Which the Eye Fears: Writing with Read-Blindness Explains Massive Activations in Transformers

Swagatam Mukhopadhyay; Vishal Vivek Saley; Vraj Parikh; Mausam

cs.CLSep 28, 2026

Which the Eye Fears: Writing with Read-Blindness Explains Massive Activations in Transformers

Authors: Swagatam Mukhopadhyay, Vishal Vivek Saley, Vraj Parikh, Mausam

Organizations: PsiDagger · IIT Delhi

Abstract

Massive activation features (MAs) in Transformers are extreme-value residual-stream features that persist across layers despite the model's ability to suppress them. Why do they survive? Our investigation using an operator-level mechanistic analysis of attention and feed-forward (FFN) blocks reveals that these blocks systematically ignore MA coordinates while reading, but not while writing; creating a read-write asymmetry that blocks corrective feedback while allowing continued accumulation. We find that both attention and feed-forward layers have this read-blindness, and contribute to the emergence and persistence of MAs. To validate prior work that hypothesized that FFN's amplification abilities is the primary reason for MAs (Sun et al., 2026), we analyze the model checkpoints during learning. Contrary to our expectation, read-blindness emerges before FFN amplification, suggesting that it acts upstream in the MA mechanism. We further contribute gradient analysis to link this behavior to surprising asymmetries in the loss landscape, concluding that the model actively maintains this read-blindness. Finally, we find that removing read-blocking at different locations induces compensatory shifts elsewhere, but MAs still persist.

Figures & tables

	Llama 135M	Llama 1.28B	Llama 2.56B	Qwen3 1.7B
Massive coordinates $\|\mathcal{M}\|$	10	10	5	7
Maximum activation magnitude	280	$2.27\!\times\!10^{3}$	$3.92\!\times\!10^{3}$	$2.23\!\times\!10^{3}$
Input availability:
Attention normalized input, $\tilde{X}$	0.58 (-0.60)	0.68 (-0.73)	0.55 (-0.64)	0.58 (-0.37)
FFN normalized input, $\tilde{Y}$	0.56 (-0.65)	0.56 (-0.13)	0.50 (-0.63)	0.54 (-0.48)
Read side weight view:

Table 1: Erasure results. Cells report the median

{\mathrm{era}}_{\mathcal{M}}

with rank-biserial

r_{b}

against the full complement

\neg\mathcal{M}

in parentheses. Bold cells have

r_{b}>0.95

; stars denote one-sided Mann–Whitney tests of

{\mathrm{era}}_{\mathcal{M}}>{\mathrm{era}}_{\neg\mathcal{M}}

with

p<0.05

Signed alignment	Llama 135M	Llama 1.28B	Llama 2.56B	Qwen3 1.7B
Attention, $X_{k}$ with $\Delta_{k}^{\mathrm{attn}}$	+0.03 (+0.45)	+0.09 (+0.29)	+0.02 (-0.27)	+0.06 (+0.16)
FFN, $Y_{k}$ with $\Delta_{k}^{\mathrm{ffn}}$	+0.05 (+0.57)	+0.25 (+0.59)	-0.02 (-0.20)	+0.10 (+0.14)

Table 2: Signed-alignment results. For each coordinate,

S

is median-aggregated across layers. Cells report the median

S_{\mathcal{M}}

with rank-biserial

r_{b}

comparing

\mathcal{M}

and

\neg\mathcal{M}

in parentheses.

Diagnostic	value $\mathcal{M}$	$r_{b}$
$\lVert U_{k}\rVert_{F}$	7.74	+0.78
$\lvert\lambda_{\star}^{(k)}\rvert$	2.19	+0.98
$\lvert\lambda_{\star}^{(k)}\rvert/\lVert S_{k}\rVert_{F}$	0.39	+1.00
Within-set $\lvert\cos(s_{\star})\rvert$	0.53/0.46	—

Table 3:

s_{\star}

amplifier diagnostics for Llama 1.28 Base. Per-coordinate entries are medians over

\mathcal{M}

and a deterministic, size-matched

\neg\mathcal{M}

; the final row is the layer-mean of the within-set median pairwise absolute cosine.

r_{b}>0

means larger values on

\mathcal{M}

Figure 1: FFN read-blindness precedes amplifier specialization. We assign massive and control coordinates at the final checkpoint and trace these fixed sets through 21 checkpoints from $0$ to $20$ k training steps. (a) Maximum-over-layer FFN erasure for eventual massive coordinates (orange) and non-massive controls (gray). (b) Maximum-over-layer leading amplifier gain for the same sets. Lines and shaded regions show medians and interquartile ranges; vertical dotted lines mark $t_{\rm sep}$ , the first checkpoint at which the rank-biserial effect reaches $r_{b}\geq 0.95$ . FFN erasure separates at $2.3$ k steps, before amplifier gain at $4.5$ k steps. (c) Final-checkpoint amplifier norm $\|U_{k}\|_{F}$ versus the inverse participation ratio of the corresponding FFN write row $W_{\text{down}}[k,:]$ . Colored points are massive coordinates, with size denoting massiveness and color denoting attention-input erasure; gray points are non-massive coordinates.

