PaperlayerView exact revision on arXiv

Source-linked AI summary

Words Speak Louder Than Order: A Behavioral Evaluation of Gemma 4

Amanda Fitch

arXiv:2609.30716v1cs.CLcs.AI

TL;DR

The paper examines how a model prioritizes conflicting documents when source framing and reading order compete. Using a counterbalanced behavioral evaluation, it finds framing has a stronger effect, while primacy varies with wording and text repetition. These results describe surface behavior, not the internal circuits behind it.

  • Problem

    When models read conflicting documents, overlapping vocabulary, order, and tone biases make it challenging to measure which source they prioritize, a question relevant to reliable tools.

    Source: §1 · pp. 2–3

  • Method

    The authors use a four-way counterbalanced evaluation of short, single-turn prompts to isolate source framing and reading position while canceling natural vocabulary biases.

    Source: §1 · p. 3, §2.1 · p. 5, §3 · p. 10

  • Results

    In direct competition, source framing shifted answers by +1.054 nats versus −0.265 for reading position, a +0.774-nat item-by-item difference; primacy varied by at least 5x with wording.

    Source: §4 · p. 16

  • Takeaways & Limitations

    A fixed positional-bias value is not supported: its magnitude depends on prompt wording, and verbatim repetition increases primacy while wording variation reduces it.

    Source: §4 · p. 16

    The findings are behavioral observations; identifying the responsible neural circuits would require internal activation patching, which is outside the study’s scope.

    Source: §4 · p. 16

Abstract

from arXiv · show

When a language model receives two conflicting documents as input, how does it decide which one to prioritize? Does it rely on how the sources are framed or the presentation order of the documents? We evaluated this behavior on Google's pre-trained Gemma 4-e4b model across a targeted behavioral suite (n = 13 items, 784 forward passes in short, single-turn contexts) using a completely counterbalanced experimental design. This setup allowed us to mathematically isolate the specific effects of source framing and reading position, while ensuring the model's natural vocabulary biases were canceled out. Across ten test conditions, we discovered the following: 1. Source framing heavily overpowers reading position. When directly competing, the semantic framing of a source (such as presenting it as an official guideline or a fresh update) had a significantly stronger impact on the model's final answer than the presentation order of the document. 2. The model favors the first document it reads, but this bias is highly variable. While the model consistently demonstrated a primacy effect (preferring the first document presented), the actual strength of this bias fluctuated by at least a factor of 5 based solely on the surface wording. 3. Overall structural repetition, not short copy-cues, drives positional bias. The model's preference for the first document is not a mechanical reaction to short, repetitive trigger phrases, such as "is [Answer]". However, the primacy effect does increase significantly when the two competing documents are structurally identical, using word-for-word verbatim templates. Introducing variation in the overall wording between the two sources reduces this positional bias.

A Fully Counterbalanced Targeted Behavioral Study

In Gemma 4, source framing outweighed presentation order by a statistically significant 4-to-1 margin when the two competed directly. The reported findings are behavioral observations about text processing.

A Fully Counterbalanced Targeted Behavioral Study

  • In direct competition, semantic framing outweighed presentation order by a statistically significant 4-to-1 margin.Figure 1 compares the magnitude of the two effects.Source: Figure 1 · p. 2
  • The paper characterizes its results as behavioral observations about text processing.Code, raw data logs, and prompt templates are publicly available on GitHub.Source: A Fully Counterbalanced Targeted… · p. 2
  • Code, raw data logs, and prompt templates for all experiments are publicly available on GitHub.Source: A Fully Counterbalanced Targeted… · p. 2

1. Introduction

The paper asks how models prioritize conflicting documents when vocabulary preferences, reading order, and source tone can all shape their answers. It introduces a counterbalanced evaluation to isolate these influences and notes the stakes for reliable and secure use.

