Everyday social norms are rules of thumb with exceptions. "Setting a fire is wrong" holds until you learn the person is camping in the cold. Then it may flip back if the campsite is in dry grass during a burn ban. Reasoning of this kind is called non-monotonic, or defeasible: adding a fact can overturn a conclusion you had already reached.
Multi-agent debate means several language-model instances argue over an answer for a few rounds before settling on one. It is increasingly proposed for exactly these judgments. But debate has two documented weaknesses. Agents in an initial majority almost never move, and agents tend to concede to confident peers. In a domain full of exceptions, both weaknesses push the same way. The default answer is the one models have memorized from web text, so it tends to form the majority, and the agent that noticed the exception is the one that gets talked out of it.
This note proposes to test that prediction directly and to fix it if it holds. The test is a procedurally generated benchmark of social scenarios with layered exceptions. For each scenario we control how deep the chain of exceptions goes and how familiar the surface wording is, meaning how likely the scenario is to resemble text a model has seen in training. The fix is a debate protocol in which agents argue about which exceptions apply rather than about the final verdict. The verdict is then computed from the accepted exceptions instead of by vote. The core claim is that ordinary debate helps on defeasible social reasoning only for shallow, unfamiliar items, and actively hurts when the default is a memorized platitude. Separating "finding exceptions" from "deciding the verdict" removes that harm.
What exists and what is missing
The pieces exist separately, and none of them answers this question.
BoardgameQA generates defeasible problems from proof trees with rule priorities. It shows single models collapse as conflicts deepen: fine-tuned T5 XXL reaches about 90% at depth 1 and falls to about 45–50% at depth 3. It also finds an "Unknown bias," a reluctance to say a conflict is unresolved. Its rules are formal ("Rule 2 is preferred over Rule 1") rather than social norms, it uses no debate, and it is a static release. We reuse its proof-tree construction but render it as social narrative, and we add a familiarity control.
Can LLM Agents Really Debate? runs debate on synthetic Knight–Knave–Spy puzzles to avoid contamination. Debate lifts size-4 accuracy from 17.33% to 69.33%, and agents in an initial majority hold their stance in over 97% of cases. Those puzzles are monotonic: no added fact ever overturns a conclusion. That paper motivates our worry about majority stickiness but cannot reveal what happens when the majority is the memorized default.
DyVal builds evaluation items dynamically from directed acyclic graphs. It shows models that claimed strong static scores falling to near zero, and GPT-4 losing 13.6% under reversed presentation order, but it is confined to formal tasks. A Cognitive Battery for Foundation Models generates items procedurally, including a "Dynamic Rule Override" task that tests whether a model can suppress a highly predictable completion. It evaluates single models only and defers its scores.
On the social side, three static datasets are the obvious sources and are presumably exposed to pretraining: δ-ROT, SCRUPLES and Social IQa. δ-ROT supplies something we need: an empirical finding that moral shifts are driven mainly by intent, participant role and vulnerability. We use those three as the axes along which exceptions are generated. The debate work that does touch social questions measures process quality without verifiable ground truth: A Diagnostic Study of Multi-Agent LLMs for Real-World Debates uses policy simulations, and A Checks-and-Balances Framework uses a fixed two-agent dyad over static corpora. Neither can say whether a debate got an exception right.
Core Claim
We make two predictions about ordinary debate. Under plain debate, the rate at which a correct non-default verdict is lost during deliberation rises with surface familiarity, holding the logical structure fixed. Perplexity-based pruning of arguments makes this worse. That pruning comes from Multi-Agent Debate with Memory Masking, which removes high-perplexity responses between rounds and gains 9 points on MATH. The reason it should backfire here is that a correct exception argument is, almost by definition, the less predictable text.
We also predict a fix. A protocol in which agents jointly maintain a ledger of candidate exceptions, attack or support each one, and leave the verdict to a deterministic resolver should eliminate most of the familiarity gap. A ledger-based protocol should gain over self-consistency at matched compute only at exception depth two and beyond. If a single agent filling the same ledger matches the multi-agent version, the honest conclusion is that the gain comes from structure, not from debate. The experiments are designed so that this outcome is reportable rather than hidden.
