PaperlayerView exact revision on arXiv

Source-linked AI summary

Cross-Prompt Reuse of Entity-Replacement Responses

Jinkun Chen, Xuemin Yu, Sijia Han, Vlado Keselj

arXiv:2611.00352v1cs.CLcs.AI

TL;DR

The paper asks how much of an entity-replacement response transfers between factual questions, beyond whether answers change. It transfers source-question changes into a target question and measures the resulting answer-contrast shift; reuse reproduces some overall change, but not reliably the relation-selective component on small plain-format checkpoints.

  • Problem

    Prior prompt-transfer work and factual-recall accounts do not establish how much of a receiving question’s response is reproduced, including its relation-specific component.

    Source: §1 · p. 1

  • Method

    The authors add aligned entity-replacement changes from source questions to a target and compare predicted with measured log-probability shifts between the two relations’ published answers, given known answer prefixes.

    Source: §1 · p. 1, §3.1 · p. 4

  • Results

    Other-relation hidden reuse removes 38–51% of no-change error on plain-format Qwen and Llama; Gemma reaches 58–69% with BOS restored, while selectivity remains unreliable on small plain-format checkpoints.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Overall response reproduction and relation selectivity can diverge: substantial transfer may reproduce a shift shared by both questions without reliably reproducing their relation-specific shifts.

    Source: §1 · p. 2, §5 · p. 9

    Evaluation covers six instruct checkpoints across three families with size and training lineage confounded, and new replacements remain within largely familiar relation and template families.

    Source: §7 · p. 10

Abstract

from arXiv · show

With model weights fixed, we measure how much of an entity-replacement response transfers between English factual questions, beyond whether answers change. We add the hidden-state change from two source questions to a target question about the same entity, comparing the predicted log-probability shift between two relations' published answers, given known answer prefixes, with the measured one. Hidden reuse injects the aligned change at 50% depth before the target's remaining layers; direct reuse adds source output-score changes. With other-relation sources, hidden reuse removes 45% and 49% of no-change error on 0.5-1B Qwen and Llama. Gemma-2-2B removes 22% and passes one of three comparisons in plain text without its default beginning-of-sequence (BOS) token. A later protocol restoring it yields 58-68%, passing every comparison on three cohorts. Larger checkpoints, chat format and new entities show replication and format- or cohort-dependent failures. On small plain-format checkpoints, other-relation hidden reuse mainly reproduces the shift both relations' questions share, without reliably lowering relation-selective error. With target-relation sources, hidden reuse's post hoc advantage over direct reuse is 0.2-0.9 nats outside chat format, unconfirmed and not replacement-specific. The same relation-neutral change at both questions' entity tokens lowers relation-selective error more than random or other-replacement changes, indicating receiving-question dependence.

1 Introduction

The paper measures how much of an entity-replacement response transfers between factual questions, including whether the transferred change reflects each question’s relation. Other-relation hidden reuse reproduces part of the response, but relation-selective reproduction is unreliable on small plain-format checkpoints.

1 Introduction

  • Entity replacement makes the target’s own response measurable, extending beyond whether an interchange changes an answer to how much of the target’s signed contrast change transfers.Source: §1 · p. 1
  • The evaluation compares predicted with measured shifts in the log-probability contrast between two relations’ published answers, using known answer prefixes.The contrast separates a shift shared by both questions from shifts toward each question’s own relation answer.Source: §1 · p. 1
  • Hidden reuse injects aligned source changes at 50% depth for the target’s remaining layers to read out, whereas direct reuse adds source output changes.The routes use the same two source questions and similarity weights.Source: §1 · pp. 1–2
  • Figure 1 contrasts hidden reuse with zero prediction, other-relation hidden reuse, and direct reuse; its answer-mass shares measure contrast reproduction, not answer correctness or free generation.Source: Figure 1 · p. 2, §1 · p. 2
  • Other-relation hidden reuse removes 38–51% of no-change error on plain-format Qwen and Llama; Gemma reaches 58–69% with BOS restored, while selectivity remains unreliable on small plain-format checkpoints.Gemma passes every comparison with BOS restored; without it, it passes at most one of three comparisons per cohort.Source: §1 · p. 2
  • The same relation-neutral change at both questions’ entity tokens lowers relation-selective error more than random or other-replacement changes, indicating receiving-question dependence.Source: §1 · p. 2
  • Span-aligned addition outperforms the best tested single-vector construction on five of six checkpoints, and four under pooled correction across 223 tests.Transferred changes up to 50% depth remove somewhat less error than the target’s own change at those tokens.Source: §1 · p. 2

2 Related Work

Related work studies transferring recorded changes, representations, and edits across prompts, alongside accounts of factual recall and intervention. This paper distinguishes its measure of transferred signed response from answer agreement and motivates controls for perturbation effects.

