acceptodds
Under review as a conference paper at ICLR 2027

CORE-FV: FROM MARGINAL TO CONDITIONAL HEAD SELECTION FOR FUNCTION VECTORS

Abstract

Function vectors provide a compact mechanism for extracting and transferring task-specific behaviors from in-context demonstrations by aggregating the contributions of a small subset of attention heads. Existing approaches typically identify these heads using marginal importance scores, but evaluating heads independently overlooks interactions among them, causing redundant or complementary heads to be selected suboptimally. We introduce CORE-FV (Conditional Ordering through Removal Evaluation for Function Vectors), a conditional head selection framework that evaluates each attention head in the context of the remaining intervention set. CORE-FV performs an adaptive backward refinement procedure that iteratively removes heads whose absence minimally affects the induced task behavior, producing an interaction-aware ordering of heads. Across eight in-context learning transformations and multiple open-weight language models, CORE-FV generally outperforms a matched Average Indirect Effect baseline, achieving average accuracy gains of up to 11.84 percentage points. Moreover, on composite tasks requiring sequential or parallel transformations, CORE-FV substantially outperforms marginal head selection, improving average accuracy from 41.81% to 87.34% on Llama-3.1-8B. These results highlight that head contributions can be inherently context-dependent, and that interaction-aware selection provides a more effective foundation for constructing compact and algebraically composable task representations.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, function vectors, in-context learning, attention heads, causal interventions, head selection

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_UpUyOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

In earlier work on creating function vector such as work by Todd et al., 2024 selects heads for function vectors based on the individual head importance and neglect the interaction between heads. The paper proposes a novel method Conditional Ordering through Removal Evaluation for Function vectors (CORE-FV) to select heads such that the interaction between heads are captures into the function vectors. CORE-FV works as a deletion based method where it starts with the full set of heads, delete the least-affecting-the-task heads, re evaluate the heads in the new remaining head context and continue this cycle. The paper shows that the CORE-FV (function vectors) does well on composite tasks where interaction between heads is required. The proposed method has also been evaluated against earlier baselines on multiple model scale and seeds.

Critical strengths
  • I think CORE-FV is an interesting idea and well motivated on making function vector much richer capturing dependency of attention heads which allows the models to do well on composite tasks and will be very valuable to the community in continuing the work on function vector for more complex tasks.

  • The results are convincing on both the Function Vectors benchmark as well as on composite tasks. The average gains over AIE are beyond seed standard deviation and consistent across all model scale except MIstral-7B-v0.3 where CORE-FV is closer to AIE. On composite tasks CORE-FV gains are clearly evident improving over AIE relatively by ~40% on GPT-J-6B and over 100% relatively on Llama-3.1-8B.

  • The paper also discusses the feasibility of the method and introduces group deletion to bring O(H^2) evaluation down to O(H) evaluation.

  • The paper is very well written, easy to follow with sufficient details on method, experimental setup and results.

Critical weaknesses
  • Does group deletion of heads change the quality of the heads? What is the trade-off between doing single-head deletion versus group head deletion? Which one was used in the analysis?

  • In Table 3, each earlier work in Todd et al., 2024 didn’t use a coefficient, and as CORE-FV uses one, it is hard to compare whether the gains are due to the coefficient or a different head selection method.

Questions for authors
  • For composite tasks (table 2), what is the reasoning behind the model choices for evaluation?
2 comments
AuthorsOpenReview ↗

Additional Ablations, Coefficient Analysis, and Expanded Model Evaluation

We thank the reviewer for their thoughtful feedback and positive assessment of our work. Below, we address the two concerns and the question about model choices, highlighting the additions to the revised manuscript.

W1. Trade-off between Adaptive Group Deletion and Single-Head Deletion

Adaptive group deletion was used in all main analyses.

To directly assess its effect on selection quality, we added a controlled comparison with single-head deletion in Section 4.5, covering GPT-J-6B and Llama-3.1-8B on Antonymy and English--French across three seeds. Adaptive deletion reduces runtime by 16--19$\times$ on GPT-J and 44--48$\times$ on Llama, while maintaining comparable accuracy. This supports our default schedule as a substantial computational saving with modest, task-dependent accuracy changes. Time constraints limited this additional experiment to two tasks, and we are happy to extend the single-head ablation to all tasks in the final version.

