PaperlayerView exact revision on arXiv

Source-linked AI summary

Target-Token Asymmetry Can Masquerade As Depth Specialization

Danlin Li, Miao Li, Xu Han, Jiecheng Lu

arXiv:2610.19344v1cs.CLcs.LG

TL;DR

Cross-corpus layer-ablation studies can mistake differences in target-token composition for depth specialization. This paper decomposes early–late damage interactions by target type, tests shared-mixture summaries, and screens risk using tokenizer information. In its main case, whitespace carries the positive contrast, and removing it reverses pooled estimates across the reported high- and medium-risk pairs.

  • Problem

    Layer-ablation comparisons across corpora average over potentially different target-token mixtures, complicating interpretation of a larger early-layer effect as task specialization.

    Source: §1 · pp. 1–2

  • Method

    The paper decomposes cross-corpus early–late damage interactions by target type, recomputes them under shared target mixtures, and evaluates a tokenizer-only risk screen.

    Source: §1 · pp. 2–3

  • Results

    In the main case, only whitespace has a positive interaction; removing whitespace reverses the pooled effect, and exclusion reverses every positive pooled estimate in high- and medium-risk pairs.

    Source: §1 · p. 3, §7 · p. 9

  • Takeaways & Limitations

    Cross-corpus depth claims should report target tokens, decompose effects by type, and check whether a mixture-matched summary retains the sign.

    Source: §7 · p. 10

    The primary claim concerns H=3 mean-blend top-1 CAI; the position and control-accuracy analyses cover narrower settings, and positive H=1 mean-blend content cells remain unsplit.

    Source: §7 · p. 9

Abstract

from arXiv · show

Layer-ablation studies often perturb early or late layers of a language model, compare the damage to prediction on two corpora, and read a larger early-layer effect on one corpus as early-layer specialization for its task. Such a comparison averages over the tokens that each corpus asks the model to predict, and these target tokens differ: math solutions, for example, ask for far more whitespace and numbers than encyclopedia prose. We show that this difference can create the appearance of depth specialization. We split a cross-corpus early--late damage interaction by target-token type and recompute it under a target mixture shared by both corpora. Between GSM8K and Wikitext, a positive interaction replicates across five model instances. In the three fully decomposed models, only whitespace has a positive interaction. Removing whitespace reverses the pooled effect in all three, and matching the full target mixture reverses it in two. The diagnosis holds under position matching, adversarial reassignment of token types, and a continuous readout. The risk can also be anticipated from the tokenizer alone. A screen computed before any model runs places every held-out pair it scored in the region its label predicts, and on the pairs it rates at high or medium risk, removing whitespace reverses every positive pooled estimate. A smaller content-word effect one token ahead follows the share of content targets that the unablated model already predicts. A cross-corpus depth effect therefore supports an interpretation of task specialization only after the target-token mixture is reported, the effect is decomposed by token type, and a mixture-matched summary keeps its sign.

1 INTRODUCTION

Cross-corpus layer-ablation contrasts can reflect differences in the target tokens each corpus asks the model to predict, not only differences in task-relevant depth. The paper decomposes the contrast by target type and tests shared-mixture summaries and a tokenizer-only risk screen; its main case’s positive contrast is concentrated on whitespace.

1 INTRODUCTION

  • Math and prose corpora differ in target-token composition, so perturbations affecting whitespace or numbers can shift pooled accuracy even when effects on words are alike.Because the score averages over all predicted targets, a replicated pooled contrast alone cannot distinguish depth effects from composition effects.Source: §1 · pp. 1–2
  • The study defines CAI as a cross-corpus early-versus-late damage interaction, decomposes it by target type, and recomputes summaries under a shared mixture.It also tests whether tokenizer-only quantities can screen corpus-pair risk before model evaluation.Source: §1 · p. 2
  • In the main case, whitespace targets carry the positive interaction: other target types are non-positive, and removing whitespace reverses the pooled sign.Position matching, adversarial token-type reassignment, and a continuous readout support the whitespace finding.Source: §1 · p. 3
  • Across high- and medium-risk pairs with positive pooled estimates, excluding whitespace reverses every estimate, while held-out pairs scored by the screen fall in its predicted region.The screen uses tokenizer information and is computed before model runs.Source: §1 · pp. 2–3
  • The proposed reporting protocol decomposes effects by target type, tests mixture-matched summaries, and splits a remaining positive cell by whether the unablated model predicts its targets.For a one-token-ahead content-word effect under a second perturbation, that split traces the effect to differences in the share of content targets already predicted.Source: §1 · p. 3