2 Related Work

  • Activation and function vectors transfer recorded prompt-pair differences [Turner et al., 2024; Panickssery et al., 2024], while entity interchange and patching study related but different transfer questions [Huang et al., 2024; Wang and Xu, 2025].The cited interchange work measures answer agreement and probabilities, unlike this paper’s share of the target’s signed contrast change.Source: §2 · p. 3
  • Factual-recall accounts predict mid-depth transfer: subjects are enriched before attributes are extracted [Geva et al., 2023], and subject lookups can be context-independent.Other work describes additive relation-agnostic and relation-specific components [Chughtai et al., 2024; Merullo et al., 2024].Source: §2 · p. 3
  • Prior findings that random norm-matched vectors can shift distributions and content-free directions can move behavior motivate random and different-replacement controls [Turner et al., 2024, App. G; Malla et al., 2026].Source: §2 · p. 3

3 Setting and Response Comparisons

The section defines a controlled comparison of entity-replacement responses across factual questions, using answer contrasts to measure how much a target question reproduces. It tests hidden-state and output-score reuse across models, cohorts, and response-comparison protocols.

3 Setting and Response Comparisons

  • The answer studies used three cohorts of new-entity blocks, totaling 99 blocks, with four targets per block and disjoint inputs within each cohort.They used Qwen2.5-0.5B-Instruct, Llama-3.2-1B-Instruct, Gemma-2-2B-it, and same-family 7B, 8B, and 9B replications.Source: §3 · pp. 3–4, Table 1 · p. 5

3.1 Response reuse

  • Hidden reuse injects aligned source hidden-state changes at 50% depth and runs the target’s remaining layers, while direct reuse adds weighted source output-score changes.The answer studies use suffix interpolation and score each candidate prefix with its own source capture.Source: §3.1 · p. 4

3.2 Answer contrasts

  • Answer contrasts compare replacement-induced changes in the log probability masses of the target-relation and other-relation published answers, scored as complete strings.The method evaluates known answer prefixes and compares predicted with observed contrast changes.Source: §3.2 · p. 4
  • In Qwen 0.5B’s original plain-format evaluation, hidden reuse removed 45.2% of the no-change error (2.869 nats for zero prediction; 1.572 for hidden reuse).The comparison measures contrast reproduction, not answer correctness or free generation.Source: §3.2 · p. 5
  • Half the sum of the two swapped-question changes isolates relation-selective movement, while half their difference isolates the shift shared by both questions.Selective error measures the distance between the predicted and observed selective components.Source: §3.2 · p. 5
  • Confirmatory comparisons were pre-specified and Holm-corrected within families, while descriptive and post hoc analyses were reported without confirmatory status.The tests assess recurring block-level effects; a failed test does not establish equivalence, and cross-cohort tests assume block independence despite reused relation labels.Source: §3.2 · p. 5

4 Cross-Prompt Response Reuse

The section tests how much an entity-replacement response transfers between factual questions, how that depends on source relation and encoding, and where in the target computation reuse remains effective. Results show substantial answer-score reuse in several settings, but weaker or unconfirmed relation-selective and target-relation advantages.

4.1 Answer-conditioned response reproduction

  • 45–51% of no-change error is removed on Qwen and Llama, versus 14–22% on Gemma in the original plain-format cohort.With BOS restored, Gemma removes 58–69% across three cohorts and passes every comparison; chat-format comparisons all pass, while fresh-cohort results vary by checkpoint.Source: Table 2 · p. 6, §4.1 · p. 6
  • Excluding targets whose answers share a word with an entity name brings plain-format Gemma's pooled shares near zero, while changing other pooled shares by at most five points.Without those targets, Qwen and Llama comparisons still pass, but Gemma-2-2B's sole original pass no longer does.Source: §4.1 · p. 6
  • 88–97% of full-vocabulary KL divergence is reduced at the answer cue by hidden reuse, versus 44–87% by other-relation direct reuse.Greedy generation reproduces edited answer strings for 5–23% of targets within 16 tokens, but its construction and scope differ from candidate scoring and do not isolate autoregressive error accumulation.Source: §4.1 · pp. 6–7

