- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper has explored the massive activations in HLA LLMs. There are two activation patterns related to the full-attention layer placement, i.e., PAS and ISP. The training analysis shows that PAS and ISP emerge early in pre-training, and the output gating reduces their magnitude but does not eliminate their layer-wise structure. Furthermore, the paper also introduces a mechanism where the model writes anomalous activations that turn the corresponding token into an attention sink. Experiments show that removing the four largest PAS coordinates from each full attention input significantly degrades average task accuracy.
Critical strengths
-
This paper reveals two architectural coordinates in HLA LLMs, i.e., PAS and ISP. These two coordinates are related to the massive activations in HLA LLMs.
-
The experiments cover five linear attention architectures, and the observations are conducted based on open-source models including Kimi Linear, Qwen, Nemotron, and Zamba.
-
The paper presents an interesting observation, i.e., removing a few PAS coordinates is harmful to the performance of the LLM. The observations provide useful evidence for the functional significance of these coordinates.
Critical weaknesses
-
The write-sink-cancel mechanism is primarily validated using fixed coordinates within a representative GDN model. The latency cancellation observed in ISPs is largely inferred from the cross-layer phenomena. Because the maximum feature is re-selected at each layer in some analyses, it cannot be confirmed that the exact same information or coordinate is being propagated throughout.
-
Although the study covers a range of models from 1.2B to 397B parameters, the primary evidence consists of a small number of representative inputs and trajectory plots for the first token. It lacks the comprehensive statistics, confidence intervals, and token-level comparisons found in the M-A-P model suite.
-
The experiments lacked sufficient controls, such as magnitude-matched non-PAS coordinates, other tokens, or random coordinates. As a result, the performance drop might stem partly from the removal of extreme coordinates themselves instead of being specific to the PAS structure.
-
The proposed method requires an initial standard forward pass to obtain a reference trajectory and save intermediate activations. However, since it simultaneously modifies these activations and restores endpoints, it is difficult to pinpoint exactly which operation yields the benefits. Given the marginal overall improvement and performance degradation observed in some tasks, it cannot yet be considered a retrieval-augmented method ready for direct deployment.
Questions for authors
Please refer to the Critical Weaknesses part. I am glad to raise my score if my concerns are well addressed.
3 comments
Response to Reviewer A5hk (Part 2/2)
We continue with the suppression controls and connection ablations.
Q3. Do magnitude-matched or random non-PAS deletions produce the same loss?
We examine whether the observed degradation is specific to PAS coordinates. In the same M-A-P hybrids, even the four largest eligible non-PAS coordinates fail to match the PAS deletion L2 magnitude within ±5% (0/4,096 FA events per model). Thus, a magnitude-matched non-PAS four-coordinate set is unavailable at the tested sites. We compare zeroing the first token's four largest-magnitude coordinates at each FA input, replacing them with their channel means over the unedited prompt, and zeroing four random coordinates outside the local top four. Recall is mean accuracy over FDA, SWDE, Based SQuAD, and Based NQ; commonsense is mean accuracy over PIQA and ARC-Easy (equal task weights). PPL is WikiText-103 continuation perplexity (lower is better).
| Model / metric | Normal | Top4 zero | Mean fill | Random4 zero |
|---|---|---|---|---|
| 340M Recall mean (%) | 28.91 | 4.49 | 21.88 | 29.30 |
| 340M Commonsense mean (%) | 62.85 | 58.33 | 59.05 | 62.72 |
| 340M PPL | 14.61 | 51.97 | 19.96 | 14.61 |
| 1.3B Recall mean (%) | 47.07 | 28.91 | 21.88 | 47.07 |
| 1.3B Commonsense mean (%) | 72.80 | 57.92 | 65.22 | 72.80 |
| 1.3B PPL | 9.50 | 14.73 | 14.88 | 9.50 |
Mean replacement also causes substantial degradation without zeroing: recall decreases by 7.03/25.20 points, commonsense accuracy drops, and PPL increases. Random-four deletion, by contrast, remains close to normal performance. These results provide additional evidence for the importance of PAS, as discussed in §4.3 and Appendices F.4–F.6.
