acceptodds
Under review as a conference paper at ICLR 2027

Target-Token Asymmetry Can Masquerade as Depth Specialization

Abstract

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.

Primary area:
interpretability and explainable AI
Keywords:
Large Language Models, Mechanistic interpretability, Target-token composition

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Av4WOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper examines whether differences in target-token composition can distort layer analyses and produce misleading evidence of depth specialization. It introduces a cross-corpus contrast between early- and late-layer ablations, breaks it down by token type, and reevaluates it using shared target mixtures. In its comparison, whitespace targets drive the positive pooled contrast: removing them reverses its sign in all three fully decomposed models, while matching to a reference mixture reverses it in two. The paper also presents a tokenizer-only risk screen and a reporting protocol.

Critical strengths

S1. The paper identifies a relevant measurement confound. It clearly explains why replication across models does not establish task specialization when the evaluated corpora share the same differences in token composition.

S2. The central empirical finding undergoes extensive and rigorous scrutiny. The evaluation spans several model families and corpus pairs, with checks covering position matching, token-classification rules, continuous readouts, and perturbation strengths.

S3. The proposed reporting protocol is practical. Recording target types, reporting conditional effects, and comparing shared-mixture estimates can improve interpretability analyses using existing prediction data.

Critical weaknesses

W1. The main finding relies on a restricted setting. The analysis uses next-token logits to score tokens three positions ahead, while late-layer mean-blend causes the model to predict whitespace at a high proportion of GSM8K positions. Although other horizons and operators are examined, their patterns differ. Stronger evidence connecting this diagnosis to standard next-token prediction and task performance would clarify its practical relevance.

W2. Correctness matching requires cautious interpretation. Reweighting by unperturbed correctness (Section 5) may remove meaningful capability differences, so attenuation after matching alone does not rule out task-related depth specialization.

Questions for authors

See weakness

2 comments
AuthorsOpenReview ↗

Reply to Reviewer Av4W

Thank you for the careful reading and positive assessment. We address your two concerns below, using the new measurements summarized in the general response.

W1: standard next-token prediction and task performance.

We extended the mean-blend analysis to H=1–8 and compared both operators at H=1 on GSM8K/Wikitext. The mean-blend sweep uses five model instances, 300 sequences per corpus and 2,000 sequence-bootstrap draws. The H=1 mild-resample analysis uses 250 sequences per corpus.

  • H=2–8, mean-blend: in 34 of 35 model–horizon combinations, the pooled CAI has a pointwise 95% interval above zero and the directly measured non-whitespace CAI has an interval below zero. Whitespace exclusion reverses all 34 positive pooled estimates. The exception is Qwen2.5 at H=2, whose pooled interval contains zero.
  • H=1, mild-resample: after matching target type and unablated prediction correctness, the initially positive pooled CAI reverses in Qwen2.5, has a 95% interval containing zero in three instances, and remains positive but smaller in Llama-3.2.
  • H=1, mean-blend: the negative pooled CAI remains negative after the same adjustment in all five instances, with every interval below zero.

The audit therefore remains informative at ordinary next-token prediction, although the relevant composition factor and outcome depend on the intervention. Persistence after this adjustment does not rule out other composition effects.

We did not measure generation-based task accuracy. Our evidence concerns token-level contrasts under the tested interventions, not a demonstrated bias in end-task performance. Establishing that connection requires a separate evaluation.

W2: matching on unperturbed correctness.

We agree that matching on unperturbed correctness can remove meaningful capability-related variation. Attenuation after matching alone therefore does not rule out task-related depth specialization. We interpret the matched estimates as sensitivity checks and retain the unadjusted results.

To examine the mechanism beyond attenuation, we inspect the initially correct content-word predictions in the H=1 mild-resample series. For each corpus, the removal gap is the fraction of these predictions lost under the early band minus the fraction lost under the late band. The product below multiplies the mean of the two removal gaps by the difference in the control-correct shares of content targets. These diagnostics use the paper's sequence-averaged convention.