4.3 Standard constructions on the testbed

  • Span-aligned addition beats the best tuned single-vector construction on all but Llama-3.1-8B under family-wise correction and on four checkpoints under pooled correction.The best construction is selected on the original cohort and evaluated alongside span addition on 64 already-studied new-entity blocks.Source: Table 3 · p. 7, §4.3 · p. 9

4.1 Answer-conditioned response reproduction

  • With target-relation sources, hidden reuse's 0.20–0.90-nat advantage over direct reuse outside chat format is unconfirmed and not replacement-specific.A pre-specified test failed on all six checkpoints, and different replacements gained more from hidden reuse than the correct replacement in the reported plain and plain+BOS comparisons.Source: §4.1 · p. 7
  • 9–38% of observed variance is relation-selective on original plain-format blocks, while hidden reuse cuts shared error by 34–62% on small checkpoints.Selective-error intervals are not wholly positive on small plain-format checkpoints; on two larger checkpoints, same-replacement reuse lowers selective error by 34–36% in fresh cohorts, compared with −9 to 5% for different-replacement reuse.Source: §4.1 · p. 7
  • 60–83 of 99 blocks show lower selective error from an identical relation-neutral entity-row change, which beats random and other-replacement changes on all six checkpoints.For the identical-change tests, random perturbations and the other replacement's neutral change lower selective error by at most 0.22 nats.Source: §4.1 · pp. 7–8

4.2 Depth and row position of answer reuse

  • 100% of error reduction at d = 0 declines to 46–70% at d = 1/2; at half depth, omitting answer-prefix rows preserves most of the aligned-addition share.Up to half depth, transferred entity rows remove somewhat less error than the target's own entity-row patch.Source: Figure 2 · p. 8, §4.2 · pp. 8–9

5 Discussion

The discussion separates substantial but incomplete overall response fidelity from relation-selective transfer, which is unreliable on small plain-format checkpoints. It also examines how source choice and injected-change construction affect the comparisons.

5 Discussion

  • Overall fidelity is substantial but incomplete, while relation selectivity is not reliably reproduced on small plain-format checkpoints.Other-relation transfer mainly reproduces the shift shared by both questions.Source: §5 · p. 9
  • Target-relation sources substantially narrow the hidden-over-direct gap except for plain-format Gemma, while the residual advantage remains unconfirmed and not replacement-specific.The identical relation-neutral change retains selective effects that depend on the receiving question, beyond random and other-replacement controls.Source: §5 · p. 9
  • Aligned changes on all rows beat entity-row-only changes at 50% depth and the best tested single vectors on five of six checkpoints.Scaling single vectors to span-aligned row norms narrows the gap, but directions, coverage, and prefix information also vary, so row structure is not isolated.Source: §5 · p. 9

6 Conclusion

The conclusion finds that entity-replacement responses are partly portable across questions, while emphasizing limits in relation selectivity and construction-dependent comparisons. The measurements quantify response magnitude, selectivity, and construction dependence.

6 Conclusion

  • Other-relation hidden reuse removes 38–51% of no-change error on plain-format Qwen and Llama, and 58–69% on Gemma with BOS restored.Results include format- and cohort-dependent exceptions.Source: §6 · p. 9
  • On small plain-format checkpoints, other-relation hidden reuse mainly reproduces the shared shift without reliably lowering relation-selective error.The identical relation-neutral change can nevertheless produce selective effects that depend on the receiving question.Source: §6 · p. 9
  • With three cohorts pooled, target-relation sources substantially narrow the hidden-over-direct advantage except for plain-format Gemma; the residual advantage is unconfirmed and not replacement-specific.Source: §6 · p. 9
  • Span-aligned addition beats the best tested single vectors on five of six checkpoints, four under pooled correction.Together, the measurements quantify the magnitude, relation selectivity, and construction dependence of cross-question response reuse.Source: §6 · p. 9

7 Limitations

The limitations constrain interpretation by identifying design, evaluation, and statistical boundaries. The appendices describe the scoring and reuse constructions, cohorts, and supporting analyses.

