ChronoState: Hidden Elapsed-Time Conditioning for Temporal-State Action Selection in Frozen-Backbone Language Models

Many automated systems, such as security agents or cache managers, must make decisions based on both symbolic state and elapsed wall-clock time. While prompt-based timestamps are common, they can be unreliable, easily manipulated, or buried in long contexts. This paper investigates whether a frozen-backbone language model can instead use a hidden, non-token scalar representing elapsed time to inform its decisions. The authors introduce ChronoState, a benchmark where symbolic state is provided in the prompt while elapsed time is injected directly into the model's residual stream via a gated FiLM (Feature-wise Linear Modulation) mechanism.

The researchers utilize a Qwen2.5-3B-Instruct model with its backbone weights frozen. They encode elapsed time as a 31-dimensional vector of sinusoidal and logarithmic features. This vector is injected into the model's hidden states using gated FiLM modulation, which is initialized to zero to ensure the model initially behaves like the original pretrained backbone. A small LoRA (Low-Rank Adaptation) surface is used to map these conditioned states to specific temporal actions. The authors test the model's causal reliance on this hidden scalar by performing interventions, such as zeroing or shuffling the time input, to see if the model's decision-making changes accordingly.

The study demonstrates that the hidden chronometric injection (CI) channel allows the model to achieve high accuracy (approx. 0.93) on temporal-state tasks. Causal ablations, such as shuffling the time input, show that the model's performance collapses, confirming that it is indeed relying on the injected scalar rather than just memorizing prompt patterns. However, the authors note that a baseline using standard prompt-injected timestamps performs even better (approx. 0.99). Furthermore, while the model generalizes well to new durations and templates, it struggles to transfer to entirely new task families, such as complex quota-reset logic.