Progressive Cramming: Reliable Token Compression and What It Reveals

Token cramming compresses sequences into learned embeddings with near-perfect reconstruction, but fixed token budgets and 99\% accuracy thresholds leave it unclear whether residual errors reflect optimization failures or fundamental limits. We introduce progressive cramming, which grows the target prefix token-by-token, stopping only when reconstruction is no longer achievable within a fixed optimization budget. Progressive trajectories occupy low-dimensional structure in embedding space. Prepending a crammed embedding causes a moderate but consistent accuracy drop on multiple-choice benchmarks even with the original prefix in context, and collapses capability almost entirely under generative evaluation. Causal attention-knockout interventions trace this degradation to the embedding's interactions in the model's early layers. These results position progressive cramming as a tool for studying compression limits and show that perfect reconstruction - achievable through brittle steering rather than transferable semantics - is insufficient for meaningful compression.

AI language models can squeeze a passage of over a thousand words into a single compact "code" and then recite it back almost perfectly. This looked like remarkable compression - but does the code actually store the text's meaning, or just exploit a quirk of how the model computes? We developed progressive cramming, which builds up the stored passage one word at a time and stops exactly when the model can no longer reproduce it, pinpointing where compression succeeds or fails. We then tested whether a model holding such a code could still answer simple questions about the text, and traced how the code steers the network inside. The answer was surprising: even with perfect playback, the code preserves no usable meaning - it mostly hijacks the model's internal "attention". Our work matters because perfect reconstruction is a misleading goal: a method can score flawlessly while encoding nothing the model can actually use.