acceptodds
Under review as a conference paper at ICLR 2027

From Post-Training Plasticity to Function-Targeted Adaptation in Large Language Models

Abstract

Post-training is central to developing and enhancing the capabilities of pretrained large language models (LLMs), yet its effects on their internal functional organization remain poorly understood. Inspired by experience-driven plasticity in biological intelligence, we investigate post-training plasticity: how the functional organization of attention heads evolves as LLMs acquire or enhance capabilities through post-training. We find that post-training primarily strengthens functional structures already present in pretrained models. This observation motivates Function-Targeted Adaptation (FTA), a selective fine-tuning paradigm that translates findings on post-training plasticity into targeted parameter updates. FTA identifies attention heads associated with a target capability and restricts updates to their corresponding weights, allowing us to examine whether existing functional structures provide a shared basis for both capability enhancement and knowledge suppression. We evaluate this hypothesis through skill learning and knowledge unlearning, assessing target adaptation alongside retention of non-target capabilities. Across multiple models and adaptation objectives, FTA updates at most around 10% of attention heads yet matches or exceeds full fine-tuning in target adaptation, while substantially reducing interference with non-target capabilities and retaining performance close to that of the original model. These results connect a mechanistic understanding of post-training plasticity with selective model adaptation, demonstrating how existing functional structures can guide both learning and unlearning.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
post-training, functional fine-tuning, large language model, interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CoUhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how the functional organization of attention heads changes during post-training and uses these observations to propose Function-Targeted Adaptation (FTA).

First, the authors use probing and gradient-based attribution to identify attention heads associated with specific capabilities, including mathematical reasoning, harmlessness, and knowledge unlearning. By tracking these heads across checkpoints of one-epoch fine-tuning runs on Llama-3.1-8B, they find that many of the most important heads remain among the top-ranked heads, although their importance scores change during training. They also observe that heads associated with other capabilities remain relatively stable.

Based on these findings, the authors propose FTA, which identifies capability-relevant attention heads in the base model and restricts fine-tuning to their query, key, value, and output projections (~1% of the model parameters). The method is evaluated on four models (Llama-3.1-8B, Llama-3.2-3B, OLMo-2-7B, and OLMo-2-13B) for mathematical reasoning, harmlessness, and honesty under SFT and DPO, as well as knowledge unlearning on TOFU. It is compared against full fine-tuning, LoRA, P2D, and a random-head control. The results suggest that FTA can achieve comparable or better target-task performance than the baselines while better preserving performance on ten non-target benchmarks.

Critical strengths
  1. The general idea of the paper is interesting, especially the connection between understanding how models change during post-training and leverating that information to guide adaptation.

  2. The authors evaluate FTA across different model families, model sizes, capabilities, and training objectives. They consider not only improvements on the target task but also the preservation of non-target capabilities, evaluating performance across ten additional benchmarks.

  3. I found the comparison with randomly selected attention heads insightful. The results suggest that the benefits of FTA come not only from restricting the number of updated parameters but also from selecting which heads to update. The comparison with parameter-matched LoRA is also useful in this regard. Although I have some concerns about the reliability of the experimental comparisons, I think these ablations provide some evidence supporting the main idea of the paper.

Critical weaknesses
  1. The paper appears to report only one training run per configuration. While the authors analyze variability across sampled generations (Table 14), this does not capture variability across independent training runs. Reporting results across multiple seeds, at least for the main comparisons, would help establish how robust the improvements are.

  2. Probing and gradient-based attribution can identify heads associated with a capability, but do not establish a causal role. Ablating the selected heads could help validate this interpretation. Additionally, it is unclear whether the observed stability of top-k heads reflects the strengthening of existing functional structures or simply the limited changes expected from one epoch of fine-tuning.

  3. The paper would benefit from comparing the proposed probing-based selection with other ways of identifying parameters relevant to a task. For example, could a simpler gradient-based importance measure computed directly from the adaptation objective achieve similar results? The authors already use gradient-based attribution without a probe for unlearning (Appendix A.1), making this comparison particularly relevant.

  4. I would have liked a clearer discussion of how the findings in Section 3 differ from previous observations, particularly given that the post-training plasticity analysis is presented as one of the main contributions. It would also be interesting to discuss how these findings relate to Park et al. (2025), which reports newly recruited attention heads during post-training.

  5. The authors primarily measure forgetting using ROUGE-L on the forget set, but a low ROUGE-L score does not necessarily indicate that the underlying knowledge has been removed. Although the authors acknowledge this limitation, additional unlearning metrics would help support the conclusions about knowledge suppression.

Questions for authors
  1. Could the authors clarify the discrepancies between the reported results? For example, the target average for LoRA (r=8) on Llama-3.1-8B under mathematical reasoning SFT is 78.4 in Table 1, but 44.2 in Tables 5 and 17. Similarly, the LoRA results for Llama-3.2-3B in Table 8 repeat the individual target benchmark scores reported for Llama-3.1-8B in Table 1.

  2. Several baselines perform considerably worse than the original model after fine-tuning. For example, on Llama-3.1-8B under mathematical reasoning SFT, the reported results for full fine-tuning, LoRA, and P2D are all below the base model (80.4). There are also cases in the alignment experiments where full fine-tuning causes very large drops in non-target performance. Could the authors provide more details about the hyperparameter search spaces and selection criteria used for each baseline? This would help clarify whether the results reflect limitations of the adaptation methods or differences in optimization stability.

