Grad-ELLM: LLM Attribution Method

• Grad-ELLM is a method that provides per-token attribution for decoder-only LLMs by integrating gradient sensitivity with spatial attention.
• It leverages layer-wise linear decomposition and channel-wise gradient weighting to efficiently compute token contributions without extensive perturbation.
• The framework introduces π-Soft-NS and π-Soft-NC metrics for a rigorous evaluation of attribution faithfulness, addressing computational and methodological challenges.

Grad-ELLM (Gradient-based Explanations for Decoder-only LLMs) is a specialized input-attribution framework designed to provide faithful and efficient explanations for the outputs of decoder-only transformer-based LLMs. Developed to address the limitations of generic attribution methods on modern autoregressive architectures, Grad-ELLM combines sensitivity analysis through gradients with internal transformer attention mechanisms to yield step-wise, per-token interpretability that is both computationally tractable and empirically robust (Huang et al., 6 Jan 2026).

1. Motivation and Limitations of Model-Agnostic Attribution

The predominance of large autoregressive LLMs has prompted demand for transparent and faithful attribution methods, specifically those assignable on a per-token, per-step basis. Conventional model-agnostic approaches—such as LIME, vanilla saliency, Integrated Gradients, and DeepLIFT—suffer critical drawbacks when applied to decoder-only transformers:

• Transformer Architectural Blindness: They neglect the role of structured components like self-attention and feed-forward layers, treating the network as a monolithic function and disregarding the compositional flow of information.
• Computational Inefficiency: Many require $\mathcal{O}(n^2)$ or more forward passes per input, as in perturbation-based schemes that mask or re-sample tokens individually, leading to intractability on long contexts.
• Non-natural Perturbation Regimes: Techniques that employ “hard” perturbations—e.g., token deletion or outright masking—may produce off-manifold inputs never encountered during model training, resulting in misleading faithfulness scores.

Grad-ELLM circumvents these issues by exploiting the operational structure of the transformer decoder, maintaining alignment with the model's native computation flow (Huang et al., 6 Jan 2026).

2. Attribution Mechanism in Grad-ELLM

The Grad-ELLM method synthesizes gradient-based channel importance and spatial attention scores at every generation step, leveraging internal transformer representations without necessitating architectural modification.

Layer-wise Linear Decomposition

At each generation step for output token $y_t$, the model computes a logit vector $\ell_t$ conditioned on all prior inputs and outputs. Grad-ELLM decomposes the logit via a first-order Taylor expansion as:

$$\ell_t \approx \sum_{k=0}^{N-1} \mathrm{LP}(o^{(k)}_t)$$

where $o^{(k)}_t$ is the output of the $k$-th attention block and $\mathrm{LP}$ denotes the final linear projection.

Channel-wise Gradient Weighting

For each layer $k$, the partial logit $\ell_t^{(k)}$ is locally linearized:

$$f_k(o_t) \approx w^{(k)} \cdot o_t, \quad \text{with}\quad w^{(k)} = \frac{\partial \ell_t}{\partial o_t^{(k)}}$$

These $w^{(k)}$ weights represent channel sensitivities, evaluating the marginal effect of each channel on the target logit.

Spatial Scoring via Attention

The attention output at each layer is itself a weighted sum over all contextual tokens. For input token $i$ in layer $k$:

$$h_i^{(k)} = \mathrm{ReLU}\left(\sum_{c=1}^d w_c^{(k)} \cdot \lambda_i^{(k)} \cdot v_{i,c}^{(k)}\right)$$

where $\lambda_i^{(k)}$ are normalized attention weights (optionally “loosened” to [0,1] to temper softmax peaking) and $v_{i,c}^{(k)}$ is the $c$-th component of the value vector for token $i$.

Aggregating $h_i^{(k)}$ over layers yields attribution heatmaps $H_i$ tracing input-token influence on each output at every generation step.

3. Faithfulness Metrics: $\pi$-Soft-NS and $\pi$-Soft-NC

Classical faithfulness evaluation in attribution employs “hard” token deletions/insertion, measuring impact on model outputs, but is vulnerable to metric inflation if retention probabilities differ across methods. Grad-ELLM generalizes “soft” perturbation metrics to ensure fair, distribution-controlled comparisons by introducing the $\pi$-Soft-NS (“sufficiency”) and $\pi$-Soft-NC (“comprehensiveness”) metrics.

