acceptodds
Under review as a conference paper at ICLR 2027

Identifying Support Knowledge Representation for Large Language Models

Abstract

Attention head pruning provides structured compression for large language models, yet identifying which heads carry target-domain capabilities remains challenging. To initiate the exploration of target-domain knowledge, we define the Support Knowledge (SK), which is the learned patterns of information selection and integration that support target-domain predictions. Based on the definition of SK, we further propose the Activation-Calibrated Margin (ACM) to determine their retention priority. The ACM combines intra-layer rankings of the token competition margin, defined as the difference between the top two attention logits, and the head output magnitude. During scoring, both statistics are collected in a single forward pass over calibration data without training or gradient computation. We then probe expanding candidate sets per layer, allocating a fixed pruning budget across layers based on marginal increases in loss and prediction disagreement proportion for each additional candidate. On domain datasets, ACM better maintains language modeling quality and prediction agreement with the full model than existing baselines, especially under higher pruning ratios, while achieving the highest accuracy in domain question-answering experiments among compared pruning methods.

Primary area:
interpretability and explainable AI
Keywords:
Knowledge Discovery, Data Mining, LLM

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_x8ecOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ACM, a training-free attention-head pruning method motivated by “Support Knowledge” for domain-specific predictions. It combines attention-logit margins and activation magnitudes with probing-based layer-wise budget allocation. Experiments evaluate domain performance, calibration stability, and corpus transfer.

Critical strengths
  1. The paper addresses domain-preserving head pruning without retraining. Its combination of head scoring and layer-wise budget allocation shows encouraging language-modeling and Qwen downstream results.

  2. Component ablations, magnitude-controlled experiments, and calibration analyses help assess the method.

Critical weaknesses
  1. Comparative advantages and cross-model robustness remain uncertain. FLAP is the only prior structured-pruning baseline; Appendix H omits its bias compensation, while its main-experiment configuration is unclear. Existing ablations do not compare scores under a common allocator or matched per-layer budgets.

  2. Practical efficiency and total pruning costs are insufficiently evaluated. The paper reports modest input-processing speedups and parameter reductions on one model, without end-to-end generation latency, decoding throughput, or timing variability. The complete pipeline’s wall-clock time and peak memory remain unreported, leaving the overall resource tradeoff unclear.

Questions for authors
  1. Baseline coverage and attribution: Could the authors include comparisons with recent structured-pruning methods, such as SlimLLM [1] or Týr-the-Pruner [2], or explain their applicability to this setting? Comparisons should account for differences in pruning units, calibration/search budgets, and weight reconstruction or compensation. Is FLAP’s original compensation used in the main experiments? How much of ACM’s advantage remains when alternative scores receive the same allocator or identical per-layer pruning budgets?

  2. Cross-model robustness: What explains the substantial downstream reversals in Table 8, particularly the FormationEval result for Llama-3.1-8B at 30% pruning? Do the existing results support a narrower conclusion about the model families or pruning regimes where ACM is effective?

  3. Efficiency: Can the authors report absolute prefill and decoding latency, end-to-end generation time at specified input/output lengths, and the hardware and measurement protocol? What are the wall-clock time and peak memory of the complete scoring-and-allocation pipeline relative to FLAP?

  4. Calibration protocol and overhead: Please specify calibration/probing sizes and sample selection, and clarify whether QA evaluation questions and their answers or explanations were excluded. Appendix C describes language-modeling splits, but item-level separation for QA remains unclear. Please also report the wall-clock time, peak memory, and storage requirements of calibration and probing for ACM and competing baselines. Accounting for these upfront costs, under what deployment workload do the inference savings yield a net efficiency benefit?

