Interwhen: Timing in Dynamic Systems

• Interwhen is a framework that formalizes decision-making for intervention timing in dynamic systems by integrating control theory, wireless sensing, and large language models.
• It employs methodologies such as neural MDP policies, temporal windowing, and sequential disclosure to balance performance, latency, and resource costs.
• Applications range from intermittent control in networked systems to verifiable reasoning in LLMs, offering near-optimal tradeoffs and enhanced system reliability.

Interwhen denotes a set of formal, technical approaches to the question “when” to intervene, synchronize, act, reveal, or communicate within dynamic decision-making systems—spanning control theory, wireless sensing, and LLMs—so as to optimally trade off process performance against resource constraints, latency, or verifiability. The Interwhen principle emerges wherever the timing of intermediate actions (transmissions, disclosures, or verifications) profoundly shapes systemic reliability and efficiency.

1. Core Formulations and Domains of Interwhen

Interwhen arises in distinct but structurally parallel forms across several fields:

• Networked Control: The scheduler in an intermittently controlled plant must decide at which instants to send full plant state to the controller, thereby mediating between communication costs and control quality (Aggarwal et al., 12 Apr 2025).
• Wireless Perception: In multisensory wireless systems, Interwhen quantifies how disparate sensory and digital events are fused via “Temporal Windows of Integration” (TWI), manipulating when events are perceived as simultaneous or causally ordered (Popovski, 2024).
• Autoregressive Model Reasoning: For LLMs, Interwhen formalizes when internal (private) reasoning should be disclosed into the irreversible public token stream, balancing latency (“silence tax”) against premature commitments (Wei et al., 5 May 2026).
• Test-time Verifiable Reasoning: The Interwhen framework enables adaptive, verified intervention within a model’s reasoning trace, determining at what points to check, correct, and steer outputs for soundness and efficiency (Bhat et al., 5 Feb 2026).

A unifying theme is that the “when” of intermediate system events—sensor transmissions, reasoning disclosures, or verification checks—must be controlled by explicit, policy-driven mechanisms grounded in rigorous stochastic, combinatorial, or RL-based optimization.

2. Formal Approaches: Scheduling, Disclosure, and Verification

Each domain expresses Interwhen as a structured decision problem:

Intermittent Control (Networked Systems)

• System modeled as a discrete-time stochastic linear process:

$$x_{k+1} = A x_k + B u_k + w_k,\quad a_k \in \{0,1\}$$

Scheduler decides whether to transmit the current state $x_k$ (paying penalty $\lambda$), or to skip. The controller operates with partially nested information.

• The scheduling policy reduces to a Markov decision process (MDP) in the controller's estimation error $e_k = x_k - \hat{x}_k$, with per-step cost:

$C(e, a) = \|e\|^2_\Gamma + \lambda a$

where $\Gamma$ emerges via Riccati recursion under certainty-equivalent control (Aggarwal et al., 12 Apr 2025).

• Deep Q-learning (InterQ) approximates the optimal Q-function over continuous error domains, yielding an actionable “send/skip” decision that optimally trades off control fidelity against network load.

Multimodal Perceptive Networks (Temporal Windows)

• Events from multiple inputs are timestamped using:

$s(t) = \lceil t / W \rceil$

for window size $W$; any events within the same TWI are treated as simultaneous.

• Probabilities of simultaneity and causality violation are governed by misalignments relative to TWIs, and tradeoffs are formalized via tail bounds and stochastic latency models. The correct choice of $W$ calibrates simultaneity error, causal reversals, and throughput (Popovski, 2024).

LLM Disclosure Policies

• For LLMs, a two-channel sequence is imposed:

$$\tau = (c_1, z_1, \dots, c_K, z_K),\quad c_k \in \{R, A\}$$

where R denotes private reasoning and A denotes public disclosure.

• The core question—“when to speak”—is framed as a sequential decision policy trained via entailment-aligned supervised fine-tuning and reinforced via RL, optimizing for the Pareto tradeoff between accuracy and content-latency (Wei et al., 5 May 2026).

Test-Time Verified Reasoning

• Interwhen prompts the model to emit verifiable partial states $x_k$0. Each $x_k$1 is parsed and checked by a verifier $x_k$2 (symbolic or LLM-based).
• Upon violation, the system intervenes by inserting feedback and continuing. Two main regimes:
  • Self-verification (k-Stable): Early-stops when subsequent answer proposals agree.
  • External verification: Enforces correctness with domain-specific checking at each step, guaranteeing “soundness” by construction (Bhat et al., 5 Feb 2026).

3. Algorithmic and Systemic Insights

Networked Control: InterQ and Threshold Structures

• InterQ employs a neural MDP policy, where Q-functions yield almost-ellipsoidal switching curves: estimation errors within a calculable ellipsoid mean “do not send,” while errors outside trigger transmission. This threshold structure is derived directly from Bellman-comparison and is data-driven, robust, and nearly optimal compared to standard baselines (periodic, event-triggered) (Aggarwal et al., 12 Apr 2025).