2 MEASURING A DEPTH CONTRAST BY TARGET TYPE

The paper measures cross-corpus depth contrast as a difference of differences in intervention damage, then conditions and reweights the contrast by target-token type. It also defines a diagnostic interaction for assessing how strongly a setting singles out whitespace.

2 MEASURING A DEPTH CONTRAST BY TARGET TYPE

  • CAI compares early- and late-band excess damage on a focal corpus against a reference corpus; positive CAI means the early band causes more excess damage on the focal corpus.The study uses the first and last layer thirds as the early and late bands, respectively.Source: §2 · p. 3
  • The primary setting blends each band’s residual-stream outputs toward the same layer’s corpus mean with α = 0.2, and scores top-1 predictions three positions ahead.The content-word analysis instead uses H=1 mild-resample; the target margin supplies a continuous readout.Source: §2 · p. 3
  • A deterministic rule classifies decoded target tokens as whitespace, number, punctuation, function word, or content word, and conditional CAI replaces accuracy with per-sequence accuracy on one type.A leading space does not change a word’s class, and tokens mixing letters and digits count as content words; a cell requires at least 10 sequences per corpus.Source: §2 · p. 3
  • Mixture summaries weight conditional CAIs using one shared target distribution, while pooled CAI uses each corpus’s own target mixture.Whitespace-excluded summaries set its weight to zero and renormalize the remaining eligible types; intervals are 95% sequence-resampling percentile intervals.Source: §2 · p. 4
  • The diagnostic interaction contrasts the whitespace-minus-content CAI gap in the primary setting with that gap under H=1 mild-resample.A large positive value marks a pair whose primary-setting contrast is concentrated on whitespace; the tokenizer-only screen uses whitespace-rate and characters-per-token gaps before model evaluation.Source: §2 · p. 4

3 THE POOLED EFFECT COLLAPSES TO WHITESPACE

For GSM8K versus Wikitext, the pooled CAI is small and positive, but the whitespace cell is much larger; excluding whitespace reverses the estimate in all three models. The whitespace contrast is tied to late-band gains on GSM8K and remains positive under replication and position controls.

3 THE POOLED EFFECT COLLAPSES TO WHITESPACE

  • Late-band mean-blend raises GSM8K whitespace accuracy from about 0.02 to 0.79–0.95, while changing Wikitext whitespace accuracy by at most 0.16.The whitespace cell’s late part is positive in every model and its early part is negative; the late band often predicts a single space on GSM8K.Source: §3 · pp. 4–5
  • The whitespace cell exceeds the small positive pooled CAI by more than an order of magnitude, while the whitespace-excluded estimate is negative in all three models.The reference-mixture estimate is negative for Qwen and Llama but positive for Gemma.Source: §3 · p. 4, Table 1 · p. 5
  • The diagnostic interaction is positive in all five same-sequence audit instances, ranging from +0.52 to +0.81, with no non-positive bootstrap draw.All relative-position quartiles and the position-matched estimate remain positive in each model; the effect also appears both inside and well beyond the final answer clause.Source: §3 · p. 5

4 SCREENING CORPUS PAIRS BEFORE A MODEL RUNS

The tokenizer-only screen assigns corpus-pair risk labels before interventions, using tokenizer statistics to anticipate regions of model-output behavior. Its scored held-out outcomes generally match those predictions, while whitespace exclusion reverses every positive pooled estimate among high- and medium-risk pairs.

