Conformity Mitigations in Large Language Models Lie on a Single Resistance-Receptivity Frontier

# Conformity Mitigations in Large Language Models Lie on a Single Resistance–Receptivity Frontier
Source: [https://arxiv.org/html/2608.11247](https://arxiv.org/html/2608.11247)
Zafar Hussain Aarhus University zafar@cas\.au\.dk &Kristoffer Nielbo Aarhus University kln@cas\.au\.dk

###### Abstract

Recent advances in language models have enabled collaborative settings in which multiple models leverage one another’s capabilities, iteratively improving, transforming, and extending each other’s outputs\. Each agent sees what the others assert before it answers, so peer opinion competes with the model’s own parametric knowledge, and a wrong majority can overturn an answer the model would otherwise get right\. We measure that displacement in 23 open\-weight models, 19 conditions, and three datasets, yielding more than a million graded responses\. A unanimous wrong majority reverses 22\.8% of a model’s correct MMLU answers, 54\.8% on GPQA and 71\.0% on SimpleQA, and 84–89% of the reversed answers match the peers’ answers\. Existing mitigations aim to increase Resistance, the rate at which a model keeps its correct answer under this pressure, which is only half of what a collaborating agent needs\. We pair it with Receptivity, the rate at which a model adopts a correct peer answer after initially answering incorrectly\. We score six methods on both axes, four drawn from prior work and two of our own\. Each gains Resistance only by losing Receptivity, and their means fall on a single Resistance–Receptivity frontier withR2R^\{2\}between 0\.80 and 0\.90\. Reflection, the strongest published method, gains 7\.9 points of MMLU Resistance and gives up 15\.3 of Receptivity\. Reasoning is the one exception\. On GPQA and SimpleQA it trades like the rest, but on the MMLU subjects whose answers a model can derive for itself it raises Resistance by 7\.2 points and Receptivity by 9\.6 at once, the only intervention we find that improves both\.

Conformity Mitigations in Large Language Models Lie on a Single Resistance–Receptivity Frontier

Zafar HussainAarhus Universityzafar@cas\.au\.dkKristoffer NielboAarhus Universitykln@cas\.au\.dk

## 1Introduction

Multi\-agent systems rely on interactions in which language models assess and refine one another’s responses\. While these interactions can improve collective reasoning, they also create opportunities for models to drift away from initially correct answers under pressure\. Asch documented the same susceptibility in people, who often abandon a correct judgment once a unanimous group contradicts them\(Asch,[1951](https://arxiv.org/html/2608.11247#bib.bib2)\)\. We study conformity in language models and ask what mitigation costs rather than just what it fixes\.

Our design changes nothing about a question except what the peers say about it\. Each model first answers alone, then answers again after being told that several peer models unanimously agree on a different answer\. Holding the peer script fixed across conditions keeps the pressure controlled and repeatable\. Averaged over 23 open\-weight models, a unanimous wrong group makes a model abandon its own correct answer on 22\.8% of broad\-knowledge multiple\-choice questions, and the rate rises to 71\.0% on free\-form factual recall\. The abandoned answers do not scatter over the remaining options but concentrate on the one named by the group \(Section[5\.1](https://arxiv.org/html/2608.11247#S5.SS1)\), so peer pressure alters the answer itself rather than confidence in it alone \(Figure[1](https://arxiv.org/html/2608.11247#S1.F1)a\)\. A single miscalibrated agent can, therefore, pull a whole group off a correct answer and leave that group less reliable than any of its members\.

![Refer to caption](https://arxiv.org/html/2608.11247v1/results/plots/fig0_overview.png)Figure 1:Two verbatim exchanges \(gemma\-3\-27b\-it, MMLU\)\. \(a\) Unanimous wrong peers turn a correct answer into a wrong one\. \(b\) Unanimous correct peers turn a wrong answer into a correct one\.Two recent studies established the phenomenon in language models and proposed the first mitigations\(Zhuet al\.,[2025](https://arxiv.org/html/2608.11247#bib.bib1); Wenget al\.,[2025](https://arxiv.org/html/2608.11247#bib.bib18)\)\. We reproduce both results and test every mitigation that reduces conformity to incorrect peers\. Conformity is only half of what happens when a model meets its peers\. A model that never yields never adopts a wrong answer, and it never takes a right one either\. Figure[1](https://arxiv.org/html/2608.11247#S1.F1)b shows what a conformity rate misses, a question the model gets wrong on its own and corrects once the peers agree on the right answer\. In a system built so that agents catch each other’s errors, losing that correction is a cost\. We measure*Receptivity*, the rate at which a model accepts a correct peer answer after an initial mistake, alongside*Resistance*, the rate at which it keeps a correct answer under incorrect peer pressure\. Each method in this paper is scored on both\.

Scoring both axes places a method on a plane rather than a line, and that plane turns out to be nearly one\-dimensional\. All six methods we test fall on a single downward\-sloping line across all datasets, in the pooled average, and within almost every model taken alone\. A model facing a unanimous group has only its own answer and the peers’ to work with, so an instruction can shift which of the two it leans toward but cannot make that leaning selective, and favoring its own answer rejects wrong peers and right ones alike\. The line binds Anchored Reconsideration, which we wrote specifically to avoid a fixed lean, as tightly as it binds the methods that came before it\.

One condition breaks the pattern, within limits\. A model asked to*reason*through a question before answering gains a third input, a derivation that owes nothing to the peers and that it can check for itself\. On the MMLU subjects whose answers this pool reliably derives, reasoning is the only intervention that raises both quantities at once\. Where the derivation is unreliable, as on GPQA and SimpleQA, it behaves like any other method\. The second axis, the frontier that six developed methods share, and the condition under which one method rises above it are what this paper contributes\.

## 2Related Work

Two studies established that language models conform to peer models and proposed the first mitigations\.Zhuet al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib1)\)adapted the Asch paradigm and introduced Devil’s Advocate and Question Distillation, andWenget al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib18)\)introduced BenchForm together with Empowered Persona and Reflection\.Baltajiet al\.\([2024](https://arxiv.org/html/2608.11247#bib.bib10)\)observe the same yielding for live multi\-agent collaboration, where agents conform under peer pressure and lose hold of an assigned persona, which bears directly on the persona\-based mitigation we test\. None of the three asks what mitigation costs when the peers are right\. Our open\-weight pool scores every method against correct and incorrect peers alike\.

Both halves of the result were initially established in human subjects\. Asch showed that a unanimous majority makes a subject abandon a correct judgment\(Asch,[1951](https://arxiv.org/html/2608.11247#bib.bib2)\)\.Allen and Levine \([1969](https://arxiv.org/html/2608.11247#bib.bib4)\)proved that breaking the unanimity is not by itself what frees the subject, since a partner giving the subject’s own answer lowered conformity on every item type they tested, while a more extreme dissenter lowered it only on some\.Sperberet al\.\([2010](https://arxiv.org/html/2608.11247#bib.bib5)\)supply the cost side, arguing that agents who depend on communication cannot protect themselves by growing uniformly harder to move, because the closure that prevents misinformation also prevents what is worth learning\.Willis \([1965](https://arxiv.org/html/2608.11247#bib.bib6)\)names the two responses that this implies, judging each peer claim on its merits \(independence\) and refusing to move \(anticonformity\)\. The distinction is invisible to a study that only presents wrong peers, since the two responses differ solely in what they do when the group is right, which is the case our design adds\.

Models yield to human authority even when their answers are correct\(Sharmaet al\.,[2023](https://arxiv.org/html/2608.11247#bib.bib11); Perezet al\.,[2023](https://arxiv.org/html/2608.11247#bib.bib22)\), and targeted training data reduce this sycophancy\(Weiet al\.,[2023](https://arxiv.org/html/2608.11247#bib.bib12)\)\. That pressure is human rather than peer, and the cost of holding firm again goes unmeasured\. Multi\-agent debate work asks a related question, whether models reasoning together outperform a single model reasoning alone, and typically assumes peers argue in good faith\(Duet al\.,[2024](https://arxiv.org/html/2608.11247#bib.bib13); Lianget al\.,[2024](https://arxiv.org/html/2608.11247#bib.bib14)\)\.Khanet al\.\([2024](https://arxiv.org/html/2608.11247#bib.bib9)\)find that non\-expert judges grow more accurate after hearing expert debaters, more so once debaters are optimized for persuasiveness, a gain that rests on the judge’s willingness to be moved and disappears once that willingness is suppressed\. Our results support this concern, since peer agreement carries no information on truth\.

A model whose parametric knowledge disagrees with information in its context has to decide which to trust\(Longpreet al\.,[2021](https://arxiv.org/html/2608.11247#bib.bib15); Xieet al\.,[2024](https://arxiv.org/html/2608.11247#bib.bib16)\), and weighs that context by relevance rather than by the credibility markers a human would use\(Wanet al\.,[2024](https://arxiv.org/html/2608.11247#bib.bib8)\)\. Peer conformity is the social form of the same decision, the conflict carried out by other agents rather than by a document\. Step\-by\-step reasoning improves accuracy on tasks that admit a derivation\(Weiet al\.,[2022](https://arxiv.org/html/2608.11247#bib.bib21)\)and owes nothing to peers, butTurpinet al\.\([2023](https://arxiv.org/html/2608.11247#bib.bib7)\)shows that a suggested answer can steer a chain of thought without appearing in it\. Only a correct derivation counts as evidence, and that is what bounds where reasoning improves both axes\. We add the cost dimension this literature has left out, measured on a pool broad enough to separate competence from scale\.

## 3The Resistance–Receptivity Frontier

Stopping a model from following incorrect peers also stops it from following correct ones, so a mitigation has to be scored on both cases\. The two scores come from pools built by a baseline pass in which each model answers each question individually\. The questions it gets right form its*resist pool*and the questions it gets wrong form its*adopt pool*, so change is measured against the model’s isolated answer, as is standard in this literature\(Zhuet al\.,[2025](https://arxiv.org/html/2608.11247#bib.bib1)\)\. In the resist pool, the peers unanimously assert an incorrect answer, and we record Resistance, the share of questions on which the model keeps its correct answer, and conformity, the share on which it adopts the peers’ answers\. The two are not exact complements, since a model sometimes lands on a third answer or abstains, cases that account for 2–9% of the pool, and we report them separately\. In the adopt pool, the peers unanimously assert the correct answer, and Receptivity is the share of questions on which the model switches to it\.

Each method occupies one point on a plane of Receptivity against Resistance\. Methods that land on a common downward line differ only in how much of one quantity they trade for the other\. We call that line a*Resistance–Receptivity frontier*, and its existence changes the test that a mitigation has to pass\. Every point on the line lowers conformity by giving up Receptivity, so a reduction in conformity is not on its own evidence that a method has improved anything, and the question becomes whether a method rises above the line\. Rising above requires new information rather than new wording\.

The conditions we test are divided into two kinds, and only one of them can provide that information\. A*snap*condition demands an immediate committed answer, so it can only reweight the model’s prior answer against the peers’ assertion, whereas a*deliberative*condition asks for a derivation first and so adds a new input\. Reweighting should slide a method along the frontier, while evidence that correlates with correctness should lift it off the line, and only where the model can actually produce such evidence\. These are the two predictions the results test, one for the five snap methods and one for the deliberative condition\.

## 4Experimental Setup

#### Models\.

Our pool holds 23 open\-weight, instruction\-tuned models from 19 families, listed in Appendix[A](https://arxiv.org/html/2608.11247#A1)\. Total parameters range from 14B to 70B and active parameters from 2\.4B to 70B\. The pool covers three architecture types: 12 dense, 9 mixture\-of\-experts, and 2 hybrid state\-space models, released between 2023 and 2026\. Models are served one at a time with vLLM at temperature 0 on a single GB10 workstation\. Where a model exposes a thinking mode, we disable it in every snap condition so that the committed answer is a single token\. The deliberative condition is treated separately in Section[5\.4](https://arxiv.org/html/2608.11247#S5.SS4)\.

#### Data\.

Three datasets cover three answering settings, so that no finding rests on a single format or difficulty band\. MMLU provides broad\-knowledge multiple choice over 14,042 questions\(Hendryckset al\.,[2021](https://arxiv.org/html/2608.11247#bib.bib17)\), GPQA\-Main supplies hard science multiple choice over 427 questions\(Reinet al\.,[2024](https://arxiv.org/html/2608.11247#bib.bib19)\), and SimpleQA supplies free\-form factual recall over 4,326 questions\(Weiet al\.,[2024](https://arxiv.org/html/2608.11247#bib.bib20)\)\. In the multiple\-choice sets, the wrong peer answer is a fixed option determined by anmd5hash of the question ID, which spreads it evenly across the options\. The same hash shuffles the GPQA options, so the correct letter never sits in a fixed position\. Multiple\-choice answers are graded by exact letter match\. For SimpleQA, we generate one plausible but incorrect answer per question with the GPT\-4o\-miniOpenAI \([2024](https://arxiv.org/html/2608.11247#bib.bib44)\), a model outside the subject pool, and the answers are graded by a neutral judge, a local Llama\-3\.3\-70B\-InstructMeta \([2024](https://arxiv.org/html/2608.11247#bib.bib43)\)\.

