Decoupling Internal Representational Changes and Causal Importance in Fine-Tuned Large Language Models
Lingfang Li, Procheta Sen, Shubham Das, Danushka Bollegala
cs.AI
Quick verdict
Well-controlled multi-model study: FT's representational change decouples from EAP causal importance; component overlap predicts cross-task interference.
Summary
This paper asks whether the layers of an LLM that change most during full fine-tuning are the same layers that causally matter for task performance. Across four models (GPT-2 Small, Qwen2-0.5B, Llama-3.2-1B, Llama-2-7B) and three task families (sentiment classification on SST-2/Yelp, extractive QA on SQuAD/CoQA, En-Fr machine translation on KDE4/Tatoeba), the authors compare layer-wise attention-pattern KL divergence and logit-lens probing before/after fine-tuning against Edge Attribution Patching (EAP) importance over the top-400 edges. They report: (1) EAP-important components concentrate in a few layers (normalized entropy 0.40-0.74 vs 0.90-1.00 for attention change); (2) the changed and important layers largely differ, so representational change magnitude is only weakly related to causal importance; (3) across 24 dissimilar task pairs per model, overlap of EAP components correlates positively with performance degradation on the second task (Spearman rho=0.5-0.7, p<=0.005), and freezing shared components during fine-tuning recovers 2.05-84.34% of the lost performance. Validation includes activation patching of top-400 versus random edges (Wilcoxon p<0.05) and corruption-token robustness (rho=0.71-0.72). Main limitations: three task families, a single language pair for MT, models up to 7B, layer-alignment evidence based mainly on entropy comparisons, and correlations over 24 points without confidence intervals or multiple-comparison control.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- EAP-identified task-relevant components are localized in a few layers (normalized entropy 0.40-0.74), unlike the much flatter distribution of attention-pattern changes during fine-tuning.
- Layers with the largest attention/representational changes during full fine-tuning do not coincide with the layers hosting the most causally important components, so the magnitude of internal change is only weakly related to causal importance.
- Across 24 dissimilar task pairs per model, overlap of EAP components between tasks correlates positively with performance degradation on the second task after fine-tuning on the first (Spearman rho=0.5-0.7, p<=0.005).
- Freezing the shared EAP components while fine-tuning on Task 1 recovers 2.05-84.34% of the lost Task 2 performance (Table 7), providing intervention-based evidence that shared components mediate interference.
- QA tasks show the highest induction-head overlap with task-important heads in three of four models, peaking at 16.66% for GPT-2 Small on CoQA (Table 1).
Strengths
Broad, controlled experimental scope
Four model families (GPT-2 Small 124M to Llama-2-7B), three task families x two datasets each, full fine-tuning with complete hyperparameters (Table 5) and trends replicated across all models in Appendix L (Figs 12-14), supporting generality of the three headline findings.
Interventional validation beyond attribution scores
Activation patching of top-400 EAP edges vs 400 random edges (Table 3, 20 seeds, Wilcoxon signed-rank p<0.05; e.g., Llama-3.2 MT KDE4: 82% vs 44%), plus the freeze-shared-components intervention (Appendix J, Table 7) that upgrades the correlational overlap-degradation claim in Sec 5.3 into causal evidence.
Layered robustness checks
Quantified localization statistic
Weaknesses
Headline decoupling claim is under-quantified
Thin statistics for the overlap-degradation correlation
Possible task-similarity confound
Task-specific corruption strategies complicate cross-task comparison
Freeze intervention coverage and reporting
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks two linked questions about full-parameter fine-tuning of decoder LLMs: (1) are the layers whose attention patterns and representations change most during fine-tuning the same layers that contain the causally important components for task performance, as identified by Edge Attribution Patching (EAP)? (2) does overlap in EAP-identified components across tasks drive cross-task performance transfer? Experiments cover four models (GPT-2 Small 124M, Qwen2-0.5B, Llama-3.2-1B, Llama-2-7B) and three task families x two datasets each (SST-2/Yelp sentiment, SQuAD/CoQA extractive QA, KDE4/Tatoeba En-Fr MT), with hyperparameters in Appendix F (about 14 A100 GPU hours). Three claims carry the paper: EAP-important components localize in few layers; changed layers and important layers diverge; and component overlap predicts second-task degradation, backed by a freeze intervention.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does 'decoupling representational change and causal importance' mean in this paper?
The authors measure how much each transformer layer's attention patterns and representations change during full fine-tuning, separately from which components (attention heads, MLPs, edges) Edge Attribution Patching identifies as causally needed for task performance. Their central finding is that these two pictures do not coincide: the layers that change most during fine-tuning are generally not the layers containing the most causally important components.
Which models and tasks were studied?
Four models - GPT-2 Small (124M), Qwen2-0.5B, Llama-3.2-1B and Llama-2-7B - each fully fine-tuned on six datasets spanning three task families: sentiment classification (SST-2, Yelp), extractive question answering (SQuAD, CoQA), and English-to-French machine translation (KDE4, Tatoeba).
Does sharing important components between tasks help transfer?
Not in their setting. For dissimilar task pairs (e.g., classification vs. generation), higher overlap of EAP-identified components correlated with larger performance drops on the second task after fine-tuning on the first (Spearman rho=0.5-0.7), and freezing the shared components during fine-tuning recovered part of the lost performance (2.05-84.34% across nine tested pairs).
Is the code and data available for this paper?
The paper states that code is available at github.com/LingfangLi/repr-change-vs-causal-importance. All datasets are public benchmarks, models are open weights, and Appendix F documents the full fine-tuning setup (about 14 GPU hours on a single A100 80GB).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure axis values were reconstructed from embedded text only; the per-layer distributions in Figures 2, 3, 11 and 12-14 could not be visually verified, so the qualitative 'profiles look different' judgments rely on the authors' description.
- Table 6 formatting in the extracted text is lossy (arrow placement, missing separators), introducing some ambiguity in the sign of individual Perf Delta entries.
- Review is based on arXiv v2 (23 Sep 2026) text extraction; Appendix references to Figure 10 (GPT-2 correlation rho=0.5) lack accompanying p-values in the extracted text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.