Perceptive Wireless: TWI as Time Aggregation

• TWI is analogous to the perceptual fusion window in neuroscience. Adjusting $x_k$3 offers fine-grained control over perceived time (“granularity of now”) and system-level resource allocation. Tight synchronization, throughput, and minimization of false causal orderings depend on careful window selection. Analytical bounds connect jitter, delay, and required window size for application-specific constraints (Popovski, 2024).

LLM Reasoning: Interleaved Generation

• The SxS scheme trains the model to emit “think” or “speak” tokens by aligning disclosure with entailment—tokens are only disclosed when their correctness is supported by prior private reasoning. RL with group-normalized rewards ensures accuracy is maintained, while token-level measurements (Average Inter-Response Wait, AIRW) quantify latency improvements (Wei et al., 5 May 2026).

Verifiable Reasoning: Meta-Prompted Trace Extraction

• Meta-prompting ensures the LLM marks verifiable steps within its free-form output. Sequential monitors check these steps and intervene on errors, achieving state-of-the-art performance in both early stopping and external verification scenarios. Notably, token overhead for verified correctness is much lower than generate-test or structured beam approaches (Bhat et al., 5 Feb 2026).

4. Empirical Findings and Comparative Performance

Control Systems

• The InterQ policy, trained via deep Q-learning, strictly dominates periodic and event-triggered baselines across rates $x_k$4 and disturbance types, producing superior Pareto fronts of control cost versus communication rate. Its almost-ellipsoidal switching rules are empirically validated (Aggarwal et al., 12 Apr 2025).

Reasoning Models

• Interwhen’s k-Stable early stopping reduces token usage by nearly 30% without any loss in accuracy, outperforming entropy- or confidence-based early stoppers. External verification with stepwise checking delivers up to +10 pp accuracy gains over best-of-N and Tree-of-Thoughts approaches, at only 1.3–1.5× the token cost (far below the 4–6× cost in baselines) while guaranteeing 100% soundness (if the process terminates) (Bhat et al., 5 Feb 2026).
• In SxS-style LLMs, learned disclosure policies yield Pareto-optimal tradeoffs: for instance, accuracy on AIME25 increased from 73.8% (standard CoT) to 80.0% with SxS, while AIRW decreased from 21 316 to 8 519; on GPQA, accuracy increased from 19.0% to 49.3% (Wei et al., 5 May 2026).

System/Domain	Baseline Acc.	Interwhen Acc.	Relative Token Cost	Soundness
Maze (External)	89.23%	98.74%	1.26×	100%
SpatialMap	82.71%	86.13%	1.22×	100%
GameOf24	96.82%	96.99%	1.42×	100%

All results from (Bhat et al., 5 Feb 2026).

5. Implementation and Deployment Strategies

• Verifier Integration: Domain-specific verifiers are packaged as Python or Z3 scripts; when not available, LLM critics serve as generic checkers (Bhat et al., 5 Feb 2026).
• Stateless Monitoring: Output streaming allows monitor intervention with minimal memory or compute overhead. Meta-prompted formatting ensures rapid extraction of verifiable states without model retraining.
• Adaptive Intervention: For safety-critical deployments, ensemble verifiers or multi-stage verification strategies are advocated. Timeouts, fallback policies, and safeguards address corner cases and prevent system hangs.
• No Model Retraining: Interwhen's meta-prompt+monitor design works directly with off-the-shelf LLMs, requiring only prompt engineering, not fine-tuning (Bhat et al., 5 Feb 2026).

6. Systemic Limitations, Extensions, and Open Questions

• Restriction to Verifiable Steps: Effectiveness depends on the existence of meaningful, intermediate states that can be checked mechanically or with LLM critics. Completely open-ended tasks may not admit such checkpoints (Bhat et al., 5 Feb 2026).
• Prompt Impact on Distribution: Injected feedback or termination tokens can shift the LLM’s output space, necessitating careful prompt design.
• Dynamic Scheduling and Windowing: For control and wireless systems, adaptive policies for window size ($x_k$5) or scheduling rates responsive to real-time network load or sensor uncertainty remain an open problem (Popovski, 2024).
• Integrations with ML Policy Learning: A plausible extension is combining sequential verification with process-reward models to handle soft-verifiable or gradable properties, as well as integration with multi-agent or tool-augmented workflows (Bhat et al., 5 Feb 2026).
• Completeness and Correction Guarantees: Theoretical analysis of whether sequential intervention is sufficient to universally steer a model toward correct outputs remains open.

7. Theoretical and Practical Significance

Interwhen frameworks codify a unifying principle: control over “when” to transmit, reveal, or act is fundamental to efficiency, accuracy, and trustworthiness in modern technical systems. Across control, wireless, and LLM reasoning, policy-driven, adaptive timing of intermediate interventions can provide near-optimal cost-performance tradeoffs while facilitating rigorous verification in high-stakes environments. The modular, model-agnostic structure of Interwhen (notably as public code—https://github.com/microsoft/interwhen) enables direct adoption in diverse reasoning, planning, and control domains (Bhat et al., 5 Feb 2026).