Research Program
The generator. Each item starts from a small defeasible theory. A default norm (for example, "taking someone's medication without asking is wrong") is followed by a chain of $k$ exceptions. Each exception overturns the one before it and is drawn from the intent, role and vulnerability dimensions. The depth $k$ runs from 0 to 3.
Two further item types sit alongside the main chain. Distractor items contain a detail that looks like an exception but does not bear on the norm, such as a role that is mentioned but irrelevant. Undetermined items contain two exceptions of comparable weight pointing in opposite directions, so the correct answer is "it depends."
An LLM renders the theory into a short narrative. A different model parses the narrative back into the theory, and items where the parse disagrees are discarded. Crowd annotators then label a stratified sample. Templates where human majority judgment disagrees with the formal label are dropped. Following SCRUPLES, we treat residual human disagreement as signal rather than noise: low-agreement templates go to a separate split scored against the human label distribution.
The familiarity manipulation. We use the same theory skeleton with two kinds of seed action. Familiar seeds are actions drawn from δ-ROT, SCRUPLES and Social IQa. Novel seeds are invented practices with invented but internally consistent norms, in the spirit of the battery's alien-grammar items. Familiar and novel items therefore share structure exactly and differ only in surface wording.
As a continuous measure, we also score each item's default phrasing by n-gram frequency in FineWeb. A fresh draw of items can be produced for every evaluation, so the benchmark itself cannot be memorized.
The defeater-ledger debate. In each round, every agent must post entries to a shared ledger. An entry either names a specific sentence of the scenario as an exception to a specific prior rule, or attacks an existing entry. An attack can claim the cited detail is irrelevant, outweighed, or itself overturned.
Agents never vote on the verdict. After a fixed number of rounds, which we cap at three because Peacemaker or Troublemaker finds sycophantic decay outpacing gains beyond three, each entry is kept or rejected by a majority of agents. The verdict is then computed from the surviving exception chain by standard argumentation semantics. An unresolved tie yields "undetermined."
The design intent is that peer pressure can still act on individual entries, but there is no single verdict-level stance for a majority to sit on. Removing a correct exception requires an explicit, inspectable attack.
Baselines
The naive reference point is a single model with chain-of-thought prompting. It establishes the depth curve and the familiarity gap for each model. We expect both effects to appear already here, consistent with BoardgameQA's report that PaLM 540B with chain-of-thought fails out of the box on prioritized conflicts.
The compute-matched ensemble is self-consistency: $N$ independent chain-of-thought samples with a majority vote, where $N$ is chosen so total generated tokens equal those of the debate conditions. This is the strongest cheap competitor. Pitre et al. report that high-process debates recover from an initially wrong majority 74% of the time, whereas voting by construction cannot.
Plain debate follows FORD, the strongest established protocol in the file on social commonsense. FORD stages stance selection, refutation and concession, then a judge, and reaches 81.63% average accuracy across seven commonsense benchmarks with ChatGPT and Davinci-003.
The strongest recent method is memory-masked debate from Multi-Agent Debate with Memory Masking, in both its subjective form (agents mark peers' answers correct or incorrect) and its objective form (perplexity pruning). It reports MATH accuracy rising from 71.2% to 80.2% with Qwen2.5-Math-7B and a 6.6% gain on AIME. We also run the balanced peacemaker/troublemaker persona mix with a central judge from Yao et al., which cut disagreement collapse on CommonsenseQA from about 81% to 41%. For all debate methods on defeasible social reasoning, there is no reported result found on Lacuna, so their numbers here will be new.
The key ablation is the single-agent ledger. One model fills the same exception ledger over the same number of passes, and the same resolver computes the verdict. A second ablation, the multi-agent ledger with a vote on the verdict instead of the resolver, isolates whether removing the verdict-level vote is what matters.
All conditions use Qwen2.5-7B, Qwen3-32B and LLaMA 3.3-70B, the models used in the cited debate studies, on identical item draws. Each condition gets the same prompt-tuning budget on a separate development draw and the same token ceiling.
The ledger protocol counts as beating the field under three conditions together. First, on depth 2–3 items it exceeds compute-matched self-consistency and masked debate by a margin outside seed variance. Second, it reduces the familiar-minus-novel accuracy gap relative to plain debate. Third, it does not raise the rate of accepting distractor exceptions.
