A new technique for shrinking AI model files also happens to make them harder to sabotage.
Researchers introduced Seed-Q, a method for compressing the weights of large language models that reconstructs them from a compact random seed instead of storing every number directly. Unlike earlier seed-based approaches, Seed-Q uses a lightweight linear-feedback shift register to generate weights and spends more of its limited bit budget on the weights that matter most to accuracy, squeezing the rest harder. The decoder figures out on its own which weights get the extra bits, so there is no need to store extra metadata or run a calibration pass beforehand. In tests across several LLMs, Seed-Q matched the accuracy of the existing SeedLM method using fewer bits, and beat it outright at the same 4 bits per weight.
The real hook here is security, not just file size. Because each seed regenerates a whole cluster of weights, a single bit-flip attack - the kind used to corrupt model parameters in memory - corrupts many weights at once instead of one quietly, making tampering far easier to spot. That is a meaningful upgrade for anyone shipping models to edge devices or hardware they do not fully control, where physical attacks on memory are a real threat, not a hypothetical.
The team also built the scheme into an ASIC accelerator with only modest extra hardware cost, suggesting this is not just a paper trick - though "modest overhead" is doing some quiet lifting until someone benchmarks it against a production chip.