Silent Dissent: LLM Agents That Yield to the Majority Still Represent Their Original Premise

Ziang Ni; Peng Zou

cs.CLOct 2, 2026

Silent Dissent: LLM Agents That Yield to the Majority Still Represent Their Original Premise

Authors: Ziang Ni, Peng Zou

Organizations: Sun Yat-sen University Guangzhou, China

Abstract

Multi-agent debate is increasingly used to reach consensus among LLM agents, yet agents often yield to a unanimous majority. When an agent changes its answer, has it changed its mind or only its statement? We study this with two-hop factual questions whose intermediate entity (the bridge, e.g. the country in "the capital of the country where the Sagrada Familia is located") is never stated by anyone. Scripted peers, in the role of Asch's confederates, unanimously assert a wrong answer taken from another fact with a different bridge. At the moment the agent answers, we read the bridge from its residual stream with the Jacobian lens (J-lens) and, for comparison, the logit lens. In pre-registered tests on held-out facts with four open-weight models, agents of Qwen3.5-4B, Qwen3.6-27B and Gemma-4-E4B-it that gave in still represented their original bridge in the pre-registered layers below the output (hit@100 above a control entity: 0.85, 0.22 and 0.24), where the logit lens rarely ranked it among the top 100 tokens (0.00-0.06). These agents also represented the bridge behind the peers' answer, beyond a mention baseline. A pre-registered addendum hid the agent's earlier answer or removed it: agents that gave in still represented their original bridge in all four models (0.43, 0.29, 0.37 and 0.25 with the answer hidden), including Llama-3.1-8B-Instruct, which barely did so with its answer in view (0.03). The premise can thus be computed from the question alone while the agent states the majority's answer. Hiding the earlier answer also changed conformity: Qwen3.5-4B gave in on 89% of questions instead of 8%. In exploratory interventions, injecting the bridge's J-lens direction brought agents back to their original answer only in the two Qwen models. Stated consensus in multi-agent debate can thus overstate agreement. We also report the negative results of our pre-registered program.

Figures & tables

Panel of Figure 1 — Figure 1. The paradigm (illustrative, not a recorded transcript; plain format, abridged). Neither bridge is written anywhere in the transcript. Peers can instead attack the first hop (“The country where the Sagrada Familia is located is Italy. The capital city of Italy is Rome.”) or the second hop (right bridge, wrong capital). Transcript of a two-round debate and the lens readout at the agent's answer position.

Model	Rule	Band / $L$	Cat.	Items	ans.	hop 1	hop 2
					Gave in (%)
Qwen3.5-4B	v1	23–30 / 32	L, B, U	718/668	8	8	0
Qwen3.6-27B	v2	52–55 / 64	L, B, P c	514/476	39	49	6
Gemma-4-E4B	v1	32, 34–40 / 42	L, B, U	533/476	25	52	24
Llama-3.1-8B ∗	v2	17–23 / 32	L, B, P n	570/542	90	97	8

Table 1. Setup on the test split. Band /

L

: pre-registered layers / number of layers. Items: test items / items entering round 1 (answered correctly in round 0). Gave in: share of agents stating the peers’ answer in round 1 with three peers. Categories (first entity

\to

bridge

\to

answer): L landmark

\to

country

\to

capital; B person

\to

birth city

\to

country; U university

\to

city

\to

country; P c , P n person

\to

undergraduate university

\to

its city or country.

Model	H1 retention	H2 peers’ bridge	H3 hop 2 $-$ hop 1	H4 J-lens $-$ logit lens
Qwen3.5-4B	0.852 [0.782, 0.917], $n{=}54$	0.373 [0.261, 0.491], $55{\times}661$	untestable ( $0{\times}53$ )	0.792 [0.722, 0.857]
Qwen3.6-27B †	0.223 [0.174, 0.280], $n{=}168$	0.126 [0.080, 0.173], $172{\times}448$	0.323 [0.115, 0.533], $12{\times}213$	0.223 [0.174, 0.280]
Gemma-4-E4B-it	0.237 [0.175, 0.299], $n{=}120$	0.211 [0.163, 0.258], $120{\times}475$	0.555 [0.476, 0.631], $112{\times}244$	0.192 [0.137, 0.248]
Llama-3.1-8B-Instruct †	0.034 [0.020, 0.048], $n{=}485$	0.027 [0.014, 0.041], $487{\times}540$	0.400 [0.260, 0.533], $44{\times}521$	0.030 [0.018, 0.043]

