acceptodds
Under review as a conference paper at ICLR 2027

Who Grades the Edit? Frozen and Re-fit Readouts for Position Interventions in RoPE Models

Abstract

Rotary position embeddings (RoPE) encode relative token offsets, yet a linear readout of the residual stream tracks absolute position. Existing position-bias mitigations intervene on this coordinate. We ask whether an edit disrupts that readout or removes the information it reads. We evaluate upstream edits using a probe frozen before the edit and a probe re-fit on the edited activations. In Gemma, blocking the attention-sink value reduces position by for the frozen probe but only for the re-fit probe. The frozen score falls below zero even though most position information remains linearly recoverable. After removal of the fitted leading direction, re-fit probes retain – of baseline position across four base models. A separate behavioral audit asks whether editing a fitted position basis changes position-biased retrieval. On Llama-Instruct multi-document QA, sequence-wide basis removal flattens the lost-in-the-middle dip beyond a matched decoder-null control. The control also changes behavior, so the relevant quantity is the paired difference: (95% CI: ). This finding does not establish information erasure and is limited to the tested model, task, layer, basis, and decoder; Qwen's behavioral audit stops because a control fails. Evaluating a position edit requires distinguishing readout disruption, remaining recoverability, and a separately controlled behavioral effect.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
mechanistic interpretability, positional representations, rotary positional embeddings (RoPE), causal interventions, linear probing, long-context language models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_VTHUOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies whether position interventions in RoPE models actually remove position information or merely disrupt a previously trained readout. It compares frozen and re-fit position probes, analyzes redundancy and attention-head contributions, and separately evaluates behavioral effects on position-biased retrieval.

Critical strengths
  1. This paper is transparent about failed controls and evidential limitations.
Critical weaknesses
  1. This manuscript reads more like an experiment report to me. Its highly templated and repetitive presentation makes it difficult to follow.

  2. The large frozen / re-fit discrepancy in Gemma is presented as a main finding, but its simplest explanation is not examined. Since out-of-sample R2R^2 is highly sensitive to shifts in prediction offset and scale, the result may simply reflect trivial probe miscalibration. Without intercept-only and affine recalibration baselines, it remains unclear whether full probe refitting is necessary, limiting the interpretation and novelty of Contribution 1.

  3. For Contribution 3, the only interpretable behavioral evidence comes from Llama on a small sample. The matched control also changes behavior, and overall QA performance does not clearly improve. The Qwen and Gemma audits provide no independent confirmation. Therefore, I do not find the behavioral evidence convincing.

Questions for authors

See Weaknesses.

1 comment

Thank you for the comments on presentation, recalibration, and behavioral evidence.

[W1. Presentation.]

We have revised the Introduction and Contributions around the sink-value readout comparison and controlled behavioral contrast. The four-level structure remains, with primary results in Secs. 4 and 7, detailed procedures and secondary audits in the appendices, and interpretation and limitations grouped in Sec. 8.

[W2. Intercept and affine recalibration.]

We fitted the recalibration and full-refit baselines on 128 documents and evaluated them on 128 separate documents. For Gemma sink-value blocking, reference, intercept-corrected, affine-corrected, and full-refit R2R^2 are −16.954-16.954, 0.5960.596, 0.6550.655, and 0.8200.820. Thus, an offset shift explains much of the extreme frozen failure.

With clean probes fitted at the evaluation context length, full refitting still exceeds affine recalibration by 0.1610.161 [0.152,0.170][0.152,0.170] in Gemma and 0.1450.145 [0.137,0.153][0.137,0.153] in Llama base models. These nominal 95% intervals use 20,000 paired document-bootstrap resamples conditional on fitted probes. The advantage persists across three Gemma fitting cohorts sharing the same 128 evaluation documents. Contribution 1 now separates the substantial calibration effect from this remaining refit advantage (Sec. 4, Table 2; Appendix E).

[W3. Behavioral evidence, active control, and generalization.]

We evaluated Llama-Instruct and Qwen-Instruct on the same 256 held-out QA questions, with 128 separate fitting questions. Treatment fixes all 32 basis coefficients to their prompt means at prompt positions only; its control preserves a separately fitted scalar position readout while matching support and realized displacement.