• Soft Masking: Each input token $x_i$ is zeroed independently with probability $1 - \tilde{s}_i$, where the attribution score $s_i$ is transformed via an $\alpha$-exponent to enforce $\frac{1}{m}\sum_i \tilde{s}_i = \pi$ (target proportion retained).
• Evaluation: Faithfulness is measured by the normalized Hellinger distance in output distributions: a sweep across $\pi \in [0.05, 0.95]$ generates $\pi$-Soft curves, summarized via area-under-curve (AUC) for quantitative assessment.

This approach eliminates spurious performance variance due to differing mean retention rates, enabling rigorous, apples-to-apples method comparison (Huang et al., 6 Jan 2026).

4. Experimental Evaluation and Results

Experiments assessed Grad-ELLM on sentiment classification (IMDb, SST2), yes/no question answering (BoolQ), and open-generation tasks (TellMeWhy, WikiBio), utilizing LLaMA-7B and Mistral-7B as backbone models.

• Baselines: Raw attention, vanilla saliency, Input$\times$Grad, Integrated Gradients, DeepLIFT, Layer-GradCAM, Value Zeroing, and Random.
• Metrics: $\pi$-Soft-NS and $\pi$-Soft-NC AUCs, classical insertion/deletion, and qualitative heat maps.

Quantitative Highlights

Model	$\pi$-Soft-NS AUC	$\pi$-Soft-NC AUC	Best Baseline $\pi$-Soft-NS / NC
LLaMA-7B	0.401	1.115	Random (0.379/1.078); DeepLIFT/Saliency (0.339), Input$\times$Grad (0.974)
Mistral-7B	0.383	0.491	Random (0.548)

• Grad-ELLM leads on $\pi$-Soft-NS and $\pi$-Soft-NC (LLaMA), with notable qualitative clarity in attribution heatmaps (“very little positive” review in IMDb, “10/10 :)” in SST2).
• On Mistral-7B, Random dominates $\pi$-Soft-NC, suggesting that grouped-query/sliding-window attention may induce uniformity in token dependencies (Huang et al., 6 Jan 2026).

Classical insertion/deletion metrics are more sensitive to the density of attribution maps, with Grad-ELLM ranking behind sparse baselines under these tests.

5. Computational Complexity and Applicability

Grad-ELLM requires one backward pass per generation step to compute $\partial \ell_t / \partial o^{(k)}$, alongside extraction of precomputed attention maps. The resulting per-token time and space complexity is $\mathcal{O}(Nd(m+t))$, significantly lower than perturbation-based schemes. However, this method presupposes white-box access to the model’s internals and is thus not applicable to black-box or API-limited deployments.

Grad-ELLM's approach is highly amenable to implementation on open-source transformer models but not directly compatible with closed-source or proprietary systems where access to attention maps and gradients is restricted.

6. Limitations and Future Directions

Grad-ELLM’s performance degrades in scenarios demanding extremely sparse attribution (e.g., top-$k$ selection), as its denser “loosened” heatmaps may underperform on traditional insertion/deletion metrics. Remediation via thresholding or top-$p$ filtering can selectively increase map sparsity. The framework exclusively addresses faithfulness in the causal sense—effect on model outputs—and does not assess human-centric plausibility or interpretability.

Future extensions proposed include adaptation to instruction-tuned and multimodal decoder-only architectures, integration with global causal-tracing techniques, and optimization of the attention-loosening transform to balance interpretive sparsity and noise robustness.

7. Impact and Broader Significance

Grad-ELLM demonstrates that incorporating both the channel-wise gradient sensitivity and spatial attention structure of decoder-only transformers enables attribution methods that are more aligned with the model's underlying computation, outperforming both naïvely attention-based and generic model-agnostic baselines in terms of faithfulness. Its methodology refines evaluation practice for interpretability research through equitable, parameter-controlled faithfulness metrics, and establishes a new basis for explanatory techniques tailored to advanced autoregressive LLMs. As the landscape of foundation models evolves, methods like Grad-ELLM will inform future directions in mechanistic interpretability and diagnostic toolkits for black-box generative models (Huang et al., 6 Jan 2026).