PaperlayerView exact revision on arXiv

Source-linked AI summary

Prompt framing governs LLM default following in collective-action

Eladio Montero-Porras, Axel Abels, Tom Lenaerts

arXiv:2610.03253v1cs.GT

TL;DR

It remains unclear whether pre-filled defaults merely inform LLM choices or systematically shift them in consequential interfaces. The paper measures default-induced changes in choice distributions across one-shot CPR and TPG games. Defaults attract probability mass in both, but the strength of that pull depends on wording and the model–game pairing.

  • Problem

    It remains unclear whether pre-filled defaults systematically shift LLM choices, even as models make or recommend decisions through interfaces that present suggested values.

    Source: §1 · pp. 1–2

  • Method

    The study measures default-value probability mass against no-default baselines across wording and default conditions in one-shot CPR extraction and TPG contribution games.

    Source: §2.3 · p. 5, §2.7 · p. 6, §4 · p. 12

  • Results

    Pre-filled defaults increase probability mass on their value in both games, while default pull varies with wording and model–game pairing.

    Source: §4 · pp. 12–13, §6 · p. 14

  • Takeaways & Limitations

    Because the same model can shift from low to high default pull under trivial rewording, single-prompt evaluations cannot characterize default susceptibility as a stable model property.

    Source: §4 · p. 13, §6 · p. 14

    The seven tested systems expose token log probabilities and are not a representative sample of frontier models, so the study makes no prevalence claims.

    Source: §5 · p. 14

Abstract

from arXiv · show

Large language models are increasingly deployed as agents that make or recommend decisions on behalf of users, often operating through interfaces that pre-fill suggested values or default options. Whether models treat such defaults as merely informational or as suggestions that systematically alter their choices remains unclear. We study default deference in two one-shot social dilemmas: a common-pool resource (CPR) extraction game and a threshold public-good (TPG) contribution game. We measure how defaults shift each model's choice distribution relative to its no-default baseline, across default values, wordings, and action-space granularities. We find that pre-filled defaults pull probability mass on the default value in both games, but the magnitude depends strongly on wording: the same model can show high pull under one formulation and near-zero pull under another. Permission-style wording reduces default pull in both games, more strongly in CPR than in TPG. Default pull is weaker in coarse action spaces, and conflict defaults attract more mass than agreement ones. These results indicate that default deference depends on the model, the wording of the interface, and the structure of available choices. For agentic systems, evaluating model behaviour without controlling the surrounding choice architecture can miss an important source of behavioural variation.

1 Introduction

LLM choices in strategic settings can depend on contextual framing, but whether pre-filled defaults systematically shift initial choices across collective-action problems remains an open question. The study tests default deference across two dilemmas and finds that wording can strongly alter it.

1 Introduction

  • LLMs increasingly make or recommend decisions through interfaces whose wording, option structure, and salience can shift behavior without changing the underlying model.Prior strategic-game research also finds that LLM outputs depend on contextual information, prompts, opponents, and interaction structure.Source: §1 · pp. 1–2
  • A pre-filled default is a value or option selected unless the agent actively overrides it, and may operate through perceived recommendation, reduced effort, or selective attention.The paper asks whether such interface features systematically alter an LLM agent’s chosen action.Source: §1 · p. 2
  • The study asks whether a pre-filled option shifts initial choice probabilities across strategic environments and whether that pull depends on dilemma structure, wording, baseline inclination, or action-space granularity.Existing evidence shows LLM choices can respond to prompts, information salience, and framing, but the effects of small wording changes within a default frame remain unclear.Source: §1 · pp. 2–3
  • The CPR extraction and TPG contribution games provide contrasting collective-action settings for testing whether default-induced shifts track a general tendency to follow pre-filled options or depend on strategic structure.The study measures probability mass on the default relative to a no-default baseline; prior human studies report default effects in CPR and public-good settings, with weaker and less consistent sensitivity in threshold games.Source: §1 · pp. 2–3
  • Permission-style wording suppresses default following and instruction-style wording amplifies it across both games, with the same model shifting from resistant to deferential under trivial rewording.The findings suggest default susceptibility is not a stable model property, but depends on the surrounding decision scaffold.Source: §1 · p. 3

2 Experimental Design

The design compares one-shot extraction and contribution games across action-space granularities, default values and wordings, using seven models with token-logprob access. Default pull is measured against each model’s no-default choice distribution.