InstanceRemoval gap, GSM8K / WikitextMean gap × control-accuracy differenceContent-word CAI
Qwen2.5-7B0.024 / 0.028+0.0127+0.0103
Llama-3.2-3B0.041 / 0.030+0.0091+0.0142
Gemma-2-2B0.031 / 0.033+0.0091+0.0086
OLMo-2-7B0.048 / 0.046+0.0126+0.0148
Qwen3-1.7B0.037 / 0.037+0.0113+0.0125

The baseline accuracy difference is observed directly: the unablated model predicts a larger share of GSM8K content targets correctly. The removal gaps are similar across corpora in four instances, with a larger difference in Llama-3.2. The product is comparable in magnitude to the content-word CAI, but it is a descriptive calculation, not an exact decomposition or proof that the baseline difference is unrelated to capability. Llama-3.2 also retains a smaller positive pooled contrast after matching, as reported under W1.

Two further checks help delimit the interpretation:

  • The adjustment does not eliminate every contrast. The negative H=1 mean-blend contrast persists in all five instances. This shows that matching need not remove a contrast, not that it selectively removes only artifacts.
  • Input-only strata retain a smaller positive content-word contrast. Matching on token repetition and leading-space status uses no model predictions. The reweighted estimates remain positive, with 95% intervals above zero in four of five instances (Table 47). This residual deserves further investigation and is not itself proof of task specialization.

Our conclusion is that the pooled contrast is sensitive to which targets are evaluated and how predictable they are before intervention. The audit makes that sensitivity explicit. Neither attenuation nor persistence after matching, on its own, establishes whether the model has task-dependent depth specialization.

Thanks to the author's effort in addressing my concerns. I don't have any further questions now. And I will maintain my score.

Reviewer_ST1SOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper claims that whitespace prediction confounds prior work on identifying at which depth a certain capability is encoded in language models. Specifically, the work examines whether apparent depth specialization reflects differences in the tokens being predicted across datasets. Using early- and late-layer interventions, the authors find that whitespace targets drive the main GSM8K/Wikitext contrast, and that accounting for target-token composition can reverse the result. They also propose a tokenizer-based screening rule and recommendations for cross-dataset comparisons.

Critical strengths
  • The paper raises awareness that future work needs to pay attention to whitespaces and punctuation when conducting depth specialization experiments
  • Results show that using whitespace as target token vs the actual answer makes a difference
Critical weaknesses
  • The paper introduces a potential problem, but it is not argued convincingly that this is actually a problem in other papers. Notably, for
  • The paper reports on applying their own methodology to the DoLa dataset, but does not replicate the depth analysis of that paper. At this example, it is hard to judge that this would have been a problem for the methodology used in DoLa
  • before and after reading the paper, my opinion is unchanged that it is probably not the best idea to conduct analysis based on white space prediction in a math dataset. The paper quantifies that, but it is little surprising that it makes a difference.
Questions for authors

see weaknesses

1 comment
AuthorsOpenReview ↗

Reply to Reviewer ST1S

Thank you for the review. We address your concerns below. The central distinction is that we start with a score pooled over all evaluated targets, not a whitespace-only test, and ask which targets support its interpretation as a task-dependent depth effect.

1. Relevance to existing depth analyses.

Our experiments identify a risk in interpreting pooled token-level scores. They do not show that particular prior papers are wrong. Layer-intervention studies by Gromov et al. (2025), Nepal et al. (2025), Sun et al. (2025) and Song et al. (2026) motivate the question of task-dependent depth. Most use task accuracy, which our experiments do not directly audit. CAI is our diagnostic contrast, not an established evaluation protocol.

We examine a specific inference: that a cross-corpus early–late contrast in pooled accuracy reflects task-dependent layer specialization. Decomposing the contrast by target type and recomputing it under a shared target mixture tests how much that interpretation depends on which tokens each corpus asks the model to predict.

New measurements address whether the result is restricted to the original H=3 setting:

  • H=2–8, mean-blend: across five model instances on GSM8K/Wikitext, using 300 sequences per corpus and 2,000 sequence-bootstrap draws, 34 of 35 model–horizon combinations have a pooled CAI with a pointwise 95% interval above zero and a directly measured non-whitespace CAI with an interval below zero. Whitespace exclusion reverses all 34 positive pooled estimates. The exception is Qwen2.5 at H=2, whose pooled interval contains zero.
  • H=1, ordinary next-token prediction: under mild-resample, matching target type and unablated prediction correctness reverses the positive pooled contrast in Qwen2.5, leaves intervals containing zero in three instances, and retains a smaller positive contrast in Llama-3.2. Under mean-blend, the negative contrast remains negative, with its interval below zero, in all five instances after the same adjustment.

