- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper mainly focuses on the formation mechanism of massive activations in LLMs. Specifically, this work identified a spatially fixed and direction-specific MAGC within the input to early-layer spike FFNs. When the value of this channel is sufficiently large or small, the SwiGLU FFN output asymptotically approaches a quadratic weighted combination of a few columns of the corresponding matrix. Furthermore, this work also validates this mechanism through random inputs, single-channel interventions, and theoretical derivations across six LLMs of varying scales and model families, further demonstrating that modulating the MAGC can significantly alter attention sinks.
Critical strengths
-
This work presents the MAGC mechanism and locates massive activations from empirical phenomena to parameter-level causal chains between single input channels, SwiGLU gating, and a handful of down-projection columns.
-
The theoretical analysis and experiments are well combined.
-
Consistent structures are observed on models with 1B-32B parameters. Besides, MAGC demonstrates potential application value via the attention sink intervention.
Critical weaknesses
-
The most comprehensive MAGC substitution experiments, coefficient concentration analyses, and output collinearity analyses focus primarily on LLaMA-3.2-1B. For the other models, only visualizations based on a limited number of random sequences are provided, without systematic reporting of intervention success rates, effect sizes, variance, or cross-dataset stability. As a result, the existing evidence is insufficient to support the central conclusion that MAGC is a universal mechanism across different model families.
-
This work locates the MAGC by analyzing the difference in average FFN inputs between the first and second tokens and identifies spike FFNs using a manually set threshold. However, it fails to demonstrate the robustness of these results regarding token selection, prompts, random seeds, or threshold values. Especially in the case of QwQ-32B, the authors had to skip several channels exhibiting the largest differences and manually select a channel that was not a massive-activation channel. This indicates that the identification rule does not consistently apply across all models and undermines the claim that a unique, fixed MAGC exists for every model.
-
Theorem 4.1 describes the general asymptotic behavior of an arbitrary input channel as . However, MAGC values are finite in practice, e.g., approximately 8.8 for LLaMA-3.2-1B. The paper does not provide error bounds, approximation accuracy, or trigger thresholds for finite values of , nor does it theoretically derive why the coefficients associated with MAGC are so highly concentrated. This key property is primarily established through empirical observation.
Questions for authors
Please refer to the Critical Weaknesses part.
2 comments
Thank you for the response. I have decided to maintain my rating of 2. The following concerns still affect my assessment.
-
The cross-model intervention results still lack per-model sample counts, success rates, effect sizes, and variability across inputs. Consequently, it is difficult to assess how reliably the intervention induces massive activations across the evaluated models. The columns reported in Table 1 are selected partly for their alignment with the observed output, which also limits the value of the appendix plots as independent evidence for the proposed mechanism.
-
Identifying the MAGC in QwQ-32B still requires an additional filtering step among the ten highest-ranked channels. The sensitivity of the identification procedure to token choice and the spike-FFN thresholds remains unclear. In the Qwen experiment, the single-channel intervention failed to suppress existing massive activations. This leaves open whether a single fixed channel can account for naturally occurring massive activations across the evaluated models.
-
At natural MAGC values, the cosine similarities for QwQ-32B and Ministral-3-3B are 0.642 and 0.499, respectively, indicating limited directional agreement in these two models. Cosine similarity is insensitive to the scaling factor, so it leaves the accuracy of the predicted output magnitude untested, including on the dimensions that exhibit massive activations. The finite range over which the quadratic approximation accurately reproduces those activations therefore remains unclear.
Response to Reviewer KFDu
Thank you for your comments on the generalizability of MAGC on other models. The coefficient concentration analyses and output collinearity analyses are included in the corresponding subfigures for each model. For example, Figure A6(c) shows that the coefficient vector (α) is concentrated, while Figures A6(d)–(f) show the collinearity between the columns of the spike FFN and its output.
To evaluate the generalizability of the MAGC substitution experiments, we conduct the same experiments across different models. The results show that adjusting the MAGC of non-spike tokens in the corresponding gating direction consistently induces massive activations.
Thank you for raising these important questions. We have further examined the stability of our MAGC identification procedure across different samples.
To improve the robustness of identification, we now calculate the channel statistics by averaging over 100 samples, rather than relying on a single sample. We find that, for all models except QwQ-32B, the top-1 channel identified by the argmax criterion is exactly the same across all 100 samples (100/100). This provides strong evidence that the identified MAGC is stable across different input samples for these small models. We acknowledge that this identification procedure is less robust for the larger QwQ-32B model. This limitation suggests that identifying MAGC in larger models may require further investigation, and we will clarify this point in the revised manuscript.
We also acknowledge that our experiments cannot rule out the possibility that multiple input channels jointly contribute to massive activations. Empirically, the results on the focal LLaMA-1B model provide evidence for the uniqueness of MAGC, as manipulating the identified channel can both induce and suppress MAs. However, for Qwen2.5B, manipulating a single channel can induce MAs in non-spike tokens, but cannot suppress MAs in existing spike tokens. This suggests that multiple input channels may contribute to the formation or maintenance of MAs.
Theoretically, if two input channels have similar or aligned FFN parameter structures, such that activating either channel leads to a concentration of α, both channels could potentially induce the same MA phenomenon. Therefore, our findings do not establish that MAGC is the only input channel capable of producing MAs. Rather, our experiments demonstrate that at least one identifiable input channel, i.e., the MAGC, can control the occurrence of MAs.
Thank you for your comment. As suggested, we compared the actual spike-FFN output with the Theorem 4.1 approximation x2y∗, for all six models.
Specifically, we feed 20 natural English sentences (without BOS token) to each model. For every spike token (the first token), we take its real input h to the spike FFN and its real output y, and let x=h(i∗) be the natural MAGC value. We compute the cosine similarity between y and the Theorem 4.1 approximation $x^2 y^* $, where $y^* $ is computed from the weights alone.
Table. Cosine similarity between the real output y and the Theorem 4.1 approximation x2y∗ (for non-spike tokens, we compute the mean of all non-spike tokens).
In four of the six models (LLaMA-3.2-1B, LLaMA-2-7B, Qwen2.5-1.5B, and DeepSeek-R1-8B), the direction predicted by Theorem 4.1 from the pretrained weights closely matches the actual spike-FFN output at natural MAGC values, with cosine similarities above 0.98. In contrast, the outputs of non-spike tokens are nearly orthogonal to the predicted direction. For QwQ-32B and Ministral-3-3B, the cosine similarities are relatively lower (0.64 and 0.50, respectively), but remain substantially higher than those of non-spike tokens (0.14 and 0.13).
These results suggest that the asymptotic direction captures a component of the actual spike-FFN output, which is consistent with the asymptotic nature of Theorem 4.1.