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

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

This paper introduces ChronoState, a compositional temporal-state benchmark and chronometric injection (CI) interface for studying whether a frozen-backbone language model can compose a system-supplied, non-token elapsed-time scalar with visible symbolic state to select temporal actions. The central claim is intentionally narrow: under direct supervision, a hidden elapsed-time channel can support high-accuracy action selection from symbolic temporal state, without claiming autonomous time tracking, subjective time perception, or architectural superiority over prompt timestamps.

The problem setting is defined as: "Let fθ be an autoregressive language-model backbone with frozen parameters θ. Let x denote the visible prompt, s denote symbolic state described in the prompt, and τ ∈ R≥0 denote elapsed monotonic seconds since a system-defined reference event. The model receives τ externally from a trusted clock or controller; hidden means hidden from the user-visible token sequence, not from model computation."

The chronometric injection interface encodes elapsed seconds into a 31-dimensional fixed feature vector: χ(τ) = [sin(τ/T1), cos(τ/T1),..., sin(τ/T15), cos(τ/T15), log(1+τ)] T, with divisors T ∈ 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 4096, 16384, 65536, 86400, 604800 seconds. This encoding is injected into the residual stream of a frozen Qwen2.5-3B-Instruct bf16 backbone through gated FiLM modulation: h′l = hl + αl ⊙ (γl ⊙ hl + βl), where γl and βl are learned projections of χ(τ), and αl is a learned per-channel gate. Zero initialization (αl = 0, Wlγ = 0, bγl = 1, Wlβ = 0, bβl = 0) ensures the model initially behaves exactly like the frozen backbone. The trainable surface includes 35 chrono injection sites, 145 LoRA modules (rank 8 on q/k/v/o projections and lm head), totaling 9,577,472 of 3,095,516,160 parameters (0.309%).

The ChronoState benchmark covers six temporal families: cache expiration (REUSE/REFRESH based on TTL), job completion (WAIT/RETRIEVE based on expected duration), deadline feasibility (WAIT/DEADLINE MISSED based on remaining slack), quota reset (WAIT QUOTA/RETRIEVE based on reset interval), session staleness (REUSE/ASK CONFIRM based on staleness threshold), and multi-constraint task management with a fixed priority policy: "1. if an active deadline predicate is infeasible, output DEADLINE MISSED; 2. else, if an active quota predicate has not reset, output WAIT QUOTA; 3. else, if an active staleness predicate is stale, output ASK CONFIRM; 4. else, if an active cache predicate is expired, output REFRESH; 5. else, if an active job predicate is complete, output RETRIEVE; 6. else, if an active cache predicate is still valid, output REUSE; 7. otherwise output WAIT."

Training uses 12,000 train examples, 2,000 validation, 4,000 standard-test, and 2,000 each for held-out splits (template, duration, composition, family) per seed, across three seeds (78,000 total records). Training elapsed-time values are 10, 30, 300, 1800, 7200, 43200, 86400, 604800 seconds, while held-out-duration values are 60, 900, 3600, 10800, 21600, 172800, 345600 seconds. Training uses 8,000 steps, learning rate 10−4, AdamW, weight decay 0.01, chunk length 512, forced-choice A/B/C/D next-token softmax loss, and log-probability scoring.

Main results (Table 3): Hidden-time CI reaches 0.9305 ± 0.0134 accuracy and 0.9410 ± 0.0103 balanced accuracy on the standard test. No-time control drops to 0.5511 ± 0.0042, shuffled-time control drops to 0.3323 ± 0.0097, and LoRA-only (no chrono channel) reaches 0.5513 ± 0.0095. However, the fair prompt+LoRA timestamp baseline is substantially stronger at 0.9893 ± 0.0052 accuracy and 0.9920 ± 0.0039 balanced accuracy. Vanilla no-time and vanilla prompt-timestamp baselines (without trainable surface) are near chance at 0.3731 and 0.3738 respectively.

Causal analysis via shuffled-time intervention (Table 4): Across seeds, hidden-time accuracy is 0.9378, 0.9388, 0.9150, while shuffled-time accuracy falls to 0.3243, 0.3295, 0.3430. Wrong-state consistency (WSC) is 0.9115, 0.9230, 0.8835, meaning when the injected scalar is wrong, the model tends to recompute the action from the wrong scalar rather than ignore it. Conflict examples (visible timestamp wrong, hidden scalar true) show scalar-follow of 0.9065 ± 0.0130 and prompt-follow of 0.3323 ± 0.0072, demonstrating hidden-channel dominance but not safe disagreement handling.

Generalization results (Table 5): Hidden-time CI remains strong on held-out templates (0.9357 ± 0.0109), held-out durations (0.8375 ± 0.0055), and held-out multi-constraint compositions (0.8652 ± 0.0438). However, the held-out-family split (pure quota, excluded from training except inside multi-constraint examples) reaches only 0.5065 ± 0.0559 accuracy and 0.5695 ± 0.0496 balanced accuracy, with many true RETRIEVE cases are predicted as WAIT QUOTA. This is identified as the main generalization failure and limits any claim of broad predicate abstraction.

Linear diagnostic baselines (Table 6) show that a linear model over χ(τ) plus engineered state features reaches 0.9230 on standard, 0.9265 on template, 0.8220 on duration, 0.8138 on composition, and 0.7137 on family, while the rule oracle reaches 1.0000 on all splits. This clarifies that ChronoState is not intrinsically unsolvable and that the learned model does not abstract pure quota as well as the engineered diagnostic baseline.