Treatment-minus-control dip changes are −0.0242-0.0242 [−0.0312,−0.0177][-0.0312,-0.0177] in Llama and −0.0190-0.0190 [−0.0284,−0.0098][-0.0284,-0.0098] in Qwen (97.5% modelwise paired question-bootstrap intervals, Bonferroni-adjusted for the two primary contrasts). Both runs pass technical and specified control checks (Sec. 7, Table 4).

Llama's control remains active despite meeting the equivalence band. Qwen's treatment-identity 95% interval includes zero ([−0.0330,+0.0253][-0.0330,+0.0253]). Neither treatment meets the minimum-effect rule, which requires a treatment–identity 95% interval entirely below −0.065-0.065. The accuracy comparisons establish neither a corresponding dip effect nor an overall gain (Appendix G.1, Table 24).

The prompt-only Qwen comparison extends the treatment-decoder-null contrast to a second model. It uses different data and intervention support from the earlier sequence-wide audit, whose halted verdict is unchanged (Appendix G.8). Independent behavioral confirmation in Gemma remains absent.

Reviewer_8FkdOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1(−1)
Soundness
2
Read review
Paper summary

This paper investigates whether interventions targeting absolute-position representations in RoPE-based language models genuinely remove positional information or merely disrupt a particular readout. The authors compare frozen and re-fitted linear probes and extend the analysis to representation-level interventions, attention-head ablations, and downstream retrieval behavior. Experiments across four base models show that positional information often remains linearly recoverable after targeted interventions, despite substantial degradation in frozen-probe performance. The paper also reports a behavioral effect of position-subspace removal on lost-in-the-middle bias in Llama-Instruct, although the effect does not meet the authors’ prespecified minimum-effect criterion. The paper highlights an important distinction between readout disruption, information recoverability, and behavioral effects, but the strength and generality of its empirical conclusions remain limited.

Critical strengths
  1. Important research question. The paper highlights a fundamental limitation of evaluating representation interventions solely through frozen probes: a decline in readout performance does not necessarily imply that the underlying information has been removed. Distinguishing readout disruption from information recoverability is conceptually useful for mechanistic interpretability and representation editing.
  2. Broad and structured empirical investigation. The study examines multiple models and intervention types, covering frozen versus re-fitted probes, positional subspace removal, attention-head ablations, and downstream behavioral evaluations. This provides a broader empirical picture than relying on a single probing metric or intervention.
  3. A connection between diagnosis and optimization. Rather than limiting the contribution to post hoc analysis, the authors use target-class proximity to guide the search objective and local curvature to adapt optimization step sizes. The ablation experiments comparing support-only, curvature-only, and combined variants help illustrate the respective contributi
  4. Careful qualification of empirical claims. The authors explicitly report failed controls, inconclusive behavioral results, implementation deviations, and exploratory analyses. They generally distinguish statistically significant effects from results that satisfy their prespecified decision criteria, which improves the transparency of the study.
Critical weaknesses
  1. Limited robustness of the central finding. The most striking frozen-versus-re-fitted probe discrepancy is concentrated in Gemma, while the Formation experiments use only 16 text windows and lack document-level uncertainty estimates for the primary comparison. The paper also does not adequately isolate the contribution of probe refitting from simpler calibration changes, such as intercept or affine recalibration. Consequently, the robustness and generality of the central finding remain insufficiently established.
  2. Inconclusive and narrow behavioral evidence. The principal behavioral result is obtained on Llama-3.1-8B-Instruct in a multi-document QA task. Although the treatment–control contrast is statistically significant, the treatment does not satisfy the prespecified minimum-effect criterion. The Qwen audit is halted because the complement-placebo control fails, and the remaining evidence does not establish a robust cross-model behavioral effect.
  3. Insufficient control of intervention-specific confounds. Several interventions change multiple factors simultaneously. For example, context-window manipulations jointly alter token count, recency, and visible content, while the attention-head controls match layer composition and head count but not output norm or residual displacement. Moreover, independently fitted positional subspaces exhibit limited alignment, raising concerns about the uniqueness and stability of the identified representation. These limitations weaken the mechanistic attribution of the observed effects.
  4. Limited methodological novelty and incomplete linkage between representation and behavior. The frozen-versus-re-fitted evaluation builds on established ideas in probing and concept erasure, and the paper does not yet establish a sufficiently distinct methodological advance beyond their application to positional interventions. In addition, the representation and behavioral audits are largely separate: the behavioral experiments do not directly establish whether the information remaining recoverable after an intervention is the information used by the model during retrieval. This limits the paper’s ability to support a unified mechanistic account.
Questions for authors