Table 2. Pre-registered hypotheses on the test split: estimate, 95% bootstrap interval and number of records (H2, H3: agents that gave in

\times

comparison group). H1: original-bridge readout of agents that gave in; H2: their peer-bridge readout minus that of mention-condition agents; H3: original-bridge readout under second- minus first-hop pressure; H4: H1 under the J-lens minus under the logit lens, paired. Every interval excludes zero. † Band chosen by rule v2 for a J-lens advantage on development data, so H4 tests whether the advantage persists in test agents that gave in. For Qwen3.6-27B the logit lens reads 0.000 on these records, so H4 equals H1. With intervals clustered by bridge and a Holm correction, all but Llama’s H2 remain supported (§ 5 ).

Figure 2. Original-bridge readout in the pre-registered band (hit@100 minus control; mean, 95% CI, records under each label) for agents answering alone in round 0, agents whose peers agreed, agents that held under answer pressure, and agents that gave in. Round 0 and agree include all items; held and gave in split the pressure items by outcome, so the conditions are different item sets. Within each condition the two lenses are compared on the same records. Four panels, one per model, with J-lens and logit-lens readouts across four conditions.

Model	Gave in (%): vis./hid./abs.	O1 hidden	O3 absent	O2 hidden $-$ absent	O4 J-lens $-$ logit
Qwen3.5-4B	8 / 89 / 62	0.434 [0.402, 0.467], $n{=}591$	0.500 [0.462, 0.541], $n{=}408$	$-0.066$ [ $-0.117$ , $-0.015$ ]	0.383 [0.349, 0.417]
Qwen3.6-27B	39 / 73 / 77	0.290 [0.241, 0.339], $n{=}321$	0.281 [0.240, 0.330], $n{=}339$	0.009 [ $-0.060$ , 0.075]	0.278 [0.231, 0.327]
Gemma-4-E4B-it	25 / 36 / 43	0.365 [0.307, 0.421], $n{=}170$	0.212 [0.166, 0.259], $n{=}203$	0.154 [0.081, 0.230]	0.215 [0.171, 0.259]
Llama-3.1-8B-Instruct	90 / 71 / 60	0.254 [0.220, 0.292], $n{=}382$	0.391 [0.348, 0.433], $n{=}320$	$-0.136$ [ $-0.192$ , $-0.081$ ]	0.205 [0.175, 0.238]

Table 3. Addendum: round 1 with the agent’s round-0 answer hidden or absent (three peers, answer pressure, test split). Gave in: share of agents stating the peers’ answer, with the answer visible (the main runs), hidden and absent. O1, O3: original-bridge readout of agents that gave in under hidden and absent (J-lens, 95% bootstrap interval, records); O2: hidden minus absent; O4: J-lens minus logit lens under hidden, paired. Supported: O1, O3 and O4 in every model, O2 only for Gemma.

Figure 3. Original-bridge readout per layer (answer pressure, three peers, round 1, test split) for agents that gave in (solid) and agents whose peers agreed (dashed), under the J-lens (blue) and the logit lens (orange). Shaded: the pre-registered band (for Qwen3.6-27B and Llama chosen for a J-lens advantage on development data); the last layer is the model output. Four panels of per-layer readout curves, J-lens and logit lens, for agents that gave in and agents whose peers agreed.

Figure 4. Injection into agents that gave in (exploratory): share whose most likely next token is the first token of the original answer after adding

\alpha\lVert h\rVert d

along each direction at one band layer, pooled over three band layers. Bands: 95% Wilson intervals counting each agent once. Four panels of dose-response curves.