  3. Could the authors clarify how the elbow criterion is used to determine the number of selected heads, k? Although Figure 4 provides the ranked importance curves, the procedure is not described in enough detail to reproduce the selection. It would also help to report the selected values for each model and capability in a table. Additionally, which value was used for harmlessness: k=116, as reported in Section 3.2 and Appendix A.3, or k=107, as reported in Tables 6 and 7?

  4. Does the efficiency comparison in Table 5 include the full cost of head selection, including activation extraction, probe training, and attribution? If not, could the authors report these costs to provide a more complete picture of FTA’s efficiency?

6 comments
AuthorsOpenReview ↗

Response to Reviewer

Dear Reviewer,

We deeply appreciate the time and effort you have dedicated to reviewing our work. Below, we address the questions you raised.

Q. Reporting results across multiple seeds, at least for the main comparisons, would help establish how robust the improvements are.

Following the reviewer's suggestion, we conducted three independent training runs with different random seeds for the main experimental comparisons, reporting the mean and standard deviation of target and non-target performance.

The results in Table 3 show that FTA achieves the best or tied-best mean target performance in 9 out of 13 settings with significant (p < 0.05) or marginally significant (p < 0.1) improvements in several comparisons. In particular, FTA consistently outperforms the baselines on mathematical reasoning and achieves strong results on honesty-related tasks, although some settings exhibit greater variability.

For non-target capabilities, FTA demonstrates significantly better preservation on Llama-3.1-8B Math SFT than all baselines (p < 0.01), while maintaining broadly comparable performance to P2D in several other settings, with some trade-offs.

Overall, these additional analyses strengthen the empirical evidence for FTA's robustness across training runs. We will incorporate these results and clarifications into the revised manuscript.

Table 1. Target and non-target performance (mean±std over three seeds).

ModelSettingTarget fullTarget LoRATarget P2DTarget OursNon-Target fullNon-Target LoRANon-Target P2DNon-Target Ours
Llama-3.1-8BMath SFT71.6±0.179.1±0.366.9±0.682.4±1.256.4±1.563.2±1.360.7±1.868.4±0.1
Math DPO73.7±1.477.9±0.576.2±1.379.0±1.760.4±4.964.2±1.064.5±1.666.5±0.1
Harmless SFT94.7±1.296.0±3.580.0±14.293.3±5.152.4±2.663.1±0.664.6±0.263.8±0.5
Harmless DPO100.0±0.0100.0±0.0100.0±0.0100.0±0.043.8±3.563.4±1.161.7±0.562.6±0.7
Honesty SFT26.4±1.929.7±1.937.8±2.246.4±3.438.2±4.265.2±0.265.3±0.264.6±0.1
Honesty DPO15.3±5.834.5±0.747.4±9.739.0±10.76.8±1.665.2±0.365.2±0.364.6±0.4
Unlearning (GradDiff)0.4±0.521.8±14.327.1±22.20.4±0.732.2±27.962.5±1.749.1±18.356.3±4.8
OLMo-2-13BMath SFT73.2±0.970.0±0.371.3±0.578.9±3.857.3±0.561.2±0.663.2±0.261.2±0.2
Math DPO60.1±2.762.3±0.863.8±0.274.6±10.954.6±0.854.5±0.660.8±0.160.8±1.0
Harmless SFT78.3±9.957.7±5.769.3±6.077.3±6.763.7±0.665.4±0.065.7±0.164.8±0.5
Harmless DPO100.0±0.0100.0±0.097.3±4.699.0±1.760.9±1.264.8±0.165.0±0.263.3±1.2
Honesty SFT56.9±6.448.0±0.753.3±1.457.3±0.661.6±1.165.3±0.065.4±0.163.9±1.4
Honesty DPO59.8±8.156.3±2.355.3±2.068.9±3.461.4±1.065.3±0.165.5±0.163.0±1.9

Table 2. Statistical analysis of Ours vs. others. bold: p < 0.1

ModelSettingTarget Ours vs. fullTarget Ours vs. LoRATarget Ours vs. P2DNon-Target Ours vs. fullNon-Target Ours vs. LoRANon-Target Ours vs. P2D
Llama-3.1-8BMath SFT0.00190.0174<0.00050.00250.00990.0087
Math DPO0.00760.19020.04530.08190.02810.0811
Harmless SFT0.65750.75160.12070.00720.09890.9529
Harmless DPO------0.00460.82120.0758
Honesty SFT0.00130.00220.01390.00420.99020.9935
Honesty DPO0.02080.27120.8144<0.00050.94420.9442
Unlearning (GradDiff)0.50.06090.08630.13580.92820.2851
OLMo-2-13BMath SFT0.05740.02750.03590.0010.50.9999
Math DPO0.07040.09460.11410.00070.00090.5
Harmless SFT0.55360.00950.09950.03670.91340.958
Harmless DPO0.79230.79230.29880.03520.9190.935
Honesty SFT0.4619<0.00050.01230.04630.88730.898
Honesty DPO0.09070.00430.00380.14320.91470.925

Q. Causal validation of capability-associated heads and uncertainty about whether top-k head stability persists beyond one epoch of fine-tuning.