#### Conditions\.

Every model and dataset passes through the same 19 conditions, listed in Appendix[B](https://arxiv.org/html/2608.11247#A2)\. Seven of them run once, a no\-peer baseline, five steps of rising peer pressure, and one Receptivity probe; the remaining twelve are the six mitigations at both polarities\. A conformity gradient raises the pressure in steps, from one wrong peer, to a wrong majority with one correct dissenter, to a unanimous wrong group whose assertion is bare, briefly reasoned, or stated with high certainty\. A separate Receptivity probe presents a unanimous correct group\. Each mitigation runs twice, once against a unanimous wrong group and once against a unanimous correct group\. Four of the six mitigations come from prior work, Devil’s Advocate and Question Distillation fromZhuet al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib1)\)and Empowered Persona and Reflection fromWenget al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib18)\), and all four are snap methods that require an immediate committed answer\. We add two of our own\. Anchored Reconsideration is a fifth snap method, asking the model to decide on the merits and to move off its own answer only for a nameable reason, favoring neither the majority nor its prior answer\. Reasoning\-first is the deliberative condition, presenting the same peer block and asking the model to work through the question step by step before committing, with a budget of 2,048 tokens\. All conditions are single\-pass, matching how an agent answers inside a live system, and prompt templates are given in Appendix[B](https://arxiv.org/html/2608.11247#A2)\.

#### Reasoning\.

Free generation costs far more than a single committed token, so the reasoning condition runs on samples, drawing 300 questions on MMLU, a fixed set of 300 on SimpleQA, and the full pools on GPQA\. Method comparisons in Section[5](https://arxiv.org/html/2608.11247#S5)use*matched*questions, with every snap condition re\-scored on exactly the question IDs drawn for that model’s reasoning sample\. Appendix[C](https://arxiv.org/html/2608.11247#A3)repeats every comparison on the full pools, where no conclusion changes\. Aggregate numbers are macro\-averages that weight each model equally\. We estimate uncertainty with a cluster bootstrap over 10,000 draws that resamples the 23 models with replacement, and report 95% percentile intervals\. Method effects are per\-model differences relative to the no\-intervention condition\. The frontier is an ordinary least\-squares fit to the six snap\-answer points \(five snap mitigations and the no\-intervention baseline\), and the reasoning gap is the vertical distance of the reasoning point from that line, bootstrapped the same way\.

## 5Results

We first characterize the pressure itself, showing that conformity scales with the number of peers that assert a wrong answer and that the displaced answers concentrate on the option those peers named \(Section[5\.1](https://arxiv.org/html/2608.11247#S5.SS1)\)\. We then ask what predicts it and find that how much a model gives up is set by what it can verify for itself rather than by how large it is \(Section[5\.2](https://arxiv.org/html/2608.11247#S5.SS2)\)\. Against this background, all mitigations converge along a single Resistance–Receptivity frontier \(Section[5\.3](https://arxiv.org/html/2608.11247#S5.SS3)\), a threshold that the deliberative condition precisely surpasses on the questions that this pool is capable of deriving \(Section[5\.4](https://arxiv.org/html/2608.11247#S5.SS4)\)\.

### 5\.1Conformity is large, graded, and targeted

Table 1:Conformity under each level of peer pressure, as the percentage of resist\-pool answers switched to the peers’ answer, averaged over the 23 models\. Bootstrap intervals are given in Appendix[D](https://arxiv.org/html/2608.11247#A4)\.Conformity increases with every increment of peer pressure on all three datasets \(Table[1](https://arxiv.org/html/2608.11247#S5.T1)\)\. One wrong peer reverses 11\.2% of a model’s correct MMLU answers and a unanimous group of four reverses 22\.8%, and the same step raises the rate by half again or more on GPQA and SimpleQA\. The number of peers who assert the wrong answer, rather than the emphasis that any one of them places on it, is what drives the rate\.

![Refer to caption](https://arxiv.org/html/2608.11247v1/results/plots/fig1_premise.png)Figure 2:Outcome composition on the resist pool as peer pressure grows, mean over 23 models, one panel per dataset\. Bands from the bottom are the correct answer kept, the peers’ wrong answer adopted \(conformity, value printed\), some other wrong answer, and abstention\. The five conditions are the five rows of Table[1](https://arxiv.org/html/2608.11247#S5.T1)\.Figure[2](https://arxiv.org/html/2608.11247#S5.F2)accounts for the rest of the resist pool\. The share on which a model keeps its correct answer falls at every step, the share on which it adopts the peers’ answer rises to replace it, and the two remaining outcomes, a third wrong answer and an abstention, together stay below 9% throughout\. Pressure converts held answers into adopted ones rather than producing confusion or refusal, and it does so in nearly every model taken on its own \(Appendix[H](https://arxiv.org/html/2608.11247#A8)\)\.

The displaced answers concentrate on the peers’ option\. Under a unanimous wrong group, 83\.6% of the abandoned MMLU answers and 89\.4% of the abandoned GPQA answers match the option the peers named\. SimpleQA offers no option set at all, and 87\.9% still reproduce the peers’ string exactly\. Generic prompt sensitivity would account for a model at temperature 0 changing its answer, but not for the answer it changes to\.

Three features of the gradient affect subsequent mitigations\. The first reproduces in models the dissenter effectAsch \([1951](https://arxiv.org/html/2608.11247#bib.bib2)\)found in people, since a single correct voice inside an otherwise wrong majority lowers conformity by 7\.5 points on MMLU and 16\.0 on SimpleQA, an effect that Section[5\.3](https://arxiv.org/html/2608.11247#S5.SS3)traces to what the dissenter says rather than to the broken unanimity\. The second and third run against the intuition that a more forceful peer is a more persuasive one\. Asserted certainty persuades*less*than a plain statement on every dataset and in at least 21 of the 23 models, while a one\-line rationale persuades slightly more, by 2\.1 points on MMLU\. Models read a stated reason as evidence and asserted confidence as grounds for suspicion, so what changes an answer is the content a peer offers rather than the conviction it professes\.

### 5\.2Conformity tracks competence, not scale

What a model knows predicts how many of its correct answers a wrong group reverses, whereas a model’s size exhibit not predictive power in this regard\. Under a unanimous wrong group, conformity runs from 1\.8% to 83\.7% across the pool on MMLU and from 10\.1% to 97\.6% on GPQA \(Appendix[A](https://arxiv.org/html/2608.11247#A1)\), and the two ends of that range are not the two ends of the parameter count\. Figure[3](https://arxiv.org/html/2608.11247#S5.F3)sets the same 23 rates against both candidate explanations\. Conformity decreases as baseline accuracy increases \(r=−0\.64r\{=\}\{\-\}0\.64,p=0\.001p\{=\}0\.001, on MMLU;r=−0\.49r\{=\}\{\-\}0\.49,p=0\.017p\{=\}0\.017, on GPQA\) and is flat against active parameters over a pool spanning 2\.4B to 70B \(r=\+0\.01r\{=\}\{\+\}0\.01on MMLU,\+0\.10\+0\.10on GPQA and\+0\.18\+0\.18on SimpleQA, none significant\), so the null on size reflects an absent relation rather than a narrow range\. On SimpleQA, the association between accuracy and conformity is weak\(r=−0\.17r\{=\}\{\-\}0\.17\) and sensitive to the two models at the upper end of the narrow accuracy range\. This lack of association is what recall ought to look like, since knowing more facts is no help in checking any one of them\.

![Refer to caption](https://arxiv.org/html/2608.11247v1/results/plots/figc_competence.png)Figure 3:Per\-model conformity under a unanimous wrong group against baseline accuracy \(left\) and active parameters \(right, log scale\), with least\-squares fits\.Resistance is also not a fixed trait of a model\. Figure[4](https://arxiv.org/html/2608.11247#S5.F4)places each model’s three rates side by side, and the ordering they produce is not a single ordering\. The two multiple\-choice sets rank the pool almost identically \(Spearmanρ=0\.97\\rho\{=\}0\.97\), but SimpleQA, where Resistance is weakest and least tied to accuracy, reorders it substantially \(ρ=0\.74\\rho\{=\}0\.74against MMLU and0\.650\.65against GPQA\)\. Yi\-1\.5\-34B\-Chat is the steadiest model on both multiple\-choice sets and still gives up 63\.0% of its correct SimpleQA answers, so steadiness on one benchmark guaranties nothing on another\. Architecture, denoted as D/M/H in the figure, does not delineate a line through the pool either, as the differences among the three groups are smaller than the variability within each group \(Appendix[E](https://arxiv.org/html/2608.11247#A5)\)\.

![Refer to caption](https://arxiv.org/html/2608.11247v1/results/plots/fig4_permodel.png)Figure 4:Per\-model conformity under a unanimous wrong group on MMLU, GPQA and SimpleQA\. Models are ordered by their mean rate across the three datasets, with the most conforming at the top\. D/M/H tag dense, mixture\-of\-experts and hybrid state\-space architectures\.The relation is much stronger between subjects than between models\. Among the 57 MMLU subjects, each averaged across all 23 models, conformity decreases with subject accuracy atr=−0\.92r\{=\}\{\-\}0\.92\(p<10−23p\{<\}10^\{\-23\}\)\. Conformity is higher on GPQA chemistry \(62\.5%\) than on physics \(52\.3%\) or biology \(39\.5%\), and higher on SimpleQA questions whose answer is a number \(77\.5%\) or a date \(74\.4%\) than on those naming a person \(59\.8%\)\. Conformity is highest where the answer gives the model nothing it can check for itself\.

Lucky guesses inflate these rates on the multiple\-choice sets, since an option picked at random enters the resist pool alongside an answer the model knows, roughly 12% of correct answers at MMLU’s mean accuracy and up to 57% at GPQA’s\. We therefore restrict the pool to answers a model also kept under a single wrong peer\. Conformity on those answers is 14\.7% on MMLU and 34\.4% on GPQA, against 73\.4% and 84\.0% on the remainder, and SimpleQA divides the same way\. The accuracy relation holds on the restricted pool \(r=−0\.62r\{=\}\{\-\}0\.62on MMLU,−0\.40\-0\.40on GPQA\)\. Guessing therefore accounts for part of the pooled rates but not for the effect itself\. An answer kept under one wrong peer is the best evidence available here that the model knew it rather than guessed it, and a seventh of those answers on MMLU and a third on GPQA still switch once four peers contradict them\.

### 5\.3Six methods, one frontier

Resistance and Receptivity already move in opposite directions across the 23 models without any intervention\. Resistance increases with baseline accuracy \(r=\+0\.67r\{=\}\{\+\}0\.67on MMLU,\+0\.51\+0\.51on GPQA\) while Receptivity decreases \(r=−0\.42r\{=\}\{\-\}0\.42on MMLU,−0\.32\-0\.32on GPQA\), and the pattern survives a control for architecture reported in Appendix[E](https://arxiv.org/html/2608.11247#A5)\. What lets a model keep a correct answer against a wrong group also makes it slower to accept a correct answer from a right one, so the mitigations inherit the trade rather than introduce it\.

![Refer to caption](https://arxiv.org/html/2608.11247v1/results/plots/fig2_frontier.png)Figure 5:The Resistance–Receptivity frontier on matched questions\. Each point is one condition, averaged over the 23 models\. The dashed line is the least\-squares fit through the six snap\-answer points\. All three panels share one Receptivity axis and one Resistance axis\. Coordinates are listed in Appendix[C](https://arxiv.org/html/2608.11247#A3)\.Five snap mitigations, built around five different mechanisms, land on one line together with the no\-intervention baseline \(Figure[5](https://arxiv.org/html/2608.11247#S5.F5)\)\. The least\-squares fit through the six snap\-answer points hasR2R^\{2\}of 0\.88 on MMLU, 0\.80 on GPQA and 0\.90 on SimpleQA\. Averaging across heterogeneous models is not what produces it, since fitted within a single model the slope stays negative for 22 of 23 models on MMLU and for all 23 on the other two datasets\. Each method chooses where on the line to stand and none stands above it\.