[1] Jialong Guo, Xinghao Chen, Yehui Tang, and Yunhe Wang. “SlimLLM: Accurate Structured Pruning for Large Language Models.” ICML, 2025. https://proceedings.mlr.press/v267/guo25a.html

[2] Guanchen Li et al. “Týr-the-Pruner: Structural Pruning LLMs via Global Sparsity Distribution Optimization.” NeurIPS, 2025. https://proceedings.neurips.cc/paper_files/paper/2025/hash/db2cbf43a349bc866111e791b58c7bf4-Abstract-Conference.html

2 comments

Thank you for your careful review and helpful suggestions.

Method comparison and downstream performance. In the main experiments, FLAP uses a query-head adaptation of its fluctuation score, without the bias compensation used in the original method. As described in Section 4.1 (ref to line 307 of our manuscript), ACM* retains ACM scores but globally prunes the lowest-scoring heads to decouple head scoring from budget allocation. To compare the scoring methods, we select heads for pruning by globally ranking the scores from ACM* and FLAP separately. ACM* achieved both a lower PPL ratio and higher top-1 prediction agreement in 30 of 48 settings.

On FormationEval, Llama-3.1-8B achieved accuracies of 0.403 with our ACM and 0.564 with FLAP when 30% of query heads were pruned. On PetroBench, the corresponding accuracies were 0.497 and 0.290 for the same model and pruning rate. We have not separately analysed the cause of this difference. We thank the reviewer for recommending SlimLLM and Týr-the-Pruner, and will consider including them in comparisons in a future version.

Efficiency and calibration settings. We evaluated Llama-3.2-3B on an A800 with 20% of query heads pruned, a batch size of 1, and an input length of 256 tokens. Generation was fixed at 128 tokens. Each condition used 2 warm-up runs and 7 measured runs. The median generation times for the original model before and after the pruning experiments were 2.143 and 2.135 seconds, respectively. The corresponding times were 2.204 seconds for FLAP and 2.173 seconds for our ACM. The appendix also reports prefill speedups.

For the downstream QA, calibration questions were separated from evaluation questions. Calibration text was extracted only from the questions.

Thank you for their time and effort. However, my main concerns remain insufficiently addressed; I therefore maintain my current rating.

Reviewer_VCjBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces the concept of Support Knowledge (SK) to describe domain-specific factual knowledge that large language models rely on when generating responses in specialized fields. The authors propose ACM, a training-free scoring method that combines logit margin and output magnitude signals to identify SK-related attention heads without requiring model retraining or access to internal activations beyond standard inference outputs. Through extensive evaluation across seven models ranging from 1.7B to 27B parameters and four datasets spanning petroleum, medical, legal, and general domains, the paper demonstrates that ACM outperforms existing head identification methods in both accuracy and stability. The work further provides practical guidance on domain-matched calibration data selection and shows that SK-relevant heads remain consistent across different model architectures and scales, offering actionable insights for efficient domain adaptation of large language models.

Critical strengths

I think the strongest part is the ACM scoring itself — combining margin and activation signals is a straightforward idea but it works, and the ablation (Fig. 4) does a decent job showing each signal alone isn't enough. The experiments are also extensive for this type of work: 7 models from 1.7B to 27B across 4 domain datasets. The head selection consistency results (Table 9) are a nice addition too, which show the method isn't totally sensitive to which calibration subset you pick.

Critical weaknesses
  1. OR logic may cause false positives. The formula is essentially a probabilistic OR: a head scores high if either signal is strong. However, some heads are naturally high-activation due to processing frequent tokens (e.g., positional markers, punctuation) without encoding domain knowledge. ACM may retain these "generically active" heads as important, and the paper does not analyze this false positive rate.

  2. OR logic may also cause false negatives. Heads that contribute through moderate but stable signals on both dimensions can be outranked by heads with one extreme signal driven by noise. For example, a head with accidentally high margin but low activation may rank above a genuinely important head whose contribution is evenly distributed across both signals.

  3. The paper only evaluates on text-only LLMs and does not test on multimodal LLMs (MLLMs), leaving the generalizability of ACM to architectures with vision encoders and cross-modal attention unverified.

