Progressive Cramming: When One Embedding Stores 1,500 Tokens, Is the Model "Understanding" or Taking a Shortcut?
The Setting: A Counterintuitive Finding
Imagine compressing an entire Wikipedia article—1,500 tokens—into a single embedding vector, then feeding that vector to a frozen LLM. Something remarkable happens: the model reproduces the entire article verbatim with less than 1% error.
This is the result shown in the 2025 "Token Cramming" paper by Kuratov et al. A 1,500x compression ratio, far beyond the ~10x lossless compression of traditional encoders. It looked as though an untapped "information jungle" was hiding in the Transformer's embedding space.
But researchers at FusionBrain Lab asked an unsettling question: does the model truly "understand" the compressed information? Or has it merely found some shortcut that makes reconstruction look like understanding?
Their July 2026 paper Progressive Cramming (arXiv: 2607.21231) offers a sobering answer: cramming's high-fidelity reconstruction is likely not "semantic encoding" but "fragile attention hijacking."
Fixing the Experimental Tool First: Progressive Compression
The original cramming method had a methodological problem: it used a fixed token budget and a 99% reconstruction accuracy threshold. 99% sounds high—but how severe is the remaining 1%?
The researchers uncovered a hidden catastrophe: that 1% of errors is concentrated almost entirely in the earliest 0-1 token positions. Under teacher forcing this looks harmless—99% of positions are correct. But once you switch to autoregressive generation, disaster strikes: the first token is wrong → all subsequent tokens are generated from a wrong prefix → the entire sequence rapidly collapses. The result: even with 95-99% average convergence during training, greedy-decoding convergence at inference is nearly 0%.
It's like copying answers with a teacher pointing at the right ones (teacher forcing)—fast and accurate. But once the teacher leaves (autoregressive generation), one wrong first character derails everything.
To fix this, the authors propose Progressive Cramming:
1. Start with 1 token; optimize the embedding until reconstruction is 100% perfect. 2. Expand to 2 tokens, warm-starting from the previous embedding. 3. Keep expanding one token at a time, only on top of a successful step. 4. Stop when perfect reconstruction is no longer achievable.
The benefit: a clear success/failure boundary. Not "99% correct," but either 100% reconstruction or a definite failure point.
Finding 1: Optimization Trajectories Are Low-Dimensional
First finding: the progressive cramming optimization trajectory occupies a low-dimensional manifold in embedding space.
The embedding dimension is 2,048–4,096 (Llama-3.1-8B), yet 99% of the trajectory's variance is explained by just 30–100 PCA components. Crammed embeddings don't wander the high-dimensional space—they travel along a narrow "track," a structured subspace of the model's computational geometry. Cramming succeeds because it finds this track.
Finding 2: Compressed Embeddings Become "Attention Black Holes"
The more critical finding comes from attention analysis: the compressed embedding captures 20–60% of attention mass in middle layers, acting as a massive attention sink that pulls in attention from surrounding tokens.
This sounds like good news—the compressed embedding matters, so the model attends to it. But causal analysis shows: this attention concentration is not the model "using" the compressed information; it's the compressed embedding "hijacking" the model's attention.
Using "attention knockout" experiments (manually blocking the compressed embedding's attention connections at certain layers):
- Blocking early-layer attention to the compressed embedding: downstream capabilities recover.
- Blocking late-layer attention: almost no effect.
- Likelihood evaluation (comparing option likelihoods): even with the original prefix still in context, prepending a compressed embedding causes a consistent but moderate accuracy drop.
- Generative evaluation (model generates answers): capability collapses almost completely.
- Its conclusions rest mainly on models like Llama-3.1-8B. Whether other architectures (e.g., state-space models like Mamba) exhibit the same "attention hijacking" mechanism remains open.
- The "fragile attention manipulation vs. semantic encoding" dichotomy may be oversimplified—the reality could be a mixture; the causal analysis is strong but doesn't fully rule out partial semantic encoding.
- Progressive compression itself is costly (token-by-token optimization), and warm-starting may introduce its own biases (each step's solution depends on the previous step's local optimum).
Yet late layers are exactly where attention mass is largest. Conclusion: attention concentration is a symptom of cramming, not the cause of its success. The real capability damage comes from early-layer interactions, which carry relatively little attention mass.
Finding 3: Reconstruction Success ≠ Capability Preservation
If cramming were genuine semantic encoding, a compressed prefix should preserve downstream capabilities—the information is all there, and the model should be able to read it out.
Testing HellaSwag and ARC-Easy:
This contrast is crucial. If the compressed embedding truly encoded semantics, generative mode should exploit it too. Its collapse shows: the model is not "reading" the compressed information; it is being "guided" by it in a fragile way—and that guidance shatters once teacher forcing is removed.
Finding 4: Compression Capacity Grows with Depth and Width
Cramming capacity is not unlimited—it is bounded by the frozen reconstructor's ability. Truncating the model to its first N layers (with a short fine-tune to repair the truncation point), the researchers found: compression capacity grows monotonically with retained depth and model width.
This is itself counter-evidence against pure semantic encoding: if cramming encoded semantics, it shouldn't depend so heavily on reconstructor depth—semantic information should be readable within the first few layers.
Rethinking Cramming: Not Encoding, but Manipulation
Assembling the findings, the paper reaches a sobering conclusion:
> High-fidelity reconstruction reflects not dense semantic encoding, but the ease of "manipulating" the model through its attention mechanisms.
Analogy: cramming doesn't write information into the model's "memory book" for it to read; it finds a switch that directly triggers the model's "conditioned reflexes." Reconstruction looks perfect, but the model isn't "understanding"—it's been hijacked by a carefully optimized embedding that steers it along a specific trajectory toward the correct token sequence.
This distinction matters for practical applications of compressed embeddings. If cramming were semantic encoding, you could compress long contexts into a few embeddings and the model would still use them. But if it's attention manipulation, any use that deviates from the training distribution causes collapse—exactly what the generative evaluation shows.
Methodological Contribution: Stop Using the 99% Threshold
Beyond the scientific findings, the paper exposes the blind spot of the "99% reconstruction accuracy" standard. Under a fixed budget + 99% threshold, you see "nearly perfect"—but you don't see that the 1% of errors sits in the most lethal positions (the earliest tokens), causing total collapse under autoregressive generation.
This fits a broader pattern: 99% token-level accuracy can mask 100% generation-level failure. Progressive compression's value is that it exposes those fatal errors by demanding 100% reconstruction with token-by-token expansion, giving a definitive answer to whether compression actually works.
An Honest Assessment
The paper has limitations:
A Deeper Lesson
The paper raises a more general question: the gap between "appearing to work" and "actually working" in AI research may be far larger than we assume. How many signals—99% reconstruction rates, high benchmark scores, post-RLHF "politeness"—reflect genuine semantic understanding versus some form of attention hijacking or reward shortcut?
Progressive Cramming offers a method: when you see an "almost perfect" result, ask where the remaining 1% lies and how lethal it is. The answer may change what "perfect" means.
---
Paper: https://arxiv.org/abs/2607.21231 Code: https://github.com/FusionBrainLab/progressive_cramming HTML version: https://arxiv.org/html/2607.21231v1 Authors: FusionBrain Lab