Which Direction Does the Evidence Point? Unbounded Beliefs with Many Alternatives

A classic result in social learning says that a line of people, each deciding after seeing what earlier people chose, eventually settles on the right answer only if private evidence is unbounded. That means someone, occasionally, gets evidence strong enough to overrule any crowd. Every version of this result in the literature we reviewed is binary: two states of the world, two actions. With ten answer options, "strong enough evidence" stops being a single number. Evidence can be overwhelming against one wrong option and nearly useless against another. We propose that the correct many-option version of the unbounded-beliefs condition is directional. Learning succeeds only if the true state can be separated, with arbitrarily strong evidence, from every group of wrong states that point to the same action. Wrong herds therefore persist exactly along the weakest direction of the evidence, not wherever an agent's confidence is lowest.

The claim can be tested with language models. Take a chain of language-model agents answering ten-option MMLU-Pro questions, each shown the answers of the agents before it. For the same question, seeding a wrong herd on a distractor the model barely separates from the truth should produce a lasting cascade. Seeding it on a distractor the model easily rejects should collapse, even though the model's overall confidence in the correct answer is identical in both cases. A single-number account of evidence strength cannot predict that difference. If the directional account can, it gives both a new theorem shape and a cheap diagnostic for when sequential multi-agent systems will lock onto wrong answers.

Core Claim

Let the world be in one of $M$ states $\theta \in \{1,\dots,M\}$. Each agent sees a private signal $s$ drawn from density $f_\theta$ and picks one action from a finite set. Write $\ell_{\theta\theta'}(s) = \log f_\theta(s)/f_{\theta'}(s)$ for the pairwise log-likelihood ratio, the evidence $s$ gives for $\theta$ over $\theta'$. Call a set $W$ of wrong states action-sharing if some single action is optimal whenever belief is concentrated on $W$. We conjecture that asymptotic learning holds under full observation of history if and only if, for every true $\theta$ and every action-sharing set $W \not\ni \theta$,

$$ P_\theta\Big(\min_{\theta' \in W} \ell_{\theta\theta'}(s) > t\Big) > 0 \quad \text{for every } t, $$

so that some signal beats all of $W$ at once by any margin. The binary theorem in Preliminary results on social learning with partial observations, which ties asymptotic learning to unbounded beliefs and gives a $1-\beta$ accuracy ceiling under bounded signals, is the special case $M=2$. There, the condition reduces to one tail.

The genuinely many-option question is whether pairwise unboundedness suffices. Pairwise unboundedness means each wrong state can be beaten separately. The joint condition asks for signals that beat several wrong states at once. The two differ when several wrong states share an action. A related consequence of the conjecture concerns hedge actions, such as an "undetermined" option that is best at mixed beliefs. When one pair is weakly separated, a hedge action should widen the set of beliefs where cascades form.

Research Program

The work has three parts.

The first part is theory. We will prove or refute the conjecture using the martingale-convergence argument behind the binary theorem, carried out on the belief simplex rather than an interval. Beliefs in Decision-Making Cascades is a caution here. It studies binary Gaussian cascades with deliberately distorted priors, and it states that the recursive belief updates explode for $M \ge 3$. We avoid tracking the full recursion. We only characterize which limiting public beliefs are stable, which reduces the problem to the tail condition above.

Social Learning and Distributed Hypothesis Testing supplies the key intuition. In a simultaneous, truthful network of $M$ hypotheses, that paper shows the learning rate is a centrality-weighted sum of pairwise divergences. We carry the same pairwise view into the sequential, self-interested setting it does not cover, where rational agents copy predecessors. We keep a finite action set and full visibility of history. The $K \ge M$ memory conjecture of On Learning with Finite Memory is out of scope.

The second part is measurement. We treat a language model's answer distribution under resampling as its private signal. Resampling means varying option order, paraphrase, and sampling seed. For each question with correct option $c$ and each distractor $j$, we estimate how the pairwise log-ratio of $c$ over $j$ is distributed across these draws. The quantity of interest is its upper tail: how often a draw gives a large margin for $c$ over $j$. This yields a "directional resistance profile" for each question-distractor pair.

We then run sequential chains. Each agent is shown the answers of the previous $k$ agents. We seed the first few with a chosen wrong option and record whether later agents recover the correct answer.

The third part is intervention. The theory predicts where cascades persist, so a revealer can be targeted. A revealer is an agent that ignores history and answers from its own evidence. We deploy revealers only on question-distractor pairs with weak directional tails, and compare this against the untargeted decaying schedule $p_t = c/t$. That schedule is shown optimal in the binary model of How fragile are information cascades?.

Recent LLM-cascade work motivates the empirical setting. Steering the Herd models a mediator that steers binary cascades by tuning signal precision, and reports that language-model agents underreact to confirming information and overreact to surprise. Bandit Social Learning under Myopic Behavior proves permanent-failure bounds for two arms and flags that three or more arms may multiply the number of traps. Our directional condition is a concrete hypothesis about which traps matter.