4 SCREENING CORPUS PAIRS BEFORE A MODEL RUNS

  • Before model evaluation, the screen uses target-whitespace-rate and characters-per-token gaps to assign a risk label, then scores labels against regions computed from model outputs.The input rule and pair selection were fixed before intervention outcomes were computed.Source: §4 · p. 5, Table 2 · p. 6
  • The held-out evaluation is exploratory because output thresholds were set with calibration rows in view.Scored assignments remain unchanged over a wide range of cutoffs.Source: §4 · p. 6
  • All nine scored held-out rows fall in their label-predicted region, and all six prospective DoLa region predictions agree with observed outcomes.Eight of the nine held-out rows also meet the interval criterion; the exception is Gemma on AGNews/Wikitext.Source: §4 · p. 7
  • The high-risk GSM8K/FACTOR-News pair has positive pooled CAI in every model but negative mixture summaries, whereas the low-risk StrategyQA/FACTOR-News intervals all contain zero.For MATH and DROP, six interactions have intervals excluding zero; all ARC intervals include zero.Source: §4 · p. 7
  • All 14 positive pooled rows from high- and medium-risk pairs reverse sign when whitespace is excluded; reference reweighting makes 11 of those 14 negative.Table 3 reports pooled and whitespace-excluded estimates for every decomposed pair.Source: §4 · p. 7, Table 3 · p. 8

5 ROBUSTNESS ACROSS DEFINITIONS AND READOUTS

Robustness checks vary the operator, prediction horizon, token classification, and readout. Across these checks, the H=3 mean-blend effect remains concentrated on whitespace, while a smaller H=1 content-word effect tracks control accuracy.

5 ROBUSTNESS ACROSS DEFINITIONS AND READOUTS

  • At H=3, mean-blend is concentrated on whitespace, while mild-resample is near zero for both target types; the only positive content cell in the operator grid is H=1 mild-resample.The grid cells use samples of different sizes and are compared one cell at a time.Source: §5 · p. 8
  • An adversarial token map dilutes the whitespace effect by 8.2–8.7 times, yet whitespace remains the largest positive H=3 cell in every model and the only positive cell in two.At H=1 mean-blend, content-word CAI is positive in all three models when restricted to tokens without leading whitespace.Source: §5 · p. 8
  • Under mean-blend, whitespace target-margin CAI ranges from +1.23 to +3.37 logits with every interval excluding zero; five of six content estimates are negative.The remaining content estimate, Gemma at H=1, is positive with an interval including zero.Source: §5 · p. 8
  • Across five instances, the H=1 mild-resample content-word CAI ranges from +0.0086 to +0.0148, but reweighting by control-pass correctness yields estimates from −0.0028 to +0.0031 with intervals including zero.The unablated model predicts a larger share of GSM8K content targets, and the operator removes more correct predictions under the early band than the late band.Source: §5 · pp. 8–9

6 RELATED WORK

The discussion frames target-token composition as a source of aggregation bias in cross-corpus depth attribution. The analysis complements work on intervention and metric validity by partitioning targets under a fixed intervention.

6 RELATED WORK

  • Target-token composition is a concrete source of aggregation bias in cross-corpus depth attribution, where population mixtures can alter pooled contrasts.The decomposition changes the task interpretation attached to the pooled contrast.Source: §6 · p. 9
  • Partitioning targets under a fixed intervention complements studies that vary interventions or faithfulness metrics (Zhang & Nanda, 2024).In this analysis, partitioning concentrates the replicated H=3 contrast on whitespace and leaves all four non-whitespace conditional point estimates non-positive.Source: §6 · p. 9

7 DISCUSSION AND LIMITATIONS

The replicated pooled depth contrast is concentrated in whitespace targets, and excluding whitespace reverses positive estimates across high- and medium-risk pairs. The paper recommends reporting target mixtures, decomposing effects by target type, and checking mixture-matched summaries before interpreting cross-corpus depth effects as specialization.