Table 2:Paired per\-model change in Resistance and Receptivity against no intervention \(percentage points, matched questions\), averaged over the 23 models\. Snap methods above the rule, the deliberative one below\. \[Z\]Zhuet al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib1)\), \[W\]Wenget al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib18)\)\. Bootstrap intervals are given in Appendix[D](https://arxiv.org/html/2608.11247#A4)\.Table[2](https://arxiv.org/html/2608.11247#S5.T2)gives the magnitudes\. Every snap method with a significant Resistance gain also has a significant Receptivity loss\. Reflection is the strongest prior method, gaining 7\.9 points of MMLU Resistance against 15\.3 of Receptivity, close to two points given up for every point won on both multiple\-choice sets, a cost that has not previously been measured\. SimpleQA compresses the Receptivity axis, where a 9\.4% baseline accuracy leaves 91% of items in the adopt pool, and every method lands between 86\.3 and 94\.3%, so the trade runs almost entirely through Resistance\.

Devil’s Advocate sits closest to the no\-intervention point in Figure[5](https://arxiv.org/html/2608.11247#S5.F5)and is the most instructive of the five\. Its MMLU Resistance gain has an interval that includes zero, and it lowers Resistance outright for 14 of 23 models there and 12 of 23 on GPQA \(Table[9](https://arxiv.org/html/2608.11247#A5.T9)\)\. The contrast with the correct dissenter of Section[5\.1](https://arxiv.org/html/2608.11247#S5.SS1)explains why\. Devil’s Advocate puts a second wrong answer where that voice was, and disagreement without content does not steady a model, an orderingAllen and Levine \([1969](https://arxiv.org/html/2608.11247#bib.bib4)\)also report in people\. The result identifies what makes a dissenter effective rather than rebutting the original proposal\.

Anchored Reconsideration is the most informative row of Table[2](https://arxiv.org/html/2608.11247#S5.T2), since we wrote it to gain on both axes and it does not\. Asking for a decision on the merits aims at rational updating rather than firmness, yet it gains 4\.9 points of MMLU Resistance against 10\.7 of Receptivity and lands between Reflection and Empowered Persona\. A confident wrong peer and a confident right one look identical inside the exchange, so any lean the prompt induces applies to both\. A point above the line requires information from outside the exchange, which is what the last experiment introduces\.

### 5\.4Where reasoning improves both axes

Reasoning\-first is the only condition under which both Resistance and Receptivity increase\. Across the full MMLU pool, it increases Resistance by 2\.6 points and Receptivity by 2\.2, neither individually significant, although the Receptivity estimate is the only one in Table[2](https://arxiv.org/html/2608.11247#S5.T2)that is positive\. The frontier gap combines both axes into one number, the Resistance a condition reaches minus the Resistance the snap line of Figure[5](https://arxiv.org/html/2608.11247#S5.F5)predicts at the same Receptivity, so a positive gap means higher Resistance than the snap methods reach at that level of Receptivity\. Reasoning\-first has a gap of\+3\.3\+3\.3points on MMLU, an interval whose lower bound does not admits improvement, and gaps of−0\.6\-0\.6and−1\.9\-1\.9points on GPQA and SimpleQA, which leave it on the line with the snap methods\. All three hold under the leave\-one\-out, leave\-two\-out and leave\-three\-out resamplings of the pool \(Appendix[G](https://arxiv.org/html/2608.11247#A7)\)\. The condition does something the other five do not, and it does it on one dataset out of three\.

Table 3:Reasoning against no intervention on MMLU, split by whether the question can be worked out \(%, mean over the 23 models, matched questions\)\. The 20 derivable and 37 recall subjects are listed in Appendix[F](https://arxiv.org/html/2608.11247#A6)\.The dataset average is what weakens it, because MMLU asks two kinds of question\. Table[3](https://arxiv.org/html/2608.11247#S5.T3)splits its 57 subjects by whether an answer can be worked out, mathematics and the physical sciences against the subjects that reduce to recall\. We drew that line after the cross\-dataset pattern was visible, so we checked it twice\. Interleaved halves of each list recover the derivable\-minus\-recall difference within two points of the full estimate, excluding zero in both \(Appendix[F\.2](https://arxiv.org/html/2608.11247#A6.SS2)\), and a rater given only the written criterion, with no knowledge of the study, reproduced the assignment on 56 of the 57 subjects, Cohen’sκ=0\.96\\kappa\{=\}0\.96\(Appendix[F\.1](https://arxiv.org/html/2608.11247#A6.SS1)\)\. The boundary follows from the rule rather than from the pattern it produced\.

On derivable subjects, reasoning raises both axes, by 7\.2 and 9\.6 points, the only intervention to do so with intervals excluding zero, while on the recall subjects it moves neither \(\+0\.9\+0\.9and−1\.3\-1\.3\)\. The difference is not in computation or competence\. Both halves receive the same 2,048\-token budget, and the derivable subjects are the harder half for this pool at 66\.0% mean accuracy against 79\.2%, so a gain that followed how well a model answers would have shown up on the recall subjects instead\. What the derivable subjects offer is a route to the answer that does not pass through the group, and reasoning raises both axes where that route exists\.

The same requirement explains the two datasets where the gap is zero\. What bounds the gain is not whether a question admits a derivation, but whether the model can complete one\. SimpleQA offers nothing to derive, and GPQA admits derivations that this pool, at 36\.9% mean accuracy, rarely completes, where a wrong derivation is not evidence of correctness\. Reasoning therefore rises above the frontier on the 20 subjects that give a model an independent check on its own answer, and sits on the line everywhere else we measure\. What decides which of the two happens is the task and the model together, not the wording of the prompt\.

## 6Discussion

The frontier is a property of the task, visible in the baselines before any method is applied\. As the model’s own check on an answer weakens from MMLU to GPQA to SimpleQA, baseline Resistance falls from 74\.9% to 41\.1% to 17\.8% while Receptivity rises from 59\.6% to 70\.0% to 94\.3%, which is what an agent holding no evidence of its own should do\. The six methods move a model along a trade\-off the no\-intervention baselines show\. The failure is not that models weigh their peers badly, but that the terms on which they weigh them are set by what the model can verify\.

A derivation changes those terms because it can supply evidence beyond the peers’ assertion that correlates with correctness\. A wording cannot, which is why the reasoning condition rises above the line only on the subjects where the derivation can be completed\. Retrieval or a tool call is the stronger form of the same idea, since the check it returns does not depend on the model completing the derivation itself, and it should clear the frontier on tasks where reasoning does not\. We have not tested that prediction, and it is the natural next step\.

Until that test is run, a system built today still faces the choice the frontier forces on it\. Where the task admits a derivation the agent can complete on its own, arithmetic, code execution, a formal proof, routing the agent through that derivation before it sees its peers clears the frontier, for the same reason reasoning does on the MMLU subjects that permit it\. Where no such derivation exists, the system stays on the line, and the only real decision is which side of it to occupy\. A single confident wrong agent can pull a whole group off a correct answer, which argues for Resistance, and agents that fail to catch each other’s genuine mistakes lose the benefit multi\-agent systems are built to provide, which argues for Receptivity\. Whichever side a designer picks should be reported on both axes, because an anti\-conformity instruction buys its Resistance by making an agent deaf to right peers as well as wrong ones\. Two of our findings sharpen that choice\. A dissenting voice is worth including only when it happens to be correct, and making an agent sound more confident does not make its peers more likely to believe it\.

## 7Conclusion

Measured on one axis, conformity mitigations look like fixes\. Measured on two, they are positions on a single Resistance–Receptivity frontier that holds across three datasets and six methods, and that binds the instruction we wrote against it as tightly as the methods before it\. Conformity is highest where a model has no independent means of checking its answer, and that is also where every mitigation reduces to a trade\. Deriving the answer before committing is the one condition that improves both axes, and it does so only on the subjects where this pool can complete the derivation\. That boundary is the finding\. A model breaks with the crowd where it can check the answer without the crowd, and a system that needs both Resistance and Receptivity has to supply that check rather than another instruction to stand firm\.

## 8Limitations

Our peers are scripted rather than generated by live models, which gives control over what is asserted at the cost of naturalism\. Each condition uses one prompt template, so robustness to paraphrase goes unmeasured, and the design does not include a content\-free placebo, so the separation of social adoption from generic prompt sensitivity rests on where the displaced answers land \(Section[5\.1](https://arxiv.org/html/2608.11247#S5.SS1)\)\. All conditions are single\-pass, which puts multi\-round debate out of scope and makes our Question Distillation and Reflection single\-pass renderings of methods originally stated with more turns, chosen for comparability\. Multi\-turn versions might sit elsewhere on the same frontier\. Models are instruction\-tuned and run at temperature 0, and behavior under sampling may differ\.

The reasoning condition runs on samples, as its intervals reflect, and its SimpleQA resist side pools only 598 responses, since few SimpleQA questions are answered correctly alone\. The lucky\-guess analysis stratifies by behavior rather than by answer\-token probability, which these runs did not retain\. The derivable\-recall split is assigned by hand at the subject level \(Appendix[F](https://arxiv.org/html/2608.11247#A6)\) and after the pattern was visible, so a question\-level division could sharpen or soften the contrast\. Both checks in Section[5\.4](https://arxiv.org/html/2608.11247#S5.SS4)reuse the subjects, models, and dataset that produced it, which leaves a new pool, or a fourth dataset with its own derivable and recall subjects, as the confirmatory test\.

SimpleQA grading depends on a neutral judge model, external to the pool and larger than every subject, which we have not validated against human grading\. Two automated checks limit that risk\. The judge agrees with a verbatim\-match heuristic on 98\.6% of the 91,816 of 99,475 baseline responses whose gold answer is long enough to test, and the disagreements are nearly all the gold string appearing incidentally rather than as the stated answer, which the heuristic cannot detect and a judge reading the full response can\. Of the 4,326 planted distractors, one is flagged invalid by the generation\-time filter, none duplicates the gold answer, and the two near\-duplicates a substring check surfaces are not genuine matches\. Neither check replaces human grading, but together they make an artifact large enough to move the reported SimpleQA rates unlikely\. Within these limits, both central results hold on every dataset and across the full pool\.