Thus, the audit is informative beyond one horizon, but the relevant composition factor and outcome depend on the intervention. Matching alone does not rule out meaningful capability differences, nor do these results establish that all depth effects are compositional.

2. What the DoLa-related experiments establish.

You are right that using datasets from DoLa does not reproduce its depth analysis. Our introduction also misdescribed its candidate-layer comparison. For LLaMA-7B, DoLa's Table 10 uses the same lower bucket for GSM8K, StrategyQA and FACTOR, whereas TruthfulQA uses a higher bucket. We apologize for this inaccurate comparison.

Our GSM8K/FACTOR-News and StrategyQA/FACTOR-News runs evaluate our tokenizer screen on additional corpus pairs. They do not test DoLa's layer-selection procedure and therefore do not establish whether DoLa itself is affected. Its observations about token-dependent behavior across layers motivate examining target types, but are not evidence of this confound in DoLa's own method.

3. Further evidence supporting our original conclusion.

We welcome the opportunity to provide further evidence for our original conclusion. The submitted results show that a positive pooled depth contrast can replicate across models yet reverse when whitespace targets are excluded. Our analysis starts with a score over words, numbers, punctuation and whitespace together, not a whitespace-only reasoning test. The additional horizon experiments build on this evidence and show that the sign pattern extends beyond H=3 under the tested mean-blend settings.

For the canonical H=3 mean-blend comparison (Table 1):

ModelPooled CAIDirectly measured non-whitespace CAI
Qwen2.5-7B+0.0253−0.0253
Llama-3.2-3B+0.0428−0.0329
Gemma-2-2B+0.0421−0.0155

All six 95% intervals exclude zero. Non-whitespace targets make up more than 90% of GSM8K targets in each model, yet their combined contrast has the opposite sign to the aggregate. The four non-whitespace target-type point estimates are also all non-positive. Thus, agreement across models does not by itself show that the aggregate reflects a broad task-related depth effect.

The paper also reports positive pooled contrasts on base checkpoints and after equalizing perturbation norms across corpora (Tables 48 and 49). These checks alone do not reveal which targets support the pooled sign.

The practical contribution is to make that interpretation testable: report target-type contributions and mixture-matched estimates before interpreting a pooled cross-corpus contrast as evidence of task-dependent depth specialization. The tokenizer-only screen helps identify comparisons that warrant this inspection. The lesson is not simply to avoid predicting spaces, but to check what an aggregate depth measurement is actually measuring.

Reviewer_T9s4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper raises the question of whether methods used to identify layer specialization are affected by confounders, in particular, the proportion of whitespace tokens in the corpus that model is asked to predict. The authors define a measure called the contextual ablation interaction: it quantifies how much perturbing early versus late layers affects prediction accuracy on this focus corpus relative to a reference corpus. They analyze how CAI changes depending on the type of target token and find for one particular setting that whitespace tokens are responsible for a large percentage of the observed effect, i.e., its the whitespace tokens that make it appear that there is a layer specialization effect. The paper also proposes a screening rule based on differences in the comparison corpora's whitespace rates and characters per token to identify potentially confounded comparisons

Critical strengths
  • The paper is overall clearly written and well motivated (from an intuition perspective, its lacking in citations to works for which these results would be relevant). The mechanism causing the confounder is well explained
  • The audit demonstrates the sensitivity of a measurement protocol to subtle artifacts: raising awarenes of these types of confounders is always beneficial to the community.
Critical weaknesses
  • It's unclear how prevalent the tested (rather specific) measurement actually is and therefore, how relevant the paper's main result actually is. The strongest results are shown for target prediction of a token three positions ahead under "mean-blending", but is the normal protocol for this type of analysis? No works that use this setup are cited and the authors don't provide a justification for these choices for their analysis setting. If I understood results correctly, then whitespace is not a significant confounder (or is a much weaker one) when the next token is the object of prediction or under the "mild resample" intervention protocol. In short, the paper raises an interesting point but I don't see any implications of it
  • The empirical evidence supporting the tokenizer screening rule, which would be a nice check/tool for people doing these types of analyses, is weak
Questions for authors
  • I didn't understand how this is an example of SImpson's paradox. Can you clarify?
2 comments
AuthorsOpenReview ↗

Reply to Reviewer T9s4

Thank you for the thoughtful review. Below we address your three concerns and summarize the relevant new measurements.

How common is this measurement, and why H=3 under mean-blend?