To assess the causal relevance of the identified heads, we conducted additional inference-time ablation experiments on Llama-3.1-8B for mathematical reasoning. Specifically, we perform head ablation by scaling each selected attention head's output by a small factor (ϵ=0.001\epsilon=0.001), comparing the effects of ablating the top-k capability-associated heads against an equal number of randomly selected heads. As shown in the table below, ablating the identified heads causes substantial performance degradation across all four benchmarks, whereas ablating random heads results in considerably smaller declines. These results provide additional causal evidence for the functional involvement of the selected heads.

Table 3. Performance comparison between masking the top-kk heads and kk random heads on Llama-3.1-8B for math reasoning.

MethodGSM8KSVAMPASDivGSM+Math Avg.
Base81.9088.2084.0067.5080.40
Random-kk74.9880.8080.3657.7273.46
Top-kk55.6559.2074.0841.8557.69

To further investigate the stability of top-k heads, we extended our post-training plasticity analysis to three epochs (300% training progress) of full-parameter fine-tuning. As shown in Figure 4 of Rebuttal PDF, head stability persists beyond a single epoch, while changes in head importance remain concentrated among initially important heads (AUROC ≥\geq 0.73), indicating continued adaptation of existing functional structures. Furthermore, the relatively low cross-domain overlap suggests that these changes remain capability-specific. These observations are consistent with our original one-epoch findings, demonstrating that the identified patterns of functional plasticity persist over extended fine-tuning rather than being limited to a single epoch.

Q. Comparison of probing-based head selection with direct gradient-based importance measures.

Our current experiments already include P2D, a task-sensitive head selection method that identifies important attention heads by measuring weight-distribution shifts following a lightweight proxy fine-tuning process. To further address this concern, we conducted an additional comparison with direct gradient-based head selection.

The results show that FTA outperforms direct gradient-based selection by 13.2 points on target tasks (81.0% vs. 67.8%) and 5.0 points on non-target tasks (68.3% vs. 63.3%). These results further support the effectiveness of probing-based capability localization for selective adaptation.

Table 4. Comparison between probing-based head selection and direct gradient-based head selection on Llama-3.1-8B for math reasoning.

BenchmarkOursDirect GradientΔ
Target
GSM8K82.871.7−11.1
SVAMP89.472.6−16.8
ASDiv83.471.4−12.0
GSM+68.455.5−12.9
Avg81.067.8−13.2
Non-Target
ARC-C81.979.4−2.5
LogiQA256.147.7−8.4
HumanEval+63.464.6+1.2
CRUX30.024.6−5.4
MMLU63.259.4−3.8
CSQA76.575.6−0.9
OBQA83.282.4−0.8
SIQA73.569.2−4.3
SQuAD71.148.6−22.5
BoolQ84.081.1−2.9
Avg68.363.3−5.0

Q. Distinction from prior studies on post-training plasticity, particularly Park et al. (2025).

Several prior studies have investigated how LLM components, particularly attention heads, evolve during fine-tuning.

Zhao et al. [1] employ gradient-based attribution to analyze attention-head activation patterns, showing that SFT selectively activates task-specific heads. Wang et al. [2] investigate fine-tuning through circuit analysis and find that functional circuits retain high node similarity despite substantial changes in their connections.

Park et al. [3] examine post-training dynamics through circuit analysis, focusing on emergent reasoning heads and their evolution under different training paradigms. A key methodological distinction lies in how head involvement is measured: Park et al. identify emergent heads based on discrete membership in sparse circuits, whereas we track continuous probe-based importance scores across training checkpoints. Under the former criterion, a head that already contributes to the baseline model but is not selected into its sparse circuit may be classified as emergent once it enters the post-training circuit.

Importantly, Park et al. also report that approximately two-thirds of baseline attention-head nodes remain active after distillation (Section 4.1), while GRPO primarily reuses existing computational pathways. These observations are consistent with our findings on the persistence of pre-existing functional heads. Our analysis further characterizes non-uniform changes in head importance and cross-capability stability, revealing how capability-associated functional structures evolve throughout post-training.

Overall, we systematically characterize head-level functional plasticity, and translate these insights into FTA for selective parameter updates. We will clarify these connections and distinctions in the revised manuscript.

[1] Zhao, Y., et al. Analyzing the rapid generalization of SFT via the perspective of attention head activation patterns.

[2] Wang, X., et al. Towards understanding fine-tuning mechanisms of LLMs via circuit analysis.

[3] Park, Y., et al. Thinking sparks!: Emergent attention heads in reasoning models during post training.

Q. More unlearning metrics beyond ROUGE-L.

To provide a more comprehensive evaluation, we follow the standard TOFU evaluation protocol in [1] and report additional metrics, including answer probability and Truth Ratio, across the Forget Set, Retain Set, Real Authors, and World Facts.

As shown in the following table, FTA achieves a near-zero forget-set answer probability and the highest forget-set Truth Ratio, indicating strong suppression of target knowledge. Meanwhile, FTA achieves the highest retain-set ROUGE-L score and answer probability among all methods, along with superior performance on Real Authors and World Facts, demonstrating better preservation of unrelated knowledge. These complementary metrics provide stronger evidence for FTA's effectiveness in balancing targeted knowledge suppression and knowledge retention.

Table 5. More results on TOFU.

