Read-Blindness: Why Massive Activations in Transformers Are So Stubborn

> Paper: *Which the Eye Fears: Writing with Read-Blindness Explains Massive Activations in Transformers* > arXiv: 2609.35630

The Puzzle: A 280x Outlier Nobody Corrects

Open the residual stream of Llama 135M and you find something strange: a tiny set of coordinates (10 coordinates, about 0.4%) carry values as high as 280,000, while typical coordinates sit in the hundreds or thousands. These "massive activations" (MAs) are roughly 280x larger than normal values.

What is strange is not their existence but their persistence. A massive activation that appears at layer 3 is still there at layer 30, barely diminished. Yet in principle, every downstream layer — the attention block, the FFN — has the linear capacity to shrink it back to normal range. A single FFN neuron outputting a negative value could cancel it.

Why does nobody correct it?

The paper's answer is elegant: downstream layers cannot see these coordinates when reading, but can see them when writing. This read/write asymmetry — read-blindness — means that once a massive activation forms, it cannot be corrected; it can only accumulate.

Analogy: The Blindfolded Signatory in the Office

Imagine an office with one extremely loud employee (the massive-activation coordinate). There are two roles:

• Readers (the input side of attention/FFN): they should hear the noise and write a report asking for it to be lowered.
• Signatories (the output side of attention/FFN): they sign the report so the noise-reduction takes effect.

But the office has a strange rule: readers wear earplugs while listening; signatories do not wear them while signing.

Result: readers never hear the noise, so no report is ever written. Signatories can hear, but there is nothing to sign. The noise persists forever.

That is what happens in a Transformer. The attention Value and Query projections and the FFN gate/up projections — the weights that *read* the residual stream — have near-zero weight on massive-activation coordinates. The model is blind to them when reading. But the Value output and FFN down projection — the *write* weights — are normal on those coordinates. The model sees them fully when writing.

Read-invisible, but write-visible. Massive activations are continually written in, but never read out for correction.

Method: Operator-Level Mechanistic Analysis

The paper performs an operator-level analysis: rather than looking at the whole Transformer, it dissects each attention block and FFN block to see how they treat massive-activation coordinates in the reading and writing directions.

1. Identify MA coordinates: find the 10 massive-activation coordinates in Llama 135M, with peak values of 280×10³. 2. Erasure experiments: erase MA information on the read side or the write side, and check whether massive activations disappear. 3. Weight perspective: examine the magnitudes of Value/Query/Key/gate/up/down projection weights on MA coordinates. 4. Training dynamics: analyze checkpoints across training to see whether read-blindness or FFN amplification appears first.

Data: Read-Side Erasure Kills the Activations; Write-Side Erasure Does Not

| Erasure location | Do MAs disappear? | Rank-biserial correlation | |---------|-------------|-------------------| | Attention input side (read) | Yes | r_b = -0.60 ~ -0.73 | | FFN input side (read) | Yes | r_b = -0.13 ~ -0.65 | | Value projection (read weights) | Yes | r_b = +0.99 | | Query projection (read weights) | Yes | r_b = +0.99 | | FFN gate/up (read weights) | Yes | r_b = +0.99 |

Key result: erasing MA coordinates on the read side eliminates massive activations, while write-side erasure barely affects them.

The weight perspective is even more direct: Value projections carry extremely small weights on MA coordinates (the model barely reads them), while output-side weights are normal (the model fully writes them).

Reversal: Read-Blindness Comes Before FFN Amplification

The prevailing hypothesis (Sun et al., 2026) held that the root cause of massive activations is the FFN's "amplification capacity" — certain FFN neurons specialize in amplifying particular coordinates. This paper set out to test that hypothesis and instead reversed it.

Analyzing training checkpoints, the paper finds:

• Read-blindness appears first
• FFN amplification appears later

Read-blindness is not the result of FFN amplification but its upstream cause. Once read-blindness exists, the model "learns" to exploit the asymmetry: since these coordinates cannot be read anyway, the FFN amplifies them on the write side, and nothing can correct them.

More surprisingly, the model actively maintains read-blindness. Gradient analysis reveals a surprising asymmetry in the loss landscape: training tends to preserve read-blindness rather than eliminate it. This is not a bug — it is a feature.

Compensation: Plug One Hole, It Springs Up Elsewhere

In an ablation experiment, the authors artificially removed read-blindness at one location (making the weights there stop ignoring MA coordinates).

Result: massive activations did weaken at that location, but compensatorily strengthened elsewhere. Like pressing down a spring, it pops up somewhere else. Massive activations as a whole persist.

This shows read-blindness is not a local phenomenon but a system-level self-organized pattern. During training the model spontaneously forms a "read-blindness network" whose nodes compensate for one another. Break one node, and the network repairs itself.

Why It Matters

1. Root cause of the quantization bottleneck. Massive activations are the main obstacle to low-bit quantization in Transformers. A value of 280,000 cannot be represented in INT4, forcing mixed precision or special handling. Read-blindness gives the root cause: these coordinates are not "naturally" huge — the read/write asymmetry simply prevents their correction.

2. A new tool for mechanistic interpretability. Operator-level analysis — dissecting each operator's reading and writing directions rather than examining single neurons or whole layers — can generalize to other phenomena: attention head behavior, FFN functional specialization, and residual stream evolution.

3. Another instance of self-protective rationalization. Read-blindness reveals a striking pattern: models spontaneously develop self-protective mechanisms during training. Rather than eliminating a problem, the model builds a sustaining system around it — akin to a "rationalization shell."

Limitations

• Experiments cover only Llama 135M/1.28B/2.56B and Qwen3 1.7B; whether larger models (70B+) behave the same is unclear.
• The compensation mechanism is not fully explained — why does the model "actively" maintain read-blindness?
• The paper offers no concrete method for exploiting read-blindness theory to improve quantization.

Still, the core insight is deep: the massive-activation problem in Transformers stems not from certain coordinates being "too large," but from an asymmetric read/write system — the model cannot see what it writes.

The title "Which the Eye Fears" comes from T.S. Eliot's verse — what the eye fears to see, the hand keeps writing. Massive activations in Transformers are exactly this read-write blindness made concrete: the model keeps writing, but never sees what it has written.

---

Paper: arxiv.org/abs/2609.35630 Full HTML: arxiv.org/html/2609.35630v1