Figure 2: Directional curvature on $\mathcal{M}$ versus $c\mathcal{M}$ in the 1.28B base model (30 batches, 64 directions). Read-side slices are stiffer on $\mathcal{M}$ ; write-side slices are sloppier. Error bars are bootstrap $95\%$ confidence intervals.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

	Llama 135M	Llama 1.28B	Llama 2.56B	Qwen3 1.7B
Massive coordinates $\|\mathcal{M}\|$	10	10	5	7
Maximum activation magnitude	280	$2.27\!\times\!10^{3}$	$3.92\!\times\!10^{3}$	$2.23\!\times\!10^{3}$
Input availability:
Attention normalized input, $\tilde{X}$	0.52 (-0.40)	0.52 (-0.24)	0.52 (-0.56)	0.76 (+0.19)
FFN normalized input, $\tilde{Y}$	0.51 (-0.41)	0.80 (+1.00) *	0.65 (+0.93) *	0.80 (+0.68) *
Read side weight view:

Appendix

Table 4: Null occupancy results. Cells report the median

{\eta}_{\mathcal{M}}

with rank-biserial

r_{b}

in parentheses. Bold cells have

r_{b}>0.95

; stars denote one-sided Mann–Whitney tests of

{\eta}_{\mathcal{M}}>{\eta}_{\neg\mathcal{M}}

with

p<0.05

	$W_{V}$ Frozen	$W_{V}$ Reparam	$W_{\mathrm{up}},W_{\mathrm{gate}}$ Frozen
Massive coordinates $\|\mathcal{M}\|$	7	10	11
Maximum activation magnitude	$2.08\!\times\!10^{3}$	$2.12\!\times\!10^{3}$	681
Input availability:
Attention normalized input, $\tilde{X}$	0.51 (-0.52)	0.56 (-0.57)	0.61 (-0.17)
FFN normalized input, $\tilde{Y}$	0.58 (+0.70) *	0.52 (-0.28)	0.57 (-0.17)
Read side weight view:

Appendix

Table 5: Llama 1.28B Erasure results for intervention settings. Bold cells have

r_{b}>0.95

; stars denote one-sided Mann–Whitney tests of

{\mathrm{era}}_{\mathcal{M}}>{\mathrm{era}}_{\neg\mathcal{M}}

with

p<0.05

	$W_{V}$ Frozen	$W_{V}$ Reparam	$W_{\mathrm{up}},W_{\mathrm{gate}}$ Frozen
Massive coordinates $\|\mathcal{M}\|$	7	10	11
Maximum activation magnitude	$2.08\!\times\!10^{3}$	$2.12\!\times\!10^{3}$	681
Input availability:
Attention normalized input, $\tilde{X}$	0.57 (+0.30)	0.54 (-0.13)	0.55 (+0.02)
FFN normalized input, $\tilde{Y}$	0.89 (+1.00) *	0.80 (+1.00) *	0.68 (+0.77) *
Read side weight view:

Appendix

Table 6: Llama 1.28B Null occupancy results for intervention settings. Bold cells have

r_{b}>0.95

; stars denote one-sided Mann–Whitney tests of

{\mathrm{era}}_{\mathcal{M}}>{\mathrm{era}}_{\neg\mathcal{M}}

with

p<0.05

Property	Llama 135M	Llama 1.28B	Llama 2.56B	Qwen 1.7B
Number of Transformer layers ( $L$ )	10	16	27	28
Residual-stream dimension ( $d_{\mathrm{model}}$ )	768	2,048	2,560	2,048
Feed-forward dimension ( $d_{\mathrm{ff}}$ )	1,536	8,192	7,680	6,144
Number of query heads ( $n_{h}$ )	8	32	40	16
Number of key–value heads ( $n_{kv}$ )	8	32	40	16
Attention-head dimension ( $d_{h}$ )	96	64	64	128

Appendix

Table 7: Llama ( Touvron et al., 2023 ) and Qwen3 ( Team, 2025 ) Architectures of the models used in our experiments.

Figure 3: Emergence of $\mathcal{M}$ discrimination during pre-training of the 1.28B model (21 checkpoints, $0$ – $20$ k steps; per-coordinate maximum across 16 layers). Solid colored and dashed gray lines show medians and interquartile ranges for $\mathcal{M}$ ( $n=10$ ) and sampled $c\mathcal{M}$ controls ( $n=100$ ), respectively. The dotted line marks $t_{\rm sep}$ , the first checkpoint with rank-biserial effect $>0.95$ . FFN read-blindness preceeds its amplifier.

