| --- |
| language: |
| - en |
| license: cc-by-4.0 |
| pretty_name: DebateLedger |
| tags: |
| - multi-agent-debate |
| - llm-evaluation |
| - mmlu-pro |
| - collapse |
| - correction |
| viewer: false |
| task_categories: |
| - question-answering |
| --- |
| |
| # DebateLedger |
|
|
| **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](https://lixin.ai/DebateLedger/) · [arXiv](https://arxiv.org/abs/2609.35279) · [OpenReview](https://openreview.net/forum?id=E8FfL8c7XE) · [Code (GitHub)](https://github.com/LiXin97/DebateLedger) · [Gated tier](https://huggingface.co/datasets/XINLI1997/DebateLedger-gated) |
| |
| This is the open data tier of DebateLedger: probe and debate traces for |
| evaluating multi-agent LLM debate by its transitions. Three copies of one model |
| answer a multiple-choice question and debate for three rounds; each run is |
| recorded with the answers of every agent in every round, so collapses (a correct |
| initial majority that ends wrong) and corrections (a wrong initial majority that |
| ends correct) can be counted and interventions replayed on the saved debates. |
| The primary cohort has 6,925 MMLU-Pro debates with 253 collapses. |
| |
| ## Files |
| |
| Files use repository-relative paths. Downloading this dataset into the root of |
| a clone of the [code repository](https://github.com/LiXin97/DebateLedger) |
| places them where the rebuild scripts expect them: |
| |
| ```bash |
| git clone https://github.com/LiXin97/DebateLedger && cd DebateLedger |
| pip install -U huggingface_hub |
| hf download XINLI1997/DebateLedger --repo-type dataset --local-dir . |
| ``` |
| |
| | Files (`abc_exp/results/`) | Content | Records | |
| |---|---|---:| |
| | `debate_traces_{gemini_3-flash, openai_gpt-5.4-mini, vllm_llama-3.1-8b, vllm_phi-4-mini, vllm_qwen3-4b, vllm_qwen3-8b}.jsonl` | Primary MMLU-Pro debate cohort (six models), per-round answers | 6,925 debates | |
| | `per_debate_r1_features.jsonl` | Round-1 feature matrix derived from the primary cohort | 6,925 debates | |
| | other `debate_traces_*.jsonl` (50 files) | Extension and response-period debates: further models, reasoning-mode and private-revision checks, mixed-model panels, GSM8K | 7,636 debates | |
| | `sa_causal_*.jsonl` (44 files) | 8-probe screen: one record per model, question and agent with the initial answer and the eight probe replies | 35,664 records | |
| | `cross_benchmark_*.jsonl` (16 files) | GPQA, TruthfulQA and ARC-Challenge stress checks | 1,033 debates | |
| | `block0_*`, `block1_*`, `block3_*` | Early pilot blocks | 1,705 records | |
| | `r5_gemini_alpha_panel_*.jsonl` | Three-day closed-API repeatability panel | 190 records | |
| | `gate_rerun/*.jsonl` (10 files) | Probe scores of Llama-3.1-8B, Phi-4-mini, Qwen3-4B and Qwen3-8B collected again for the probe-gated freeze and the channel panel: default condition on the debate questions, all three conditions on the 200-question pool | 24,335 records | |
|
|
| Also included: the aggregate tables and audits behind the paper |
| (`abc_exp/results/*.json|md|csv|tex`), the datasheet |
| `abc_exp/results/DEBATE_EVAL_ARTIFACT_DATASHEET.md`, `croissant.json` (Croissant |
| and Responsible AI metadata with SHA-256 digests of the core files and record sets for |
| the trace files), and the |
| reproducibility card `README_REPRO.md`. |
|
|
| ## Record fields |
|
|
| Primary-cohort debate traces, one JSON object per debate: |
|
|
| | Field | Meaning | |
| |---|---| |
| | `question_id`, `correct_label` | MMLU-Pro question and its correct option letter | |
| | `initial_answers`, `initial_majority`, `initial_correct` | The three agents' answers before debate, their majority, and whether it is correct | |
| | `round_traces` | One entry per debate round: `answers`, `majority`, `majority_changed` | |
| | `final_answers`, `final_answer`, `final_correct` | Answers and majority after the last round | |
| | `collapsed`, `corrected` | Transition labels: correct to wrong, wrong to correct | |
| | `agent_flips`, `n_agent_flips`, `total_cost`, `model` | Answer changes per agent, API cost, model identifier | |
|
|
| Newer debate traces store the same information per agent (`initial_states`, |
| `round_traces` with `agent_id`, `answer` and `round_num`) and the labels under |
| `outcome` (`collapse`, `correction`, `agent_outcomes`), together with backend, |
| temperature and timestamp metadata. Probe traces (`sa_causal_*`) hold |
| `initial_answer`, `initial_correct`, the eight entries of `probe_results` |
| (`strength`, `social`, `alt_answer`, `post_answer`, `revised`, token usage and |
| cost) and the derived flip rates (`alpha_total`, `alpha_social`, `alpha_solo`). |
|
|
| ## What was removed |
|
|
| Model-generated text is not part of this tier. The `reasoning` fields (agent |
| responses in each debate round) and the `initial_response_prefix` field (probe |
| responses) are set to `null`. Parsed answers, correctness labels, round |
| structure, token counts and costs are unchanged, which is enough to rebuild the |
| transition tables, the family-level screen analysis and the replay summaries. |
| The primary cohort was recorded with answers only, so nothing was removed from |
| it. The full text of the 57 files that contained it, of the thinking-off rerun, and the very-strong |
| (convince-wrong) and social-pressure probe templates, are in the |
| [gated tier](https://huggingface.co/datasets/XINLI1997/DebateLedger-gated). |
|
|
| In the gate-rerun probe files (`abc_exp/results/gate_rerun/`) the model output (`initial_text` and `probe_results[].text`) is `null`; that text is not released in either tier. |
|
|
| No model weights are included. |
|
|
| ## Versions |
|
|
| - `alpha-tot-v1.0.4`: adds the gate-rerun probe scores (`abc_exp/results/gate_rerun/`) and the results of the probe-gated freeze and the 22-model channel panel. Earlier traces are unchanged. |
| - `alpha-tot-v1.0.3`: adds the N=9 sensitivity results. Traces are unchanged. |
| - `alpha-tot-v1.0.2`: aggregate results now read every row at the paper's endpoint, Round 3 majority with alphabetic ties. |
| Traces are unchanged: rows with a logged fourth round still store its majority in `majority_answer`, and |
| `abc_exp/r3_unified/rebuild_r3.sh` recodes a working copy before rebuilding. Adds the thinking-off rerun of the four |
| Qwen3.5/3.6 rows in `abc_exp/results/thinking_off_rerun/`. |
| - `alpha-tot-v1.0.1` (metadata only): `croissant.json` describes the trace files record by record |
| (debate traces, 8-probe screen traces, cross-benchmark rows, probe-panel rows, and the |
| transcript-free Round 1 feature matrix). The data files are the same as in `alpha-tot-v1.0`. |
| - `alpha-tot-v1.0`: first public release (2026-09-28). |
|
|
| ## Citation |
|
|
| ```bibtex |
| @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} |
| } |
| ``` |
|
|
| ## License |
|
|
| CC BY 4.0. Questions come from public benchmarks (MMLU-Pro, GPQA, GSM8K, |
| ARC-Challenge, TruthfulQA) and remain subject to their licenses. |
