Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.37157

What Survives Distillation? A Mechanistic Interpretability Study of GPT-2 and DistilGPT-2

Nabeelah Faumi·Deshan Sumanathilaka
View original source ↗Read source document ↗Export metadata ↓

Abstract

Knowledge distillation compresses large language models into compact students that preserve aggregate metrics, yet the mechanistic consequences of depth-reduction distillation remain largely uncharacterized. This work applies linear CKA, logit lens analysis, attention pattern decomposition, sparse autoencoder feature statistics, and activation patching to compare GPT-2 and its six-layer student DistilGPT-2 across two canonical circuits: the Indirect Object Identification (IOI) circuit and the Greater-Than (GT) circuit. The results reveal a compression-collapse dichotomy: the IOI circuit suffers complete mechanistic collapse (DistilGPT-2: 35.2% accuracy vs. 94.8%, zero name-mover preservation), while the GT circuit survives with 87.7% task performance recovery through active remapping of the teacher's layer-9 comparison head into student layer~4 (mean Pearson r = 0.945). These findings demonstrate that circuit compressibility governs mechanistic survival under distillation, and that aggregate metrics are insufficient diagnostics for circuit integrity.

Explore related subjects

cs.CL

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Nabeelah Faumi, Deshan Sumanathilaka. 2026-10-27. What Survives Distillation? A Mechanistic Interpretability Study of GPT-2 and DistilGPT-2. https://arxiv.org/abs/2610.37157

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

HateSense: Hierarchical Sentiment Analysis Framework for Hate Speech Detection: Implementing Binary and Multiclass Classification Strategy

Hate speech detectors must separate hateful posts from the far more common offensive but non-hateful ones, and two popular ideas for doing so are hierarchical (coarse-to-fine) classification and sentiment features. We revisit our earlier transformer-based framework, which combines both ideas, and test each component in a controlled study on two public benchmarks (Davidson et al., 24,783 tweets; HateXplain, 19,229 posts) with five random splits, validation-based tuning, and paired bootstrap tests. With linear models over word and character n-grams, the dominant-class-first hierarchy used in our framework lowered macro-F1 by 1.1-1.5 points relative to a flat classifier (p <= 0.007), and adding VADER sentiment features changed Davidson macro-F1 by at most 0.6 points, never significantly in the class-weighted settings. Class-weighted training was the only intervention with a large effect (+2.0 to +3.5 macro-F1 and +5.9 to +9.1 hate-class F1), and sentiment cannot separate the two abusive classes because hateful and offensive tweets are almost equally negative (mean VADER compound -0.35 vs. -0.29). Cross-dataset transfer cut macro-F1 to 43-53, and a reclaimed slur doubled the false hate rate of a logistic-regression detector, showing that in-domain scores overstate real-world reliability.

cs.CL↗
arXiv2026

EffiBench-X: A Multi-Language Benchmark for Measuring Efficiency of LLM-Generated Code

Existing code generation benchmarks primarily evaluate functional correctness, with limited focus on code efficiency and often restricted to a single language like Python. To address this gap, we introduce EffiBench-X, the first multi-language benchmark designed to measure the efficiency of LLM-generated code. EffiBench-X supports Python, C++, Java, JavaScript, Ruby, and Golang. It comprises competitive programming tasks with human-expert solutions as efficiency baselines. Evaluating state-of-the-art LLMs on EffiBench-X reveals that while models generate functionally correct code, they consistently underperform human experts in efficiency. Even the most efficient LLM-generated solutions (Qwen3-32B) achieve only around \textbf{62\%} of human efficiency on average, with significant language-specific variations. LLMs show better efficiency in Python, Ruby, and JavaScript than in Java, C++, and Golang. For instance, DeepSeek-R1's Python code is significantly more efficient than its Java code. These results highlight the critical need for research into LLM optimization techniques to improve code efficiency across diverse languages. The dataset and evaluation infrastructure are submitted and available at https://github.com/EffiBench/EffiBench-X.git and https://huggingface.co/datasets/EffiBench/effibench-x.

cs.CL↗
arXiv2026

FullFront: Benchmarking MLLMs Across the Full Front-End Engineering Workflow

Front-end engineering involves a complex workflow where engineers conceptualize designs, translate them into code, and iteratively refine the implementation. While recent benchmarks primarily focus on converting visual designs to code, we present FullFront, a benchmark designed to evaluate Multimodal Large Language Models (MLLMs) \textbf{across the full front-end development pipeline}. FullFront assesses three fundamental tasks that map directly to the front-end engineering pipeline: Webpage Design (conceptualization phase), Webpage Perception QA (comprehension of visual organization and elements), and Webpage Code Generation (implementation phase). Unlike existing benchmarks that use either scraped websites with bloated code or oversimplified LLM-generated HTML, FullFront employs a novel, two-stage process to transform real-world webpages into clean, standardized HTML while maintaining diverse visual designs and avoiding copyright issues. Extensive testing of state-of-the-art MLLMs reveals significant limitations in page perception, code generation (particularly for image handling and layout), and interaction implementation. Our results quantitatively demonstrate performance disparities across models and tasks, and highlight a substantial gap between current MLLM capabilities and human expert performance in front-end engineering. The FullFront benchmark and code are available in https://github.com/Mikivishy/FullFront.

cs.CL↗