acceptodds
Under review as a conference paper at ICLR 2027

Query Transfer Can Revise Rules Without Preserving the Task

Abstract

When a model follows a revised instruction, the unchanged requirements must still hold. If the destination country changes, the budget and distance ranking still apply. We copy attention queries from a final-rule restatement into a prompt that must resolve the revision. These queries govern how the model reads its context. Query transfer improves revision in Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct, as direct restatement also does. Paired scoring nevertheless exposes correct answers lost within the aggregate gains. A separate last-token query intervention tests source-score reconstruction between two prompts that both restate the final rule. In Qwen, adding record states to query transfer brings a monitored answer margin closer to its source value but lowers selection accuracy. Most new errors copy an incorrect source answer. Correct-source counterexamples reveal a second failure because the monitored margin omits an eligible competitor in the requested category. Queries from prompts without records also shift scores toward changed requirements, but these effects do not consistently improve complete execution. Complete-candidate scoring locates the missed ranking comparison, and paired evaluation identifies the successful answers an intervention fails to preserve.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, activation patching, attention query transfer, instruction revision, constraint satisfaction, intervention evaluation, large language models

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_LAZ3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies instruction revision — when one requirement changes (e.g., destination country), the unchanged ones (budget, ranking) must still hold — via causal interventions on attention queries. Queries from a source prompt that restates the final rule are transferred into a recipient prompt at recipient RoPE coordinates. On controlled selection tasks with exact ground truth (six-record tables; category + threshold + ranking), the authors find: (1) query transfer improves revision accuracy comparably to direct restatement on Llama-3.1-8B and Qwen2.5-7B, but paired per-table scoring shows it breaks some previously correct answers, a loss invisible in aggregate accuracy; (2) adding record K/V improves reconstruction of a monitored source margin in Qwen yet can worsen selection, because a two-candidate margin omits eligible competitors — full-candidate scoring separates copied source errors from correct-source reversals; (3) record-free source queries shift scores toward changed requirements but do not reliably yield complete execution, with spillover across requirement axes and retained answer-letter preferences. Evaluation uses exact McNemar tests, Holm correction within and across families, and table-level bootstrap intervals.

Critical strengths
  1. Methodological rigor in evaluating causal interventions. Paired preservation (Eq. 6), the complete-candidate margin (Eq. 4), the mutually exclusive error taxonomy, and global multiplicity correction form a template that addresses known weaknesses in how patching results are reported. The finding that aggregate gains conceal broken previously-correct answers is shown cleanly.
  2. A well-isolated negative result. In the Figure 1/Figure 4 example, record-K/V transfer cuts the D-F reconstruction error from 9.21 to 0.54 logits while flipping the answer from correct D to incorrect C through the omitted competitor C; the correct-source reversal analysis (Table 1) shows this is not a one-off.
Critical weaknesses
  1. The writing obscures the contribution enough to limit its impact. The paper's value is methodological — it asks the community to adopt evaluation practices — and that depends on legibility. As written, even a sympathetic reader must reconstruct the experimental matrix from scattered definitions.
  2. Narrow scope and no positive mechanistic account. Two 7-8B models, one synthetic six-record task, one intervention family (suffix/last-token queries). The results are predominantly cautionary: no preservation advantage over simply restating the rule in the prompt, Llama and Qwen disagree on whether record-state matching helps reconstruction, and record-free effects "do not consistently improve complete execution." A constructive takeaway about how revision is implemented — e.g., which query content carries the retained requirements — is missing, so the net contribution is a warning plus a protocol rather than an advance in understanding.
Questions for authors
  1. Query transfer and restatement tie on full-cohort revision accuracy (63/96, 69/96) and damage counts. Is there any setting — subset, metric, or budget — where the internal intervention is strictly preferable to restating the rule? If not, what is the practical motivation for query transfer beyond diagnosis?
  2. How do the repair/damage conclusions change on the complement of the capability subset (the tables where the unmodified model already fails)?
  3. Do reconstruction-vs-selection discrepancies persist with more records/categories, where the candidate-omission failure should be more prevalent?
  4. (Future work) A per-head analysis of which transferred query content encodes the retained threshold/ranking requirements would turn the diagnostic framing into a positive result.
Reviewer_o9r8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies attention-query transfer for instruction revision in controlled selection tasks. It measures repaired answers, newly introduced errors, and the relationship between targeted score reconstruction and complete task correctness.

Critical strengths
  • The evaluation tracks both repairs and newly introduced errors. Direct restatement and random-query controls help assess the intervention, while paired outcomes reveal losses hidden by aggregate accuracy.
  • Concrete examples show improved score reconstruction alongside an incorrect final answer, including cases with a correct source answer. These examples provide useful evidence about the limits of targeted reconstruction metrics.
Critical weaknesses
  • The relationship between intervention selection and evaluation is unclear. Appendix D.2 reports selecting query positions using 19 Llama examples and 33 Qwen examples. Appendix B.1 describes the evaluation cohorts but does not explain whether they overlap with these selection samples. The procedure for choosing the layer range is also unclear. These gaps make it difficult to assess whether selecting the intervention settings could have inflated the reported gains. The authors should clarify sample overlap, the settings considered, and when the final choices were fixed.
Questions for authors
  • How do the intervention-selection samples in Appendix D.2 relate to the evaluation pools in Appendix B.1? Please clarify whether evaluation examples informed the choice of query positions or layer ranges, and when these settings were fixed. If evaluation examples were used for selection, could you provide a limited validation on examples unused during selection?
Reviewer_EpzyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates whether transferring attention queries from a prompt that explicitly restates the final instruction helps a language model resolve an instruction revision while preserving unchanged constraints. Query transfer improves revision accuracy, but paired evaluation reveals that it also breaks some previously correct answers. The core contribution is an evaluation argument: targeted score recovery, complete-answer correctness, and preservation of previously correct behavior measure different properties and should be assessed together.

Critical strengths

The paper addresses a meaningful evaluation problem. Distinguishing successful revision from the preservation of unchanged requirements is useful for both instruction following and mechanistic interpretability.

Critical weaknesses
  1. The experimental scope of the paper is limited. Experiments in the paper cover two similarly sized models on six-record synthetic tasks within a single user message. In the highlighted Qwen compatibility experiment, 23 of 28 newly introduced errors copy an incorrect source answer; only two correct-source cases combine improved reconstruction with newly incorrect selection. These counterexamples do not support the generality of the methodology started in the paper.
  2. In Qwen before-->after, record K/V improves reconstruction and also raises correctness from 306/384 to 327/384. In Llama, reconstruction gets worse on average while correctness rises in one direction. The abstract’s framing that “adding record states ... brings a monitored answer margin closer ... but lowers selection accuracy” holds for only one of four model-direction settings. The paper acknowledges disagreement “in either direction,” but this mostly shows that the metric is weakly related to accuracy. It does not show a systematic failure mode.
  3. The data points supporting the main claim around the correct-source reversal phenomenon occur in 2 conditions for Qwen after-->before and 0 for Qwen before-->after, out of 63 and 295 eligible conditions (Table 1, lower panel). Llama has 2 per direction. The text says “Both Qwen reversals occur on different tables,” and the Figure 1 example is one of them. Thus the reversal claim seems to be supported by a tiny number of samples.
Questions for authors

Suggestions for future work - Test at least one larger model and one natural-language or non-synthetic selection task to show the dissociation is not an artifact of 7–8B models on templated tables.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.