Figure 4: FFN write-row concentration vs. amplifier gain (left) and $s_{\star}$ mass on $\mathcal{M}$ (right). Color denotes attention-input erasure, size denotes massiveness $R$ , and grey points are non-MA coordinates.

Figure 5: Curvature and radial loss-gradient projection for $W_{\rm down}[k,:]$ in deep layers $L\in\{9,12,15\}$ . Negative $\nabla\mathcal{L}_{\rm CE}\cdot\hat{w}$ means that the gradient vector points inward, so the loss-descent direction points outward. Strong-amplifier $\mathcal{M}$ coordinates concentrate in the negative-curvature, negative-gradient quadrant, unlike the matched non-massive coordinates.

Figure 6: Local first-order AdamW pressure on activation magnitude in the Llama 1.28B model. Lines show the batch mean and shaded regions show the batch 10th–90th percentile spread. Orange denotes $\mathcal{M}$ and dashed blue denotes the matched non-massive control. The full AdamW step increases the magnitude of $\mathcal{M}$ while leaving the control near zero. This effect is dominated by the saved optimizer state. The current gradient conditioned on that state is smaller, and weight decay acts in the opposite direction.

	R=2	R=5	R=10	R=20
Massive coordinates $\|\mathcal{M}\|$	15	10	6	4
Maximum activation magnitude	$2.27\!\times\!10^{3}$	$2.27\!\times\!10^{3}$	$2.27\!\times\!10^{3}$	$2.27\!\times\!10^{3}$
Input availability:
Attention normalized input, $\tilde{X}$	0.68 (-0.66)	0.68 (-0.73)	0.62 (-0.74)	0.61 (-0.77)
FFN normalized input, $\tilde{Y}$	0.56 (-0.14)	0.56 (-0.13)	0.56 (-0.04)	0.56 (-0.12)
Read side weight view:

Appendix

Table 8: Erasure results at different MA selection thresholds for Llama 1.28B base model. Cells report the median

{\mathrm{era}}_{\mathcal{M}}

with rank-biserial

r_{b}

in parentheses. Bold cells have

r_{b}>0.95

; stars denote one-sided Mann–Whitney tests of

{\mathrm{era}}_{\mathcal{M}}>{\mathrm{era}}_{\neg\mathcal{M}}

with

p<0.05

Explore similar work

Massive Activations Are Architecturally Robust: A Controlled Scratch/Commitment Residual Stream Test
Jun 17, 2026Maruthi VemulaResidual Stream Model Activations
University of North Carolina at Chapel Hill
Sparse Inter-Layer Dependencies of Transformer FFN Neurons
Jul 13, 2026Johannes Knittel, Hanspeter PfisterTransformer Attention Feed-Forward
Harvard University
Training, Reading, and Editing Legible Transformers
Jul 9, 2026Mark OskinTransformer Architectures Model Activations
School of Computer Science and Engineering University of Washington

Jun 17, 2026, cs.LGJ/K move · Enter open · S save

Massive Activations Are Architecturally Robust: A Controlled Scratch/Commitment Residual Stream Test

Maruthi Vemula

University of North Carolina at Chapel Hill

Residual Stream Model Activations Outliers

Open paper AskPDF arXiv

Trained transformers reliably develop massive activations, a small number of hidden dimensions whose magnitude is far above the median and which concentrate on the sequence-start token. Whether these outliers are a removable artifact of the residual stream's overloaded read and write role, or instead a functional necessity, is actively debated. We test the artifact hypothesis directly, with an architectural intervention. Our architecture, Ledger Residuals, splits the residual stream into a mutable scratch stream (Deliberation) that intermediate computation may freely overwrite and a protected, decode-only accumulator (Commitment) that holds the representation the model reads out. If massive activations exist only because one stream is forced to be both scratchpad and answer, then a dedicated answer channel should remove the need for them. We find that it does not. In matched-loss language models at the 160M and 290M scales, the model rebuilds the canonical fixed-dimension, start-token outlier inside the protected channel. The rebuilt feature is smaller in magnitude than in a standard transformer but more sharply concentrated on the start token, and a stronger sparsity penalty makes it more persistent and more concentrated still, rather than removing it. Massive activations therefore look architecturally robust: they re-emerge in whichever representation the model decodes from, which is what we would expect if they are functional rather than incidental. We release our architecture and measurement code.

Abstract

Explore similar work

Massive Activations Are Architecturally Robust: A Controlled Scratch/Commitment Residual Stream Test

Sparse Inter-Layer Dependencies of Transformer FFN Neurons

Training, Reading, and Editing Legible Transformers