7 DISCUSSION AND LIMITATIONS

  • Report each corpus’s target-type proportions and characters per token, then log target IDs, report conditional cells, and split remaining positive cells by unablated correctness.The paper presents these checks as reusing predictions already collected.Source: §7 · p. 9
  • A positive CAI has a task-level interpretation only when the corpora pose comparable target-token prediction tasks; the reported main claim is specifically about H=3 mean-blend top-1 CAI.The paper notes that its position analysis covers canonical-pair whitespace and its control-accuracy split covers H=1 mild-resample, leaving positive H=1 mean-blend content cells unsplit.Source: §7 · p. 9, §A.1 · p. 14

A.2 TARGET-TOKEN CLASSIFIER

  • The audit classifies each target token into whitespace, number, punctuation, function word, or content word using ordered rules on its decoded string.The vocabulary-level map is applied to the target token at each evaluated source position.Source: §A.2 · p. 15

A.3 BANDS, HORIZONS, AND OPERATORS

  • The canonical audit uses first- and last-third layer bands, H=3 mean-blend for the aggregate effect, and H=1 mild-resample for calibration and the content-token cell.Both interventions blend each band’s block outputs at every position; donor sequences for mild-resample come from the same corpus and are deranged.Source: §A.3 · p. 15

C.1 CONDITIONAL CAI CELLS

  • The positive H=3 mean-blend effect is isolated to whitespace: non-whitespace cells are non-positive, and excluding whitespace reverses the estimate in all three primary models.Across positive pooled estimates in high- and medium-risk pairs, whitespace exclusion reverses every estimate.Source: §7 · p. 9, §C.1 · p. 16

C.4 TARGET-TYPE WEIGHTS

  • GSM8K has higher H=3 whitespace and number target rates than Wikitext, so their pooled scores average over different target-token mixtures.Whitespace can also encode structure in code, markup, and tables, making source code a boundary case for the screen.Source: §7 · p. 9, §C.4 · p. 18, Table 7 · p. 18

C.5 MIXTURE DIAGNOSTICS

  • Reference-mixture reweighting reduces the estimate to near zero or reverses its sign, while the whitespace-excluded estimate is negative in all three primary models.The reference-mixture summary changes weights on measured conditional cells; it does not remove the whitespace cell.Source: §7 · p. 9, §A.1 · p. 14, §C.5 · p. 19

D EARLY/LATE MECHANISM DECOMPOSITION

  • Late-band mean-blend carries the whole positive whitespace cell in every model, lifting GSM8K whitespace accuracy from near zero to most targets, with a gain several times its Wikitext change.On GSM8K, the late band predicts a space at 0.70 to 0.91 of all positions, compared with 0.09 to 0.10 under control, and predicts whitespace about as often at non-whitespace targets.Source: D EARLY/LATE MECHANISM DECOMPOSITION · p. 19

E POSITION AND ANSWER-DISTANCE CONTROLS

The whitespace effect persists under controls for relative position and distance from GSM8K’s final answer span. Position reweighting and answer-distance comparisons preserve positive contrasts.

E POSITION AND ANSWER-DISTANCE CONTROLS

F.1 AUDIT CELLS

The audit finds a strong whitespace-specific H=3 mean-blend contrast across five instances, while other interventions and readouts show that the pattern depends on target type, horizon, and measurement. Target conditioning also changes the interpretation of a non-intervention layerwise readout.