2.1 Games

  • In the CPR, four players choose extraction from 0–30 tokens under declining returns; in the TPG, four players must contribute at least 60 or forfeit their contributions.The games are one-shot and differ in action direction, payoff shape, and motivational frame, so observed game differences cannot be attributed to any one of those axes.Source: §2.1 · pp. 3–4

2.2 Granularity manipulation

  • The coarse condition restricts the 0–30 action range to {0, 15, 30}; the fine condition retains all 31 integer options, with only the answer-format line changing.The manipulation tests whether concentrating baseline probability on three focal values changes default following.Source: §2.2 · pp. 4–5

2.3 Default manipulation and wording variants

  • Each non-baseline prompt appends a fixed lead-in naming the pre-filled value and one of four suggestion sentences ranging from neutral pre-fill to system recommendation.The variants include permissive wording that emphasizes override choice and instructive wording that invokes the system as the source of the suggestion.Source: §2.3 · p. 5

2.4 Default values

  • CPR defaults are 0, 11, 18, 23, 26, and 30; TPG defaults are 0, 15, and 30, each tested under all four wordings.The CPR values span abstention, social optimum, Nash equilibrium, exploitative and indefensible defaults, and collapse; the TPG values represent free-riding, fair share, and maximal commitment.Source: §2.4 · p. 5

2.5 Models

  • Seven models with exposed token-logprob access are tested because the primary measure requires each model’s choice distribution, not just a single answer.The model set is determined by that method requirement.Source: §2.5 · p. 5

2.6 Elicitation and coverage

  • Token log probabilities are used to reconstruct answer probabilities, while cells with top-five valid-answer coverage below 0.90 are excluded from analyses requiring a fully observed distribution.Digit-by-digit tokenization is handled by combining per-position probabilities, and excluded cells are not imputed.Source: §2.6 · p. 6

2.7 Metrics

  • Net pull measures probability added to the default value over baseline; defaults matching a model’s most probable baseline choice are agreement cells, and other defaults are conflict cells.Wasserstein pull is a robustness check that credits probability mass moving toward the default without landing exactly on it, but is undefined for agreement cells.Source: §2.7 · pp. 6–7

3 Results

Pre-filled defaults attract probability mass in both dilemmas, but the extent and direction of pull vary with wording, action-space granularity, and the model–game pairing.

3.1 Pre-filled defaults shift output distributions toward the suggested value in both games

  • Defaults attract probability mass in both games: CPR pull is positive at every tested value, while TPG responses are heterogeneous and near-bimodal across models.In TPG, some models show high pull while Gemma4-31B and gpt 5.4 show near-zero pull.Source: §3 · p. 7, §3.1 · pp. 7–8, Figure 2 · p. 8

3.2 Permission framing suppresses default pull; instruction framing amplifies it in CPR only

  • Permission wording V2 produces the lowest pull in both games; V1 and V3 exceed V0 in CPR, but TPG shows no matching pattern.In TPG, V1 is 0.433 versus V0 at 0.412, while V3 is 0.321.Source: §3.2 · p. 8, Table 1 · p. 9, Figure 3 · p. 9

Default pull is weaker in coarse action spaces; conflict cells attract more pull than agreement cells