Why not just…?
Why not just use self-consistency voting, which is cheaper and immune to sycophancy? Voting aggregates the initial distribution of answers. If that distribution is dominated by a memorized default, voting returns the default. That is the specific failure we hypothesize, so voting is the primary compute-matched comparison rather than a straw man. If voting matches the ledger protocol on deep, familiar items, then nothing about debate is needed, and we will report it that way.
Why not just have one model write out the exceptions and run the same solver? This is the real threat. The single-agent ledger ablation exists to settle it. BoardgameQA's depth collapse suggests a single model misses exceptions it never considers. Debate's plausible contribution is recall: different agents surfacing different candidate exceptions. We will measure ledger recall of the true exception chain directly. If single-agent recall is equal, the paper's contribution narrows to the benchmark, the familiarity finding, and the structured-resolution result, and the claim about debate is withdrawn.
Why not just evaluate on a held-out or paraphrased split of SCRUPLES or δ-ROT? Paraphrase changes wording but keeps the scenario, and its memorized default, recognizable. It also gives no control over exception depth or verifiable ground truth. DyVal's near-zero scores for models that excelled on static sets show how much static scores can hide. Our familiar-versus-novel contrast at fixed structure measures the very effect paraphrasing cannot separate. We will still report on static SCRUPLES and Social IQa as an external sanity check.
Why not just fine-tune a small specialist, as δ-ROT did? The δ-ROT student scores 0.86 automatic validity against GPT-4's 0.77. That figure measures generating plausible single-step exceptions, not resolving chains of them. We include a model fine-tuned on δ-ROT as a single-agent baseline on depth-1 items. If it matches the ledger protocol at depth 2–3 on novel items, the multi-agent machinery is unjustified.
Proposed Experiments
The first decisive check is the familiarity gap. We define it as
$$\Delta_m = \mathrm{Acc}_m(\text{novel}) - \mathrm{Acc}_m(\text{familiar}),$$
computed per method $m$ on items whose correct verdict is non-default, at each depth. The prediction is that $\Delta$ is larger for plain debate than for single-model chain-of-thought, and larger still for objective masking. Regressing per-item correctness on FineWeb n-gram frequency should show the same pattern continuously. If $\Delta$ is near zero everywhere, the contamination story is wrong and we say so.
The second check is a defeater-retention analysis. We restrict to debates where at least one agent starts with the correct non-default verdict while the majority holds the default. Within that set, we report the fraction ending correct, alongside Pitre et al.'s engagement and responsiveness measures. For the masked variants, we also compare the perplexity of correct exception arguments with that of default arguments. A higher perplexity for correct exceptions would confirm that perplexity pruning censors exactly the arguments that matter.
The third check is the head-to-head at matched tokens across depths 0–3, plus the two ablations above. We report accuracy, the undetermined-item F1 (testing whether debate relieves BoardgameQA's Unknown bias), the distractor-acceptance rate, and distributional fit on the low-agreement split.
A protocol that only helps by accepting more exceptions would show up immediately as a higher distractor-acceptance rate. That over-correction is the failure we most want to rule out.
References
- BoardgameQA: A Dataset for Natural Language Reasoning with Contradictory Information
- Can LLM Agents Really Debate? A Controlled Study of Multi-Agent Debate in Logical Reasoning
- DyVal: Dynamic Evaluation of Large Language Models for Reasoning Tasks
- A Cognitive Battery for Foundation Models: Theory-Grounded Benchmarks for Attention, Learning, Metacognition, Executive Function, and Social Cognition
- What Makes it Ok to Set a Fire? Iterative Self-distillation of Contexts and Rationales for Disambiguating Defeasible Social and Moral Situations
- SCRUPLES: A Corpus of Community Ethical Judgments on 32, 000 Real-Life Anecdotes
- Social IQa: Commonsense Reasoning about Social Interactions
- A Diagnostic Study of Multi-Agent LLMs for Real-World Debates
- A Checks-and-Balances Framework for Context-Aware Ethical AI Alignment
- Multi-Agent Debate with Memory Masking
- Peacemaker or Troublemaker: How Sycophancy Shapes Multi-Agent Debate
- Examining Inter-Consistency of Large Language Models Collaboration: An In-depth Analysis via Debate