  4. Turning continuous scores into ranks throws away how much stronger one head is vs another. With only ~15 heads per layer the rank steps are coarse (~0.07), so a head with a genuinely huge signal can end up right next to one with a barely higher signal. The paper doesn't check whether this ranking step actually hurts head selection quality.

  5. Lacking some relevant work on knowledge injection & attention: [1] Yu, B., Huang, W., & Liu, K. (2026, March). SR-KI: Scalable and Real-Time Knowledge Integration into LLMs via Supervised Attention. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 40, No. 41, pp. 34486-34494). [2] Li, R., Xu, J., Chen, Y., Chen, X., Yang, Y., Wang, J., ... & Zhang, X. Y. (2026). Generation-Augmented Generation: A Plug-and-Play Framework for Private Knowledge Injection in Large Language Models. arXiv preprint arXiv:2601.08209. [3] Chen, T., Tan, Z., Gong, T., Wu, Y., Chu, Q., Liu, B., ... & Yu, N. (2024, November). Llama slayer 8b: Shallow layers hold the key to knowledge injection. In Findings of the Association for Computational Linguistics: EMNLP 2024 (pp. 5991-6002).

Questions for authors

See weaknesses.

2 comments

Thank you for your careful review and helpful suggestions.

General-purpose heads and false positives. Domain prediction relies on both domain information and general language processing capabilities. Heads that process common content, such as punctuation, may also be useful for domain prediction. Therefore, a head's general function alone does not establish that retaining it is a false positive. Our ACM aims to retain computations that support prediction in the target domain. We evaluate their retention through domain prediction performance after pruning.

Combination rule and false negatives. As defined in Section 3.1 (ref to line 203 of our manuscript), a larger raw margin corresponds to a lower transformed boundary score. Therefore, a head does not receive a higher ACM retention score simply because its raw margin is high. We compared our ACM with a multiplicative combination that places greater emphasis on the joint contribution of both signals. As reported in Table 11 in the appendix (starting at line 1473), in 142 of 144 paired settings, our ACM achieved both a lower PPL ratio and higher top-1 prediction agreement. This result supports the overall effectiveness of the current combination.

Multimodal models. This paper focuses on domain prediction and attention head pruning in text-only language models. Applying ACM to multimodal models requires re-examining head scoring and budget allocation for the specific architecture.

For models with separate cross-attention, our ACM currently defines candidate positions using causal prefixes, so valid positions need to be redefined according to the cross-modal attention mask. The attention rules and output formats of vision encoders also differ from those of autoregressive language models. If they are pruned, it is also necessary to specify how the final task outputs will be used to measure pruning damage in each layer.

For architectures that feed visual representations into a language model, our ACM can in principle be applied to self-attention heads in the language backbone, but its effectiveness remains to be validated. Similarity between visual regions and the mixture of visual and textual candidates may change the relationship between low margins and head importance. A large number of visual tokens may also affect the weight of different positions in the score. In addition, evaluation needs to consider both prediction agreement before and after pruning and visual task accuracy, and measure the effect of language-head pruning on overall inference efficiency.

Therefore, validating applicability to multimodal models is currently beyond the scope of this paper.

Within-layer ranking. As explained in Section 3.2 (ref to line 221 of our manuscript), to combine the margin and output magnitude, which have different numerical scales, our ACM first computes their within-layer ranks and then determines head retention priorities. Existing ablation results show that the combination based on these two rankings effectively preserves prediction performance in the evaluated settings.

Related work. We thank the reviewer for recommending these works. SR-KI, Generation-Augmented Generation and Llama SLayer primarily study knowledge injection or the enhancement of domain capabilities. Our work focuses on computations in pretrained models that support domain prediction and evaluates their preservation through attention head pruning. Therefore, these works differ from our ACM in their research questions to some extent. We will cite these works in the revised related work section to clarify their relationship to our ACM.

Thank you for the detailed response and for running the additional analyses. While I appreciate the authors' efforts, my main concerns remain only partially addressed. Taking all the discussions into account, I have decided to maintain my current rating.