MethodForget Prob.↓Forget Truth Ratio↑Retain Prob.↑Retain Truth Ratio↑Retain ROUGE-L↑Real Authors Prob.↑Real Authors Truth Ratio↑World Facts Prob.↑World Facts Truth Ratio↑
Full2.5×10⁻⁶⁷±3.5×10⁻⁶⁷34.7±10.621.5±30.432.0±9.223.4±32.737.7±19.039.9±23.540.5±20.542.5±24.3
LoRA3.4±0.737.9±2.515.5±4.248.4±0.037.5±1.441.4±2.441.0±3.151.6±0.854.8±0.1
P2D17.7±1.844.9±0.655.2±0.550.0±0.951.5±3.947.8±5.850.4±7.554.8±4.557.8±4.7
Ours1.3×10⁻⁴±1.8×10⁻⁴52.7±3.471.5±0.550.0±1.670.2±2.352.9±0.759.2±1.557.5±2.362.6±2.4

[1] Maini, P., et al. Tofu: A task of fictitious unlearning for llms.

Q. Could the authors clarify the discrepancies in the reported results, for LoRA (r=8r=8) under Math SFT on Llama-3.1-8B and Llama-3.2-3B?

Thank you for identifying these inconsistencies. For Llama-3.1-8B under Math SFT, the LoRA (r=8) results are 81.3 (GSM8K), 84.9 (SVAMP), 81.5 (ASDiv), and 65.8 (GSM+), yielding an average of 78.4, as reported in Table 1.

The corrected results for Llama-3.2-3B LoRA (r=8) are presented in the table below. FTA achieves the highest performance across all four benchmarks, with an average accuracy of 72.2%, outperforming full fine-tuning and LoRA. We will correct these reporting errors in Appendix and ensure consistency across all tables in the revised manuscript.

Table 6. Corrected Math SFT results on Llama-3.2-3B, including the updated LoRA (r=8) results.

MethodGSM8KSVAMPASDivGSM+Avg.
Base53.964.465.051.058.6
Full FT65.083.080.147.568.9
LoRA (r=8)44.172.673.541.357.9
FTA (Ours)69.184.280.355.372.2
Reviewer_BGEKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how attention heads change during LLM post-training and proposes Function-Targeted Adaptation (FTA). The authors report that important head sets remain largely stable during post-training, motivating selective adaptation of existing task-associated components. Experients include full fine-tuning, LoRA, P2D, and random-head adaptation, with target-task metrics, ten non-target benchmarks, and ablations on head selection, selection ratios, probe seeds, and LoRA ranks. These results connect a mechanistic understanding of post-training plasticity with selective model adaptation, demonstrating how existing functional structures can guide both learning and unlearning.

Critical strengths
  1. A practical connection between functional analysis and selective adaptation: FTA uses task-associated internal representations to determine where parameter updates occur. It offers a concrete way to translate interpretability analysis into adaptation decisions, with encouraging reported results on mathematical reasoning and honesty.

  2. Broad evaluation of adaptation and retention. The paper considers multiple model sizes, two model families, several capabilities, and both learning and unlearning. Evaluating target performance alongside ten non-target benchmarks, together with random-head controls, makes the study more informative than target-task evaluation alone.

  3. These results connect a mechanistic understanding of post-training plasticity with selective model adaptation, demonstrating how existing functional structures can guide both learning and unlearning.

Critical weaknesses
  1. The evidence does not establish the claimed functional mechanism or selective knowledge removal. Probe attribution on complete question–response sequences measures label predictability, but the paper does not sufficiently exclude response-format or stylistic cues, or establish causal involvement of the selected heads. Retraining probes across checkpoints further complicates comparisons of importance magnitudes. Similarly, low forget-set ROUGE-L demonstrates suppressed answer reproduction but does not establish knowledge removal; direct TOFU retain-set evaluation is missing. The reported generation-sampling repetitions also do not establish robustness across independent training runs.

  2. Limited evidence of generalization across model families. The expriments only include Llama and OLMo, evaluation on other model familys, such as Qwen or Deepseek, would help establish whether the proposed selection criterion generalizes well.

  3. The random-head control does not clearly isolate functional relevance from structural differences. The main text states that the random baseline is matched in trainable-parameter count and update structure, while the appendix primarily describes matching the number of selected heads. Under grouped-query attention, equally sized head sets can select different numbers of shared K/V groups and therefore expose different parameter budgets and affect different numbers of unselected query heads. The paper also does not specify whether the random baseline matches FTA’s layer distribution. Without documenting these controls, the observed advantage cannot be attributed as confidently to functional relevance alone.

Questions for authors

Check the weakness. I will raise my score if you can solve my questions

5 comments
AuthorsOpenReview ↗

Response to Reviewer

Dear Reviewer,

Thank you very much for your time and effort in reviewing our paper. We have thoughtfully addressed each question you raised and hope our responses alleviate your concerns.

Q. The evidence does not establish the claimed functional mechanism or selective knowledge removal. Probe attribution on complete question–response sequences measures label predictability, but the paper does not sufficiently exclude response-format or stylistic cues, or establish causal involvement of the selected heads. Retraining probes across checkpoints further complicates comparisons of importance magnitudes. Similarly, low forget-set ROUGE-L demonstrates suppressed answer reproduction but does not establish knowledge removal; direct TOFU retain-set evaluation is missing. The reported generation-sampling repetitions also do not establish robustness across independent training runs.

We address each concern with additional analyses and experiments.

