Pruning via Causal Attribution Preserves Reasoning Performance in Large Language Models

summary

The gist

Large language models excel at multi-step reasoning but incur substantial inference cost; CAP introduces Causal Attribution Pruning (CAP)…

In short

Causal Attribution Pruning (CAP) is a training-free method to prune large language models by identifying critical attention heads. CAP measures each head's causal impact on reasoning tasks using a small calibration set, converting these scores into weight importance values. This guides fine-grained pruning that preserves reasoning performance better than traditional methods at moderate sparsity levels.

Key concepts

Causal Attribution Score
This score measures how much the model's performance (specifically token-level cross-entropy) increases when a specific attention head is temporarily masked during inference. It directly quantifies the functional contribution of that head to solving reasoning problems.
Head-Level Causal Attribution
Instead of using simple correlation, CAP calculates the expected performance degradation caused by masking individual attention heads on a small set of reasoning problems. This allows the method to identify which specific heads are functionally critical for complex tasks like math or commonsense reasoning.
Weight-Level Importance Values
The paper converts the calculated head-level causal scores into importance values for individual weights in the projection matrices. These values determine which scalar parameters should be pruned during inference, linking the functional attention mechanism to the actual model weights.

Terminology used across episodes

This episode discusses

The paper

Pruning via Causal Attribution Preserves Reasoning Performance in Large Language Models · Read on arXiv

Amogh Sheth, Andrew Lin, Biruk Assefa, Yi Wen Huang, Yuhao Ge

Edison Academy Magnet School · Massachusetts Institute of Technology · State University of New York College at Plattsburgh · The University of Texas at Austin

Large language models (LLMs) excel at multi-step reasoning but incur substantial inference cost. We introduce Causal Attribution Pruning (CAP), a training-free method that identifies critical attention heads by measuring their causal impact on reasoning tasks and uses these head-level scores to guide fine-grained weight pruning. For each attention head, CAP estimates the expected performance degradation when the head is masked during forward passes on a small calibration set of reasoning problems. These causal scores are then converted into weight-level importance values for the corresponding projection matrices. Unlike magnitude-only or activation-based criteria, CAP's interventional measurement directly captures each head's functional contribution, yielding relative accuracy gains of up to 61% over Wanda on ARC-Challenge at 20% sparsity. We evaluate CAP on GSM8K, StrategyQA, and ARC-Challenge using Llama-3-8B-Instruct and Mistral-7B-Instruct at 10%, 20%, and 50% sparsity. At moderate sparsity (10-20%), CAP improves over Wanda in most model-benchmark configurations. with especially large gains on ARC-Challenge for Llama-3. Our results suggest that attention-head-level causal attribution can better preserve reasoning performance on downstream benchmarks than correlational pruning criteria at equivalent sparsity, while remaining limited by coarse MLP attribution at 50% sparsity.

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Pruning via Causal Attribution Preserves Reasoning Performance in Large Language Models".

Jane: Here's the gist of the paper. Large language models excel at multi-step reasoning but incur substantial inference cost; CAP introduces Causal Attribution Pruning (CAP), a trainingfree method that identifies critical attention heads by measuring their causal impact on reasoning tasks and uses these head-level scores to guide fine-grained weight pruning Pruning redundant reasoning tokens The core idea of CAP is to measure the causal contribution of each attention head to reasoning performance and use these measurements to guide weight-level pruning at inference time.

Tom: By estimating the expected performance drop when individual heads are masked on a small calibration set, CAP moves beyond correlational proxies to quantify which heads are functionally critical for downstream tasks.

Jane: This method formalizes attention-head-level causal attribution for pruning by measuring the expected increase in token-level cross-entropy when a head is temporarily masked during inference.

Tom: How the training signal is built The calibration dataset Dcal is constructed by sampling from three reasoning benchmarks: GSM8K (grade-school math word problems), StrategyQA (multi-hop commonsense reasoning), and ARC-Challenge (science questions requiring world knowledge and inference).