F.1 AUDIT CELLS

  • All five same-sequence audit interactions are positive, with no non-positive draw among 10,000 bootstrap resamples; the H=3 mean-blend whitespace cell exceeds its content cell by more than 0.51 in every instance.The H=1 mild-resample whitespace and content cells agree within 0.011.Source: §F.1 · pp. 23–24, Table 15 · p. 23

F.2 WHAT THE INTERACTION SUBTRACTS

  • The diagnostic interaction compares the H=3 mean-blend whitespace-minus-content gap with the corresponding H=1 mild-resample gap.It measures the separation of the primary-setting whitespace cell from its content cell relative to the reference setting.Source: §F.2 · p. 24

G.1 OPERATOR, HORIZON, AND TARGET-TYPE GRID

  • Across the operator-by-horizon grid, whitespace concentration is largest for H=3 mean-blend, while H=3 mild-resample is near zero for both target types.The grid reports top-1 CAI cells for GSM8K/Wikitext; the two operators’ estimates use different sample sizes.Source: §G.1 · p. 24, Table 16 · p. 24, Figure 6 · p. 25

G.2.1 LOG-PROBABILITY CAI TARGET-TYPE AUDIT

  • With log-probability CAI, the pooled result reverses for Qwen but remains positive for Llama and Gemma before reference reweighting; content, punctuation, and function-word cells are negative in all three.The number cell is positive for Llama and Gemma.Source: §G.2.1 · p. 26, Table 17 · p. 26

G.2.2 CONTINUOUS TARGET-MARGIN AUDIT

  • All six target-margin whitespace intervals exclude zero, while five of six content point estimates are negative and the remaining Gemma H=1 interval includes zero.Target margin is the gold token’s logit minus its largest competitor, allowing continuous changes even when the gold token remains top-1.Source: §G.2.2 · p. 26, Table 18 · p. 26

G.2.3 NON-CAI LAYERWISE LENS AUDIT

  • In the non-intervention logit-lens readout, H=1 content-target late-layer contrasts exceed the pooled estimate in all three models, while H=3 contrasts are near zero.This layerwise top-1 domain difference is averaged over the final third of layers and lies outside the CAI estimand.Source: §G.2.3 · p. 26, Table 19 · p. 26

G.2.4 ACTIVATION PATCHING WITH TARGET-TYPE UNCERTAINTY

Across naturalistic corpus pairs, target-token composition remains relevant when the audit changes interventions and extends to held-out pairs. The tokenizer-only screen prospectively stratifies measured risk, while code shows why whitespace frequency alone is not enough.

G.2.4 ACTIVATION PATCHING WITH TARGET-TYPE UNCERTAINTY

  • Same-domain patching is positive for content and non-whitespace targets in all three primary models, while cross-domain signs vary by model.The check uses H=1 target log probability; Figure 6c shows the intervention with top-1 correctness for whitespace and content-word targets.Source: §G.2.4 · p. 27, Table 20 · p. 27

H PRIOR-PATTERN TAXONOMY

I CONDITION ACCOUNTING FOR THE COVERAGE TABLE

  • The coverage accounting includes 41 distinct conditions across 13 pairs and five instances after adding audit-only and screen-only cases.Table 23 reports pooled and whitespace-excluded H=3 CAI intervals for ten expansion pairs.Source: I CONDITION ACCOUNTING FOR THE… · p. 28, Table 23 · p. 29

J.1 PROSPECTIVE HELD-OUT DOMAIN-PAIR EXPANSION

  • All nine scored held-out assignments fall in their predicted risk regions, though Gemma on AGNews/Wikitext is an exception to the interval prediction.The exception’s interval is above zero despite point estimates meeting the predefined low-risk criteria.Source: §J.1 · pp. 30–31

J.2 REASONING-PAIR EXTENSION

  • Across 12 model–pair rows, the whitespace-rate gap and pooled CAI have Spearman ρ = 0.9371, with agreement in 60 of 66 pairwise orderings.MATH and DROP yield six positive interactions with intervals excluding zero; ARC’s three interaction intervals include zero.Source: Table 27 · p. 32, §J.2 · p. 32

J.3 DOLA-MOTIVATED CORPUS-PAIR AUDIT

  • All six prospective DoLa-motivated outcomes fall in their predicted regions: GSM8K/FACTOR-News shows the high-risk pattern, while StrategyQA/FACTOR-News intervals contain zero.The audit measures the corpus pairs using the paper’s interventions, not DoLa’s layer selection or decoding gains.Source: §J.3 · pp. 32–33, Table 29 · p. 33

