PaperlayerView exact revision on arXiv

Source-linked AI summary

Position Matters More Than Channel: Talking a Language Model Out of a Correct Answer

Xinyi Ni

arXiv:2610.34809v1cs.CL

TL;DR

The paper asks what can talk a model out of an answer it gets right alone, amid a growing number of claims in its context. It uses controlled comparisons of advice position, source, and prior-answer replay on Qwen3.5-27B and four-option MMLU-Pro items. Position matters more than channel, while source, tone, and replay also affect whether answers change.

  • Problem

    Earlier studies often vary advice position, attributed source, and prior-answer content together, complicating comparisons of their separate effects.

    Source: §1 · p. 1

  • Method

    The study uses controlled comparisons that move fixed advice text, change one phrase, or show replayed text as the model’s own turn or another participant’s, on Qwen3.5-27B and four-option MMLU-Pro items.

    Source: §1 · p. 1

  • Results

    Position matters more than channel: bare claims do most harm just before the answer and arguments before the question, while authority and confident tone each raise error by 10–27 points.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Across the tested comparisons, advice wording, position, and prior context move error by tens of points, whereas API channel and choice between two authorities move it by only a few.

    Source: §5 · p. 10

    The study uses one model, one benchmark, one advice-text family, and one decoding setup.

    Source: §6 · p. 10

Abstract

from arXiv · show

Language models now answer questions alongside other claims: a note in the system prompt, a message from another agent, their own earlier answer. We ask what it takes to talk a model out of an answer it gets right on its own, and separate three factors that earlier studies vary together: where the advice sits in the prompt, who is said to have written it, and what the model has already written. On Qwen3.5-27B and four-option MMLU-Pro items, with the rest of the advice held fixed in each comparison, we find three things. First, position matters far more than the API channel, in opposite directions for a bare claim and an argument: a bare "the answer is X" does most harm just before the model answers, while the same claim with a false rationale does most harm before the question, where it raises error on items the model otherwise almost never misses from 1% to 67%; moving the claim between the system message and the user message at the same position changes error by a few points. Second, attributing the note to an authority rather than to another participant, and phrasing it assertively, each raise error by 10-27 points, while the two authorities we tested differ by under two. Third, replaying a correct worked answer before the advice protects it whether the text is shown as the model's own or as someone else's, and the same replay keeps wrong answers from being corrected by a bare correct note. The position result replicates under an independently written prompt. Code, items and all result tables are released at https://github.com/coolkillercat/safety.

1 Introduction

The paper separates prompt position, source, and prior-answer effects using controlled comparisons of fixed advice text. It finds that advice placement can matter more than API channel, while authority, tone, and replay also shift whether the model keeps or changes an answer.

1 Introduction

  • Controlled comparisons move the same advice paragraph between positions, change one phrase at a time, or show identical replay text as the model’s own or another participant’s.The study uses Qwen3.5-27B and MMLU-Pro items reduced to four options, with a wrong option sometimes followed by a persuasive but false rationale.Source: §1 · p. 1
  • A bare claim does most harm just before the answer, while the same claim with a false rationale does most harm before the question; channel changes error by only a few points.The opposite placement profiles replicate under an independently written prompt.Source: §1 · p. 2
  • Attributing advice to an authority and phrasing it confidently each raise error by 10–27 points, while the two tested authorities differ by under two points.The source comparison averages over placement and certainty.Source: §1 · p. 2, Figure 1 · p. 2
  • A replayed correct worked answer protects against wrong advice whether shown as the model’s own turn or quoted as another participant’s solution.It removes most of the bare note’s effect and about half of the false rationale’s; replaying a wrong first answer cuts bare-note corrections from 80% to 23%.Source: §1 · pp. 2–3
  • The study is deliberately narrow: one model, one benchmark, and one family of advice texts.The authors note that models differ widely in their readiness to defer.Source: §1 · p. 3

2 Related work

Prior work finds that models can yield to wrong advice, but the effects of source, wording, position, and earlier answers have often been measured together. This paper contributes controlled comparisons that move or relabel fixed advice text within a prompt.