Explore similar work

Sep 29, 2026physics.soc-ph

Multi-agent discussion gains less when dissent is withheld

Multi-agent systems of LLMs add discussion to majority voting and are therefore expected to be more capable. However, empirical reports conflict on whether discussion improves accuracy or leads to an incorrect consensus. Here, we introduce a parsimonious model that explains when discussion improves accuracy and when it ends in an incorrect consensus, built from four behaviors repeatedly observed in LLM agents: (1) withholding dissent, (2) internalizing a stated answer, (3) reconsidering after seeing dissent, and (4) correcting toward the correct answer. The model shows that discussion can overturn an incorrect initial majority only when the withholding rate

c

is below a critical rate

c^* = γ/(γ+ a)

, set by the net correction rate

γ

and the internalization rate

a

. We estimate these rates from conversation logs with a Bayesian method and place LLM teams relative to

c^*

. As the model predicts, the gain from discussion shrinks as withholding rises, across LLMs and on a hidden profile benchmark, HiddenBench, and MedEInst. Instructing agents not to withhold dissent increases this gain. Turning reasoning off also increases the gain, because reasoning raises the internalization rate

a

and keeps agents from reconsidering a minority answer. These findings reconcile the conflicting reports and identify when discussion outperforms majority voting.

Chand Sahil Mansuri, Xin Wang, Mengying Li +4

School of Systems Science and Industrial Engineering, Binghamton University, Binghamton, NY, USA · School of Management, Binghamton University, Binghamton, NY, USA · IBM T. J. Watson Research Center, Yorktown Heights, NY, USA

Disagreement Majority Voting

Ask PDF

Sep 7, 2026cs.AI

What Does Multi-Agent Debate Actually Change?

Multi-agent debate, in which several LLMs exchange arguments before producing an answer, raises a basic question: does expressed disagreement reflect changes in the members' own positions? No single signal can settle this question, so we organize the analysis around five questions: (A) does the debater say it disagrees; (B) does its reply text actually argue; (C) does its own position change after each debate turn; (D) how much, quantitatively, does the position change; and (E) how do members' final positions compare with their initial ones? We evaluate two- and three-member committees on 50 curated opinion questions from GlobalOpinionQA, using same-model, same-family, and mixed-family configurations with friendly, neutral, and hostile instructions assigned at each turn. (A) Tone changes reported agreement: the share of replies reporting strong agreement is 70.7-96.3% under friendly instructions, compared with 9.5-19.3% under hostile ones, varying with model choice. (B) Self-reports broadly align with text judgments, but consistency varies by agreement level and model choice. (C) Position changes depend on the interaction: after a peer's leaning-disagree reply, members reporting strong agreement switch options more often than those reporting leaning disagreement. (D) For open-weight members, the probability of the option a member already holds stays near saturation, even after a peer's pushback, while endorsement of its earlier position text drops after a peer's argument relative to neutral filler, more for Qwen3.8-27B than Inkling. (E) The selected option is unchanged from members' initial to final positions in over 90% of comparisons in every configuration. Taken together, expressed disagreement need not translate into position revision, either within individual exchanges or over a complete debate.

Chen Qian

Multi-Agent Debate Large Language Model Decisions Disagreement

Ask PDF

May 30, 2026cs.CL

Not All Flips Are Conformity: Decomposing Stance Convergence in Multi-Agent LLM Debate

Multi-agent debate (MAD) is a promising strategy for improving LLM reasoning, but when agents converge on a shared answer, it is unclear whether that convergence reflects genuine deliberation or social compliance. We show that the conventional answer flip rate conflates three distinct mechanisms: spontaneous instability, stance-induced conformity, and reasoning-induced persuasion. Our three-source decomposition framework isolates each through controlled counterfactual conditions. In the primary MMLU-Pro setting, 37% of agent-question observations change under self-reflection alone, while robustness tests show substantial model-dependent instability across GPQA-Diamond and three model families; strict conformity is 29% in the primary setting and remains predominantly harmful across model replications (57-77% correct-to-wrong). A controlled information-gradient experiment reveals that even vacuous reasoning is associated with 20-39% error adoption among resistant agents, with reasoning-like presentation carrying substantial persuasive weight. Harmful conformity can be predicted from Round 0 features (AUC = 0.79), and risk-targeted intervention reduces it by 13.6 percentage points (p < 0.001). However, without correctness labels or self-reflection controls, reducing peer adoption does not improve accuracy, because harmful and beneficial influence cannot be distinguished.

Xiqi Hao, Zengqing Wu, Yu-Xuan Qiu +4

Beijing Institute of Technology, Zhuhai · University of Osaka · Shenzhen University

Multi-Agent Debate Debate Conformity Persuasion Answer Convergence

Ask PDF