ModelTaskAIE time (min)Adaptive time (min)Single-head time (min)Adaptive accuracy (%)Single-head accuracy (%)
GPT-J-6BAntonymy14.76 ± 2.1118.93 ± 6.71305.31 ± 39.6485.10 ± 1.2484.21 ± 0.52
GPT-J-6BEnglish--French16.81 ± 0.7121.60 ± 5.07406.54 ± 20.9483.04 ± 1.4784.78 ± 0.70
Llama-3.1-8BAntonymy35.25 ± 1.9736.41 ± 7.691,614.58 ± 61.4686.91 ± 0.4984.88 ± 0.69
Llama-3.1-8BEnglish--French40.28 ± 2.2039.13 ± 4.901,880.13 ± 132.8190.12 ± 1.1290.69 ± 1.29

Entries are mean ± sample standard deviation across three seeds. Times measure head selection only, including preparation and complete ranking, but excluding model loading and downstream tuning and evaluation. Runs use NVIDIA T4 GPUs.

W2. Contribution of Coefficient Tuning in Table 3

We agree that the original comparison did not clearly separate the contributions of head selection and coefficient tuning. We have added a fixed-α=1\alpha=1 variant to Table 3, setting all three component coefficients to one, with no coefficient tuning. Using the same 10 heads per component, it achieves 47.01 ± 3.56% average accuracy across eight tasks, compared with 43.25 ± 9.16% for component-specific coefficients and 42.88% for the published Sum FV result. These results show that CORE-FV supports composition without coefficient tuning. The comparison with published FV is approximate, since Todd et al. (2024) may use different seeds and the code for their composition experiments is not publicly available.

Several tasks show substantial variation across seeds in both the fixed-α\alpha and component-α\alpha settings. Our task-specific heads are reselected across seeds, whereas the original FV setup uses a fixed global head set; this may contribute to the variation. Refining coefficients on the composed task's validation split increases average accuracy to 64.50 ± 0.39%, with much lower variation in the average across seeds. These results show that coefficient refinement contributes substantially to the strongest composition result, so we do not attribute that gain solely to head selection. The controlled evidence for head selection comes from our matched AIE comparisons in Tables 1 and 2, where both methods use the same coefficient-selection procedure.

Q. Why Were Composite Tasks Initially Evaluated on Only Two Models?

GPT-J and Llama were our initial main models. We expanded the standard-task evaluation to additional models later, while the composite experiments initially remained limited to those two; there was no methodological reason to restrict them. In the revision, Table 2 now covers all five models, adding Qwen3-4B-Base, Mistral-7B-v0.3, and OLMoE-1B-7B. The trend holds across the added models: CORE-FV improves average composite-task accuracy by 15.44--40.27 percentage points over AIE and 24.38--35.19 points over LTV.

Thank you for addressing the questions and adding new results. I will retain my positive score.

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

This paper introduces CORE-FV, an interaction-aware attention-head selection method for constructing function vectors. Existing Function Vector methods rank heads using marginal scores such as Average Indirect Effect (AIE), which evaluate each head independently and may overlook redundancy, complementarity, and interference among heads. CORE-FV instead starts from all candidate heads and iteratively removes the heads whose absence least affects the joint intervention. The remaining heads are re-evaluated after every accepted deletion, producing a conditional head ordering. Across eight standard ICL transformations and five open-weight models, CORE-FV generally improves over a matched AIE baseline. Its advantage is especially large on composite tasks that require sequential or parallel transformations, suggesting that conditional selection better captures head interactions.

Critical strengths

CORE-FV is conceptually simple and directly compatible with the existing Function Vector pipeline, changing only the head-selection procedure.

Critical weaknesses

1.CORE-FV improves over AIE on only 23 of 40 standard model-task pairs, ties on one, and performs worse on 16. It also slightly underperforms AIE on average for Mistral-7B, so the advantage is not universal. 2.CORE-FV achieves only moderate average improvements over AIE on several settings, while introducing substantially higher head-selection costs. Unlike AIE, which scores heads independently, CORE-FV requires repeated evaluation after each deletion or deletion-group proposal.

Questions for authors

How stable is the selected head ranking across different refinement-example subsets or random seeds?

1 comment
AuthorsOpenReview ↗

CORE-FV: Accuracy Gains, Head-Selection Cost, and Stability

We thank the reviewer for their thoughtful feedback. We address the consistency of accuracy gains, head-selection cost, and stability across random seeds.

W1. Consistency and Size of Accuracy Gains

We agree that CORE-FV's advantage on the eight standard tasks is not universal (Section 4.2). However, of the 16 pairs where AIE performs better, 10 differ by less than one percentage point, with a maximum loss of 3.28 points. AIE already achieves at least 95% accuracy in nine cases. In contrast, 14 of CORE-FV's 23 wins exceed three points, including +76.47 on English--French for Llama-3.1-8B. Across five models, CORE-FV improves standard-task accuracy by 4.56 points.