7 Limitations

  • Relation labels covary with question wording, cohorts largely reuse familiar relation and template families, and eligibility is limited to equal-length English replacements.Results cover six instruct checkpoints from three families, with size and training lineage confounded; chat changes multiple format components, and some comparisons and fresh cohorts were not tested in chat.Source: §7 · p. 10, §B.2 · p. 17
  • Confirmatory studies use one interface at 50% depth, while construction comparisons jointly vary coverage, direction, and prefix information and do not test behavioral steering.Complete-answer evaluation conditions sources and targets on known published-answer prefixes; candidate groups carry little absolute probability mass and exclude aliases and formatting variants. Statistical inference assumes independent blocks or relation/template components, with possible dependencies and limited power.Source: §7 · p. 10

Appendix guide

  • The appendix guide maps Appendices C–G to the analyses supporting Sections 4.1–4.3 and identifies tables inventorying test families and correction-sensitive conclusions.Appendix A covers constructions and scores, and Appendix B covers blocks, cohorts, checkpoints, formats, and study order.Source: Appendix guide · p. 15

A Response Construction and Scores

  • Suffix alignment maps source rows positionally between the first changed and final rows, and fixed source weights use distances between normalized unedited final states.The mapping is a positional heuristic, not semantic correspondence; direct reuse does not pass source changes through the target’s remaining layers.Source: A Response Construction and Scores · pp. 15–16
  • Recovery values of 1 and 0 denote exact reproduction and the clean target’s error, respectively; negative values denote larger error.Row-norm matching fixes row magnitudes, not content.Source: A Response Construction and Scores · p. 16
  • Complete-answer scoring compares predicted and measured changes in published-answer-group contrasts, conditioning each candidate position on known preceding tokens.The clean target also receives each prefix, and source weights are fixed across candidates.Source: A Response Construction and Scores · p. 16

B.1 Blocks and eligibility

  • Each answer-study block crosses two factual entity replacements, two queried relations, and six published-question templates, with sources and targets drawn from those combinations.The introductory example and Table 5 were selected retrospectively for readable wording, not outcomes.Source: §B.1 · p. 17, Table 5 · p. 17

B.2 Answer cohorts and overlap

  • The original answer cohort has 35 disjoint four-target blocks; two fresh cohorts add 33 and 31 blocks but reuse largely familiar relation labels and templates.The fresh cohorts therefore assess new entities within familiar relation and wording families, not new-family generalization.Source: §B.2 · p. 17

B.3 Artifacts, data and computation

The study uses fixed released checkpoints, public English factual materials, and controlled input protocols. Its scope is limited to controlled analysis rather than deployment, with data, licensing, and fairness boundaries noted.

B.3 Artifacts, data and computation

  • The study tests fixed Qwen, Llama, and Gemma checkpoints at 50% depth, with later scale replications and plain, plain+BOS, and chat-format protocols.Plain-format studies omit the default BOS token for Llama and Gemma; later protocols restore it or use chat templates.Source: §B.3 · p. 17
  • Materials are public English factual questions and answers, Wikidata metadata, and public named-entity facts; no new personal data are collected.The exact ZsRE and LPAQA prompt-file licensing could not be established, so those materials require upstream attribution and terms review before repackaging.Source: §B.3 · p. 18
  • Runs use CPU inference, float32 weights, float64 scoring, and a logit-replay check requiring continuations to match clean logits within 10^-3.The authors plan to release code, identifiers, predictions, run records, and protocols, subject to source-material terms.Source: §B.3 · p. 18
  • Reported compute totals distinguish CPU time from allocated core-hours and do not infer a project-wide total.Saved-data checks run no models.Source: §B.3 · p. 18
  • The study does not conduct a dedicated offensive-content audit, and its English-only selection and limited checkpoints preclude population-level fairness claims.The authors warn that public named-entity facts may be sensitive, outdated, or biased.Source: §B.3 · p. 18

B.4 Earlier next-token studies

Earlier next-token studies tested finite-source changes against derivative and replacement controls, with outcomes varying by study. The later answer study followed a staged protocol in which prior descriptive results informed subsequent designs, but each core study had a written protocol before its own outcomes.

B.4 Earlier next-token studies

  • All 12 multi-token comparisons passed their Holm family, while only the three identity contrasts passed in the single-token study.The single-token derivative comparisons remained inconclusive across ten components.Source: §B.4 · p. 19
  • The studies compared finite-source changes with target- and source-local derivatives and with same-versus-different replacement controls.The multi-token sample contained 184 replacements across 46 relations and 45 components.Source: §B.4 · p. 19
  • Across the 40-block crossing and native-amplitude studies, all reported tests passed within their families, but the direction studies do not establish complete-answer fidelity.Five of six mean margin contrasts changed sign under alternative direction rescalings.Source: §B.4 · p. 19

B.5 Study sequence

  • A 157-pair pilot was excluded because its pair-level tests assumed independence despite recurring subjects and relations; it fixed the later 50% interface.The later descriptive all-layer scan did not retune that interface.Source: §B.5 · p. 19
  • Each core study had a written protocol before its edited-target outcomes, but later designs were informed by earlier findings and some diagnostics were post-outcome.The answer study selected other-relation sources after an earlier descriptive comparison separated the routes most clearly.Source: §B.5 · p. 19

C Prediction Fidelity, Calibration and the Answer Estimand

This section distinguishes reproduction of a measured answer-contrast change from answer correctness, and describes calibration, block-level comparisons, and a shared-versus-selective decomposition. It also qualifies interpretation by candidate scope and post hoc sensitivity analyses.