Please see the weaknesses.

1 comment

Thank you for identifying the calibration, behavioral-evidence, and attribution concerns.

[W1. Calibration and Formation robustness.]

We added recalibration baselines for sink-value blocking in Gemma and Llama base models, using 128 fitting and 128 separate evaluation documents each. For Gemma's reference decoder, intercept correction raises R2R^2 from −16.954-16.954 to 0.5960.596, explaining much of the extreme frozen failure.

With clean probes fitted at the evaluation context length, full refitting exceeds affine recalibration by 0.1610.161 [0.152,0.170][0.152,0.170] in Gemma and 0.1450.145 [0.137,0.153][0.137,0.153] in Llama. These nominal 95% intervals use 20,000 paired document-bootstrap resamples conditional on fitted probes. Two additional Gemma fitting cohorts retain this advantage on the same evaluation documents; they test fitting sensitivity, not independent evaluation replication (Sec. 4, Table 2; Appendix E).

These comparisons do not reconstruct uncertainty for the original 16-window grid (eight fitting/eight evaluation windows). Its point estimates, threshold sensitivity, and failed controls remain separate in Appendix F.

[W2. Behavioral evidence and cross-model scope.]

We evaluated Llama-Instruct and Qwen-Instruct on the same 256 held-out questions, with 128 separate fitting questions. Treatment fixes all 32 basis coefficients to their prompt means at prompt positions only. The control preserves a separately fitted scalar position readout while matching support and realized displacement.

Treatment-minus-control dip changes are −0.0242-0.0242 [−0.0312,−0.0177][-0.0312,-0.0177] in Llama and −0.0190-0.0190 [−0.0284,−0.0098][-0.0284,-0.0098] in Qwen (97.5% modelwise paired question-bootstrap intervals, Bonferroni-adjusted for the two primary contrasts). Both runs pass technical and specified control checks (Sec. 7, Table 4; Appendix G.1).

Llama's control remains active, and Qwen's treatment–identity 95% interval includes zero ([−0.0330,+0.0253][-0.0330,+0.0253]). Neither treatment meets the minimum-effect rule, which requires a treatment-identity 95% interval entirely below −0.065-0.065. Accuracy comparisons establish neither a corresponding dip effect nor an overall gain. The original Qwen sequence-wide HALT is unchanged (Appendix G.8), and independent Gemma behavioral confirmation remains absent.

[W3. Confounds and subspace stability.]

These limits remain explicit in Sec. 8: visible-set edits jointly change cardinality, recency, and content; writer controls match layer composition and head count but not output norm or residual displacement; and refitted subspaces have limited alignment and fit-sensitive effects. Accordingly, we claim neither pure counting, position-specific writing, nor a uniquely identified, fit-stable basis (Secs. 4, 6; Appendix G.7).

[W4. Novelty and the representation–behavior connection.]

Sec. 2 credits the established probing and intervention principles. Our empirical contribution separates calibration, remaining position recoverability, and controlled behavioral effects.

Recoverability is now measured immediately after the same QA interventions. After treatment, affine-corrected R2R^2 is approximately 0.0030.003 in both models. Full-refit R2R^2 is 0.4070.407 in Llama and 0.4850.485 in Qwen, versus 0.8430.843 and 0.7190.719 unedited, respectively (Table 4). Thus, the behavioral contrast coexists with reduced but remaining recoverability; this does not establish mediation. Exploratory transfer and recomputation analyses reuse 32 questions and did not identify a native reader. The separate reader-confirmation sample remains unused (Appendix H); the four audits do not establish a connected causal circuit.