## References

- 01\-ai/yi\-1\.5\-34b\-chat\.Note:[https://huggingface\.co/01\-ai/Yi\-1\.5\-34B\-Chat](https://huggingface.co/01-ai/Yi-1.5-34B-Chat)Accessed: 2026\-05\-19Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.24.23.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.18.15.1)\.
- M\. Abdin, J\. Aneja, H\. Behl, S\. Bubeck, R\. Eldan, S\. Gunasekar, M\. Harrison, R\. J\. Hewett, M\. Javaheripi, P\. Kauffmann, J\. R\. Lee, Y\. T\. Lee, Y\. Li, W\. Liu, C\. C\. T\. Mendes, A\. Nguyen, E\. Price, G\. de Rosa, O\. Saarikivi, A\. Salim, S\. Shah, X\. Wang, R\. Ward, Y\. Wu, D\. Yu, C\. Zhang, and Y\. Zhang \(2024\)Phi\-4 technical report\.External Links:2412\.08905,[Link](https://arxiv.org/abs/2412.08905)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.19.18.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.20.17.1)\.
- ai21labs \(2025\)Ai21labs/ai21\-jamba2\-mini\.Note:[https://huggingface\.co/ai21labs/AI21\-Jamba2\-Mini](https://huggingface.co/ai21labs/AI21-Jamba2-Mini)Accessed: 2026\-05\-16Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.7.6.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.16.13.1)\.
- V\. L\. Allen and J\. M\. Levine \(1969\)Consensus and conformity\.Journal of Experimental Social Psychology5\(4\),pp\. 389–399\.External Links:[Document](https://dx.doi.org/10.1016/0022-1031%2869%2990032-8)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p2.1),[§5\.3](https://arxiv.org/html/2608.11247#S5.SS3.p4.1)\.
- P\. Apertus, A\. Hernández\-Cano, A\. Hägele, A\. H\. Huang, A\. Romanou, A\. Solergibert, B\. Pasztor, B\. Messmer, D\. Garbaya, E\. F\. Ďurech, I\. Hakimi, J\. G\. Giraldo, M\. Ismayilzada, N\. Foroutan, S\. Moalla, T\. Chen, V\. Sabolčec, Y\. Xu, M\. Aerni, B\. AlKhamissi, I\. A\. Mariñas, M\. H\. Amani, M\. Ansaripour, I\. Badanin, H\. Benoit, E\. Boros, N\. Browning, F\. Bösch, M\. Böther, N\. Canova, C\. Challier, C\. Charmillot, J\. Coles, J\. Deriu, A\. Devos, L\. Drescher, D\. Dzenhaliou, M\. Ehrmann, D\. Fan, S\. Fan, S\. Gao, M\. Gila, M\. Grandury, D\. Hashemi, A\. Hoyle, J\. Jiang, M\. Klein, A\. Kucharavy, A\. Kucherenko, F\. Lübeck, R\. Machacek, T\. Manitaras, A\. Marfurt, K\. Matoba, S\. Matrenok, H\. Mendonça, F\. R\. Mohamed, S\. Montariol, L\. Mouchel, S\. Najem\-Meyer, J\. Ni, G\. Oliva, M\. Pagliardini, E\. Palme, A\. Panferov, L\. Paoletti, M\. Passerini, I\. Pavlov, A\. Poiroux, K\. Ponkshe, N\. Ranchin, J\. Rando, M\. Sauser, J\. Saydaliev, M\. A\. Sayfiddinov, M\. Schneider, S\. Schuppli, M\. Scialanga, A\. Semenov, K\. Shridhar, R\. Singhal, A\. Sotnikova, A\. Sternfeld, A\. K\. Tarun, P\. Teiletche, J\. Vamvas, X\. Yao, H\. Zhao, A\. Ilic, A\. Klimovic, A\. Krause, C\. Gulcehre, D\. Rosenthal, E\. Ash, F\. Tramèr, J\. VandeVondele, L\. Veraldi, M\. Rajman, T\. Schulthess, T\. Hoefler, A\. Bosselut, M\. Jaggi, and I\. Schlag \(2025\)Apertus: democratizing open and compliant llms for global language environments\.External Links:2509\.14233,[Link](https://arxiv.org/abs/2509.14233)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.12.11.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.17.14.1)\.
- S\. E\. Asch \(1951\)Effects of group pressure upon the modification and distortion of judgments\.InGroups, Leadership and Men,H\. Guetzkow \(Ed\.\),pp\. 177–190\.Cited by:[§1](https://arxiv.org/html/2608.11247#S1.p1.1),[§2](https://arxiv.org/html/2608.11247#S2.p2.1),[§5\.1](https://arxiv.org/html/2608.11247#S5.SS1.p4.1)\.
- K\. Bae, E\. Choi, K\. Choi, S\. J\. Choi, Y\. Choi, K\. Han, S\. Hong, J\. Hwang, T\. Hwang, J\. Jang, H\. Jeon, K\. Jeon, G\. J\. Jo, H\. Jo, J\. Jung, E\. Kim, H\. Kim, J\. Kim, J\. Kim, S\. Kim, S\. Kim, S\. Kim, Y\. Kim, Y\. Kim, Y\. Kim, E\. H\. Lee, G\. Lee, H\. Lee, H\. Lee, J\. Lee, K\. Lee, S\. Park, Y\. M\. Paik, Y\. Park, Y\. Park, S\. Seo, S\. Yang, H\. Yeen, S\. Yi, and H\. Yun \(2026\)EXAONE 4\.0: unified large language models integrating non\-reasoning and reasoning modes\.External Links:2507\.11407,[Link](https://arxiv.org/abs/2507.11407)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.11.10.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.27.24.1)\.
- baidu \(2025\)Baidu/ernie\-4\.5\-21b\-a3b\-pt\.Note:[https://huggingface\.co/baidu/ERNIE\-4\.5\-21B\-A3B\-PT](https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-PT)Accessed: 2026\-05\-17Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.17.16.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.15.12.1)\.
- R\. Baltaji, B\. Hemmatian, and L\. Varshney \(2024\)Conformity, confabulation, and impersonation: persona inconstancy in multi\-agent LLM collaboration\.InProceedings of the 2nd Workshop on Cross\-Cultural Considerations in NLP \(C3NLP\),Bangkok, Thailand,pp\. 17–31\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.c3nlp-1.2),[Link](https://aclanthology.org/2024.c3nlp-1.2/)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p1.1)\.
- ByteDance\-Seed \(2025\)Seed\-oss open\-source models\.Note:[https://github\.com/ByteDance\-Seed/seed\-oss](https://github.com/ByteDance-Seed/seed-oss)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.4.3.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.13.10.1)\.
- CohereLabs \(2024\)CohereLabs/c4ai\-command\-r\-v01\.Note:[https://huggingface\.co/CohereLabs/c4ai\-command\-r\-v01](https://huggingface.co/CohereLabs/c4ai-command-r-v01)Accessed: 2026\-05\-18Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.5.4.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.2.2.1)\.
- DeepSeek\-AI, A\. Liu, B\. Feng, B\. Wang, B\. Wang, B\. Liu, C\. Zhao, C\. Dengr, C\. Ruan, D\. Dai, D\. Guo, D\. Yang, D\. Chen, D\. Ji, E\. Li, F\. Lin, F\. Luo, G\. Hao, G\. Chen, G\. Li, H\. Zhang, H\. Xu, H\. Yang, H\. Zhang, H\. Ding, H\. Xin, H\. Gao, H\. Li, H\. Qu, J\. L\. Cai, J\. Liang, J\. Guo, J\. Ni, J\. Li, J\. Chen, J\. Yuan, J\. Qiu, J\. Song, K\. Dong, K\. Gao, K\. Guan, L\. Wang, L\. Zhang, L\. Xu, L\. Xia, L\. Zhao, L\. Zhang, M\. Li, M\. Wang, M\. Zhang, M\. Zhang, M\. Tang, M\. Li, N\. Tian, P\. Huang, P\. Wang, P\. Zhang, Q\. Zhu, Q\. Chen, Q\. Du, R\. J\. Chen, R\. L\. Jin, R\. Ge, R\. Pan, R\. Xu, R\. Chen, S\. S\. Li, S\. Lu, S\. Zhou, S\. Chen, S\. Wu, S\. Ye, S\. Ma, S\. Wang, S\. Zhou, S\. Yu, S\. Zhou, S\. Zheng, T\. Wang, T\. Pei, T\. Yuan, T\. Sun, W\. L\. Xiao, W\. Zeng, W\. An, W\. Liu, W\. Liang, W\. Gao, W\. Zhang, X\. Q\. Li, X\. Jin, X\. Wang, X\. Bi, X\. Liu, X\. Wang, X\. Shen, X\. Chen, X\. Chen, X\. Nie, X\. Sun, X\. Wang, X\. Liu, X\. Xie, X\. Yu, X\. Song, X\. Zhou, X\. Yang, X\. Lu, X\. Su, Y\. Wu, Y\. K\. Li, Y\. X\. Wei, Y\. X\. Zhu, Y\. Xu, Y\. Huang, Y\. Li, Y\. Zhao, Y\. Sun, Y\. Li, Y\. Wang, Y\. Zheng, Y\. Zhang, Y\. Xiong, Y\. Zhao, Y\. He, Y\. Tang, Y\. Piao, Y\. Dong, Y\. Tan, Y\. Liu, Y\. Wang, Y\. Guo, Y\. Zhu, Y\. Wang, Y\. Zou, Y\. Zha, Y\. Ma, Y\. Yan, Y\. You, Y\. Liu, Z\. Z\. Ren, Z\. Ren, Z\. Sha, Z\. Fu, Z\. Huang, Z\. Zhang, Z\. Xie, Z\. Hao, Z\. Shao, Z\. Wen, Z\. Xu, Z\. Zhang, Z\. Li, Z\. Wang, Z\. Gu, Z\. Li, and Z\. Xie \(2024\)DeepSeek\-v2: a strong, economical, and efficient mixture\-of\-experts language model\.External Links:2405\.04434,[Link](https://arxiv.org/abs/2405.04434)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.2.1.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.28.25.1)\.
- Y\. Du, S\. Li, A\. Torralba, J\. B\. Tenenbaum, and I\. Mordatch \(2024\)Improving factuality and reasoning in language models through multiagent debate\.InProceedings of the 41st International Conference on Machine Learning,ICML’24\.External Links:[Link](https://dl.acm.org/doi/10.5555/3692070.3692537)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p3.1)\.
- T\. Gemma, S\. E\. Abd, V\. Aggarwal, R\. Algayres, A\. Andreev, O\. Bachem, I\. Ballantyne, C\. Brick, V\. Cărbune, M\. Casbon, M\. Chaturvedi, A\. Chawla, V\. Cotruta, A\. Coucke, P\. Culliton, R\. Dadashi, L\. Dixon, M\. Elhawaty, U\. Evci, C\. Farabet, J\. Ferret, F\. Galgani, S\. Girgin, J\. Grill, M\. Grootendorst, J\. Guo, C\. Hardin, Y\. He, S\. M\. Hernandez, O\. Homburger, L\. Hussenot, J\. Ji, A\. Joulin, A\. Kamath, P\. Kassraie, O\. Lacombe, P\. Lahoti, G\. Liu, G\. Martins, L\. Martins, T\. Matejovicova, R\. Merhej, N\. Momchev, S\. Mondal, R\. Mullins, S\. R\. Panyam, S\. Pathak, S\. Perrin, A\. S\. Pinto, E\. Pot, A\. Pouget, A\. Ramé, S\. Ramos, D\. Reid, D\. Rim, M\. Rivière, K\. Roth, L\. Rouillard, O\. Sanseviero, P\. G\. Sessa, S\. Settle, D\. Sinopalnikov, S\. Smoot, P\. Stanczyk, A\. Steiner, L\. Stewart, I\. Tolstikhin, M\. Tschannen, A\. Tsitsulin, N\. Vieillard, R\. Wu, P\. Xu, H\. Yang, E\. Yvinec, B\. Zhang, L\. Zhang, J\. Zou, N\. Aagnes, A\. Abdelhamed, J\. Adamek, S\. Agrawal, S\. Agrawal, I\. Alabdulmohsin, J\. B\. Alayrac, U\. Alon, C\. Amarnath, A\. Anand, C\. Anastasiou, S\. Ariafar, F\. Aubet, K\. Axiotis, F\. Barbero, J\. Barral, A\. Bendebury, U\. Bergmann, S\. Bileschi, K\. Black, M\. Blondel, S\. Borgeaud, A\. Bražinskas, R\. Burnell, R\. Busa\-Fekete, M\. Cai, D\. Calandriello, G\. Cameron, C\. Caucheteux, R\. Chaabouni, G\. Chadha, J\. Chan, B\. J\. Chen, J\. Chen, L\. Chen, X\. Chen, D\. Cheng, T\. Chien, N\. Chinaev, Y\. Chou, Z\. Chu, B\. Coleman, P\. Consul, S\. Conway\-Rahman, S\. Crowell, D\. Cutler, V\. Dani, S\. Daruki, A\. Das, D\. Deutsch, N\. Dikkala, L\. Ding, Q\. Ding, S\. Dodhia, K\. Donhauser, T\. Doshi, A\. Dragan, A\. Druinsky, S\. Dua, Z\. Egyed, D\. Eisenbud, D\. Eppens, C\. Fan, B\. Fatemi, Y\. Fathullah, V\. Feinberg, M\. Ferev, S\. Flennerhag, T\. Fujimoto, J\. G\. Oliveira, I\. Galatzer\-Levy, J\. Gante, S\. Geisler, S\. Ghosal, A\. M\. Girgis, T\. von Glehn, A\. Go, A\. Gokhale, A\. Grills, Y\. Gu, M\. Gupta, P\. Gupta, G\. Guruganesh, R\. Hadsell, H\. Harkous, J\. Harlalka, D\. Hassabis, A\. Hauth, J\. Heyward, A\. Hosseini, C\. Hsia, I\. Hsu, X\. Huang, Y\. Huang, K\. Hui, A\. Hutter, T\. I, F\. Iliopoulos, A\. Jain, G\. Jawahar, Z\. Ji, Q\. Jin, M\. Johnson, K\. Joshi, A\. Kandoor, W\. Kang, K\. Kavukcuoglu, M\. Kazemi, K\. Kenealy, A\. Khalifa, P\. Kirk, I\. Korotkov, S\. Kothawade, V\. Kovalev, N\. Kovelamudi, A\. Kraft, R\. Kumar, V\. Kumar, H\. Kuppam, J\. Lannin, C\. Lee, S\. Lee, D\. Lepikhin, A\. Levkovitch, D\. Li, Q\. Li, V\. Liévin, E\. Lin, Z\. Lin, C\. Liu, T\. Liu, T\. Liu, X\. Liu, I\. Lobov, M\. Lunayach, M\. Ma, G\. Madan, A\. Maksai, E\. Malmi, M\. Matuszak, D\. McDuff, G\. Menghani, M\. Mikuła, D\. Mirylenka, K\. Misiunas, V\. Misra, A\. Mitran, K\. Mohamed, M\. Mukha, E\. Noland, J\. O’Donnell, B\. O’Donoghue, K\. Olszewska, B\. Orlando, W\. Pan, R\. Panigrahy, U\. Parekh, N\. Perez\-Nieves, C\. Park, E\. Paskie, L\. Peng, B\. Petrini, S\. Petrov, J\. Pfeiffer, B\. Piot, M\. Plomecka, S\. Poder, O\. Ponce, A\. Pramanik, D\. Racz, A\. Rajan, M\. Ramanovich, A\. Rao, M\. Ritter, V\. Rodrigues, E\. Rosen, M\. Rybiński, N\. Sachdeva, M\. E\. Sander, R\. Sathyanarayana, S\. Savla, S\. Schmidgall, T\. Schuster, G\. Scrivener, B\. Seguin, A\. Sellergren, A\. Severyn, I\. Shafran, D\. Shah, B\. Shahriari, Y\. Shangguan, A\. Shenoy, P\. Shenoy, R\. Shivanna, P\. Sho, L\. Spangher, W\. Stokowiec, T\. Strother, Y\. Su, Y\. Sun, M\. Sundararajan, A\. Tacchetti, M\. H\. Taege, P\. Tafti, J\. Tarbouriech, C\. Tekur, S\. Thakoor, R\. Thapa, M\. Traverse, L\. Treven, T\. Tu, C\. T\. Tung, Ç\. Ünlü, P\. Veličković, M\. P\. Venkat, S\. G\. Venkatesh, V\. Venkiteswaran, F\. Visin, A\. Vitvitskyi, K\. Vodrahalli, W\. Wang, X\. Wang, T\. Warkentin, J\. Wassenberg, J\. Wieting, C\. Wu, L\. Xiao, H\. Xu, Y\. Xu, F\. Xue, A\. Yadav, J\. Yan, A\. Yang, L\. Yang, M\. Yang, Z\. Ying, J\. H\. Yoo, M\. Zadimoghaddam, S\. Zafar, F\. Zhang, J\. Zhang, J\. Zhang, X\. Zhang, C\. Zhao, D\. Zhou, and C\. Zou \(2026\)Gemma 4 technical report\.External Links:2607\.02770,[Link](https://arxiv.org/abs/2607.02770)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.16.15.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.6.3.1)\.