C Prediction Fidelity, Calibration and the Answer Estimand

  • Primary answer comparisons measure reproduction of a change against direct, zero-prediction, and different-replacement baselines, with mean intervals and sign tests serving distinct roles.Sign tests assess how often a block favors a comparison; mean-effect intervals describe magnitude.Source: C Prediction Fidelity, Calibration and… · p. 19, Table 7 · p. 20, Table 8 · p. 20
  • Figure 4 displays block-level error reductions for same-replacement hidden reuse against direct reuse, zero prediction, and different-replacement hidden reuse.Its points encode block means, with positive and negative outcomes distinguished and pointwise 95% intervals shown.Source: Figure 4 · p. 21
  • Calibration regresses observed change on predicted change without recalibrating predictions; slope alone is insufficient because intercept and mean bias also matter.Figure 3 places observed-on-predicted calibration fits against the exact-reproduction line.Source: C Prediction Fidelity, Calibration and… · p. 20, Figure 3 · p. 21, Table 9 · p. 22
  • Good total response reproduction need not imply good relation-selective reproduction, so the section reports both components’ correlations and absolute errors.Selective-error reductions include nonpositive reductions.Source: C Prediction Fidelity, Calibration and… · p. 22, Table 12 · p. 23
  • Balanced relation-swapped pairs decompose observed contrast variance into selective variance and a shared component, but this algebraic summary does not identify hidden-state mechanisms.The selective share is not an attribution of the primary absolute-error gain.Source: Table 11 · p. 22, C Prediction Fidelity, Calibration and… · p. 22
  • The contrast covers literal published answer strings whose candidate probability mass is small, limiting interpretation as factual recall or free generation.Post hoc checks also show that hidden-route advantage can occur for a wrong replacement and is not by itself evidence of replacement-specific transfer.Source: C Prediction Fidelity, Calibration and… · p. 22, Table 13 · p. 23, Table 15 · p. 24

D Replication, Encoding and Source-Relation Comparisons

Replication spans checkpoint scales, formats, and new entities, while the new-entity cohorts largely retain familiar relations and templates. Source-relation matching narrows route gaps descriptively, and format- and cohort-specific qualifications remain important.

D Replication, Encoding and Source-Relation Comparisons

  • New-entity cohorts mostly reuse familiar relations and templates, so they test entity replication rather than broad generalization to new relation families.The section reports larger-checkpoint, chat-format, fresh-cohort, matched-route, and neutral-source comparisons, including failed comparisons.Source: D Replication, Encoding and… · p. 26
  • Matching the source relation substantially narrows the mean route gap except for plain-format Gemma, where the reduction is small.The source response and its read-out both change, and the residual matched-route advantages are descriptive rather than confirmed by the second-cohort pre-specified tests.Source: D Replication, Encoding and… · p. 26, Table 18 · p. 26
  • The second fresh cohort’s matched-route tests all failed, but that failure does not establish equivalence.The source-comparison tables distinguish cohort-specific tests from descriptive pooled summaries.Source: Table 18 · p. 26, D Replication, Encoding and… · p. 26, Table 22 · p. 28
  • Restoring only the BOS token is distinct from using a chat template, and the paper reports these formats separately rather than pooling away Gemma’s plain-format failures.Selectivity sign tests alone are insufficient because a constant push toward each question’s own answers could also pass the specified tests.Source: D Replication, Encoding and… · p. 29, Table 23 · p. 29, Table 24 · p. 30
  • In the original 35-block chat cohort, selective error reduction was 21.8% for Qwen 0.5B, 40.7% for Llama 1B, and 38.3% for Gemma 2B.The corresponding larger-checkpoint values were 37.2% for Qwen 7B, 6.7% for Llama 8B, and 52.1% for Gemma 9B.Source: D Replication, Encoding and… · p. 30

E Depth, Row Coverage and Conditional Scoring versus Generation

The section compares reuse across depth and prompt rows, and distinguishes conditional answer scoring from full-vocabulary readouts and greedy generation.