The advantage is stronger on three additional benchmarks, where vector construction and intervention are fixed and only head ordering changes:

  • Five additional ICL tasks (Appendix D, Table 16): 22 wins out of 25, with average improvements of 7.94--23.68 points across models.
  • Composite tasks (Table 2): 22 wins, two ties, and one loss, with average gains of 15.44--45.53 points.
  • Arithmetic Add-kk (Appendix B.4, Table 10): nine wins out of ten, increasing average accuracy from 55.68% to 84.41% on GPT-J and 87.07% to 97.38% on Llama.

Across all benchmarks, CORE-FV wins 76 of 100 comparisons, ties on three, and loses on 21. Its only lower model-level average is on Mistral's standard tasks (-0.50 points), while it gains 9.78 and 20.98 points on additional and composite tasks, respectively. Thus, its advantage is particularly pronounced on harder, multi-step tasks. We will clarify this in Section 4.2.

W2. Accuracy Gains and Head-Selection Cost

CORE-FV achieves substantial gains on harder benchmarks, particularly 15.44--45.53 percentage points on composite tasks. Moreover, head selection is performed offline per model--task pair, and both the resulting ranking and intervention configuration are reused across subsequent predictions. The backward search is not repeated for each prediction; thus, its computational cost is amortized.

We added Section 4.5 to quantify this overhead. Adaptive CORE-FV's selection time is only 28--29% higher than AIE on GPT-J and within 4% on Llama. Group deletion makes refinement 16--19$\times$ faster on GPT-J and 44--48$\times$ faster on Llama than single-head deletion. Despite more evaluations, shorter zero-shot queries keep total selection time close to AIE.

Q1. Stability of Attention-Head Selection Across Random Seeds

We conducted an additional analysis (Appendix, Attention Head Selection Stability and Overlap). Tables 17 and 18 report the overlap of top-KK heads across three seeds (42, 43, 44), for K∈5,10,20,40,60K \in \\{5,10,20,40,60\\}, across 11 tasks for GPT-J and Llama-3.1-8B.

Average overlap across tasks and KK values:

ModelCORE-FVAIECORE-FV--AIE
Llama-3.1-8B26.90%74.66%10.86%
GPT-J-6B33.95%77.61%15.52%

AIE exhibits substantially higher selection stability than CORE-FV. This reflects their different procedures: AIE independently assigns a marginal score to each head in a single evaluation stage, meaning that its score is unaffected by the presence or absence of other heads. In contrast, CORE-FV iteratively re-evaluates the remaining heads after each accepted deletion. Small variations can therefore affect subsequent conditional scores and deletion sequences. Moreover, multiple nearly equivalent deletion choices may exist under CORE-FV's performance-preservation criterion.

However, selection stability alone does not indicate selection quality. A method may consistently select the same but suboptimal heads. The low overlap between CORE-FV and AIE (10.86% for Llama and 15.52% for GPT-J) also suggests that useful task information can be captured by different head subsets.

Importantly, despite having substantially fewer heads shared across the three seeds, CORE-FV achieves higher accuracy on several tasks. Figures 8 and 9 construct function vectors using only heads shared across all three seeds for each KK. Nevertheless, CORE-FV outperforms AIE on Singularize ⊕\oplus Antonym and Next-item →\rightarrow Capitalize for GPT-J, and Antonym →\rightarrow French, English--French, and Next-item →\rightarrow Capitalize for Llama.

These results show that CORE-FV can identify more effective task-relevant subsets despite having fewer consistently selected heads. Thus, greater selection stability or overlap does not necessarily imply better downstream performance.

Overall, although AIE provides more reproducible rankings, CORE-FV identifies smaller, distinct shared subsets that achieve better accuracy on several tasks. This highlights the importance of evaluating both selection stability and the actual effectiveness of selected heads.

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

This paper revisits attention-head selection for function vectors. The authors argue that the standard Average Indirect Effect (AIE) criterion is marginal: heads are scored independently, so redundancy, complementarity, and interference among heads are not reflected in the resulting ranking. CORE-FV instead performs conditional backward elimination, repeatedly re-evaluating each head after other heads have been removed. Across five open-weight models and eight ICL transformations, CORE-FV generally improves over a matched AIE baseline, with particularly large gains on several models and on composite tasks.