  • Default pull is weaker with three coarse options than with the 31-option fine range in both games and for both agreement and conflict cells.The authors suggest that the fine range leaves more room for a default to attract mass because each value is less justified at baseline.Source: Default pull is weaker in coarse… · pp. 9–10, Figure 4 · p. 10
  • Conflict pull exceeds agreement pull across all game–granularity combinations, but agreement pull remains nonzero, so conflict-conditionality is partial.Fine CPR conflict pull is 0.59 versus 0.25 for agreement; the minimum agreement pull is 0.09 in coarse CPR.Source: Figure 4 · p. 10, Default pull is weaker in coarse… · p. 10

The same model can shift from low to high default pull under a trivial rewording

3.3 Default pull is game-dependent and model-specific

  • Default sensitivity also varies by game within a model: gpt 5.4 pulls strongly in CPR but barely in TPG, while Llama3.3 shows the reverse.The reported means are 0.63 for gpt 5.4 in CPR and 0.04 in TPG; Llama3.3 has 0.35 in CPR and 0.96 in TPG at V0.Source: §3.3 · pp. 11–12, Figure 5 · p. 12

4 Discussion

Across two social dilemmas, pre-filled defaults shift model output distributions, with the magnitude shaped more by wording than model identity. The findings also show that default pull varies with action-space granularity, agreement versus conflict, and the model–game pairing.

4 Discussion

  • Pre-filled defaults increase probability mass on the default in both games; permission wording suppresses pull in both, while instruction wording amplifies it in CPR but not TPG.The study reports five consistent patterns across seven models, four wording variants, and two action-space granularities.Source: §4 · p. 12
  • Default pull is weaker in coarse than fine action spaces, while conflict defaults attract more mass than agreement defaults, whose pull is still nonzero.These patterns hold in both games.Source: §4 · p. 12
  • A trivial rephrasing can shift the same model from low to high pull, and the game-specific pattern differs: gpt 5.4 is high in CPR but near-zero in TPG, while Llama3.3 reverses.The results characterize default susceptibility as an interaction between model and game, rather than one factor dominating.Source: §4 · p. 13
  • A single fixed prompt cannot establish a model’s general default robustness because measured susceptibility belongs to the prompt–model pair.Whether large, diverse prompt sweeps could yield a stable estimate remains open.Source: §4 · p. 13
  • In collective-action deployments, the same harmful default can attract substantial or near-zero probability mass under different framings, complicating evaluation and accountability.The study measures output-distribution shifts, not preferences or intentions, and does not identify their mechanism.Source: §4 · p. 13

5 Limitations

The study’s scope is limited by its model sample and single-shot design, and it does not identify the mechanism behind observed shifts.

5 Limitations

  • The seven systems expose token log probabilities, but this provider-correlated selection is not representative of frontier models and supports no prevalence claims.Other major providers were excluded for this reason.Source: §5 · p. 14
  • Because the design is single-shot, it measures distributional shifts at first contact and cannot establish whether they persist or decay over repeated interactions.Repeated-round dynamics are central to human benchmark studies.Source: §5 · p. 14
  • The predicted clean separation between agreement and conflict cells was not achieved, since agreement pull was not near zero.The design also does not identify whether shifts reflect anchoring, instruction-following, or another mechanism; Qwen3 235B was excluded due to provider decoding limitations.Source: §5 · p. 14

6 Conclusions

Pre-filled defaults shift LLM choices in two social dilemmas, with the magnitude governed more by default wording than by the model tested. Permission-style wording suppresses default pull, while instruction-style wording amplifies it.

6 Conclusions

  • Default-line wording governs choice-distribution shifts more than the tested model in two structurally distinct social dilemmas.Source: §6 · p. 14
  • Permission-style wording suppresses default pull across games and granularities, whereas instruction-style wording amplifies it.Source: §6 · p. 14
  • A trivial rephrasing can move the same model from low to high default pull, so single-prompt evaluations do not support stable per-model characterisations.Source: §6 · p. 14
  • How a suggested value is framed is a first-order determinant of output behaviour in collective-action settings, with implications for evaluating and deploying autonomous systems.Source: §6 · p. 14

S1 Full game prompts (S1)

The supplementary prompts share a blinded system prompt and add a default line in every non-baseline condition. Game instructions and response formats vary by dilemma and action-space granularity.