1. Response-Format and Stylistic Cues.

To examine potential surface-form confounders, we analyzed response patterns and length distributions across label groups for mathematical reasoning, harmlessness, and honesty. As shown in Figures 2 and 3 of the Rebuttal PDF, both label groups exhibit diverse and substantially overlapping phrasing patterns. Their response-length distributions also overlap across the full length range, with substantial representation from both labels. These observations suggest that the identified heads are unlikely to be explained solely by simple stylistic or length-related cues.

2. Causal Validation through Head Ablation.

To assess the causal relevance of the identified heads, we conducted additional inference-time ablation experiments on Llama-3.1-8B for mathematical reasoning. Specifically, we perform head ablation by scaling each selected attention head's output by a small factor (ϵ=0.001\epsilon=0.001), comparing the effects of ablating the top-k math related heads against an equal number of randomly selected heads. As shown in the table below, ablating the identified heads causes substantial performance degradation across all four benchmarks, whereas ablating random heads results in considerably smaller declines. These results provide additional causal evidence for the functional involvement of the selected heads.

Table 1. Performance comparison between masking the top-kk heads and kk random heads on Llama-3.1-8B for math reasoning.

MethodGSM8KSVAMPASDivGSM+Math Avg.
Base81.9088.2084.0067.5080.40
Random-k74.9880.8080.3657.7273.46
Top-k55.6559.2074.0841.8557.69

3. Consistency of Importance Scores Across Checkpoints.

We clarify that probes are not retrained across checkpoints. Each probe is trained once on the initial model (0% checkpoint) and subsequently kept frozen. At each checkpoint, we compute importance scores using the updated head activations and the same fixed probe. We will revise the wording in Lines 182–183 to eliminate this ambiguity.

4. Additional Knowledge Unlearning Metrics.

As acknowledged in Lines 368–370, "these results extend capability-associated selection from enhancement to targeted knowledge suppression, although low forget-set ROUGE-L alone does not establish complete knowledge removal.". To provide a more comprehensive evaluation, we follow the standard TOFU evaluation framework [1] and report additional metrics, including answer probability and Truth Ratio, across the Forget Set, Retain Set, Real Authors, and World Facts.

As shown in the following table, FTA achieves a near-zero forget-set answer probability and the highest forget-set Truth Ratio, indicating strong suppression of target knowledge. Meanwhile, FTA achieves the highest retain-set ROUGE-L score and answer probability among all methods, along with superior performance on Real Authors and World Facts, demonstrating better preservation of unrelated knowledge. These complementary metrics provide stronger evidence for FTA's effectiveness in balancing targeted knowledge suppression and knowledge retention.

Table 2. More results on TOFU.

MethodForget Prob.↓Forget Truth Ratio↑Retain Prob.↑Retain Truth Ratio↑Retain ROUGE-L↑Real Authors Prob.↑Real Authors Truth Ratio↑World Facts Prob.↑World Facts Truth Ratio↑
Full2.5×10⁻⁶⁷±3.5×10⁻⁶⁷34.7±10.621.5±30.432.0±9.223.4±32.737.7±19.039.9±23.540.5±20.542.5±24.3
LoRA3.4±0.737.9±2.515.5±4.248.4±0.037.5±1.441.4±2.441.0±3.151.6±0.854.8±0.1
P2D17.7±1.844.9±0.655.2±0.550.0±0.951.5±3.947.8±5.850.4±7.554.8±4.557.8±4.7
Ours1.3×10⁻⁴±1.8×10⁻⁴52.7±3.471.5±0.550.0±1.670.2±2.352.9±0.759.2±1.557.5±2.362.6±2.4

[1] Maini, P., et al. Tofu: A task of fictitious unlearning for llms.

5. Robustness Across Independent Training Runs.

To address this limitation, we conducted three independent training runs with different random seeds for the main experimental comparisons, reporting the mean and standard deviation of target and non-target performance.

The results in Table 3 show that FTA achieves the best or tied-best mean target performance in 9 out of 13 settings with significant (p < 0.05) or marginally significant (p < 0.1) improvements in several comparisons. In particular, FTA consistently outperforms the baselines on mathematical reasoning and achieves strong results on honesty-related tasks, although some settings exhibit greater variability.

For non-target capabilities, FTA demonstrates significantly better preservation on Llama-3.1-8B Math SFT than all baselines (p < 0.01), while maintaining broadly comparable performance to P2D in several other settings, with some trade-offs.

Overall, these additional analyses strengthen the empirical evidence for FTA's functional relevance, selective knowledge suppression, and robustness across training runs. We will incorporate these results and clarifications into the revised manuscript.

Table 3. Target and non-target performance (mean±std over three seeds).