CAI is our diagnostic contrast, not an established evaluation protocol. Our motivation comes from two lines of work:

  • Task-dependent depth analysis. Layer-intervention studies compare depth requirements across tasks, as studied by Gromov et al. (2025), Nepal et al. (2025), Sun et al. (2025) and Song et al. (2026). Most of these use task accuracy.
  • Pooled token-level evaluation. Layer-deletion and layerwise studies also evaluate predictions over naturalistic text, as in Lad et al. (2025) and Csordás et al. (2025).

These studies motivate the audit. They do not establish that our exact protocol is standard. Song et al. (2026) also report that depth conclusions depend on the metric. Our contribution is to examine how target composition affects the interpretation of a pooled cross-corpus contrast.

Our operators are partial versions (α = 0.2) of mean and resample ablations (Li & Janson, 2024). H=3 mean-blend serves as a case study of aggregation bias, not a standard test of reasoning.

To test whether the finding is specific to H=3, we evaluated H=1–8 under mean-blend across five model instances on GSM8K/Wikitext, with 300 sequences per corpus and 2,000 sequence-bootstrap draws. Across H=2–8, 34 of 35 model–horizon combinations have a pooled CAI whose pointwise 95% interval lies above zero and a directly measured non-whitespace CAI whose interval lies below zero. Whitespace exclusion reverses all 34 positive pooled estimates. The exception is Qwen2.5 at H=2: its pooled estimate is −0.0013, with an interval containing zero. The pattern is not isolated to H=3 within the tested mean-blend settings.

At H=1 under mild-resample, the positive pooled contrast is not driven by whitespace. At H=1, matching target type and whether the unablated model predicts the target correctly gives different outcomes under the two operators:

  • Mild-resample: the initially positive pooled contrast reverses in Qwen2.5, has a 95% interval containing zero in three instances, and remains positive but smaller in Llama-3.2.
  • Mean-blend: the negative pooled contrast remains negative, with its interval below zero, in all five instances after the same adjustment.

The audit thus detects sensitivity to this composition adjustment without erasing every contrast. Matching alone does not rule out meaningful capability differences. The practical lesson is to inspect target-type contributions and mixture-matched estimates before interpreting a pooled token-level contrast as evidence of task-dependent depth specialization.

The tokenizer screen.

We appreciate the reviewer's suggestion to further evaluate the tokenizer-only screen. To provide additional validation, we evaluated all 55 previously untested pairs among the twelve non-code corpora on the three primary models (165 model–pair evaluations). The evaluation plan and tokenizer-derived labels were fixed before these new audit runs.

  • Sign reversals: among evaluations whose pooled 95% interval lies above zero, the whitespace-excluded estimate is negative in 49 of 51 high- or medium-risk cases, compared with 1 of 13 low-risk cases.
  • Tokenizer gaps: across the 163 evaluations with an estimable interaction I, its Spearman correlation is 0.61 with the whitespace-rate gap and 0.72 with the characters-per-token gap.
  • Region rule: 79 of 93 high- or low-risk evaluations fall in the predicted region. None falls in the opposite region.

The reversal results support the screen's diagnostic use in this evaluation, while the finer region rule is informative but imperfect.

Simpson's paradox.

Our citation to Simpson (1951) concerns aggregation bias, not a strict reversal. We apologize for the ambiguity. A strict reversal would put every stratum opposite to the pooled sign. Here the whitespace stratum agrees with it. For Qwen2.5 on the canonical pair:

EstimateCAI
Pooled+0.0253
Whitespace (8.1% of GSM8K targets)+0.5181
All non-whitespace targets measured together (91.9%)−0.0253
Each of the four non-whitespace typesBelow zero (Table 1)

The pooled contrast uses each corpus's own target composition. It is not generally a weighted average of the conditional CAIs under one shared mixture. Reweighting those cells gives:

  • Whitespace excluded: negative in all three primary models.
  • Wikitext's mixture: negative for Qwen2.5 and Llama-3.2, positive for Gemma-2.
  • Combined mixture: +0.0039, +0.0182 and +0.0257, respectively, smaller than the pooled values.