Critical strengths
  1. The paper identifies a clear limitation of marginal head selection, and the proposed conditional deletion procedure directly addresses this limitation.
  2. The comparison with AIE is well controlled: the vector construction and intervention pipeline are held fixed while changing the head-selection rule.
  3. The ablation on non-accumulative ranking is particularly convincing. Removing conditional re-evaluation substantially degrades performance, directly supporting the main methodological claim.
  4. The gains on composite tasks are large and consistent across a broad range of head counts, suggesting that interaction-aware selection is especially useful when multiple transformations must be represented jointly.
  5. The experiments cover several model families and include additional interaction and vector-composition analyses.
Critical weaknesses
  1. The paper does not directly compare against the most closely related recent joint head-attribution approaches discussed in the related work. Since these methods already move beyond marginal AIE scoring, such comparisons would better establish the advantage of repeatedly re-evaluating heads as the retained set changes.
  2. CORE-FV requires repeated white-box intervention evaluations during backward refinement. Although the adaptive deletion schedule reduces the search cost, the practical overhead relative to one-shot attribution methods is not directly quantified, which may limit scalability to larger models.
Questions for authors

see weaknesses

1 comment
AuthorsOpenReview ↗

Joint Head-Attribution Comparisons and Computational Overhead

W1. Comparison with joint head-attribution methods

We thank the reviewer for this suggestion. We agree that comparisons beyond marginal AIE scoring are needed to assess the value of repeatedly re-evaluating heads as the retained set changes.

To address this directly, we implemented the sparse joint head-attribution method from Hu et al., Understanding In-context Learning of Addition via Activation Subspaces (AS; Hu et al., 2026), and added results across all five models to Table 1. AS jointly optimizes continuous head coefficients with an L1 sparsity penalty, then selects heads using a largest-gap threshold. It therefore goes beyond marginal scoring, but does not repeatedly re-evaluate the remaining heads after deletion.

CORE-FV outperforms AS in average accuracy across all eight ICL tasks on every evaluated model, with gains ranging from 1.08 to 7.51 percentage points across the five models. These results support the benefit of conditional re-evaluation beyond one-shot joint attribution, alongside the non-accumulative ranking ablation that directly tests this component. Full task-level results and standard deviations appear in the revised manuscript. Our evaluation also includes State Vectors (SV; Li et al., 2024) and Learned Task Vectors (LTV; Yang et al., 2026). LTV provides a complementary baseline that learns steering vectors through gradient-based optimization on task demonstrations.

We also considered Pham et al., Fast & Faithful Function Vectors, but their method cannot be directly compared under our current setup. Their prompts explicitly describe the task, and their method identifies relevant heads using instruction keywords such as "opposite" or "capital." Our prompts provide input-output examples without these task instructions, so the keywords their method relies on are absent. A comparison would therefore require modifying their method or changing our prompts; using the original setups would mix differences in prompting with differences in head selection. We clarify this limitation in the related work.

W2. Computational overhead

To quantify this overhead, we added Section 4.5, comparing AIE, adaptive CORE-FV, and single-head CORE-FV on two models and two tasks across three seeds. As shown below, adaptive CORE-FV takes approximately 28–29% longer than AIE on GPT-J, while its runtime on Llama remains within 4% of AIE. At the same time, it is 16–19× and 44–48× faster, respectively, than single-head refinement. These comparable runtimes are plausible because CORE-FV evaluates shorter zero-shot queries, whereas AIE evaluates interventions on longer corrupted in-context prompts.

Importantly, head selection is an offline interpretability procedure performed once for each task, rather than an operation repeated for every test sample or generation step. Therefore, its computational overhead does not translate into per-sample inference latency. In this context, the primary objective is to identify more effective and accurate attention heads, while selection time is a secondary practical consideration. Thus, the substantial computational savings of the adaptive schedule make it a practical choice without compromising its primary objective of improving head-selection quality.

We also clarified the complexity comparison in Section 3.3; the new timing results show how the adaptive schedule translates that computational saving into practice.

ModelTaskAIE (min)Adaptive CORE-FV (min)Single-head CORE-FV (min)
GPT-J 6BAntonymy14.76 ± 2.1118.93 ± 6.71305.31 ± 39.64
GPT-J 6BEnglish–French16.81 ± 0.7121.60 ± 5.07406.54 ± 20.94
Llama-3.1-8BAntonymy35.25 ± 1.9736.41 ± 7.691,614.58 ± 61.46
Llama-3.1-8BEnglish–French40.28 ± 2.2039.13 ± 4.901,880.13 ± 132.81

Entries are mean ± sample standard deviation across three seeds. Times measure head selection only, including preparation and complete ranking, but excluding model loading and downstream tuning and evaluation. Runs use NVIDIA T4 GPUs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.