J.4 CALIBRATION-SCREEN ROWS BY MODEL AND DOMAIN PAIR

  • The screen’s two tokenizer-only predictors separate high-, lower-, and boundary-risk pairs; characters per token stays positively related to risk across leave-one-out subsets.Whitespace rate tracks risk for prose but saturates on code, where whitespace is structural.Source: Figure 7 · p. 34, §J.4 · p. 35, §J.5 · pp. 35–36

J.6 CODE BOUNDARY DETAILS

  • Code’s high target whitespace rate mostly reflects newlines and indentation, unlike the plain-space targets in GSM8K/Wikitext.The CodeSearchNet code/docs pair shows the same qualitative boundary within one source family.Source: §J.6 · p. 36

J.8 WHITESPACE SUBTYPE AND CLASSIFIER ROBUSTNESS

The whitespace finding persists under alternative token-classification policies, although reallocating leading-whitespace tokens changes cell sizes and estimates. The appendix also describes how the one-token-ahead content effect is scaled against control accuracy.

J.8 WHITESPACE SUBTYPE AND CLASSIFIER ROBUSTNESS

  • Under the canonical classifier, every non-whitespace subclass is non-positive for the H=3 mean-blend artifact; leading-space word tokens retain their non-whitespace class after stripping.The policy analysis keeps sequences, activations, operator, horizon, and bootstrap procedure fixed while changing the vocabulary map.Source: §J.8 · pp. 36–37

J.9 VOCABULARY-POLICY INTERVENTION AUDIT

  • Across classifier policies, whitespace remains the largest positive cell under the adversarial map and is the only positive cell for Qwen and Llama.That map assigns every leading-whitespace token to whitespace; its shared positive Gemma content estimate has intervals including zero.Source: §J.9 · pp. 37–38, Table 41 · p. 38

K.1 CANONICAL H=1 CONTENT-TOKEN EFFECT-SIZE ROWS

  • The H=1 content-token effect size is reported relative to GSM8K control accuracy, the higher control level in every row.The relative percentage is 100× CAI divided by that control top-1 level.Source: §K.1 · p. 39, Table 43 · p. 39

K.2 FIVE-INSTANCE H=1 CONTENT-WORD RESULTS

At H=1 with mild-resample, positive cells occur on non-whitespace types, unlike the H=3 mean-blend pattern. For content words, reweighting by control-pass correctness removes the positive interval in every instance.

K.2 FIVE-INSTANCE H=1 CONTENT-WORD RESULTS

  • The H=1 content-word mild-resample results are reported with 95% bootstrap intervals, using 1,000 sequence-bootstrap resamples.Table 43 gives the four-row Cohen’s d result with paired per-sequence diagnostics.Source: Table 44 · p. 39

K.3 THE OTHER FOUR TARGET TYPES IN THE SAME H=1 SERIES

  • At H=1 under mild-resample, positive cells occur for four number targets, three punctuation cells, and one function-word cell, but none for whitespace.This non-whitespace pattern inverts the H=3 mean-blend ordering.Source: §K.3 · p. 39, Table 45 · p. 40

K.4 THE CONTENT CELL SPLIT BY CONTROL-PASS CORRECTNESS

  • For content words, every reweighted interval includes zero, and the paired difference from the original cell excludes zero in all five instances.Only Llama’s control-correct content cell has an interval excluding zero.Source: §K.4 · p. 40, Table 46 · p. 41
  • Reweighting content-word strata based on input alone leaves a smaller positive cell, with intervals above zero in four of five instances.The exception is Gemma; the high-frequency tertile is positive in all five instances, while no low-frequency interval is above zero.Source: §K.4 · p. 41, Table 47 · p. 42

K.5 CROSS-DOMAIN ACTIVATION PATCHING

  • In the three primary models, activation patching and a log-probability readout replace ablation and top-1 CAI, respectively; domain separation distinguishes content/non-whitespace from the H=3 mean-blend whitespace cell.Source: §K.5 · p. 42

L.1 CONTROL SUMMARY

  • Uncertainty summaries use sequence-level resampling, with the number of resamples varying by audit.The control-summary table reports how many tested primary models satisfy each criterion.Source: Table 48 · p. 43

L.2 BASE-CHECKPOINT ROBUSTNESS

Base-checkpoint pooling and perturbation-strength sweeps show positive aggregate contrasts, while the reporting checklist calls for token-level decomposition and mixture-aware summaries before broad interpretation.