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

The paper explores how absolute position is represented and used in RoPE-based LLMs. Its overall finding is that disrupting a fitted position readout often does not remove the underlying position information, and that position-related hidden states can affect retrieval behavior. The methodological lessons are mostly known, but the paper shows them systematically for RoPE-based models.

Critical strengths

The main methodological ideas are not new, but the paper shows they also hold for absolute position representations in RoPE based LLMs.

Critical weaknesses

Much of the paper reads like a technical LLM report and is hard to follow. The writing improves around Section 6, but later becomes difficult to read again.

Most of the results shown in this paper are well known, I have listed papers that I know of there are probably more.

  • Regarding frozen and refit probes, the retrainability of probes after removing predictive direction is well established in debiasing and concept erasure literature [1][2][3].
  • There is a wide literature on model being distributed and components that seem most important are not the only ones that carry such information [5][6].
  • The behavioural experiment shows that changing a position-related representation changes retrieval behavior, but prior work had already shown the same general idea [7][8].

[1] Adversarial Removal of Demographic Attributes from Text Data

[2] Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection

[3] Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Them

[4] On the Importance of Single Directions for Generalization

[5] Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned

[7] Amnesic Probing: Behavioral Explanation with Amnesic Counterfactuals

[8] Mitigate Position Bias in LLMs via Scaling a Single Hidden States Channel

Questions for authors

N/A

1 comment

Thank you for the comments on clarity and prior work.

[W1. Readability.]

We have revised the Introduction and Contributions around the sink-value readout comparison and the controlled behavioral contrast. The four-level structure remains, with primary results in Secs. 4 and 7, detailed procedures and secondary audits in the appendices, and interpretation and limitations grouped in Sec. 8.

[W2(a). Probing and concept erasure.]

We agree that suppressing a readout need not remove information; Sec. 2 now explicitly credits the cited literature. Our contribution examines what upstream position edits change, separating calibration from remaining recoverability.

For sink-value blocking, we used 128 fitting and 128 separate evaluation documents per base model. Offset correction explains much of Gemma's extreme frozen failure. With clean probes fitted at the evaluation context length, full refitting still exceeds affine recalibration by 0.1610.161 [0.152,0.170][0.152,0.170] in Gemma and 0.1450.145 [0.137,0.153][0.137,0.153] in Llama (nominal 95% paired document-bootstrap intervals, conditional on fitted probes). Thus, recalibration explains only part of the readout discrepancy (Sec. 4, Table 2; Appendix E).

[W2(b). Distributed representations and head contributions.]

Sec. 2 now credits the cited work on individual directions and attention heads. Our results characterize position representations: leading-direction removal retains 7474–88%88\% of baseline position R2R^2 across four base models, and selected heads exceed layer- and count-matched alternatives under both writer diagnostics in three. These are empirical characterizations, not claims of a new principle of distributed representation or unique writers (Secs. 5-6).

[W2(c). Behavioral interventions.]

Amnesic Probing links erasure to behavior, and Yu et al. already show that position edits affect retrieval. We test the difference from a control preserving a separately fitted scalar position readout while matching support and realized displacement.

On the same 256 held-out QA questions in Llama-Instruct and Qwen-Instruct, with 128 separate fitting questions, treatment fixes all 32 basis coefficients to their prompt means at prompt positions only. Treatment-minus-control dip changes are −0.0242-0.0242 [−0.0312,−0.0177][-0.0312,-0.0177] and −0.0190-0.0190 [−0.0284,−0.0098][-0.0284,-0.0098], respectively (97.5% modelwise paired question-bootstrap intervals, Bonferroni-adjusted for the two primary contrasts). Position remains linearly recoverable after these same edits (Sec. 7, Table 4).

The contrast remains conditional on the tested task, layer, basis, decoder, and operator: Llama's control is active, Qwen's treatment-identity interval includes zero, neither treatment meets the minimum-effect rule, and accuracy comparisons establish neither a corresponding dip effect nor an overall gain.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.