Algorithmic Persuasion with Evidence constrains the design. It shows that approximation guarantees for persuasion break when the receiver has more than two actions. We therefore do not expect a pairwise tournament reduction to be sound, and we test that reduction explicitly rather than assume it.

Baselines

The naive reference point is the single-agent, no-history answer. For open 7–70B models run under the same prompt, we measure MMLU-Pro accuracy directly. The strongest external anchor the file reports for that setting is Llama-3-70B-Instruct at 56.2% accuracy on the ten-option benchmark (MMLU-Pro). A sequential chain "beats doing nothing" only if its final accuracy exceeds this solo accuracy on the same questions.

The strongest aggregation competitor removes visibility entirely. Agents answer independently and a vote is taken, or they draft in isolation before sharing, as in All-Agents Drafting from Voting or Consensus? Decision-Making in Multi-Agent Debate. That paper reports 51.7% average voting accuracy on MMLU against 54.0% for consensus, and a 3.3% gain from All-Agents Drafting. We will rerun voting and All-Agents Drafting with the same models, the same number of agents, and the same token budget as our chains. We found no reported result on Lacuna for sequential-visibility cascades on MMLU-Pro, so we establish that number ourselves.

For the predictive claim, the main competitor is a scalar predictor of whether a chain recovers from a seeded wrong herd. We try two scalar features: the model's solo probability on the correct option, and the upper tail of its maximum confidence. The ablation that isolates our key ingredient replaces these with the directional tail against the seeded distractor, holding everything else fixed. It must raise held-out prediction quality, measured by area under the ROC curve, by a margin exceeding bootstrap noise.

For the intervention, the competitor is random revealers with $p_t = c/t$, given the same expected number of reveals. Targeted revealers win if they reach equal final accuracy with fewer reveals, or higher accuracy at equal reveals.

Why not just…?

Why not just reduce ten options to binary comparisons and apply the binary theorem at each node? This is the most natural competitor, and we test it directly. The binary reduction treats "correct versus herded option" as a two-state problem. Its prediction is that recovery depends only on the correct-versus-herd tail. Our set-wise condition departs from it when several wrong options share an action, or when a hedge option exists. We build synthetic Bayesian simulations where pairwise tails are unbounded but joint tails are bounded. If cascades there vanish as the pairwise reduction predicts, our joint condition is wrong and we will state the pairwise version as the result. The persuasion hardness results in Algorithmic Persuasion with Evidence suggest composition fails, but they concern a different problem and do not settle this one.

Why not just use the model's solo accuracy or confidence to predict which chains fail? This is a real threat for language models. Easy questions may simply resist herds on every distractor. The seeded-distractor design is the decisive test. The same question is seeded with its most-confusable and its least-confusable distractor, so solo confidence in the correct answer is constant within each pair of runs. If recovery rates do not differ across these seeds, the directional account adds nothing for language-model agents, and we will report that.

Why not just prompt agents to ignore predecessors, or hide the history? Prompting is weak evidence of a fix. FRAGILE found value-anchoring instructions largely failed to stop decision flips, with a 28.6% overall flip rate, and located framing divergence in the network's late layers. ManagerBench found that a single goal-pressure sentence dropped safety performance by over 50 points. Hiding history works, but it discards a signal that real deployments expose. It is also exactly the voting baseline we run. Our claim is that cheap, targeted independence recovers most of voting's accuracy while keeping visibility, and the comparison above measures whether it does.

Proposed Experiments

The experiments begin with synthetic Bayesian cascades. We use $M = 3$ to $10$ states with discrete signal distributions engineered so that chosen pairs, or chosen sets of states, have bounded evidence. Agents are exact Bayesians. We measure the limiting frequency of wrong cascades and which action they lock onto, and check each outcome against the set-wise condition, the pairwise condition, and the binary reduction. Adding a hedge action tests the predicted widening of the region where cascades form.

We then move to language-model chains on MMLU-Pro. For each question we compute directional resistance profiles and run chains of 20 agents seeded with 3 wrong answers. Each question is seeded twice, on its most- and least-separated distractors, and runs cover several open models. The decisive check has two parts. First, the gap in recovery between the two seeds must be large and consistent in sign. Second, on held-out questions, the directional tail must predict recovery better than both scalar predictors.

The final experiment compares the targeted revealer schedule against $p_t = c/t$, independent voting, and All-Agents Drafting at matched budgets. We report final chain accuracy against solo accuracy and against the reported 56.2% single-model anchor. As a further check, we use a synthetic three-state, two-action variant to see whether cascades onto a shared action require joint rather than pairwise weakness. That is the one place where the many-option theory says something the binary theory cannot.