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

This paper focuses on the pruning of LLMs on the target domains and proposes a new pruning method by only retaining the heads that contain the support knowledge of the target domain. More specifically, it proposes a method called Activation-Calibrated Margin (ACM) to rank the heads within each layer. The heads with a smaller margin between the top 2 attention logits or a larger output magnitude tend to be retained. Besides, the authors propose an extra method to decide the pruning ratio of each layer with a given pruning budget for the whole LLM. Experimental results show that the proposed pruning method can achieve better performance on the target domain than the evaluated pruning baselines.

Critical strengths

This proposed pruning method is reasonable. It considers both the head ranking within the layer and non-uniform pruning ratio across layers. When designing the retention scores for head ranking, the authors consider both factors on attention level and output level.

Critical weaknesses

The key critical weaknesses are listed as follows.

  1. The experimental results are not convincing enough to show the superiority of the proposed method on the pruning for target domain. Firstly, the authors only evaluate their method on the petroleum domain as the target domain. The authors should show results on more domain datasets such as the medical and legal domains. Secondly, the authors only evaluate their method with moderate pruning ratios. The authors should evaluate with more higher pruning ratios such as 50% and 60%. Thirdly, apart from random pruning, the authors only compare their method with FLAP and HeadConf. They should compare with more structured pruning baselines such as Bonsai, SlimGPT, OSSCAR and more.

  2. The benefits brought by the proposed pruning method are not obvious according to the experimental results. More specifically, the speedup of the pruned model is minor and the reduction in the model’s parameters and memory usage may also be small given the fact that only the query heads are removed with a relatively small pruning ratio.

Questions for authors

Could you provide more explanations about why a query head with a smaller margin is better to keep compared with the one with larger margin? The empirical results in Section 4.2 only show that removing low-margin heads is more harmful than removing random heads. Are there any other and strong empirical results support this assumption? Based on my understanding, it is also possible that the heads with a larger margin can contain the support knowledge since they are more confident in their token selection and thus have a higher impact on the final output of the target LLMs.

1 comment

Thank you for your careful review and helpful suggestions.

Experimental coverage. We chose the petroleum domain as the target domain because its specialized texts contain extensive domain terminology and knowledge, allowing us to evaluate the preservation of predictions in a specialized domain. Our ACM scores heads and allocates pruning budgets based on attention statistics and prediction changes on calibration text, without using rules specific to the petroleum domain. It can therefore, be applied to other specialized domains. We will consider cross-domain evaluation as future work.

Higher pruning ratios. We evaluated 50% and 60% query-head pruning on Qwen3-1.7B, Qwen3-8B, Llama-3.2-3B, and Llama-3.1-8B across three petroleum datasets. Averaged over three runs, ACM achieved the lowest PPL ratio and highest top-1 agreement in all 24 settings, compared with Random, HeadConf, FLAP, and ACM*. For example, at 60% pruning on DISKOS-QA, ACM achieved PPL ratios of 3.58 versus 10.17 for FLAP on Qwen3-8B, and 2.64 versus 37.60 on Llama-3.1-8B.

We thank the reviewer for recommending Bonsai, SlimGPT and OSSCAR, and will consider including them in comparisons in a revision.

Efficiency gains. As reported in the appendix (ref to line 1449 of our manuscript), on Llama-3.2-3B, pruning 20% and 30% of query heads reduces the model parameter count by 3.28% and 4.94%, respectively. At a batch size of 4 and an input length of 4096 tokens, the corresponding prefill speedups are 1.041 and 1.073, respectively.

The role of low-margin heads. A low margin means that the top two candidates in attention are closer, making their selection more sensitive to perturbations. In experiments with matched numbers of pruned heads per layer and stratification by output magnitude, pruning low-margin heads caused greater PPL degradation than pruning high-margin heads in 92 of 135 settings (see Table 2 in the appendix, starting at line 811). High-margin heads can also be important for prediction, so our ACM considers both margin and output magnitude.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.