L.2 BASE-CHECKPOINT ROBUSTNESS

  • The pooled H=3 estimate was positive before instruction tuning in all three base checkpoints, although these estimates pool over target types.Table 49 reports pooled H=3 top-1 CAI under mean-blend at α = 0.2 for the base releases.Source: Table 49 · p. 43, §L.2 · p. 43

L.3 PERTURBATION STRENGTH

  • Across seven perturbation strengths, H=1 had four to six positive cells per model, while H=3 had at most two; H=1 strength patterns differed across models.Qwen’s H=1 interval stopped excluding zero above α = 0.2, and Llama and Gemma estimates were strictly negative at α = 0.1.Source: §L.3 · p. 43, Table 50 · p. 44
  • 13 of 15 model–strength pooled intervals lay above zero; whitespace was positive at every α from 0.10 in every model, while directly measured non-whitespace was never positive.The non-whitespace interval was below zero in 11 of 13 pooled-positive rows; the only positive whitespace-excluded summary was Gemma at α = 0.10, where the direct interval was [−0.0016, 0.0016].Source: §L.3 · p. 44, Table 51 · p. 45

M RECOMMENDED REPORTING CHECKLIST FOR FUTURE STUDIES

  • The checklist calls for prospective risk screening, conditional effects by target type, mixture-reweighted summaries, and correctness splits for any remaining positive cells.It also recommends calibration across operators or readouts when target-token asymmetry is non-trivial.Source: M RECOMMENDED REPORTING CHECKLIST FOR… · p. 46
  • The checklist also specifies logging target-token identities and positions, reporting target-type distributions and intervention settings, and noting sequence counts and bootstrap resamples.For code, markup, tables, or other formats with structural whitespace or punctuation, it recommends documenting domain semantics.Source: M RECOMMENDED REPORTING CHECKLIST FOR… · p. 46
  • An effect surviving the audit supports a positive finding for the audited target types, whereas an effect that collapses identifies a mixture-driven aggregate.Source: M RECOMMENDED REPORTING CHECKLIST FOR… · p. 46

N REPRODUCIBILITY DETAILS

  • The submission reports numerical results and measurement protocols, but does not include research scripts, machine-readable result files, or figure-generation scripts.Source: N REPRODUCIBILITY DETAILS · p. 47

O BROADER IMPACT AND DATA CONSIDERATIONS

The paper frames the audit as a way to test whether aggregate depth effects depend on target mixture and pairs mixture correction with reporting remaining positive cells. Its position summary reports whitespace intervals excluding zero after position matching.

O BROADER IMPACT AND DATA CONSIDERATIONS

  • An intervention-based audit pairs mixture correction of the pooled estimate with target-type cells that remain positive, split by unablated-model correctness.The submission contains no research scripts, machine-readable result files, or model weights.Source: O BROADER IMPACT AND DATA CONSIDERATIONS · p. 48

P NUMERICAL SUMMARY FOR THE POSITION FIGURE

  • Figure 5 estimates are also reported numerically, and Figure 6’s operator-by-horizon grid is tabulated alongside the figure.Source: P NUMERICAL SUMMARY FOR THE POSITION… · p. 48
  • All 15 whitespace H=3 intervals excluded zero across relative-position quartiles and after position matching.Table 52 reports point estimates; full intervals, event counts, and answer-distance bins are in Appendix E.Source: Table 52 · p. 48

Related papers

Loading 2610.19344v1…