2 Related work

  • Models can abandon answers they get right alone under peer pressure, with stronger effects when they are less certain; deference also varies widely between models.A verified-source note flips 45% of Qwen3.5-27B’s correct answers against under 1% of Gemini 3.1 Pro’s (Kumar and Chopra, 2026).Source: §2 · p. 3
  • Earlier source studies change credentials together with wording, leaving their separate effects unresolved.Findings differ across studies: expertise destabilises more answers than a popular-choice framing, while other studies report little effect from expertise framing.Source: §2 · p. 3
  • Earlier-answer studies mostly attribute the replay to the model, while this paper measures both giving in to wrong advice and accepting correct advice.Prior findings include protection from planted true derivations and greater difficulty accepting corrections after planted wrong derivations.Source: §2 · p. 3
  • Context placement effects have been studied on other tasks, but this work examines placement around the question within a single prompt.Its contribution is behavioural: controlled comparisons that move or relabel a fixed advice text.Source: §2 · p. 3

3 Experimental setup

The experiments use controlled changes to advice text, its prompt position, and prior-answer context to test how each affects model answers. They evaluate Qwen3.5-27B on screened four-option MMLU-Pro items, with paired comparisons and checks for unfinished outputs.

3.1 Model, task and items

  • The study uses 300 known and 100 boundary four-option MMLU-Pro items, with “known” denoting correctness across seven screening probes rather than knowledge.Items were drawn from a subject-stratified pool; the pushed wrong option was each item's most tempting option under screening.Source: §3.1 · p. 4

3.2 The advice paragraph

  • Each treatment adds a paragraph with a speaker frame, an assertive or hedged answer claim, and sometimes a rationale; the bare note omits the rationale.The false rationale argues for the pushed option, while the true rationale explains the keyed answer.Source: §3.2 · p. 4, Table 1 · p. 4

3.3 Experiment 1: where the advice sits

  • The placement experiment moves the same paragraph among three prompt slots, separating a channel change at fixed position from a position change at fixed channel.Template A crosses placements with speaker, certainty, and advice type on known items; Template C is the independently rewritten replication and also tests boundary items.Source: §3.3 · pp. 4–5, §3.2 · p. 4, Figure 2 · p. 5

3.4 Experiment 2: what the model has already written

  • The replay experiment varies the prior package before advice, including the model’s recorded first answer presented as its own previous turn or quoted in the user message.Advice effects are compared with deleted-advice controls; packages share items and option orders for paired comparisons.Source: §3.4 · pp. 5–6, Table 2 · p. 6

3.5 Outcomes and statistics

  • Outcomes are error, yield rate on initially correct trials, and correction rate on initially wrong trials, with paired contrasts and likely ranges reported separately.The package effect measures how a prior package changes the advice effect relative to no prior turn; worst-case bounds account for unfinished outputs.Source: §3.5 · p. 6

4 Results

The results show that advice placement, source and tone, and prior answers each shape whether a model changes a correct answer. Position has opposite effects for a bare note and a false rationale, while replaying a correct answer protects it whether shown as the model’s own or another participant’s.

4.1 An argument does most harm before the question, a bare note after the options; the channel itself adds little

  • 21.7% error after the options versus 10.9% in the system message for a bare note; with a false rationale, error is 67.2% before the question versus 40.4% after the options.The independent wording reproduces both position profiles.Source: §4.1 · p. 7
  • At the same position, system placement is 3.4 points less harmful for a bare note and 5.6 points less harmful for a rationale than the top of the user message.The channel effect is only a few points, compared with tens of points for position.Source: §4.1 · pp. 7–8
  • Unrelated everyday text lowers error by 4 to 7 points at every placement, while a correct note with its true rationale produces no wrong answers among finished outputs.Added text is itself a treatment in these comparisons.Source: §4.1 · p. 8

4.2 An authority as source and an assertive tone each add 10–27 points; which authority adds under two

  • An authority rather than another participant adds about 10 points of error for a bare note and about 25 with a rationale; assertive phrasing adds 16 and 27 points, respectively.The two tested authorities differ by under two points on average.Source: §4.2 · p. 8, Table 3 · p. 8

