Research map / Architectures and efficiency / RNNs and normalization

Exploring Activation Patterns of Parameters in Language Models

AAAI 2025 · Architectures and efficiency › RNNs and normalization · approach: Layer-wise analysis of language models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LLMDcos, a gradient-based (first-order Taylor) per-parameter activation metric and cosine similarity between activation vectors, to analyze how parameter activation differs across layers and data domains in LLMs, validated via Wanda pruning and semantic similarity experiments.

Problem

Interpretability of LLMs: little understanding of which parameters are activated for inputs and how activation patterns differ across layers and across input domains.

Approach

Define parameter activation as the magnitude of the first-order Taylor term |w * dD/dw| and compare inputs via cosine similarity of concatenated activation vectors (LLMDcos), then use the findings to set layer-wise pruning ratios and to measure data relevance.

Key results

• On Llama-2-7B at 50% sparsity, the layer-reallocated Wanda(ours) lowers Wikitext2 perplexity from 6.46 to 6.38 while MMLU rises from 37.11 to 35.55 (mixed) and HellaSwag improves from 29.30 to 34.38.
• Using SIQA as the calibration set gives Llama2-7B lower Wikitext2 perplexity (6.67) than C4 (6.46 for C4 in the main table is lower; SIQA is higher) so the claim of calibration-task specialization is only partly supported by these numbers.
• LLMDcos with Llama2-7B achieves the best reported result on STS-B/SICK semantic similarity validation, with Llama2-13B relatively inferior.

Measured effects reported in the paper

Surprising findings

• Pruning with layer-wise sparsity reallocation improved perplexity but decreased BoolQ accuracy for Llama2-7B and Llama-7B.
• Llama2-13B results deviated from the 7B trends on semantic similarity validation, which the authors attribute to capabilities being more densely represented in larger models.

Limitations

• Limited GPU memory (eight A40s) prevented experiments on Llama-70B.
• The activation metric is a rough first-order approximation with no formal theoretical proof.
• Pruning-based validation shows mixed results across benchmarks, with declines on BoolQ.

Experimental setup

Datasets: C4 BoolQ SIQA HumanEval MMLU Wikitext2 PIQA HellaSwag GSM8K STS-B SICK

Benchmarks: Wikitext2 BoolQ SIQA PIQA HellaSwag MMLU STS-B SICK

Metrics: perplexity accuracy correlation with semantic similarity

Baselines: Wanda

Builds on: Wanda Taylor expansion pruning

Closest papers by meaning

• Sparsing Law: Towards Large Language Models with Greater Activation Sparsity (ICML 2025 · similarity 0.91)
• Evaluating the Zero-shot Robustness of Instruction-tuned Language Models (ICLR 2024 · similarity 0.89)
• Do Language Models Use Their Depth Efficiently? (NeurIPS 2025 · similarity 0.88)
• Adaptive Layer Sparsity for Large Language Models via Activation Correlation Assessment (NeurIPS 2024 · similarity 0.88)
• LayerIF: Estimating Layer Quality for Large Language Models using Influence Functions (NeurIPS 2025 · similarity 0.88)
• Activation-Informed Merging of Large Language Models (NeurIPS 2025 · similarity 0.88)