ModelSettingTarget fullTarget LoRATarget P2DTarget OursNon-Target fullNon-Target LoRANon-Target P2DNon-Target Ours
Llama-3.1-8BMath SFT71.6±0.179.1±0.366.9±0.682.4±1.256.4±1.563.2±1.360.7±1.868.4±0.1
Math DPO73.7±1.477.9±0.576.2±1.379.0±1.760.4±4.964.2±1.064.5±1.666.5±0.1
Harmless SFT94.7±1.296.0±3.580.0±14.293.3±5.152.4±2.663.1±0.664.6±0.263.8±0.5
Harmless DPO100.0±0.0100.0±0.0100.0±0.0100.0±0.043.8±3.563.4±1.161.7±0.562.6±0.7
Honesty SFT26.4±1.929.7±1.937.8±2.246.4±3.438.2±4.265.2±0.265.3±0.264.6±0.1
Honesty DPO15.3±5.834.5±0.747.4±9.739.0±10.76.8±1.665.2±0.365.2±0.364.6±0.4
Unlearning (GradDiff)0.4±0.521.8±14.327.1±22.20.4±0.732.2±27.962.5±1.749.1±18.356.3±4.8
OLMo-2-13BMath SFT73.2±0.970.0±0.371.3±0.578.9±3.857.3±0.561.2±0.663.2±0.261.2±0.2
Math DPO60.1±2.762.3±0.863.8±0.274.6±10.954.6±0.854.5±0.660.8±0.160.8±1.0
Harmless SFT78.3±9.957.7±5.769.3±6.077.3±6.763.7±0.665.4±0.065.7±0.164.8±0.5
Harmless DPO100.0±0.0100.0±0.097.3±4.699.0±1.760.9±1.264.8±0.165.0±0.263.3±1.2
Honesty SFT56.9±6.448.0±0.753.3±1.457.3±0.661.6±1.165.3±0.065.4±0.163.9±1.4
Honesty DPO59.8±8.156.3±2.355.3±2.068.9±3.461.4±1.065.3±0.165.5±0.163.0±1.9

Table 4. Statistical analysis of Ours vs. others. bold: p < 0.1

ModelSettingTarget Ours vs. fullTarget Ours vs. LoRATarget Ours vs. P2DNon-Target Ours vs. fullNon-Target Ours vs. LoRANon-Target Ours vs. P2D
Llama-3.1-8BMath SFT0.00190.0174<0.00050.00250.00990.0087
Math DPO0.00760.19020.04530.08190.02810.0811
Harmless SFT0.65750.75160.12070.00720.09890.9529
Harmless DPO------0.00460.82120.0758
Honesty SFT0.00130.00220.01390.00420.99020.9935
Honesty DPO0.02080.27120.8144<0.00050.94420.9442
Unlearning (GradDiff)0.50.06090.08630.13580.92820.2851
OLMo-2-13BMath SFT0.05740.02750.03590.0010.50.9999
Math DPO0.07040.09460.11410.00070.00090.5
Harmless SFT0.55360.00950.09950.03670.91340.958
Harmless DPO0.79230.79230.29880.03520.9190.935
Honesty SFT0.4619<0.00050.01230.04630.88730.898
Honesty DPO0.09070.00430.00380.14320.91470.925

Q. Limited evidence of generalization across model families.

To further evaluate the generalizability of FTA, we extended our experiments to Qwen3-4B-Base, introducing a third model family beyond Llama and OLMo. We evaluated mathematical reasoning, harmlessness, and honesty under both SFT and DPO, using the same baselines and evaluation protocol.

The additional results further support FTA's effectiveness across model architectures. For mathematical reasoning, FTA achieves the highest average target performance under both SFT (66.4%) and DPO (65.9%), outperforming full fine-tuning, LoRA, and P2D. For harmlessness, FTA achieves the highest safe-response rate under SFT (90.0%), while largely preserving non-target capabilities. For honesty, FTA achieves the highest target performance under SFT and DPO, while maintaining non-target performance close to the base model.

Overall, these results provide additional evidence that FTA's capability-guided head selection generalizes beyond the original Llama and OLMo families.

Table 5.Target and non-target performance (%) of Qwen3-4B-Base after post-training on mathematical reasoning.

ModelTrainMethodGSM8KSVAMPASDivGSM+Target AvgARC-CLogiQA2HEval+CRUXMMLUCSQAOBQASIQASQuADBoolQNon-Target Avg
Qwen3-4B-Base–base54.688.882.437.565.888.664.254.940.572.881.886.077.179.585.173.0
SFTfull48.888.381.434.663.388.764.556.740.672.881.685.676.878.383.872.9
LoRA50.588.882.435.864.488.764.754.940.873.081.985.876.879.584.873.1
P2D52.989.782.236.665.388.864.755.540.472.981.786.277.179.584.773.1
Ours54.886.781.143.166.489.064.555.540.472.981.986.077.179.585.273.2
DPOfull41.882.676.029.657.588.865.159.841.072.981.685.676.780.585.073.7
LoRA49.187.480.634.662.988.764.455.541.072.881.786.277.179.884.473.2
P2D49.887.681.334.963.488.664.459.841.072.981.385.477.079.285.273.5
Ours55.188.482.537.765.988.964.653.740.672.881.786.076.979.085.372.9

Table 6. Target and non-target performance (%) of Qwen3-4B-Base after post-training for harmlessness and honesty.