1. Introduction

  • Conflicting official and informal documents pose a prioritization question relevant to reliable tools, while overlapping biases make accurate measurement challenging (Zheng et al., 2023; Wang et al., 2023; Sclar et al., 2024).Source: §1 · p. 2
  • Source prioritization can be confounded by natural word preferences, presentation order, and authoritative or recent tone (Sclar et al., 2024; Liu et al., 2023; Zheng et al., 2023; Wan et al., 2024).Source: §1 · p. 2
  • In production RAG systems, these confounds can let outdated official material override informal updates and create security vulnerabilities through deceptive authoritative headers.Evaluators may also mistake positional or vocabulary artifacts for semantic choices.Source: §1 · pp. 2–3
  • Simple order-swapping leaves vocabulary preferences entangled with semantic choices, so the authors use a four-way counterbalanced design to isolate the influences by design.The evaluation covered 10 conditions, 13 items, and 784 forward passes in short, single-turn contexts.Source: §1 · p. 3
  • The study reports a roughly 5x spread in position-bias magnitude across templates, extending prior findings on prompt-variation sensitivity from accuracy to bias magnitude (Sclar et al., 2024).It also contrasts its label-level authority finding with Wan et al. (2024), who report that models largely ignore stylistic authority cues.Source: §1 · p. 3

2. Methods

The methods use a four-way counterbalanced design to separate source framing, serial position, and word preference, then test these effects across ten prompt conditions. The setup uses curated single-token word pairs, validation rules, and bootstrap intervals to assess effects and reproducibility.

2.1 Contrast algebra

  • Each prompt presents two blocks recommending different words, then asks the model to choose between them; four configurations vary block order and which word receives the Recent label.The counterbalanced grid separates base word preference, source framing, serial position, and a residual block-order control; word preferences cancel algebraically.Source: §2.1 · p. 4

2.2 The ten test conditions

  • The ten conditions vary presentation settings while keeping the underlying disagreement constant; answer-word roles are counterbalanced, but loaded bodies remain tied to their roles.Eight of 18 theoretical combinations were excluded as contradictory or structurally impossible.Source: §2.2 · pp. 5–6

2.3 The testing setup

  • The testing setup specifies candidate items, inclusion rules, statistical procedures, and system details for replication.Source: §2.3 · p. 7

2.3.1 Test items

  • The authors curated 15 semantically symmetric noun pairs to provide plausible interchangeable answers and reduce baseline prior differentials.Two pairs were excluded by validation filters.Source: §2.3.1 · p. 7

2.3.2 Rules

  • Items had to use single-token answers and pass checks in all ten conditions; pairs with strong word preference or unexpected control bias were flagged, and unfiltered comparisons were rerun as a robustness check.The final tables analyze the same dataset after consistent filtering.Source: §2.3.2 · pp. 7–8

2.3.3 Confidence intervals

  • BCa bootstrap intervals used 9,999 resamples; TOST equivalence tests used a ±0.15-nat bound, and equivalence was not demonstrated in any condition.This does not prove the effects are large.Source: §2.3.3 · p. 8
  • Across 8 of 10 conditions, the 90% confidence interval for positional bias falls entirely below −0.15 nats, with T4 and T5 as exceptions.Source: §2.3.3 · p. 8

2.3.4 Environment

  • The reported run used macOS 15.7.7, Python 3.12.13, and full 32-bit precision, with 784 forward passes and no random sampling.Source: §2.3.4 · p. 9

3. Results

The results compare document order and source framing across prompt formats, test whether repeated cues explain primacy, and measure how framing and position compete. Overall, the model favored the first document, but the size of that bias varied with wording.