- T\. Gemma, A\. Kamath, J\. Ferret, S\. Pathak, N\. Vieillard, R\. Merhej, S\. Perrin, T\. Matejovicova, A\. Ramé, M\. Rivière, L\. Rouillard, T\. Mesnard, G\. Cideron, J\. Grill, S\. Ramos, E\. Yvinec, M\. Casbon, E\. Pot, I\. Penchev, G\. Liu, F\. Visin, K\. Kenealy, L\. Beyer, X\. Zhai, A\. Tsitsulin, R\. Busa\-Fekete, A\. Feng, N\. Sachdeva, B\. Coleman, Y\. Gao, B\. Mustafa, I\. Barr, E\. Parisotto, D\. Tian, M\. Eyal, C\. Cherry, J\. Peter, D\. Sinopalnikov, S\. Bhupatiraju, R\. Agarwal, M\. Kazemi, D\. Malkin, R\. Kumar, D\. Vilar, I\. Brusilovsky, J\. Luo, A\. Steiner, A\. Friesen, A\. Sharma, A\. Sharma, A\. M\. Gilady, A\. Goedeckemeyer, A\. Saade, A\. Feng, A\. Kolesnikov, A\. Bendebury, A\. Abdagic, A\. Vadi, A\. György, A\. S\. Pinto, A\. Das, A\. Bapna, A\. Miech, A\. Yang, A\. Paterson, A\. Shenoy, A\. Chakrabarti, B\. Piot, B\. Wu, B\. Shahriari, B\. Petrini, C\. Chen, C\. L\. Lan, C\. A\. Choquette\-Choo, C\. Carey, C\. Brick, D\. Deutsch, D\. Eisenbud, D\. Cattle, D\. Cheng, D\. Paparas, D\. S\. Sreepathihalli, D\. Reid, D\. Tran, D\. Zelle, E\. Noland, E\. Huizenga, E\. Kharitonov, F\. Liu, G\. Amirkhanyan, G\. Cameron, H\. Hashemi, H\. Klimczak\-Plucińska, H\. Singh, H\. Mehta, H\. T\. Lehri, H\. Hazimeh, I\. Ballantyne, I\. Szpektor, I\. Nardini, J\. Pouget\-Abadie, J\. Chan, J\. Stanton, J\. Wieting, J\. Lai, J\. Orbay, J\. Fernandez, J\. Newlan, J\. Ji, J\. Singh, K\. Black, K\. Yu, K\. Hui, K\. Vodrahalli, K\. Greff, L\. Qiu, M\. Valentine, M\. Coelho, M\. Ritter, M\. Hoffman, M\. Watson, M\. Chaturvedi, M\. Moynihan, M\. Ma, N\. Babar, N\. Noy, N\. Byrd, N\. Roy, N\. Momchev, N\. Chauhan, N\. Sachdeva, O\. Bunyan, P\. Botarda, P\. Caron, P\. K\. Rubenstein, P\. Culliton, P\. Schmid, P\. G\. Sessa, P\. Xu, P\. Stanczyk, P\. Tafti, R\. Shivanna, R\. Wu, R\. Pan, R\. Rokni, R\. Willoughby, R\. Vallu, R\. Mullins, S\. Jerome, S\. Smoot, S\. Girgin, S\. Iqbal, S\. Reddy, S\. Sheth, S\. Põder, S\. Bhatnagar, S\. R\. Panyam, S\. Eiger, S\. Zhang, T\. Liu, T\. Yacovone, T\. Liechty, U\. Kalra, U\. Evci, V\. Misra, V\. Roseberry, V\. Feinberg, V\. Kolesnikov, W\. Han, W\. Kwon, X\. Chen, Y\. Chow, Y\. Zhu, Z\. Wei, Z\. Egyed, V\. Cotruta, M\. Giang, P\. Kirk, A\. Rao, K\. Black, N\. Babar, J\. Lo, E\. Moreira, L\. G\. Martins, O\. Sanseviero, L\. Gonzalez, Z\. Gleicher, T\. Warkentin, V\. Mirrokni, E\. Senter, E\. Collins, J\. Barral, Z\. Ghahramani, R\. Hadsell, Y\. Matias, D\. Sculley, S\. Petrov, N\. Fiedel, N\. Shazeer, O\. Vinyals, J\. Dean, D\. Hassabis, K\. Kavukcuoglu, C\. Farabet, E\. Buchatskaya, J\. Alayrac, R\. Anil, Dmitry, Lepikhin, S\. Borgeaud, O\. Bachem, A\. Joulin, A\. Andreev, C\. Hardin, R\. Dadashi, and L\. Hussenot \(2025\)Gemma 3 technical report\.External Links:2503\.19786,[Link](https://arxiv.org/abs/2503.19786)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.18.17.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.1.1.1)\.
- T\. GLM, :, A\. Zeng, B\. Xu, B\. Wang, C\. Zhang, D\. Yin, D\. Zhang, D\. Rojas, G\. Feng, H\. Zhao, H\. Lai, H\. Yu, H\. Wang, J\. Sun, J\. Zhang, J\. Cheng, J\. Gui, J\. Tang, J\. Zhang, J\. Sun, J\. Li, L\. Zhao, L\. Wu, L\. Zhong, M\. Liu, M\. Huang, P\. Zhang, Q\. Zheng, R\. Lu, S\. Duan, S\. Zhang, S\. Cao, S\. Yang, W\. L\. Tam, W\. Zhao, X\. Liu, X\. Xia, X\. Zhang, X\. Gu, X\. Lv, X\. Liu, X\. Liu, X\. Yang, X\. Song, X\. Zhang, Y\. An, Y\. Xu, Y\. Niu, Y\. Yang, Y\. Li, Y\. Bai, Y\. Dong, Z\. Qi, Z\. Wang, Z\. Yang, Z\. Du, Z\. Hou, and Z\. Wang \(2024\)ChatGLM: a family of large language models from glm\-130b to glm\-4 all tools\.External Links:2406\.12793,[Link](https://arxiv.org/abs/2406.12793)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.15.14.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.21.18.1)\.
- D\. Hendrycks, C\. Burns, S\. Basart, A\. Zou, M\. Mazeika, D\. Song, and J\. Steinhardt \(2021\)Measuring massive multitask language understanding\.InProceedings of the 9th International Conference on Learning Representations \(ICLR\),Cited by:[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.32.29.1),[§4](https://arxiv.org/html/2608.11247#S4.SS0.SSS0.Px2.p1.1)\.
- IBM Research \(2026\)Granite 4\.1 language models\.Note:[https://huggingface\.co/ibm\-granite/granite\-4\.1\-30b](https://huggingface.co/ibm-granite/granite-4.1-30b)Accessed: 2026\-05\-28Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.10.9.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.14.11.1)\.
- A\. Q\. Jiang, A\. Sablayrolles, A\. Roux, A\. Mensch, B\. Savary, C\. Bamford, D\. S\. Chaplot, D\. de las Casas, E\. B\. Hanna, F\. Bressand, G\. Lengyel, G\. Bour, G\. Lample, L\. R\. Lavaud, L\. Saulnier, M\. Lachaux, P\. Stock, S\. Subramanian, S\. Yang, S\. Antoniak, T\. L\. Scao, T\. Gervet, T\. Lavril, T\. Wang, T\. Lacroix, and W\. E\. Sayed \(2024\)Mixtral of experts\.External Links:2401\.04088,[Link](https://arxiv.org/abs/2401.04088)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.3.2.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.8.5.1)\.
- A\. Khan, J\. Hughes, D\. Valentine, L\. Ruis, K\. Sachan, A\. Radhakrishnan, E\. Grefenstette, S\. R\. Bowman, T\. Rocktäschel, and E\. Perez \(2024\)Debating with more persuasive llms leads to more truthful answers\.InProceedings of the 41st International Conference on Machine Learning,ICML’24\.External Links:[Link](https://dl.acm.org/doi/10.5555/3692070.3693020)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p3.1)\.
- T\. Liang, Z\. He, W\. Jiao, X\. Wang, Y\. Wang, R\. Wang, Y\. Yang, S\. Shi, and Z\. Tu \(2024\)Encouraging divergent thinking in large language models through multi\-agent debate\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 17889–17904\.External Links:[Link](https://aclanthology.org/2024.emnlp-main.992/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.992)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p3.1)\.
- S\. Longpre, K\. Perisetla, A\. Chen, N\. Ramesh, C\. DuBois, S\. Singh, B\. Lefebvre, B\. McCann, and R\. Socher \(2021\)Entity\-based knowledge conflicts in question answering\.InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 7052–7063\.External Links:[Document](https://dx.doi.org/10.18653/v1/2021.emnlp-main.565)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p4.1)\.
- Meta \(2024\)Meta llama 3\.3\.Note:[https://developer\.meta\.com/ai/docs/model\-cards\-and\-prompt\-formats/llama3\_3/](https://developer.meta.com/ai/docs/model-cards-and-prompt-formats/llama3_3/)Accessed: 2026\-05\-21Cited by:[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.30.27.1),[§4](https://arxiv.org/html/2608.11247#S4.SS0.SSS0.Px2.p1.1)\.
- mistralai \(2025\)Mistralai/mistral\-small\-3\.2\-24b\-instruct\-2506\.Note:[https://huggingface\.co/mistralai/Mistral\-Small\-3\.2\-24B\-Instruct\-2506](https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506)Accessed: 2026\-05\-22Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.14.13.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.7.4.1)\.
- moonshotai \(2025a\)Moonshotai/kimi\-linear\-48b\-a3b\-instruct\.Note:[https://huggingface\.co/moonshotai/Kimi\-Linear\-48B\-A3B\-Instruct](https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct)Accessed: 2026\-05\-16Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.20.19.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.23.20.1)\.
- moonshotai \(2025b\)Moonshotai/moonlight\-16b\-a3b\.Note:[https://huggingface\.co/moonshotai/Moonlight\-16B\-A3B](https://huggingface.co/moonshotai/Moonlight-16B-A3B)Accessed: 2026\-05\-16Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.9.8.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.22.19.1)\.
- NVIDIA, :, A\. Blakeman, A\. Grattafiori, A\. Basant, A\. Gupta, A\. Khattar, A\. Renduchintala, A\. Vavre, A\. Shukla, A\. Bercovich, A\. Ficek, A\. Shaposhnikov, A\. Kondratenko, A\. Bukharin, A\. Milesi, A\. Taghibakhshi, A\. Liu, A\. Barton, A\. S\. Mahabaleshwarkar, A\. Klein, A\. Zuker, A\. Geifman, A\. Shen, A\. Bhiwandiwalla, A\. Tao, A\. Guan, A\. Mandarwal, A\. Mehta, A\. Aithal, A\. Poojary, A\. Ahamed, A\. K\. Thekkumpate, A\. Dattagupta, B\. Zhu, B\. Sadeghi, B\. Simkin, B\. Lanir, B\. Schifferer, B\. Nushi, B\. Kartal, B\. D\. Rouhani, B\. Ginsburg, B\. Norick, B\. Soubasis, B\. Kisacanin, B\. Yu, B\. Catanzaro, C\. del Mundo, C\. Hwang, C\. Wang, C\. Hsieh, C\. Zhang, C\. Yu, C\. Mungekar, C\. Patel, C\. Alexiuk, C\. Parisien, C\. Neale, D\. Mosk\-Aoyama, D\. Su, D\. Corneil, D\. Afrimi, D\. Rohrer, D\. Serebrenik, D\. Gitman, D\. Levy, D\. Stosic, D\. Mosallanezhad, D\. Narayanan, D\. Nathawani, D\. Rekesh, D\. Yared, D\. Kakwani, D\. Ahn, D\. Riach, D\. Stosic, E\. Minasyan, E\. Lin, E\. Long, E\. P\. Long, E\. Lantz, E\. Evans, E\. Ning, E\. Chung, E\. Harper, E\. Tramel, E\. Galinkin, E\. Pounds, E\. Briones, E\. Bakhturina, F\. Ladhak, F\. Wang, F\. Jia, F\. Soares, F\. Chen, F\. Galko, F\. Siino, G\. H\. Agam, G\. Ajjanagadde, G\. Bhatt, G\. Prasad, G\. Armstrong, G\. Shen, G\. Batmaz, G\. Nalbandyan, H\. Qian, H\. Sharma, H\. Ross, H\. Ngo, H\. Sahota, H\. Wang, H\. Soni, H\. Upadhyay, H\. Mao, H\. C\. Nguyen, H\. Q\. Nguyen, I\. Cunningham, I\. Shahaf, I\. Gitman, I\. Loshchilov, I\. Moshkov, I\. Putterman, J\. Kautz, J\. P\. Scowcroft, J\. Casper, J\. Mitra, J\. Glick, J\. Chen, J\. Oliver, J\. Zhang, J\. Zeng, J\. Lou, J\. Zhang, J\. Huang, J\. Conway, J\. Guman, J\. Kamalu, J\. Greco, J\. Cohen, J\. Jennings, J\. Daw, J\. V\. Vialard, J\. Yi, J\. Parmar, K\. Xu, K\. Zhu, K\. Briski, K\. Cheung, K\. Luna, K\. Santhanam, K\. Shih, K\. Kong, K\. Bhardwaj, K\. C\. Puvvada, K\. Pawelec, K\. Anik, L\. McAfee, L\. Sleiman, L\. Derczynski, L\. Ding, L\. Liebenwein, L\. Vega, M\. Grover, M\. V\. Segbroeck, M\. R\. de Melo, M\. N\. Sreedhar, M\. Kilaru, M\. Ashkenazi, M\. Romeijn, M\. Cai, M\. Kliegl, M\. Moosaei, M\. Novikov, M\. Samadi, M\. Corpuz, M\. Wang, M\. Price, M\. Boone, M\. Evans, M\. Martinez, M\. Chrzanowski, M\. Shoeybi, M\. Patwary, N\. Mulepati, N\. Hereth, N\. Assaf, N\. Habibi, N\. Zmora, N\. Haber, N\. Sessions, N\. Bhatia, N\. Jukar, N\. Pope, N\. Ludwig, N\. Tajbakhsh, N\. Juluru, O\. Hrinchuk, O\. Kuchaiev, O\. Delalleau, O\. Olabiyi, O\. U\. Argov, O\. Xie, P\. Chadha, P\. Shamis, P\. Molchanov, P\. Morkisz, P\. Dykas, P\. Jin, P\. Xu, P\. Januszewski, P\. P\. Thombre, P\. Varshney, P\. Gundecha, Q\. Miao, R\. K\. Mahabadi, R\. El\-Yaniv, R\. Zilberstein, R\. Shafipour, R\. Harang, R\. Izzo, R\. Shahbazyan, R\. Garg, R\. Borkar, R\. Gala, R\. Islam, R\. Waleffe, R\. Watve, R\. Koren, R\. Zhang, R\. J\. Hewett, R\. Prenger, R\. Timbrook, S\. Mahdavi, S\. Modi, S\. Kriman, S\. Kariyappa, S\. Satheesh, S\. Kaji, S\. Pasumarthi, S\. Narentharen, S\. Narenthiran, S\. Bak, S\. Kashirsky, S\. Poulos, S\. Mor, S\. Ramasamy, S\. Acharya, S\. Ghosh, S\. T\. Sreenivas, S\. Thomas, S\. Fan, S\. Gopal, S\. Prabhumoye, S\. Pachori, S\. Toshniwal, S\. Ding, S\. Singh, S\. Sun, S\. Ithape, S\. Majumdar, S\. Singhal, S\. Alborghetti, S\. Ge, S\. D\. Devare, S\. K\. Barua, S\. Panguluri, S\. Gupta, S\. Priyadarshi, S\. N\. Akter, T\. Bui, T\. Ene, T\. Kong, T\. Do, T\. Blankevoort, T\. Balough, T\. Asida, T\. B\. Natan, T\. Konuk, T\. Vashishth, U\. Karpas, U\. De, V\. Noorozi, V\. Noroozi, V\. Srinivasan, V\. Elango, V\. Korthikanti, V\. Kurin, V\. Lavrukhin, W\. Jiang, W\. U\. Ahmad, W\. Du, W\. Ping, W\. Zhou, W\. Jennings, W\. Zhang, W\. Prazuch, X\. Ren, Y\. Karnati, Y\. Choi, Y\. Meyer, Y\. Wu, Y\. Zhang, Y\. Lin, Y\. Geifman, Y\. Fu, Y\. Subara, Y\. Suhara, Y\. Gao, Z\. Moshe, Z\. Dong, Z\. Liu, Z\. Chen, and Z\. Yan \(2025\)Nemotron 3 nano: open, efficient mixture\-of\-experts hybrid mamba\-transformer model for agentic reasoning\.External Links:2512\.20848,[Link](https://arxiv.org/abs/2512.20848)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.8.7.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.25.22.1)\.