The paper's four contributions are: (1) formulating hidden-time temporal-state reasoning as a composition problem over visible symbolic state s and hidden elapsed seconds τ; (2) introducing ChronoState, a forced-choice benchmark covering six temporal families; (3) evaluating chronometric injection against no-time, shuffled-time, prompt-timestamp, LoRA-only, prompt+LoRA, and vanilla controls; (4) characterizing the scope: hidden-time CI solves standard/template/duration/composition splits under direct supervision, but loses to prompt+LoRA and remains weak on pure held-out quota-family transfer.

The paper explicitly states what it does not support: "It does not show autonomous time tracking: τ is supplied externally, and a wrong value can drive a wrong action. It does not show superiority over prompt timestamps: prompt+LoRA remains substantially stronger on this supervised benchmark. It does not establish broad unseen-family transfer: pure held-out quota remains weak, suggesting incomplete abstraction of that predicate when it appears only inside multi-constraint training examples."

Limitations include: synthetic and deterministic benchmark, limited to Qwen2.5-3B-Instruct with three seeds, finite time range with interpolation-style held-outs, fixed time encoding without ablation, underdeveloped conflict handling (no safe abstention or disagreement detection), the held-out quota-family failure as the largest substantive limitation, and incomplete upstream-model revision recording. Future work should include external temporal-state environments, matched scalar and prefix baselines, larger and more diverse base models, leave-one-family-out evaluations for every predicate family, and explicit conflict-detection policies.

Improvements to AI Systems:

1. Add a hidden, non-token temporal-state input channel to frozen-backbone LLMs via gated FiLM modulation, enabling the model to reason over elapsed time without consuming token budget or altering visible prompts. The improved system can perform accurate time-dependent action selection (e.g., cache refresh, job retrieval, quota checks) even when time is not mentioned in text.

2. Implement a zero-initialized, gated injection layer (as in ChronoState) so that the frozen backbone’s behavior is unchanged until trained, allowing safe drop-in integration into existing LLM pipelines. The improved system can be fine-tuned for temporal reasoning without risk of catastrophic forgetting or output degradation on non-temporal tasks.

3. Use a multi-scale sinusoidal time encoding (with divisors from 2 seconds to 7 days) to capture both fine-grained and long-horizon temporal patterns. The improved system can generalize to unseen durations (e.g., 60s, 900s, 3600s) not seen during training, maintaining high accuracy on interpolation-style temporal extrapolation.

4. Train with a forced-choice next-token softmax loss over a small action vocabulary (A/B/C/D), enabling the model to select among competing temporal actions (e.g., WAIT vs. RETRIEVE vs. REFRESH) with high confidence. The improved system can handle multi-constraint scenarios by applying a learned priority policy over multiple active temporal predicates.

5. Add a causal intervention mechanism (shuffled-time testing) to detect when the model is relying on the hidden time channel vs. spurious prompt features. The improved system can be audited for temporal reasoning robustness, and if the hidden channel is corrupted, the model’s behavior change can be flagged for safe fallback.

6. Combine hidden-time conditioning with prompt+LoRA baselines to achieve the best of both worlds: the prompt+LoRA method already reaches 0.989 accuracy, but adding a hidden channel could improve robustness when timestamps are missing or unreliable. The improved system can maintain high accuracy even when visible timestamps are absent or contradictory.

7. Introduce a conflict-detection policy (e.g., abstain or output a “disagreement” flag) when the hidden scalar and visible timestamp disagree, addressing the observed scalar-follow dominance (0.9065) without safe handling. The improved system can avoid taking wrong actions in adversarial or noisy environments where time sources conflict.

8. Expand training to include all predicate families in isolation (not just inside multi-constraint examples) to fix the held-out quota-family failure (0.5065 accuracy). The improved system can achieve broad temporal abstraction across all six families, enabling reliable transfer to unseen temporal predicate types.

9. Use the linear diagnostic baseline (0.9230 on standard) as a lightweight fallback for resource-constrained settings, while the neural model handles complex compositions. The improved system can dynamically switch between a fast rule-based temporal reasoner and a full LLM based on available compute and task complexity.

10. Incorporate the ChronoState benchmark as a standard evaluation suite for any LLM claiming temporal reasoning abilities, enabling fair comparison across models and training regimes. The improved system can be continuously validated against six temporal families and four held-out splits to ensure no regression in temporal action selection.

Temporal decisions in language-model systems often depend on both symbolic task state and elapsed wall-clock time, such as cache expiration, job completion, quota resets, deadlines, or stale sessions. We study whether elapsed time can be supplied as a non-token, system-side scalar and composed with visible symbolic state by a frozen-backbone language model. We introduce ChronoState, a compositional temporal-state benchmark in which symbolic state appears in the prompt, elapsed seconds tau are supplied through a hidden chronometric-injection channel, and the model selects a forced-choice temporal action. Here, "hidden" means hidden from the user-visible token sequence, not from model computation. Using Qwen2.5-3B-Instruct as a frozen bf16 backbone with a 31-dimensional sinusoidal-plus-log time encoding, gated FiLM residual modulation, and a rank-8 LoRA action surface, hidden-time CI reaches 0.9305 +/- 0.0134 accuracy and 0.9410 +/- 0.0103 balanced accuracy. No-time and shuffled-time controls fall to 0.5511 +/- 0.0042 and 0.3323 +/- 0.0097, respectively, with high shuffled-time wrong-state consistency supporting causal dependence on the injected scalar within the trained distribution. Generalization remains strong for held-out templates, durations, and multi-constraint compositions, but held-out quota-family transfer is weak at 0.5065 +/- 0.0559, while a fair prompt+LoRA timestamp baseline reaches 0.9893 +/- 0.0052. Thus, ChronoState supports a narrow conclusion: hidden elapsed time can be composed with symbolic task state under direct supervision, but does not establish autonomous time tracking, broad unseen-family abstraction, or superiority over prompt-injected timestamps.