Measuring Collapse and Correction in Homogeneous-Panel LLM Debate
Paper • 2609.35279 • Published • 5
Request access to the DebateLedger gated tier
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This tier contains the full model-generated text of multi-agent debates and probe responses, together with the convince-wrong and social-pressure probe templates. Some responses argue for answers the model was instructed or pressured to adopt, including wrong ones, so the material can be reused to mislead. Access is granted for research on the evaluation, robustness and safety of language models and multi-agent systems, under the DebateLedger Gated Tier Research Use License in the LICENSE file of this repository.
Log in or Sign Up to review the conditions and access this dataset content.
DebateLedger (gated tier)
Measuring Collapse and Correction in Homogeneous-Panel LLM Debate · NeurIPS 2026, Evaluations and Datasets Track
Xin Li*, Mengbing Liu*, Chau Yuen · Nanyang Technological University · *Equal contribution
Project page · arXiv · OpenReview · Code (GitHub) · Open tier
Contents
This tier holds only what the open tier removes. Files use repository-relative paths.
abc_exp/results/*.jsonl(57 files): the same traces as the open tier with the model-generated text kept. In debate traces (debate_traces_*,smoke_*) thereasoningfield holds each agent's full response in every round; in three Gemini probe traces (sa_causal_gemini_*) theinitial_response_prefixfield holds the start of the probe response. Traces without model text are released only in the open tier. The primary six-model debate cohort was recorded with per-round answers only; its traces and the Round-1 feature matrix contain no model text and are in the open tier.abc_exp/config/prompts.py,abc_exp/scripts/run_alpha_probe_openrouter.py,abc_exp/scripts/run_alpha_probe_vllm.py: the unredacted code files, including the very-strong (convince-wrong) probe template and the social-pressure suffix, which the public code replaces with a placeholder.
Access
The authors review each request against the research-use terms and aim to respond within ten business days. A declined request receives a short reason and can be resubmitted once the scope or license issue is fixed.
Use with the code
Download the open tier first and this tier second, both into the root of a clone of the code repository. The gated files replace their open-tier and placeholder counterparts:
git clone https://github.com/LiXin97/DebateLedger && cd DebateLedger
pip install -U huggingface_hub
hf auth login
hf download XINLI1997/DebateLedger --repo-type dataset --local-dir .
hf download XINLI1997/DebateLedger-gated --repo-type dataset --local-dir .
With both tiers in place, the parser audits that re-read model text (abc_exp/experiments/parser_cell_stability_audit.py and the final-answer-tag counts in abc_exp/experiments/n14_reviewer_hardening_artifacts.py) can be rerun.
Why this tier is gated
The debate responses include arguments written for answers that the model later abandoned or was pushed toward by peers, and the probe templates are designed to make a model change a correct answer. Released openly, the text and templates could serve as a ready-made corpus for misleading models or people. The data contain no personal information; the questions come from public academic benchmarks.
Terms
Use of this tier is governed by the DebateLedger Gated Tier Research Use License (LICENSE). Questions come from public benchmarks (MMLU-Pro, GPQA, GSM8K, ARC-Challenge, TruthfulQA) and remain subject to their licenses. Model responses were generated with third-party models and APIs, and users are responsible for complying with those providers' terms. No model weights are included.
Citation
@inproceedings{
li2026debateledger,
title={Measuring Collapse and Correction in Homogeneous-Panel {LLM} Debate},
author={Xin Li and Mengbing Liu and Chau Yuen},
booktitle={The Fortieth Annual Conference on Neural Information Processing Systems Evaluations and Datasets Track},
year={2026}
}
- Downloads last month
- 10