Paper summary: Tom: Let’s go over what this "Pruning via Causal Attribution Preserves Reasoning Performance in Large Language Models" paper is actually claiming. It introduces CAP, which is training-free. The thesis is that you can find critical attention heads by measuring their causal impact on reasoning tasks and use those scores to guide fine-grained weight pruning at inference time.

Jane: So the core idea here is moving away from just using simple metrics like weight magnitude or activation-based criteria, which are correlational proxies. CAP says we need an interventional measurement—we measure what happens when we temporarily mask a head and see the expected performance drop on a small calibration set of reasoning problems.

Lu: To formalize that, they define the causal attribution score for an attention head h in layer l as the expected increase in token-level cross-entropy when h is temporarily masked during inference, which they denote as delta(hl) equals E(x,y)∼DcalhL f mask(hl)(x), y − L(f(x), y)i.

Meng: Wait, so for every single head in every layer, they have to run a simulation or estimation process on a calibration set just to get that score? That sounds like it could be slow.

Tom: They partition the calibration data Dcal into K disjoint subsets and compute K independent estimates, then take the median and subtract a multi-run intact baseline to get delta(b hl) equals mediandelta k(hl)Kk=one − delta b base <ref:2606.19350#pg1>. It’s an effort to reduce the noise from having just one small sample.

Jane: And once they have those head-level causal scores, they convert them into weight-level importance values by multiplying the causal scores with weight magnitudes, which creates a metric that is both causally informed and parameter-specific.

Lu: They then use a shift-and-scale transformation to map those raw scores into a bounded range, say from one point zero to ten point zero, which helps protect high-impact heads while marking low-importance weights as candidates for removal.

Meng: So they’re basically creating this importance factor I(hl) that scales the weight magnitudes Wq Wk Wv Wo and MLP weights, so you prune based on Sij equals Wij multiplied by that I(hregion(i,j)). That sounds like a very detailed pruning metric.

Tom: And they show that this results in relative accuracy gains of up to sixty-one percent over Wanda on ARC-Challenge when they use twenty percent sparsity. That’s the big number we usually look for when talking about performance preservation.

Jane: It also mentions that they evaluate CAP on GSM8K, StrategyQA, and ARC-Challenge using models like Llama-three-8B and Mistral-7B at different sparsity levels, showing improvements over Wanda in most configurations within the ten to twenty percent range.

Lu: The authors are also careful to state that they emphasize that CAP’s causal scoring operates at the attention head level, and MLP layers receive an approximate per-layer importance derived from their co-located attention heads, which is where they acknowledge the scope of the attribution.

Meng: So if I’m looking at deploying this, what do I need to worry about? Is it going to break something else in the model structure because you’re only focusing on attention heads?

Tom: The limitation they point out is that CAP exhibits limitations at high sparsity, specifically when coarse MLP treatment can cause catastrophic failure. They also note preliminary MoE experiments suggest weaker transfer to dynamically routed architectures.

Conclusion: Jane: So, to wrap up the discussion on this paper, it seems the authors are really pointing toward a specific sweet spot for this technique. CAP is most effective at preserving reasoning-benchmark performance under moderate compression, specifically in the ten to twenty percent sparsity range.

Tom: Right. The paper is titled "Pruning via Causal Attribution Preserves Reasoning Performance in Large Language Models," and it argues that by focusing on causal attribution of attention heads, you get a method that is both interventionally grounded and parameter-specific for pruning at inference time.

Lu: The implication is that instead of using general pruning criteria, we can target the parts of the model responsible for complex reasoning skills like math or commonsense knowledge.

Meng: For us on the engineering side, it means if you’re aiming to save compute while keeping critical reasoning intact, CAP might be a solid starting point for testing before moving into those higher sparsity regimes.

Tom: Exactly. It’s not about making the whole model smaller or faster in a vacuum; it’s about preserving specific capabilities that matter for downstream tasks like complex QA or math problems.

Jane: So, the big picture here is that this work suggests attention-head-level causal attribution can better preserve reasoning performance on those downstream benchmarks than correlational pruning criteria at equivalent sparsity levels.

More episodes

← Home