3.1 Primacy is reliable in direction, highly variable in size

  • Primacy was significant in 10/10 conditions, but its magnitude ranged from -0.172 to -0.884 nats—a roughly 5x spread caused by surface wording.In T4 and T5, the direction was reliable but the magnitude could not be distinguished from trivially small.Source: Table 4 · p. 10, §3.1 · p. 11

3.2 Primacy is not caused by a repeated copy cue

  • Short repeated cues did not cause primacy; word-for-word repetition increased it, while wording variation between blocks significantly reduced it.The strongest observed primacy, T1c = -0.884, occurred without cues; cue counts did not produce a linear dose-response pattern.Source: §3.2 · pp. 12–13, Table 5 · p. 12

3.3 Framing beats position by roughly 4 to 1

  • In the cue-free item-by-item test (T5b), framing outweighed position by +0.774 nats, while the full first-versus-last framing swing was 2.108 nats (roughly 8-to-1 odds).Removing repeated copy cues made framing significantly stronger; with cues (T5), its effect was about half as strong as in cue-free T5b.Source: §3.3 · p. 14

3.4 Authority headers reduce bias, body wording modulates how much

  • Authority headers reduced positional bias with both identical and reworded bodies, while body phrasing changed the size—but not the direction—of that reduction.The significant interaction contrast supports the modulation by body phrasing, even though C1’s confidence interval spanned zero.Source: §3.4 · p. 15, Table 8 · p. 15

3.5 Transparency and Robustness Checks

  • Headline findings kept the same direction and statistical significance in the 14-item unfiltered dataset, but T5b’s clean nulls partly depended on excluding Spec/Brief.The authors note that this removed item had been flagged in T5b itself.Source: §3.5 · p. 16

4. Discussion

In the head-to-head test, source framing moved the model’s answer more than document position. Primacy varied with wording and verbatim repetition, while the observed behavioral patterns do not identify internal mechanisms.

4. Discussion

  • +1.054 nats for source framing versus -0.265 for reading position; their item-by-item difference was +0.774 nats.These are reported for the cleanest condition in which the two forces compete directly.Source: §4 · p. 16
  • Positional bias fluctuated by at least a factor of 5 with header wording and verbatimness, so its strength is not stable across prompt templates.The paper cautions that a fixed deployment rule assigning a value to putting an important document first is unsupportable.Source: §4 · p. 16
  • Removing repeated copy cues did not weaken primacy consistently, while the strongest primacy effect occurred without cues; verbatim repetition tracked the effect instead.Cue removal increased bias under a neutral header and made no difference under an authoritative header.Source: §4 · p. 16
  • The association between word-for-word repetition and primacy is behavioral, not evidence identifying neural circuits or attention heads.Testing those mechanisms would require internal activation patching, which was outside the study’s scope.Source: §4 · p. 16

5. Limitations and scope

The reported effects have design-specific qualifications and should be read as behavioral findings from a tightly controlled, narrow evaluation. The study does not establish that its estimates generalize to other models, topics, or realistic long-context RAG systems.

5.1 Methodological caveats

  • The +1.054-nat framing effect measures headers and loaded bodies as a joint package, not their separate contributions.Separating them would require an unrun condition pairing an authority header with a neutral body.Source: §5.1 · p. 17
  • Verbatimness is structurally confounded with counterbalancing depth because identical-template cells allow only one wording assignment, unlike mixed-wording cells.The paper notes that averaging assignments in mixed-wording cells can attenuate wording-by-position interactions.Source: §5.1 · p. 17

5.2 What our results do not claim

  • The evaluation measures behavioral input–output relationships and makes no claims about internal representations or circuits.The study also does not support a general claim that semantic contradiction suppresses positional bias; that contrast was unstable and sensitive to outlier filtering.Source: §5.2 · p. 17
  • Failing to reject non-equivalence within ±0.15 nats does not establish large underlying effects; it means their negligibility could not be statistically bounded.The reported ≈3.76:1 framing-to-position ratio is descriptive of T5b, not an invariant model parameter.Source: §5.2 · p. 17
  • In T5/T5b, headers and bodies co-varied by design, so the +1.054-nat package effect cannot allocate variance between them.Verbatimness comparisons are also confounded with counterbalancing depth, so the paper reports an empirical association rather than causal independence.Source: §5.2 · pp. 17–18

5.3 Scope

  • The study tested only Gemma 4-e4b in two-document, single-word-answer, single-question-format English contexts.All test items were English project-management nouns, so generalization to other topics or vocabularies was not verified.Source: §5.3 · p. 18
  • The findings should not be automatically applied to realistic, multidocument RAG systems until formally replicated.The evaluation used short, single-turn prompts with full text fidelity rather than accounting for long-context compression, summarization, or KV-cache eviction.Source: §5.3 · p. 18

6. Reproducibility and provenance

The paper provides project and result-log locations, software and runtime details, and a disclosure about the timing and derivation of the A4 result. The log is reported as byte-identical to its archived raw-results copy.