- T\. OLMo, P\. Walsh, L\. Soldaini, D\. Groeneveld, K\. Lo, S\. Arora, A\. Bhagia, Y\. Gu, S\. Huang, M\. Jordan, N\. Lambert, D\. Schwenk, O\. Tafjord, T\. Anderson, D\. Atkinson, F\. Brahman, C\. Clark, P\. Dasigi, N\. Dziri, A\. Ettinger, M\. Guerquin, D\. Heineman, H\. Ivison, P\. W\. Koh, J\. Liu, S\. Malik, W\. Merrill, L\. J\. V\. Miranda, J\. Morrison, T\. Murray, C\. Nam, J\. Poznanski, V\. Pyatkin, A\. Rangapur, M\. Schmitz, S\. Skjonsberg, D\. Wadden, C\. Wilhelm, M\. Wilson, L\. Zettlemoyer, A\. Farhadi, N\. A\. Smith, and H\. Hajishirzi \(2025\)2 olmo 2 furious\.External Links:2501\.00656,[Link](https://arxiv.org/abs/2501.00656)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.6.5.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.12.9.1)\.
- OpenAI, S\. Agarwal, L\. Ahmad, J\. Ai, S\. Altman, A\. Applebaum, E\. Arbus, R\. K\. Arora, Y\. Bai, B\. Baker, H\. Bao, B\. Barak, A\. Bennett, T\. Bertao, N\. Brett, E\. Brevdo, G\. Brockman, S\. Bubeck, C\. Chang, K\. Chen, M\. Chen, E\. Cheung, A\. Clark, D\. Cook, M\. Dukhan, C\. Dvorak, K\. Fives, V\. Fomenko, T\. Garipov, K\. Georgiev, M\. Glaese, T\. Gogineni, A\. Goucher, L\. Gross, K\. G\. Guzman, J\. Hallman, J\. Hehir, J\. Heidecke, A\. Helyar, H\. Hu, R\. Huet, J\. Huh, S\. Jain, Z\. Johnson, C\. Koch, I\. Kofman, D\. Kundel, J\. Kwon, V\. Kyrylov, E\. Y\. Le, G\. Leclerc, J\. P\. Lennon, S\. Lessans, M\. Lezcano\-Casado, Y\. Li, Z\. Li, J\. Lin, J\. Liss, Lily, Liu, J\. Liu, K\. Lu, C\. Lu, Z\. Martinovic, L\. McCallum, J\. McGrath, S\. McKinney, A\. McLaughlin, S\. Mei, S\. Mostovoy, T\. Mu, G\. Myles, A\. Neitz, A\. Nichol, J\. Pachocki, A\. Paino, D\. Palmie, A\. Pantuliano, G\. Parascandolo, J\. Park, L\. Pathak, C\. Paz, L\. Peran, D\. Pimenov, M\. Pokrass, E\. Proehl, H\. Qiu, G\. Raila, F\. Raso, H\. Ren, K\. Richardson, D\. Robinson, B\. Rotsted, H\. Salman, S\. Sanjeev, M\. Schwarzer, D\. Sculley, H\. Sikchi, K\. Simon, K\. Singhal, Y\. Song, D\. Stuckey, Z\. Sun, P\. Tillet, S\. Toizer, F\. Tsimpourlas, N\. Vyas, E\. Wallace, X\. Wang, M\. Wang, O\. Watkins, K\. Weil, A\. Wendling, K\. Whinnery, C\. Whitney, H\. Wong, L\. Yang, Y\. Yang, M\. Yasunaga, K\. Ying, W\. Zaremba, W\. Zhan, C\. Zhang, B\. Zhang, E\. Zhang, and S\. Zhao \(2025\)Gpt\-oss\-120b & gpt\-oss\-20b model card\.External Links:2508\.10925,[Link](https://arxiv.org/abs/2508.10925)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.21.20.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.11.8.1)\.
- OpenAI \(2024\)GPT\-4o mini\.Note:[https://developers\.openai\.com/api/docs/models/gpt\-4o\-mini](https://developers.openai.com/api/docs/models/gpt-4o-mini)Accessed: 2026\-05\-14Cited by:[§4](https://arxiv.org/html/2608.11247#S4.SS0.SSS0.Px2.p1.1)\.
- E\. Perez, S\. Ringer, K\. Lukosiute, K\. Nguyen, E\. Chen, S\. Heiner, C\. Pettit, C\. Olsson, S\. Kundu, S\. Kadavath, A\. Jones, A\. Chen, B\. Mann, B\. Israel, B\. Seethor, C\. McKinnon, C\. Olah, D\. Yan, D\. Amodei, D\. Amodei, D\. Drain, D\. Li, E\. Tran\-Johnson, G\. Khundadze, J\. Kernion, J\. Landis, J\. Kerr, J\. Mueller, J\. Hyun, J\. Landau, K\. Ndousse, L\. Goldberg, L\. Lovitt, M\. Lucas, M\. Sellitto, M\. Zhang, N\. Kingsland, N\. Elhage, N\. Joseph, N\. Mercado, N\. DasSarma, O\. Rausch, R\. Larson, S\. McCandlish, S\. Johnston, S\. Kravec, S\. El Showk, T\. Lanham, T\. Telleen\-Lawton, T\. Brown, T\. Henighan, T\. Hume, Y\. Bai, Z\. Hatfield\-Dodds, J\. Clark, S\. R\. Bowman, A\. Askell, R\. Grosse, D\. Hernandez, D\. Ganguli, E\. Hubinger, N\. Schiefer, and J\. Kaplan \(2023\)Discovering language model behaviors with model\-written evaluations\.InFindings of the Association for Computational Linguistics: ACL 2023,A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 13387–13434\.External Links:[Link](https://aclanthology.org/2023.findings-acl.847/),[Document](https://dx.doi.org/10.18653/v1/2023.findings-acl.847)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p3.1)\.
- Qwen Team \(2025\)Qwen3 technical report\.External Links:2505\.09388,[Link](https://arxiv.org/abs/2505.09388)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.13.12.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.9.6.1)\.
- Qwen Team \(2026\)Qwen3\.6\-35B\-A3B: agentic coding power, now open to all\.External Links:[Link](https://qwen.ai/blog?id=qwen3.6-35b-a3b)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.22.21.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.10.7.1)\.
- D\. Rein, B\. L\. Hou, A\. C\. Stickland, J\. Petty, R\. Y\. Pang, J\. Dirani, J\. Michael, and S\. R\. Bowman \(2024\)GPQA: a graduate\-level google\-proof q&a benchmark\.InFirst Conference on Language Modeling,External Links:[Link](https://openreview.net/forum?id=Ti67584b98)Cited by:[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.3.1),[§4](https://arxiv.org/html/2608.11247#S4.SS0.SSS0.Px2.p1.1)\.
- M\. Sharma, M\. Tong, T\. Korbak, D\. Duvenaud, A\. Askell, S\. R\. Bowman, N\. Cheng, E\. Durmus, Z\. Hatfield\-Dodds, S\. R\. Johnston, S\. Kravec, T\. Maxwell, S\. McCandlish, K\. Ndousse, O\. Rausch, N\. Schiefer, D\. Yan, A\. Zou, and E\. Perez \(2023\)Towards understanding sycophancy in language models\.arXiv preprint arXiv:2310\.13548\.External Links:[Link](https://arxiv.org/abs/2310.13548)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p3.1)\.
- D\. Sperber, F\. Clément, C\. Heintz, O\. Mascaro, H\. Mercier, G\. Origgi, and D\. Wilson \(2010\)Epistemic vigilance\.Mind & Language25\(4\),pp\. 359–393\.External Links:[Document](https://dx.doi.org/10.1111/j.1468-0017.2010.01394.x)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p2.1)\.
- M\. Turpin, J\. Michael, E\. Perez, and S\. R\. Bowman \(2023\)Language models don’t always say what they think: unfaithful explanations in chain\-of\-thought prompting\.InProceedings of the 37th International Conference on Neural Information Processing Systems,NIPS ’23,Red Hook, NY, USA\.External Links:[Link](https://dl.acm.org/doi/10.5555/3666122.3669397)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p4.1)\.
- A\. Wan, E\. Wallace, and D\. Klein \(2024\)What evidence do language models find convincing?\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 7468–7484\.External Links:[Link](https://aclanthology.org/2024.acl-long.403/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.403)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p4.1)\.
- J\. Wei, N\. Karina, H\. W\. Chung, Y\. J\. Jiao, S\. Papay, A\. Glaese, J\. Schulman, and W\. Fedus \(2024\)Measuring short\-form factuality in large language models\.arXiv preprint arXiv:2411\.04368\.External Links:[Link](https://arxiv.org/abs/2411.04368)Cited by:[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.33.30.1),[§4](https://arxiv.org/html/2608.11247#S4.SS0.SSS0.Px2.p1.1)\.
- J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, B\. Ichter, F\. Xia, E\. H\. Chi, Q\. V\. Le, and D\. Zhou \(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.InProceedings of the 36th International Conference on Neural Information Processing Systems,NIPS ’22,Red Hook, NY, USA\.External Links:ISBN 9781713871088,[Link](https://dl.acm.org/doi/10.5555/3600270.3602070)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p4.1)\.
- J\. Wei, D\. Huang, Y\. Lu, D\. Zhou, and Q\. V\. Le \(2023\)Simple synthetic data reduces sycophancy in large language models\.arXiv preprint arXiv:2308\.03958\.External Links:[Link](https://arxiv.org/abs/2308.03958)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p3.1)\.
- Z\. Weng, G\. Chen, and W\. Wang \(2025\)Do as we do, not as you think: the conformity of large language models\.InThe Thirteenth International Conference on Learning Representations \(ICLR\),External Links:[Link](https://openreview.net/forum?id=st77ShxP1K)Cited by:[Table 6](https://arxiv.org/html/2608.11247#A3.T6),[Table 8](https://arxiv.org/html/2608.11247#A4.T8),[§1](https://arxiv.org/html/2608.11247#S1.p3.1),[§2](https://arxiv.org/html/2608.11247#S2.p1.1),[§4](https://arxiv.org/html/2608.11247#S4.SS0.SSS0.Px3.p1.1),[Table 2](https://arxiv.org/html/2608.11247#S5.T2)\.
- R\. H\. Willis \(1965\)Conformity, independence, and anticonformity\.Human Relations18\(4\),pp\. 373–388\.External Links:[Document](https://dx.doi.org/10.1177/001872676501800406)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p2.1)\.
- J\. Xie, K\. Zhang, J\. Chen, R\. Lou, and Y\. Su \(2024\)Adaptive chameleon or stubborn sloth: revealing the behavior of large language models in knowledge conflicts\.InProceedings of the 12th International Conference on Learning Representations \(ICLR\),External Links:[Link](https://arxiv.org/abs/2305.13300)Cited by:[§2](https://arxiv.org/html/2608.11247#S2.p4.1)\.
- X\. Zhu, C\. Zhang, T\. Stafford, N\. Collier, and A\. Vlachos \(2025\)Conformity in large language models\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 3854–3872\.External Links:[Link](https://aclanthology.org/2025.acl-long.195/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.195),ISBN 979\-8\-89176\-251\-0Cited by:[Table 6](https://arxiv.org/html/2608.11247#A3.T6),[Table 8](https://arxiv.org/html/2608.11247#A4.T8),[§1](https://arxiv.org/html/2608.11247#S1.p3.1),[§2](https://arxiv.org/html/2608.11247#S2.p1.1),[§3](https://arxiv.org/html/2608.11247#S3.p1.1),[§4](https://arxiv.org/html/2608.11247#S4.SS0.SSS0.Px3.p1.1),[Table 2](https://arxiv.org/html/2608.11247#S5.T2)\.
- J\. Zuo, M\. Velikanov, I\. Chahed, Y\. Belkada, D\. E\. Rhayem, G\. Kunsch, H\. Hacid, H\. Yous, B\. Farhat, I\. Khadraoui, M\. Farooq, G\. Campesan, R\. Cojocaru, Y\. Djilali, S\. Hu, I\. Chaabane, P\. Khanna, M\. E\. A\. Seddik, N\. D\. Huynh, P\. L\. Khac, L\. AlQadi, B\. Mokeddem, M\. Chami, A\. Abubaker, M\. Lubinets, K\. Piskorski, and S\. Frikha \(2025\)Falcon\-h1: a family of hybrid\-head language models redefining efficiency and performance\.External Links:2507\.22448,[Link](https://arxiv.org/abs/2507.22448)Cited by:[Table 4](https://arxiv.org/html/2608.11247#A1.T4.1.23.22.1),[Table 13](https://arxiv.org/html/2608.11247#A8.T13.3.26.23.1)\.

