Papers you can read. Results you can check.
Research on evidence and verification in multi-agent systems.
Manuscripts, formal proofs, code, and the recorded data behind the results.
Paper map · Supplement · Reproduce the results · Cite · Contribute
Overturning Wrong Majorities in Multi-Agent LLMs · November 2026
When agents share a mistake, their majority can be wrong. This paper studies a rule that lets checked passage evidence overturn the vote, while recording which evidence licensed each change.
Verification improves accuracy on the evaluated benchmark, but the original scored rule misses its pre-specified harm criterion. The formal result identifies an inspectable failure condition; it does not guarantee a low error rate. The paper reports the gains, failures, and tradeoff against a stronger model judge together.
| Start here | What you get |
|---|---|
| Read the paper | The rule, main results, controls, and limitations. |
| Supporting material | Methods, proofs, pooled-bound construction, and full replay tests. |
| Reproduce the results | Commands and pinned toolchains for the code, analyses, and formal proofs. |
| Source and citation | LaTeX build instructions, artifact map, and BibTeX. |
Each paper lives alongside the material needed to inspect its claims:
- Read the argument in the main paper and the supporting methods in its supplement.
- Inspect the study protocols, recorded inputs, checker labels, and reported outputs.
- Reproduce the analyses and run the implementation and proof checks using the artifact guide.
The paper map collects abstracts and links. catalogue.yaml provides the machine-readable index.
With Tectonic installed, or latexmk with pdfLaTeX, run from the repository root:
make -C papers/verify-then-vote/tex
This builds the main paper and supplement. Add the arxiv target to create a source bundle containing both standalone documents. Code, data, and proof checks have separate requirements; follow the artifact guide.
The collection is being prepared for its first public release. Study freezes record when analyses were specified; they are not earlier publications. After publication, tagged releases and the history will record corrections and revisions.
For a correction or reproduction report, open an issue with the paper, section, claim, and supporting evidence or commands. See CONTRIBUTING.md for the contribution and release process.
| Material | Terms |
|---|---|
| Code | Apache License 2.0 |
| Manuscripts and figures | CC BY 4.0 |
| Data we produced | CC BY 4.0 unless the artifact's DATA_LICENSES.md states otherwise |
| Third-party data and model outputs | Their own dataset or provider terms; EXAONE outputs are not under CC BY 4.0 |
The artifact's data and component licenses take precedence over this summary.