Q4. Which part of connection improves retrieval?
We separately evaluate reference-floor editing between FA inputs, endpoint restoration, and their combination. The metrics are as described above.
| Model / metric | Normal | Interior only | Endpoints only | Combined |
|---|---|---|---|---|
| 340M Recall mean (%) | 28.91 | 33.98 | 28.91 | 33.98 |
| 340M Commonsense mean (%) | 62.85 | 62.70 | 62.85 | 63.00 |
| 340M PPL | 14.61 | 14.60 | 14.61 | 14.60 |
| 1.3B Recall mean (%) | 47.07 | 50.59 | 47.07 | 49.80 |
| 1.3B Commonsense mean (%) | 72.80 | 72.72 | 72.80 | 72.82 |
| 1.3B PPL | 9.50 | 9.49 | 9.50 | 9.50 |
Interior editing provides the recall gains (approximately +5.1/+3.5 percentage points), with little change in commonsense accuracy and PPL. Endpoint restoration alone reproduces normal performance, while adding restoration provides no further mean recall benefit. We therefore focus on interior-only editing in §4.3 and Appendices F.1/F.4. Extending these benefits to broader tasks and settings, while further optimizing activation editing, remains an important direction toward practical deployment.
Thank you for the detailed response. The new cross-model and token-level statistics address my concern about the generality of PAS and ISP, and the ablation clarifies which edits drive the average retrieval gains in the two GDN models tested. Therefore, I have decided to raise my score to 3.
In a future revision, I encourage the authors to strengthen the evidence for delayed cancellation along fixed ISP coordinates across architectures, better isolate PAS-specific effects from perturbation magnitude, and report end-to-end inference costs including the reference forward pass.
Response to Reviewer A5hk (Part 1/2)
We thank the reviewer for the constructive questions, which help us strengthen the evidence for PAS, ISP, and the proposed write–sink–cancel process in hybrid models.
Q1. Does ISP follow the same coordinates across layers?
To distinguish coordinate persistence from a changing maximum, we track sink coordinates in the M-A-P GDN 340M and 1.3B hybrids (both 3:1). Top-1 identity measures whether adjacent residual boundaries share the largest absolute coordinate; top-4 overlap measures the fraction shared among their four largest coordinates. Anchor membership measures how often the left FA input's top coordinate remains among the top four later in the interval. The anchor is fixed once per interval, even when the maximum shifts. The natural inputs exhibit a common leading sink trajectory in these two models.
The top-1 coordinate is usually preserved, with high top-four overlap and anchor membership. These results indicate that ISP typically involves a persistent dominant coordinate or a stable set of dominant coordinates, rather than merely a changing maximum. This is consistent with our fixed-coordinate PAS observations in Appendix E.1 (Figs. 13–15). We add the ISP analysis there, providing further support for the write–sink–cancel account.
Q2. Do PAS and ISP generalize across inputs and tokens in the 12 large models?
Due to space constraints, the submission primarily presented representative activation trajectories rather than comprehensive statistics for all 12 models. We now quantify PAS and ISP across five natural domains, comparing attention-sink tokens with two non-sink tokens at early and middle positions, both outside the attention top three. SSA measures pre-FA peak alignment; peak excess measures local spike prominence. ISR measures retention relative to the weaker endpoint, while absolute height measures raw plateau magnitude. The table reports five-domain averages for Qwen397B, with 95% source-cluster intervals in brackets.
Sink tokens exhibit pronounced PAS and ISP patterns across the 12 models, with absolute plateau heights exceeding both controls in every model. Non-sink trajectories generally show weaker spikes; their low, relatively flat activation curves can nevertheless yield high relative ISR despite low absolute height. These findings are consistent with the token-level observations in M-A-P. §3.4 and Appendix C.3 provide the full comparisons and intervals for all 12 models.