S1 Full game prompts (S1)

  • All game runs share a blinded system prompt, and each game has a separate user prompt with the default line appended outside the baseline condition.Source: S1 Full game prompts (S1) · p. 18

S1.1 System prompt (all games)

S1.2 CPR user prompt (fine granularity, 0–30)

  • In fine-grained CPR, four group members extract from a shared resource; group payoff depends on total extraction and is distributed in proportion to individual extraction.Source: S1.2 CPR user prompt (fine… · p. 18

S1.3 CPR user prompt (coarse granularity, {0, 15, 30})

S1.4 TPG user prompt (fine granularity, 0–30)

  • In this group task, four members each start with 50 ECU in a Private Account; if total contributions reach at least 60 ECU, each keeps 100%, otherwise each keeps 10% (losing 90%).Source: S1.4 TPG user prompt (fine… · p. 19

S1.5 TPG user prompt (coarse granularity, {0, 15, 30})

S1.6 Default line (appended to all non-baseline conditions)

  • Non-baseline prompts append a numeric default and one of four suggestion sentences; TPG substitutes “contribution” for “extraction,” and baseline prompts omit the default line.Source: S1.6 Default line (appended to all… · p. 19

S2 Complete per-model framing tables (S2)

Table S1 presents per-model framing details for net pull and w pull at conflict defaults, alongside exclusions and a warning about unstable values.

S2 Complete per-model framing tables (S2)

S3 Complete w pull tables (S3)

The supplementary w-pull table reports values for all cells, but flags unstable agreement cells and corroborates headline results only where conflict-cell values are stable.

S3 Complete w pull tables (S3)

S4 Per-default framing figures (S4)

Permission wording V2 suppresses default pull across CPR default values, and the framing effect also appears under an alternative pull measure. The effect varies substantially across models, especially in TPG.

S4 Per-default framing figures (S4)

  • V2 suppresses net pull at all six CPR defaults, including the social-optimum and abstain values, making the wording effect valence-blind.Figure S1 shows the per-default pattern.Source: S4 Per-default framing figures (S4) · p. 20, Figure S1 · p. 21
  • Under V2, w pull also falls at CPR d=18, indicating that the distribution moves away from the default rather than merely showing weaker attraction.Figure S2 shows the same framing effect under this alternative measure.Source: S4 Per-default framing figures (S4) · p. 20, Figure S2 · p. 22
  • The V2 net-pull trough appears in most individual CPR models, while TPG responses are more heterogeneous, with some models showing near-zero pull across conditions.Qwen3 235B CPR panels are empty because of baseline coverage exclusion.Source: Figure S3 · p. 23

S5 Human comparison

The human comparison concerns first-contact choices in the CPR game: models shift more toward the pro-social default but show a comparable, slightly smaller shift toward the self-serving default. The mean-shift ratio puts human and model changes on a common scale, while the single-shot design cannot establish whether model effects persist over repeated interactions.

S5 Human comparison

  • The mean-shift ratio compares the treatment-minus-control mean change with the maximum possible movement toward the default, using extraction units.Zero indicates no movement; one indicates landing on the default; values above one indicate overshooting it.Source: S5 Human comparison · p. 24
  • For the pro-social default, models shifted −5.51 points on average (ratio 1.10), versus humans’ −1.50 points (ratio 0.44), overshooting the default anchor.The comparison is limited to first-contact behaviour.Source: S5 Human comparison · p. 24
  • For the self-serving default, models shifted +3.34 points (ratio 0.59), slightly less than humans’ +5.38 points (ratio 0.63).These are first-contact shifts in the matched CPR paradigm.Source: S5 Human comparison · p. 24
  • The single-shot design cannot establish whether model distributional effects persist or decay over repeated interactions.The human effect decays over subsequent rounds.Source: S5 Human comparison · p. 24
Loading 2610.03253v1…