage_comparison.png

data_supplement/analysis/direction_agreement.csv

data_supplement/analysis/length_and_score.pdf

data_supplement/analysis/length_and_score.png

data_supplement/analysis/length_differences.csv

data_supplement/analysis/operations.csv

data_supplement/analysis/paired_differences.pdf

data_supplement/analysis/paired_differences.png

data_supplement/analysis/primary_comparisons.csv

data_supplement/analysis/primary_means.csv

data_supplement/analysis/returned_label_comparisons.csv

data_supplement/analysis/sample.csv

data_supplement/analysis/scores.csv

data_supplement/analysis/secondary_comparisons.csv

data_supplement/analysis/settings.csv

data_v3/analysis/configuration_scores.csv

data_v3/analysis/coverage_rates.csv

data_v3/analysis/dimension_contributions.csv

data_v3/analysis/paired_comparisons.csv

data_v3/analysis/retained_operations.csv

data_v3/analysis/subgroup_scores.csv

data_v3/analysis/task_differences.csv

data_v3/source/draco_test.jsonl

data_v3/stage1_solo_summary.csv

data_v3/stage3_scores.csv

experiments/records.py

experiments/stage1_report.py

experiments/stage1_solo.py

experiments/stage2_council.py

experiments/stage3_judge.py

experiments/stage3_report.py

experiments/supplement_api.py

experiments/supplement_finish.py

experiments/supplement_judgments.py

experiments/supplement_preflight.py

experiments/supplement_run.py

experiments/test_stage1_solo.py

experiments/test_stage2_council.py

experiments/test_stage3_judge.py

experiments/test_stage3_report.py

experiments/test_supplement.py

requirements.txt

(55 additional files not shown) You must enabled JavaScript to view entire file list.

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.