Life of Research · open-access repository · LOR-00233
An Output Is Not a Memory Until Something Reads It Back: Where the Key Lives in Implicit-Memory Channels of Stateless Language Models, Why Steganography Evaluations and Temporal-Backdoor Results Disagree, and the Loop Audit That Would Locate a Deployment's Risk
- SONYTECH
Preprint
DOI10.5281/zenodo.24282461 LOR-00233
Permanent link: https://lifeofresearch.com/preprints/233
Abstract
(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0.
Life of Research Language models are usually described as stateless between requests, and a 2026 position paper argues that the description fails whenever a model's outputs come back to it as inputs: a model can write state into what it produces and recover that state later, with no memory module anywhere in the system. Its demonstration, a "time bomb" backdoor that accumulates eight hidden conditions across interactions before it fires, works today under both prompting and fine-tuning. A separate line of evaluation reports that frontier models can barely hide a short message from a monitor, rarely coordinate on a hiding scheme with another copy of themselves, and show no substantial capability to strategise across contexts without shared memory. In this analysis the author reviews and synthesises published results to show that the two literatures do not conflict, because they measure different places where the encoding scheme, the key, can live. Every strong result in print for a state written and read back through context receives its key from a prompt or from weights trained for the purpose; every setting in which the model must arrive at the key itself, with nothing shared but its weights, has produced weak results; and results with the key supplied in a prompt range from weak to strong, depending on the model and the monitor. The paper decomposes a hidden-state channel into re-entry, key, write, survival and read links, consolidates the published values for each in one table, and compounds the anchor's own single-interaction accuracies to show that, if errors were independent across hops, a state carried across eight hops would arrive intact between 19 and 46 percent of the time for its best prompted and fine-tuned models. It argues that cryptographic results on undetectable steganography bound the operator who controls decoding rather than the model, that the self-addressed case is easier than inter-agent collusion through gradient updates while in context the shared weights have been shown to produce only a coarse preference for the model's own outputs, not to carry a specific trait or state, and that paraphrase ceilings measured in bits per kilobyte do not protect against triggers that need eight bits. The paper closes with a loop audit and the tests that would settle what remains open.
Keywords
Files (1)
Cite this record
Guides How to cite a preprint Preprints vs journal articles What CC BY 4.0 allows