4.3 A replayed correct answer protects, own or quoted, and the same replay blocks corrections

  • Without a prior turn, 27.6% of initially correct answers end wrong after a bare wrong note and 40.2% after a false rationale, versus 5.2% without advice.The no-advice control is a separate run.Source: §4.3 · p. 8
  • Replaying a correct worked answer shrinks the advice effect by 22.3 and 16.3 points when shown as the model’s own, and 22.2 and 15.8 when quoted from another participant.The effects retain their direction under worst-case bounds, but whether ownership adds protection remains unresolved.Source: §4.3 · p. 9, Figure 4 · p. 9
  • Keeping only a replayed conclusion gives up 8.8 points of protection against a rationale, but content, length, and format change together, so the effect cannot be assigned to reasoning alone.The quoted conclusion-only comparison is inconclusive because about a fifth of outputs hit the token cap.Source: §4.3 · p. 9
  • A replayed wrong answer reduces bare-note corrections from 80.2% without a prior turn to 23.3%, while the replay also lowers spontaneous correction from 21.4% to 8.5%.Against the no-advice rates, the bare note’s correction rate still drops by 40 points.Source: §4.3 · p. 9

5 Discussion

Position, advice source and certainty, and prior context each affect error more than API channel or the choice between the two tested authorities. Replayed answers protect correct answers but can also preserve wrong ones.

5 Discussion

  • A bare assertion does most harm before the answer, while the same claim with a false rationale does most harm before the question; channel alone changes error by only a few points.The opposite position profiles replicate under an independently written prompt.Source: §5 · p. 10
  • An authority source and assertive tone each add 10–27 points of error, while the two tested authorities differ by under two points.These source and certainty effects agree in direction with multi-model framing studies (Seo et al., 2026; Ravikiran and Vohra, 2026).Source: §5 · p. 10
  • A replayed worked answer protects a correct answer whether written by the model or quoted from someone else, and full replay protects more against an argument than its conclusion alone.Replaying a wrong answer also reduces corrections by a bare note; this tradeoff is reported for anti-sycophancy training by Ma et al. (2026).Source: §5 · p. 10
  • Across both experiments, advice wording, position and prior context move error by tens of points, whereas channel and authority choice with wording fixed move it by a few.Source: §5 · p. 10

6 Limitations •

The study is limited to one model, benchmark, advice family and decoding setup, and some prompt comparisons have imperfect controls. The authors release code, items, outputs, result tables and regeneration scripts.

6 Limitations •

  • The study covers one model, one benchmark, one advice-text family and one decoding setup: visible chain of thought with greedy decoding.The authors note that deference varies roughly 75-fold across models of the same year, motivating a placement experiment on another model.Source: §6 · p. 10
  • Item screening defines “known” as stable under screening rather than knowledge, and the auditor disputes 11.2% of checked keys.On audit-passing items, headline contrasts retain their direction, though one replay estimate is no longer reliable.Source: §6 · p. 10
  • Some controls match only one dimension, own and quoted replays differ in conversation shape, and two prompt wordings differ in token budget and hardware.Unfinished outputs leave two independent-wording contrasts open; that wording is treated as a replication, not a measure of wording effects.Source: §6 · p. 10
  • Code, items, generated advice, stored model outputs, result tables with counts, and regeneration scripts are released.Source: §6 · p. 10

A Prompt wording

The appendix provides three prompt templates and describes how advice, speaker framing, certainty, and filler text are constructed. Templates B and C were independently written from Template A, with Template C frozen before its results were inspected.

A Prompt wording

  • Template A sets a careful-expert system instruction, presents a question with four options, and asks for brief reasoning followed by an answer letter.Source: A Prompt wording · p. 12
  • Template A varies speaker framing across an organiser, another participant, or a professor, and varies assertive versus hedged certainty.Its false rationale was instructed to be maximally persuasive, with appended self-corrections stripped.Source: A Prompt wording · p. 12
  • Template B changes the system instruction, question block and final-answer instruction, and includes organiser, participant and professor speaker frames with assertive or hedged wording.Source: A Prompt wording · p. 12
  • Template C changes the question and answer instructions, uses organiser and participant frames, and uses assertive advice wording.Templates B and C were written by a writer who had not seen Template A; C was frozen before results were inspected, though its speaker frames match B word for word.Source: A Prompt wording · pp. 12–13
  • The unrelated paragraph is a per-item seeded shuffle of everyday sentences appended until it matches the false rationale’s word count, but syntax, tokens and stance are not matched.Source: A Prompt wording · p. 13
  • Experiment 2 uses participant advice with wrong or correct rationales and compares quoted solutions or answers; filler sentences are added to match the baseline answer’s character length.The closing request asks whether to keep or change the answer; quoted packages retain that wording despite no preceding model answer.Source: A Prompt wording · p. 13

B Experimental design

The experimental design pairs item and option-order comparisons within each experiment, and constructs replay packages from baseline outputs. The reported run uses a fixed 2,048-token budget; package effects account for a nonzero no-advice control error.