## Appendix AThe Model Pool

Table 4:The 23 subject models, given by the exact checkpoint served and ordered by MMLU conformity\. Checkpoint names are HuggingFace repository names and Family the publishing organization that prefixes them\. The three rightmost columns give conformity under a unanimous wrong group \(%\)\. Params gives total parameters, with active parameters after the slash where the two differ\. Hybrid denotes a hybrid state\-space design; DeepSeek\-V2\-Lite\-Chat is the one model using multi\-head latent attention\.Table[4](https://arxiv.org/html/2608.11247#A1.T4)names the 23 subjects and gives each one’s conformity rate on the three datasets\.Llama\-3\.3\-70B\-Instruct\-FP8\-dynamic\(Meta\) grades the SimpleQA responses and is held out of the pool, so it shares a family with nothing it judges\. Two features of the table bear on every pooled number in the paper\. Conformity covers most of the available range on all three datasets, so the means average over models that differ from each other far more than any method moves them, and the MMLU ordering of the rows tracks the GPQA ordering but not the SimpleQA one\.

## Appendix BConditions and Prompt Templates

Two things vary between the 19 conditions, and nothing else does\. The peer block changes and one instruction is appended, while the system prompt, the question block, and the answer instruction stay fixed, as Figure[6](https://arxiv.org/html/2608.11247#A2.F6)shows with all three insertion points marked\. Condition 1 shows no peers and defines the resist and adopt pools, conditions 2–6 raise the pressure on the resist pool from one wrong peer to a unanimous wrong group of four, and condition 7 is the Receptivity probe\. The remaining 12 are the six mitigations at both polarities\. Table[5](https://arxiv.org/html/2608.11247#A2.T5)gives the exact wording of all 19\.

```
[system]
You are a knowledgeable assistant. Answer
the question as accurately as you can.
      <- Empowered Persona appends here

[user]
Question: <question>
A) <option A>    B) <option B>
C) <option C>    D) <option D>

Four other AI assistants answered this
question:
- Assistant 1: "The answer is A."
- Assistant 2: "I believe it is A."
- Assistant 3: "This is A."
- Assistant 4: "A is correct."
      <- peer block varies by condition

What is your answer?
      <- method instruction appends here
Respond with the option letter only
(A, B, C, or D).
```

Figure 6:The shared prompt, shown for a multiple\-choice question underunanimous\_barewith A as the asserted answer\. The asserted option is a fixed wrong option chosen by anmd5hash of the question ID at resist polarity and the correct option at adopt polarity\. Free\-form questions drop the option list and close with “Respond with a short factual answer only\.”#### Decoding\.

Snap conditions constrain the multiple\-choice answer to one of the four letters by structured decoding and cap free\-form answers at 96 tokens, so the committed answer carries no deliberation\. The reasoning condition generates freely up to 2,048 tokens, and the answer is parsed from the end of the generation\. The one exception is gpt\-oss\-20b, which always emits a reasoning channel and is therefore given the same 2,048 unconstrained tokens everywhere\.

Table 5:The 19 conditions\. The seven above the rule run once; the six below it run at both polarities,\_resistagainst a unanimous wrong group and\_adoptagainst a unanimous correct one, for 12 more\. “Unanimous” means four peers asserting the same answer, wrong at resist polarity and correct at adopt polarity\. The Devil’s Advocate dissenter asserts an answer differing from both the other peers’ and the correct one, so unanimity breaks without the prompt naming the correct answer\. All instructions append to the user turn except Empowered Persona, which appends to the system prompt\.

## Appendix CMethod Coordinates

Table[6](https://arxiv.org/html/2608.11247#A3.T6)gives the coordinates plotted in Figure[5](https://arxiv.org/html/2608.11247#S5.F5), first on the matched questions used throughout Section[5](https://arxiv.org/html/2608.11247#S5)and then on the full resist and adopt pools\. Matching restricts every snap condition to the question IDs drawn for that model’s reasoning sample, so the two blocks differ only in which questions they average over\. No snap value moves by more than 3\.3 points between them, and most move by less than one, and the ordering of the six snap methods survives on every axis apart from a tie between Empowered Persona and Anchored Reconsideration on SimpleQA Resistance\. Reasoning runs on samples by construction and repeats its sampled values in both blocks, so nothing in Section[5\.3](https://arxiv.org/html/2608.11247#S5.SS3)rests on the matched restriction\.

Table 6:Resistance and Receptivity, %, mean over the 23 models\. \[Z\]Zhuet al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib1)\), \[W\]Wenget al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib18)\)\.
## Appendix DBootstrap Intervals for the Main\-Text Tables

Tables[7](https://arxiv.org/html/2608.11247#A4.T7)and[8](https://arxiv.org/html/2608.11247#A4.T8)give 95% percentile intervals for every entry of Tables[1](https://arxiv.org/html/2608.11247#S5.T1)and[2](https://arxiv.org/html/2608.11247#S5.T2), both taken from the single cluster bootstrap of Section[4](https://arxiv.org/html/2608.11247#S4)\. The intervals on the conformity rates are wide because models differ widely in how much they conform; those on the method effects are narrower because each draw differences a model against itself\.

Table 7:95% bootstrap intervals for the conformity rates of Table[1](https://arxiv.org/html/2608.11247#S5.T1)\.Table 8:95% cluster\-bootstrap intervals for the paired per\-model changes of Table[2](https://arxiv.org/html/2608.11247#S5.T2)\. \[Z\]Zhuet al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib1)\), \[W\]Wenget al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib18)\)\.
## Appendix EThe Per\-Model Record

Every pooled number in this paper averages over models that differ by more than any intervention moves them, so this section gives the disaggregated record behind those numbers\. The pressure conditions are given model by model in Figure[8](https://arxiv.org/html/2608.11247#A7.F8), the figure that closes the paper, which reproduces the gradient of Section[5\.1](https://arxiv.org/html/2608.11247#S5.SS1)inside almost every panel and so rules out an averaging artifact\. What follows takes up architecture, the consistency of each mitigation across models, and the competence relation\.

#### Architecture\.

Grouping the unanimous\-bare rates by architecture gives 22\.7% for the 12 dense models \(sd 18\.1\), 24\.7% for the 9 mixture\-of\-experts models \(sd 28\.2\) and 14\.7% for the two hybrids on MMLU; 57\.8%, 54\.0% and 39\.8% on GPQA; and 72\.3%, 68\.6% and 73\.2% on SimpleQA\. No gap is large besides the standard deviation that accompanies it, the hybrid mean rests on two models, and architecture is entangled with training era, since all but two of the mixture\-of\-experts models date from 2025 or later\. We find no sign that architecture predicts conformity and regard the question as open\.

#### Method consistency\.

Table[9](https://arxiv.org/html/2608.11247#A5.T9)asks the same question of the mitigations, how often each helps an individual model rather than the pool mean\. Reflection is the most consistent, hurting Resistance in 2 models on the multiple\-choice sets and none on SimpleQA, while Devil’s Advocate improves and degrades Resistance at almost equal rates, which is what leaves its pooled gain unreliable\.