ModelValueTrainMethodSafe/HonARC-CLogiQA2HEval+CRUXMMLUCSQAOBQASIQASQuADBoolQNon-Target Avg
Qwen3-4B-Baseharmless–base23.088.664.254.940.572.881.886.077.179.585.173.0
SFTfull79.089.061.651.238.871.779.684.875.474.782.770.9
LoRA29.088.364.650.040.072.581.785.077.280.085.072.4
P2D64.088.563.948.239.572.881.783.476.480.085.171.9
Ours90.088.563.547.640.072.881.685.876.079.783.471.9
DPOfull100.087.855.250.639.870.075.581.870.075.081.468.7
LoRA99.088.164.751.841.972.681.885.075.575.885.072.2
P2D99.087.663.252.442.472.581.784.875.575.284.071.9
Ours100.088.164.543.342.572.879.085.474.977.184.171.2
honesty–base53.088.664.254.940.572.881.886.077.179.585.173.0
SFTfull45.788.762.349.441.171.078.584.672.746.976.367.1
LoRA53.788.364.352.440.972.881.885.676.879.085.072.7
P2D53.088.664.852.440.872.581.985.276.578.585.572.7
Ours55.588.564.854.940.972.781.785.877.179.084.973.0
DPOfull60.488.360.957.344.472.880.379.470.947.578.368.0
LoRA57.988.364.954.941.273.081.385.877.279.285.573.1
P2D64.088.564.856.742.673.081.586.076.778.585.873.4
Ours67.189.164.058.540.573.082.586.275.579.084.973.3
Reviewer_2y7tOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The first part of this work discusses how the functions of different attention heads evolve during post-training, demonstrating that: 1. the top-k heads during post-training have a large overlap with the top-k heads of the initial model (top-k heads means the k most capability-associated heads). More than 60% of the top-k heads at each checkpoint also belong to the initial top-k, and around 75~90% fall within the initial top-2k. 2. the changes during post-training concentrate on the initial top-k heads rather than being spread uniformly over all heads. The changes here are measured by the difference of the importance score ∣ΔI∣|\Delta I|, and the concentration is quantified by an AUROC that stays consistently above 0.5. SFT generally strengthens these heads, while unlearning tends to weaken them. 3. when post-training on one capability, the heads related to other irrelevant capabilities are largely unchanged. The second part of this work focuses on a new method named Function-Targeted Adaptation (FTA). Before post-training, one can first find the heads most associated with the specific post-training target. Then, one unfreezes only the parameters of the QKVO matrices of these chosen heads and freezes the rest of the parameters, thereby fine-tuning on the desired target. This method forms an interesting comparison with LoRA. While LoRA constrains the rank of the update, FTA constrains the physical position of the weights to be updated. The third part of this work implements a series of comprehensive experiments on two model families with four models: Llama-3.1-8B, Llama-3.2-3B, OLMo-2-13B, and OLMo-2-7B. The experiments cover two types of tasks, namely skill learning (math, honesty, harmlessness; under both SFT and DPO) and knowledge unlearning (TOFU forget05 with GradDiff). The results show that FTA achieves equal or even better target performance compared to full fine-tuning, without significantly diminishing the non-target capabilities that are usually prominently damaged by full fine-tuning. Moreover, in the unlearning task, the forget-set ROUGE-L decreases to zero while compromising general capabilities much less than full fine-tuning does. Last but not least, in the ablation, FTA performs better than randomly chosen heads in most setups, with the largest gains in math SFT and honesty. Non-target capabilities are also better preserved in most settings.

Critical strengths

First, I think the idea of this paper is simple and clear, and it connects interpretability with realistic applications. After finding that some heads are responsible for specific capabilities, the authors come up with the idea of FTA, proving that this method actually works in real LLM post-training. By choosing some specific heads, one can increase the desired capability of the model and can also suppress specific knowledge (unlearning). Second, I think this work does a good job in the comparison experiments. It has the random-head ablation. It has LoRA fine-tuning with a comparable number of parameters. It also has the ablation on the ratio of chosen heads.

Critical weaknesses

First of all, I think the most serious problem is the potential circular reasoning. If my understanding is correct, the figure 3 in the appendix is the evidence that after the post-training, the importance changes are focus on the original top-k heads (original means before post-training). However, if my understanding is correct, this is the result of FTA, which actually only trains on the original top-k heads. Therefore, it's natural that the changes are highly biased on these heads. If there is any misunderstanding, please correct me. Second, there seems to be some issue with the data. To be specific, Table 8 (p.19) Llama-3.2-3B LoRA SFT target row is identical with Table 1 Llama-3.1-8B LoRA row, which is kind of specious. Another example is the average of Llama-3.1-8B math SFT LoRA (r=8) target. The number in Table1 (78.4) is different from Table 5 and Table 17(44.2). I think they should be the same experiment thus these numbers should be identical. This discrepancy is important since this has the impact on the statement "FTA is better than LoRA".

Questions for authors
  1. Can you confirm whether Figure 2 is from full FT?
  2. Why do you use the data in Figure 3 (which is from FTA) to support the statement in Section 3?
  3. Can you also explain the data issues in the critical weakness?
4 comments
AuthorsOpenReview ↗

Response to Reviewer

Dear Reviewer,

Thank you for your thoughtful review and constructive feedback. We sincerely appreciate the time and effort you have devoted to evaluating our work. Please find our point-by-point responses to your concerns below.

Q. Can you confirm whether Figure 2 is from full FT?

Yes, we confirm that Figure 2 is based on full fine-tuning. We first investigate how the functional organization of LLMs evolves during full-parameter post-training and identify key patterns of post-training plasticity. These empirical observations subsequently motivate the development of our proposed Function-Targeted Adaptation (FTA) method in Section 4.

We will explicitly clarify the use of full-parameter fine-tuning in Section 3.2 in the revised version.

Q. About figure 3 in the appendix: post-training plasticity.

Thank you for pointing this out. We apologize for the confusion caused by mistakenly including an incorrect figure in the appendix. We have replaced it with the full fine-tuning (Full FT) results, presented as Figure 1 in the Rebuttal PDF.