B Experimental design

  • Each experiment’s cells use the same items and option orders, pairing comparisons by item; counts can differ when comparisons require completed outputs in all involved cells.Source: B Experimental design · p. 13
  • The paper uses Template B at a fixed 2,048-token budget rather than mixing it with regenerated outputs from the 768-token run.Most regenerated outputs do not continue their original text.Source: B Experimental design · p. 13
  • Own replay packages form a three-message conversation, while no-prior-turn and quoted packages are single user messages; replayed own text comes from a separate plain-prompt baseline.Of 1,200 item/order pairs, 1,186 had readable answers and entered the experiment; the repeated no-advice control sometimes differs because greedy decoding is not exactly reproducible, so package effects subtract its error.Source: B Experimental design · p. 13
  • Table 4 defines each cell as a combination of the factors varied in an experiment, with a no-advice control in the same prompt for each template.Source: Table 4 · p. 14
  • The no-advice control and advice toward the key include boundary items only for specified placements and organiser or participant sources.Source: B Experimental design · p. 14
  • The planted wrong solution argues for a wrong option other than the advised one, and some texts do not end on the intended answer but are retained and flagged.Quoted packages keep the closing request to keep or change “your answer” although no own answer precedes it; the direction of bias is unknown.Source: B Experimental design · p. 14

C Additional results

Additional results test whether the placement and replay findings survive alternative prompt wordings and controls. They also qualify the replay evidence with an unresolved quoted-conclusion comparison and show that replay can impede correction.

Experiment 1

  • Across prompt wordings, bare notes do most harm after the options, while false-rationale notes do most harm before the question; channel effects remain a few points.The second independent wording agrees in direction, with smaller effects; the independent wording also reports no errors for the true note among finished outputs.Source: Table 5 · p. 14, Table 7 · p. 15, Table 6 · p. 15

Experiment 2

  • Full replays, whether the model’s own or quoted, almost remove a bare wrong note’s effect and halve a false rationale’s, while conclusion-only replays protect less.No difference between own and quoted full replays was detected, and their effects keep their signs under worst-case bounds.Source: Table 9 · p. 16, Table 10 · p. 17
  • About a fifth of quoted-conclusion outputs were truncated, leaving worst-case bounds that include both signs for that comparison.The comparison therefore supports no directional conclusion.Source: C Additional results · p. 17

Additional controls

  • 3.9 points: on known items, the unrelated filler raises error under the false rationale relative to no prior turn, so it is not shown to be inert.The likely range is 1.3 to 6.7 points.Source: C Additional results · p. 17

Experiment 2

  • Replaying a prior answer also lowers correction rates for initially wrong answers, including the correction rate without advice.The reduction relative to the no-prior-turn rate is reliable for the bare note under every package, but not for the true rationale.Source: Table 11 · p. 17

D Item screening and option orders

The item checks document option-order overlap, the scope of the four-option conversion, and an audit of answer keys and rationales.

Item screening and option orders

  • 23% of screened items had two coinciding option orders, and the three experimental orders coincided for 16 of 400 items.Each comparison uses the option order actually shown.Source: D Item screening and option orders · p. 18
  • 1,211 distinct item files have four-option keys pointing to the original MMLU-Pro key, while three experimental items retain an “all / none” option whose meaning may change.Those three items are retained and listed in the released check.Source: D Item screening and option orders · p. 18
  • 93 of 827 audited keys were disputed, and 129 items failed at least one audit question.The accepted-key items used here are 282 of 300 known items and 79 of 100 boundary items.Source: D Item screening and option orders · p. 18

E Hardware, answer reading and release checks

This section states which hardware comparisons are limited, how final answers are read, and how unfinished outputs and released tables are handled.

Hardware

Reading the final answer

  • 58 additional outputs become readable under a laxer answer-line rule, but no Section 4 verdict changes.The paper counts an “Answer: <letter>” line only when it is the last substantive output line; otherwise the output is unfinished.Source: E Hardware, answer reading and release… · p. 18

Unfinished outputs and release checks

  • 812 of 29,964 template C outputs reached the token cap; in template B, 60 to 80 of 300 outputs with the organiser’s assertive note ended without an answer.Released tables include scheduled and readable counts, machine-readable population and handling descriptions, and validated checks.Source: E Hardware, answer reading and release… · p. 18

Related papers

Loading 2610.34809v1…