The within-type/composition decomposition (Kitagawa, 1955) is exact for token-pooled accuracy. Our sequence-averaged estimator also has a residual, which Appendix A reports explicitly.

thank you for the additional experimental results. Several of my questions/concerns were addressed. Some of the results here are interesting (e.g. that under other settings, you're also not seeing a significant effect of whitespace for a horizon of 1). My main concern about implications and applicability of the results still stand though, and so I maintain my score

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers for their careful reading. We summarize new measurements and clarify correctness matching, practical relevance, and the connection to prior work.

1. Evidence beyond H=3 and at ordinary next-token prediction.

We evaluated mean-blend at H=1–8 across five model instances on GSM8K/Wikitext, using 300 sequences per corpus and 2,000 sequence-bootstrap draws. We also analyzed H=1 mild-resample on 250 sequences per corpus. The following intervals are pointwise 95% sequence-bootstrap intervals.

  • H=2–8, mean-blend: in 34 of 35 model–horizon combinations, the pooled CAI interval lies above zero and the directly measured non-whitespace CAI interval lies below zero. Whitespace exclusion reverses all 34 positive pooled estimates. The exception is Qwen2.5 at H=2, whose pooled estimate is −0.0013 with an interval containing zero.
  • H=1, mild-resample: the positive pooled contrast is not driven by whitespace. After matching target type and whether the unablated model predicts the target correctly, it reverses in Qwen2.5, has an interval containing zero in three instances, and remains positive but smaller in Llama-3.2.
  • H=1, mean-blend: the initially negative pooled contrast remains negative after the same adjustment in all five instances, with every interval below zero.

The sign pattern extends beyond H=3 under the tested mean-blend settings. At H=1, the audit also detects sensitivity to baseline-correctness composition. We retain unadjusted and matched results. Matching can remove meaningful capability-related variation, so attenuation alone does not rule out task-related specialization. Persistence after this adjustment does not exclude other composition effects.

The within-type/composition decomposition (Kitagawa, 1955) is exact for token-pooled accuracy. Our sequence-averaged estimator also has a residual, reported in Appendix A. These findings concern token-level contrasts under the tested interventions. We did not measure generation-based task accuracy.

2. Expanded evaluation of the tokenizer-only screen.

We evaluated all 55 previously untested pairs among the twelve non-code corpora on the three primary models, giving 165 model–pair evaluations. The pair pool and tokenizer-derived labels were fixed before these new model runs. A validation run reproduces the reported numerical quantities in all 15 stored held-out evaluations within 10−610^{-6}.

For the reversal analysis, a pooled-positive evaluation has a pooled 95% interval entirely above zero. A reversal means that its whitespace-excluded point estimate is negative.

EvaluationResult
High- or medium-risk pooled-positive cases that reverse49 of 51
Low-risk pooled-positive cases that reverse1 of 13
High- or low-risk evaluations in the predicted region79 of 93

The region rule is informative but imperfect. None of the 14 misses falls in the opposite region, and 11 are Qwen2.5 evaluations. Among the 12 low-risk misses, only two have pooled intervals above zero and neither reverses after whitespace exclusion. Both high-risk misses do reverse.

These results support the screen's use in flagging whitespace-sensitive positive contrasts, not a universal region-classification rule.

3. Scope and the DoLa comparison.

We start with a score pooled over all targets, not a whitespace-only reasoning test. Decomposition reveals which targets support its apparent depth effect. CAI is our diagnostic contrast, not an established protocol. Prior depth studies motivate the question, but our experiments do not establish that their methods are confounded.

Our experiments on datasets used by DoLa evaluate our tokenizer-only screen rather than reproduce DoLa's depth analysis. We also revisited DoLa's candidate-layer selection and identified an inaccurate description in our introduction. Specifically, for LLaMA-7B, DoLa's Table 10 uses the same lower layer bucket for GSM8K, StrategyQA and FACTOR, while TruthfulQA uses a higher bucket. We will correct this description in the revised manuscript.

These runs do not establish whether DoLa itself is affected. DoLa's token-dependent layer observations instead motivate our target-type audit.

Analysis plan and manuscript updates.

The registration discloses an earlier exploratory H=1 mild-resample calculation with a different estimator. The horizon sweep and screening census had no prior outcome inspection. Their analysis plans were recorded before the registered runs.

Planned manuscript updates cover the horizon sweep, pooled H=1 results, expanded screen evaluation, corrected DoLa comparison, and these scope clarifications. The practical lesson is to report target-type contributions and mixture-matched estimates before interpreting a pooled cross-corpus contrast as evidence of task-dependent depth specialization.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.