We clarify that our post-training plasticity analysis is based on Full FT rather than FTA, where all model parameters are updated. The corrected figure further supports our observation of non-uniform functional plasticity. We will ensure that the corrected figure and its description are included in the revised manuscript.

Q. Data issue: Table 8 (p.19) Llama-3.2-3B LoRA SFT target row is identical with Table 1 Llama-3.1-8B LoRA row, which is kind of specious. Another example is the average of Llama-3.1-8B math SFT LoRA (r=8) target. The number in Table1 (78.4) is different from Table 5 and Table 17(44.2).

Thank you for identifying these inconsistencies. For Llama-3.1-8B under Math SFT, the LoRA (r=8) results are 81.3 (GSM8K), 84.9 (SVAMP), 81.5 (ASDiv), and 65.8 (GSM+), yielding an average of 78.4, as reported in Table 1.

The corrected results for Llama-3.2-3B LoRA (r=8) are presented in the table below. FTA achieves the highest performance across all four benchmarks, with an average accuracy of 72.2%, outperforming full fine-tuning and LoRA. We will correct these reporting errors in Appendix and ensure consistency across all tables in the revised manuscript.

Table: Corrected Math SFT results on Llama-3.2-3B, including the updated LoRA (r=8) results.

MethodGSM8KSVAMPASDivGSM+Avg.
Base53.964.465.051.058.6
Full FT65.083.080.147.568.9
LoRA (r=8)44.172.673.541.357.9
FTA (Ours)69.184.280.355.372.2

Dear Reviewer,

Thank you again for your thoughtful comments. We have provided detailed responses and clarifications to address your concerns. We sincerely hope our explanations help resolve your questions.

We would greatly appreciate any further feedback and would be happy to discuss any remaining concerns.

Best regards,

The Authors

Dear Authors,

Thank you for your response and clarification. However, I am even more confused after seeing your corrected results, so let me focus on one case: Llama-3.1-8B, math SFT, LoRA (r=8).

Your response says that the target average is 78.4, as reported in Table 1. However, Table 17 (and Table 5) report 44.2 for the same setting. If 78.4 is the correct value, then Table 17 now reads: FTA 81.0, and LoRA 78.4, 49.6, and 69.2 for r=8, r=16, and r=102, respectively. This non-monotonic behaviour of LoRA really confuses me. The data now suggest that fine-tuning a small number of parameters (LoRA r=8) performs similarly to FTA, and that when LoRA uses more parameters, the performance first drops significantly and then increases, which makes little sense. I think you may want to double-check the data for LoRA (r=16), where I strongly suspect another mistake. Please also confirm that the r=16 and r=102 runs use the same training and evaluation configuration as the r=8 run that gives 78.4.

So far, this rebuttal has been spent mostly on resolving data problems: copy-paste errors, inconsistencies, etc. I have to say that the paper contains so many tables and numbers that they overwhelm me, and, it seems, even the authors themselves. I understand that this is a precaution against reviewers who may say, “oh, your tests are too limited, please show that your method is still effective for more setups/LLMs.” But it really makes it hard for a reader to follow the important threads beneath all the numbers. I think a more pragmatic way is to use one simple and sharp setup to show the strongest result, so that the reader is easily convinced by your method and claims, and to put all of the supporting data, including the results on multiple models and the data less relevant to the core argument, in the appendix. (P.S. This is just a general suggestion; you don’t need to respond to this or reorganise your paper at this stage.)

I believe my concern is not resolvable within the remaining rebuttal time, and thus I keep my score.

Dear Reviewer,

Thank you for your further feedback and constructive suggestions regarding the presentation. We sincerely apologize for the confusion caused by the inconsistent results. Our previous response did not clearly explain the source of 44.2.

Both values correspond to the same setting (Llama-3.1-8B, math SFT, LoRA r=8), differing in the learning rate. At 10−510^{-5}, LoRA exhibits limited adaptation, achieving a target average of 44.2, whereas at 5×10−55\times10^{-5}, it achieves substantially better adaptation. In the submitted version, we updated Table 1 but inadvertently retained the 10−510^{-5} results for all LoRA ranks in Tables 5 and 17. Although we had identified this inconsistency, we did not explicitly clarify it in our previous response in an effort to keep the explanation concise. We acknowledge that this omission caused further confusion. We have since retrained LoRA with r=16 and r=102 at 5×10−55\times10^{-5}, using the same training data, optimizer, and training schedule as the r=8 setting. The results in the following table shows that Increasing LoRA's rank improves target performance, but non-target performance declines. None of the evaluated adaptation baselines surpasses the proposed FTA. We will update the corresponding results and clarify the experimental configurations in the revised manuscript.

Table. LoRA rank ablation on Llama-3.1-8B-Instruct under math SFT.

MethodTrainable paramsTarget AvgNon-Target Avg
base–80.467.5
Ours (49 heads)87.0M81.068.3
lora (r=8)6.8M78.467.1
lora (r=16)13.6M79.266.3
lora (r=102)86.9M79.659.4

We also appreciate the reviewer's broader suggestion on presentation, and we thank the reviewer for recognizing that our intention was to demonstrate the effectiveness of FTA across diverse settings and LLMs.

We hope this clarification helps reassure the reviewer that this reporting inconsistency does not undermine the core contribution of our work.

Thank you again for your careful assessment and valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.