Depth, Row Coverage and Conditional Scoring versus Generation

  • 51.7%, 55.3%, 68.5%, 54.5%, 48.7% and 61.9% of no-change error is removed by combining entity-row and post-entity-row predictions across the six listed checkpoints.Patching both row sets together reproduces the edit and removes 100% of error; the combined error reductions are not sums of the individual shares.Source: E Depth, Row Coverage and Conditional… · p. 31
  • The depth scan covers 35 original blocks and nine interfaces, with exact-span reuse from other-relation or target-relation sources compared against the target’s own change and direct reuse.The interface ranges from effective decoder input (d = 0) to the last layer output before final normalization (d = 1); hidden and direct reuse can differ at d = 1.Source: E Depth, Row Coverage and Conditional… · p. 31, Table 26 · p. 32
  • At d = 1/4 and d = 1/2, row-subset comparisons cover all, entity, question-and-cue, and post-entity rows across source-relation conditions.Qwen 7B’s saved scan lacks target-relation subsets and neutral sources.Source: Table 27 · p. 32
  • Full-vocabulary answer-cue readouts and greedy generation are distinct from per-prefix candidate scoring because their capture, construction, and evaluation scopes differ.The reported gap does not isolate autoregressive error accumulation, and low generation fidelity accompanies stronger conditional-score results.Source: E Depth, Row Coverage and Conditional… · p. 31, Table 28 · p. 33, Table 29 · p. 33

F Standard Patches and the Single-Vector Comparison

The section compares span-aligned reuse with tested single-vector constructions and describes the confirmatory test families and their evaluation scope.

F Standard Patches and the Single-Vector Comparison

  • 0.12–0.28 nats: the default-encoding comparison over 99 blocks favors span-aligned addition, with all six pointwise block-bootstrap intervals excluding zero.Coverage and construction change together, so this comparison does not isolate post-entity rows.Source: F Standard Patches and the… · p. 33
  • The stronger single-vector extension selects among 90 constructions on the original cohort and evaluates on two previously studied fresh cohorts, not an unseen research setting.Directions, placements, averaging and amplitudes vary, so the comparison is among tested constructions rather than an isolated row-structure effect.Source: F Standard Patches and the… · p. 33
  • The construction set varies source vectors, placements, coefficients and contexts, with the best construction selected by mean block answer error on the original 35 blocks.The selected construction is evaluated on the other 64 already-studied blocks; norm-matched constructions use magnitude information from span-aligned reuse.Source: F Standard Patches and the… · pp. 33–34
  • Confirmatory families compare later-row vectors, entity-row additions, tuned vectors, and steering-completion variants using exact sign tests with within-family Holm correction.The tables also report pooled-correction-sensitive tests and specify the favoring-block shares required for 80% power.Source: Table 30 · p. 34, Table 31 · p. 34

G Receiving-Question Dependence and Content Controls

This section tests whether identical relation-neutral entity-row changes depend on the receiving question, using knockout interventions and controls.

G Receiving-Question Dependence and Content Controls

  • Attention knockouts measure how answer-position attention to entity, question-and-cue, or pre-entity rows relates to error reduction, with outcomes varying by family.Knockouts are route interventions rather than identified heads or features, and KO-E also changes the clean answer contrast, limiting mechanistic interpretation.Source: Table 33 · p. 35, G Receiving-Question Dependence and… · p. 35
  • The same replacement’s neutral entity-row change outperforms random and other-replacement controls within the specified families, with pooled-correction exceptions.The controls use matched row norms; the other replacement’s change is mapped to the target span, and random-control sensitivity is not universally absent.Source: G Receiving-Question Dependence and… · p. 35, Table 34 · p. 36
  • On 28–32 blocks per checkpoint where span interpolation is unnecessary, NEU beats DR in 18–25 blocks; the uncorrected sign-test p exceeds 0.05 only on Qwen2.5-7B.For Llama-3.2-1B, the label-resampled interval for neutral content against zero is [−0.03, 0.37] nats despite its original within-family pass.Source: G Receiving-Question Dependence and… · pp. 35–36
  • The control analysis reports selective-error reduction from zero prediction and separately reports predicted selective-component size and full-contrast error removal.The control-comparison intervals resample blocks or relation labels; the full-metric share is not the same quantity as selective-error reduction.Source: Table 36 · p. 36, Table 35 · p. 36, Table 37 · p. 37
Loading 2611.00352v1…