6. Reproducibility and provenance

  • The project files include the code and a primary results log, which the authors certify is byte-identical to the archived raw-results copy.The cited project location and log filename are given in the passages.Source: §6 · p. 19
  • The evaluation used 784 forward passes across 10 conditions in full 32-bit precision, evaluation mode, and without random sampling.The environment details include macOS-15.7.7-arm64, Python 3.12.13, torch 2.13.0, transformers 5.14.1, numpy 2.5.1, and scipy 1.18.0.Source: §6 · p. 19
  • The A4 arm-heterogeneity value was computed from cached JSON data after the primary log was generated; displayed 0.001 differences reflect rounding.The authors state the A4 test was committed immediately after log generation and used the same run’s cached data.Source: §6 · p. 19

Appendix A — Why counterbalancing is necessary

Appendix A illustrates the size of a vocabulary-bias artifact in uncounterbalanced pilot data while stating that the pilots do not support the paper’s main claims. It also reports unusable patching cells and explains why these artifacts are presented only for transparency.

Appendix A — Why counterbalancing is necessary

  • 1.58 nats of artificial swing appeared without counterbalancing, six times the clean-test positional-bias half effect.The appendix says vocabulary preference was indistinguishable from a real positional effect in a standard test; the four-way design was intended to cancel it.Source: Appendix A · p. 20
  • The pilot patching cells were excluded from aggregate analysis because too many items failed strict quality checks.The appendix says these pilot artifacts do not meet the main-text standard and do not support the main claims.Source: Appendix A · p. 20
  • The pilot artifacts are included for transparency, not as evidence supporting the paper’s main claims.Their stated purpose is to show the size of the artifact removed from the primary data.Source: Appendix A · p. 20

Appendix B — Complete contrast index

The contrast index catalogs comparisons used to isolate framing, order, and cue effects, with values reported in nats for filtered and unfiltered datasets. Its headline margin comparison shows framing outweighing presentation position.

Appendix B — Complete contrast index

  • The index marks results significant when their 95% confidence intervals exclude zero and labels several legacy or confounded contrasts as reference-only.The semantic suppression test E1 remains unresolved.Source: Appendix B · pp. 21, 24
  • Authority headers reduced positional bias, and the cue-matched interaction showed that this attenuation depends on body wording.The C3 interaction was significant in both reported datasets: +0.315 and +0.296.Source: Appendix B · pp. 21, 23
  • +0.774 and +0.806 are the reported D1 framing-vs-position margins, each marked statistically significant.D1 is the per-item, interval-bearing version of the headline comparison; the index reports results for the filtered and unfiltered datasets.Source: Appendix B · pp. 21–22
  • A2 detected no change from removing the cue under an authority header, and pooled cue removal is flagged as invalid because its two arms respond differently.A4 tests that arm heterogeneity; A3 is retained for reference only.Source: Appendix B · pp. 22–23
  • Removing copy cues strengthened source framing by +0.539 nats in D3, while the cue-free framing effect was about twice the cued effect.D3 isolates the change in framing reliance between T5b and T5.Source: Appendix B · pp. 23–24

Appendix C — The four crossings of block order x entity-role assignment

Appendix C describes four prompt crossings that counterbalance block order and entity-role assignment, separating positional effects from vocabulary preferences. Appendix D explains the nats scale and distinguishes cued from cue-free prompts.