Table 9:Consistency of the Resistance gains of Table[2](https://arxiv.org/html/2608.11247#S5.T2)across models \(full pools\)\. Mean change against no intervention \(Δ¯\\overline\{\\Delta\}, pp\) and the number of the 23 models each method makes*worse*\.
#### Competence\.

A last split asks whether the competence relation of Section[5\.3](https://arxiv.org/html/2608.11247#S5.SS3), in which the two frontier axes already move in opposite directions before any method is applied, survives a control for architecture \(Figure[7](https://arxiv.org/html/2608.11247#A5.F7)\)\. Removing the architecture group means from both variables leaves it at\+0\.67\+0\.67and−0\.44\-0\.44on MMLU and\+0\.56\+0\.56and−0\.34\-0\.34on GPQA, so neither half is an artifact of which architectures happen to score higher\. The Receptivity half is the weaker \(p=0\.045p\{=\}0\.045on MMLU,p=0\.137p\{=\}0\.137on GPQA\), and inside a single architecture group, the largest of which holds 12 models, it is too weakly determined to stand alone, so we rely on the pooled coefficients\.

![Refer to caption](https://arxiv.org/html/2608.11247v1/results/plots/appendix/competence_by_arch.png)Figure 7:Resistance and Receptivity without intervention against baseline accuracy, with least\-squares fits\. Pearsonrris\+0\.67\+0\.67for Resistance and−0\.42\-0\.42for Receptivity on MMLU,\+0\.51\+0\.51and−0\.32\-0\.32on GPQA, and\+0\.06\+0\.06and\+0\.00\+0\.00on SimpleQA\. Color marks the axis and marker shape the architecture\. Both relations are absent on SimpleQA, where accuracy is near floor for every model\.

## Appendix FThe Derivability Split

A subject counts as*derivable*if its answers can typically be computed or derived step by step from information in the question together with general principles, which covers mathematics, calculation in physics and chemistry, formal logic, statistics, accounting, textbook economics and computer science, and as*recall*otherwise\. Table[10](https://arxiv.org/html/2608.11247#A6.T10)gives the assignment of all 57 MMLU subjects behind Table[3](https://arxiv.org/html/2608.11247#S5.T3), 20 derivable and 37 recall\. Because we sorted the subjects after the pattern was visible, two checks follow, on whether the criterion was applied consistently and whether the interaction rests on a small group of subjects\.

Table 10:Assignment of the 57 MMLU subjects for the derivability split\. Superscripts give the two halves used in the split\-half check of Appendix[F\.2](https://arxiv.org/html/2608.11247#A6.SS2), taking every other subject down each alphabetized list, so that half A of the derivable subjects \(10\) pairs with half A of the recall subjects \(19\), and half B \(10\) with half B \(18\)\.### F\.1Blinded replication of the assignment

Applying a written criterion ourselves leaves open whether we applied it consistently or bent it toward the finding\. We gave the criterion verbatim, with no other information about the paper, to a rater with no access to this document, and asked for a classification of all 57 subjects from the alphabetized list alone\. The rater agreed on 56 of 57, Cohen’sκ=0\.96\\kappa\{=\}0\.96, dividing them 21 derivable and 36 recall against our 20 and 37\. The one disagreement is astronomy, which we assigned to recall and the blinded pass to derivable, defensible either way since astronomy questions mix recalled facts with occasional calculation\. The check establishes the consistency of the rubric rather than its independence, since the blinded pass classified the same subjects of the same benchmark, and it does not substitute for the new pool or dataset that Section[8](https://arxiv.org/html/2608.11247#S8)identifies as the confirmatory check\.

### F\.2Split\-half replication of the interaction

We cut each subject list into two interleaved halves, marked in Table[10](https://arxiv.org/html/2608.11247#A6.T10), by taking every other subject down the alphabetized list, so that each half spans the same range of the alphabet and of topic rather than one half absorbing every subject with mathematics in its name\. Recomputing the derivable\-minus\-recall interaction inside each half recovers it in both, within two points of the full estimate and excluding zero on its own \(Table[11](https://arxiv.org/html/2608.11247#A6.T11)\), so no small group of subjects carries it\.

Table 11:Split\-half replication of the derivable\-minus\-recall interaction \(reasoning against no intervention, paired per model, 95% CI\)\. The interaction excludes zero in the full sample and in each half on its own\.

## Appendix GRobustness of the Reasoning Gap to the Model Pool

![Refer to caption](https://arxiv.org/html/2608.11247v1/results/plots/appendix/model_pressure_gradient.png)Figure 8:The complete per\-model record\. Each panel is one model and gives its conformity on MMLU, GPQA and SimpleQA across the five pressure conditions of Appendix[B](https://arxiv.org/html/2608.11247#A2)\(one wrong peer; three wrong with one correct dissenter; four wrong asserting bare, certain or reasoned\), that is all 345 model×\\timesdataset×\\timescondition cells that the pooled figures summarize\. Panels run from the most to the least conforming by mean rate under a unanimous wrong group, the ordering of Figure[4](https://arxiv.org/html/2608.11247#S5.F4)\. The vertical scale is shared by all panels\.Section[5\.4](https://arxiv.org/html/2608.11247#S5.SS4)pools the frontier gap over all 23 subject models, which raises a question the subject\-boundary checks do not answer, whether the null on GPQA and SimpleQA is a property of the pool or the work of a few weak models pulling the average down\. We recomputed the gap after excluding every subset of one, two or three models, 23, 253 and 1,771 subsets in all, each time refitting the six\-method snap frontier and the reasoning point on the models that remain\. No single exclusion turns both gaps positive, nor does any two\-model exclusion, the best of which still leaves GPQA at−0\.3\-0\.3\. Of the 1,771 three\-model exclusions, only 3 turn both positive, and even the best of those, dropping Yi\-1\.5\-34B\-Chat, gemma\-4\-26B\-A4B\-it and Kimi\-Linear\-48B\-A3B\-Instruct together, leaves both intervals including zero \(Table[12](https://arxiv.org/html/2608.11247#A7.T12)\)\.

Table 12:Frontier gap on GPQA and SimpleQA \(pp, 95% CI\) after removing the model subset, of the given size, that maximizes the smaller of the two gaps\.We separately checked OLMo\-2\-0325\-32B\-Instruct, the one model whose reasoning condition is known not to work as intended, since it follows the instruction to answer with the option letter alone and never produces a chain of thought, while every other reasoning response runs to 1,700–2,200 characters\. Dropping it moves the GPQA and SimpleQA gaps by\+0\.01\+0\.01and\+0\.18\+0\.18points, far short of what either search above required\. The null on GPQA and SimpleQA is a property of the pool as a whole and holds whichever handful of models one sets aside, which is what the derivability account of Section[5\.4](https://arxiv.org/html/2608.11247#S5.SS4)predicts for datasets on which this pool cannot complete a derivation\.

## Appendix HThe Complete Per\-Model Record

Figure[8](https://arxiv.org/html/2608.11247#A7.F8)disaggregates the record completely, one panel per model\. Three regularities survive the disaggregation, and one does not\. A unanimous bare majority elicits more conformity than a single dissenter in 21 of the 23 models on MMLU, 22 on GPQA, and all 23 on SimpleQA; asserted certainty lowers conformity relative to a bare assertion in 21, 21, and all 23; and task difficulty orders the three datasets within almost every panel, GPQA above MMLU for all 23 models and SimpleQA above MMLU for 22\. Peer reasoning raises conformity above a bare assertion in only 14, 12, and 16 models, so the cost of an argued wrong answer, unlike the cost of group size, is not shared across the pool\. The panels also put the between\-model spread on one scale, from 1\.8% for Yi\-1\.5\-34B\-Chat to 83\.7% for DeepSeek\-V2\-Lite\-Chat on MMLU, wider than any change a mitigation induces\.

ResourceLicense*Apache License 2\.0 — 13 models*gemma\-4\-26B\-A4B\-itGemmaet al\.\([2026](https://arxiv.org/html/2608.11247#bib.bib30)\)Apache 2\.0Mistral\-Small\-3\.2\-24B\-Instruct\-2506mistralai \([2025](https://arxiv.org/html/2608.11247#bib.bib42)\)Apache 2\.0Mixtral\-8x7B\-Instruct\-v0\.1Jianget al\.\([2024](https://arxiv.org/html/2608.11247#bib.bib25)\)Apache 2\.0Qwen3\-32BQwen Team \([2025](https://arxiv.org/html/2608.11247#bib.bib36)\)Apache 2\.0Qwen3\.6\-35B\-A3BQwen Team \([2026](https://arxiv.org/html/2608.11247#bib.bib47)\)Apache 2\.0gpt\-oss\-20bOpenAIet al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib34)\)Apache 2\.0OLMo\-2\-0325\-32B\-InstructOLMoet al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib26)\)Apache 2\.0Seed\-OSS\-36B\-InstructByteDance\-Seed \([2025](https://arxiv.org/html/2608.11247#bib.bib37)\)Apache 2\.0granite\-4\.1\-30bIBM Research \([2026](https://arxiv.org/html/2608.11247#bib.bib41)\)Apache 2\.0ERNIE\-4\.5\-21B\-A3B\-PTbaidu \([2025](https://arxiv.org/html/2608.11247#bib.bib45)\)Apache 2\.0AI21\-Jamba2\-Mini\-FP8ai21labs \([2025](https://arxiv.org/html/2608.11247#bib.bib39)\)Apache 2\.0Apertus\-70B\-Instruct\-2509\-FP8\-dynamicApertuset al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib29)\)Apache 2\.0Yi\-1\.5\-34B\-Chat01\-ai \([2024](https://arxiv.org/html/2608.11247#bib.bib48)\)Apache 2\.0*MIT License — 4 models*phi\-4Abdinet al\.\([2024](https://arxiv.org/html/2608.11247#bib.bib32)\)MITGLM\-4\-32B\-0414GLMet al\.\([2024](https://arxiv.org/html/2608.11247#bib.bib33)\)MITMoonlight\-16B\-A3B\-Instructmoonshotai \([2025b](https://arxiv.org/html/2608.11247#bib.bib40)\)MITKimi\-Linear\-48B\-A3B\-Instructmoonshotai \([2025a](https://arxiv.org/html/2608.11247#bib.bib46)\)MIT*Own license — 6 models*gemma\-3\-27b\-it†Gemmaet al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib31)\)Gemma Terms of UseNVIDIA\-Nemotron\-3\-Nano\-30B\-A3B\-BF16NVIDIAet al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib27)\)NVIDIA Open Model LicenseFalcon\-H1\-34B\-InstructZuoet al\.\([2025](https://arxiv.org/html/2608.11247#bib.bib35)\)Falcon LLM License \(TII\)EXAONE\-4\.0\-32BBaeet al\.\([2026](https://arxiv.org/html/2608.11247#bib.bib28)\)EXAONE AI Model LicenseDeepSeek\-V2\-Lite\-ChatDeepSeek\-AIet al\.\([2024](https://arxiv.org/html/2608.11247#bib.bib24)\)DeepSeek Licensec4ai\-command\-r\-v01†CohereLabs \([2024](https://arxiv.org/html/2608.11247#bib.bib38)\)CC\-BY\-NC 4\.0*Judge model*Llama\-3\.3\-70B\-Instruct\-FP8\-dynamicMeta \([2024](https://arxiv.org/html/2608.11247#bib.bib43)\)Llama 3\.3 Community License*Datasets — 3*MMLUHendryckset al\.\([2021](https://arxiv.org/html/2608.11247#bib.bib17)\)MITGPQA\-Main†Reinet al\.\([2024](https://arxiv.org/html/2608.11247#bib.bib19)\)CC\-BY 4\.0SimpleQAWeiet al\.\([2024](https://arxiv.org/html/2608.11247#bib.bib20)\)MITTable 13:Licenses for every model and dataset used in this paper\.†marks the three resources gated on Hugging Face \(gemma\-3\-27b\-it,c4ai\-command\-r\-v01, and GPQA\-Main\), which require accepting the publisher’s usage terms before download; we requested and were granted access to each under our institutional research use\.
## Appendix IAI Assistance in Research and Writing

We used AI tools to assist with code generation, debugging, data analysis, spell\-checking, formatting, and grammatical editing\. Specifically, we used Anthropic’sClaude Sonnet 5andClaude Opus 4\.8\.

## Appendix JLicenses

Table[13](https://arxiv.org/html/2608.11247#A8.T13)lists the license of every resource used in this paper: the 23 subject models, the judge model, and the three evaluation datasets\. All are used solely for inference and evaluation, consistent with each license’s terms, and no model weights were fine\-tuned, modified, or redistributed\.