Appendix C — The four crossings of block order x entity-role assignment

  • Four crossings combine block order with entity-role assignment to counterbalance both variables in each candidate word pair.The Agile/Waterfall example illustrates how the configurations swap the authoritative or recent role and which block appears first.Source: Appendix C · p. 25, Table 10 · p. 25
  • Swapping which answer word carries each semantic role cancels the model’s natural vocabulary bias, while block order measures positional bias.The resulting prompt structure is determined by combining those two assignments.Source: Appendix C · p. 25

Appendix D — Reading nats, and cued vs. cue-free prompts

  • The reference guide translates nats into odds and probabilities and gives examples of cued versus cue-free prompt wording.Cued prompts include repeated word patterns; cue-free prompts remove them.Source: Appendix D · p. 26

Nats

The appendix explains how to read measured bias in nats and how the study labels its prompt variants. Nats express the pull between two answer options, while cue status and semantic role identify prompt features.

Nats

  • A nat value is a log-probability effect size quantifying the model’s pull toward Option A over Option B.The guide also translates nat values into relative odds and implied probabilities.Source: Nats · p. 26
  • A 95% confidence interval gives the expected range of the nat value, and the guide uses it to assess measurement reliability.The preference label restates the numbers as a plain-English description of the model’s behavior.Source: Nats · p. 26

Cue vs Cue-Free Prompts

  • Prompt type distinguishes repeated-pattern cued prompts from cue-free prompts, while role labels mark authority versus recency framing.The wording template is the exact string supplied to the model, with {e} marking the injected answer word.Source: Cue vs Cue-Free Prompts · p. 27

Appendix E — Results table definitions

This appendix explains how to interpret the study’s results tables, defining positional-bias and cue-removal contrasts, framing and position effects, and checks for hidden artifacts and vocabulary bias. It also clarifies influential-factor and authority-header results, dataset variants, contrast definitions, and key behavioral terms.

3.1 Positional bias across 10 test conditions

  • The positional-bias table covers ten conditions; negative serial-position values indicate primacy, positive values indicate recency, and the measure is expressed in nats.Its confidence intervals are resampled across items, and the table also reports how many items show primacy and whether the 90% interval clears the -0.15 threshold.Source: §3.1 · pp. 28–29

3.2 Contrasts

  • Cue-removal contrasts A1 and A2 test neutral- and authority-header conditions separately; because the arms differ significantly, A3 is descriptive only and they are not pooled.A4 tests the difference between A1 and A2.Source: §3.2 · pp. 30–31

3.3 Framing and position influence

  • The framing-versus-position tests use framed headers and loaded body wording, comparing T5 with two copy cues against cue-free T5b.The cue pattern is the repeated exact word pattern, such as “is [Answer].”Source: §3.3 · pp. 31–32

3.3 Quantities for hidden artifacts

  • The reported effects were +1.054 [+0.907, +1.258] and -0.265 [-0.421, -0.137], and both are marked statistically significant because their confidence intervals exclude zero.Residual checks were near zero: Order +0.011 and Token +0.003.Source: §3.3 · pp. 32–33

3.3 Influential factors

  • D1 compares framing with position in cue-free T5b, D2 makes the same comparison with cues in T5, and D3 measures how cue removal changes framing.These contrasts separate the factors influencing the model’s decisions.Source: §3.3 · p. 33
  • All three influential-factor contrasts are statistically significant because their 95% confidence intervals exclude zero.The strict filtered dataset contains 13 word pairs that passed all sanity checks; the unfiltered sensitivity check uses 14 pairs, including one previously excluded.Source: §3.3 · pp. 33–34

3.4 Authority headers

  • The authority-header comparisons measure header effects with mixed or verbatim bodies and test whether body wording changes that effect.C1 compares mixed bodies, C2 compares verbatim bodies, and C3 is the difference between C1 and C2; the comparisons use text without repeated copy cues.Source: §3.4 · p. 34

Glossary

  • The glossary defines the study’s behavioral scope and terms, including counterbalancing, positional bias, primacy, source framing, and vocabulary bias.It distinguishes behavioral input–output evaluation from mechanistic claims about internal circuits and defines confidence intervals as 95% BCa bootstrap estimates using 9,999 resamples.Source: Glossary · pp. 35–36
Loading 2609.30716v1…