Budget-Aware Interoception

• Budget-aware interoception is defined as integrating real-time resource signals into LLM decision-making to optimize cost-performance efficiency.
• It employs techniques like control-token insertion and prompt-level budget blocks to dynamically track and adapt computational expenditures.
• Training methods, including supervised fine-tuning and RL, enable agents to strategically allocate limited resources, achieving significant accuracy and efficiency gains.

Budget-aware interoception refers to the explicit surfacing and utilization of an agent’s internal resource state—such as token, tool-call, or other computational budget—as an “interoceptive” signal during inference and decision-making. Unlike traditional approaches where resource constraints are externally enforced or at best statically encoded, budget-aware interoception equips LLMs and LLM-based agents with mechanisms to self-monitor, dynamically adapt, and strategically allocate limited inference time and resources. This paradigm yields substantial improvements in cost-performance efficiency, compositional planning, and real-time controllability in both standalone LLMs and tool-augmented agents.

1. Formalization and Motivation

Budget-aware interoception operationalizes the principle that agents should possess real-time knowledge of their own resource expenditure and remaining budget, integrating this information directly into their policy or reasoning loop. In its most direct instantiation, the resource state is represented as a “budget meter” surfaced to the model at each step—via control tokens, status blocks, or context injection—so that the agent conditions all subsequent decisions on this interoceptive input (Wen et al., 24 Aug 2025, Liu et al., 21 Nov 2025).

Motivations for budget-aware interoception include:

• Enabling fine-grained, token-level or tool-call-level spending control for cost-sensitive and latency-constrained scenarios.
• Facilitating strategic planning under hard or soft resource constraints, such as maximizing score over a sequence of tasks under a global token ceiling (Zhao et al., 7 Jan 2026).
• Overcoming performance saturation observed in static, non-budget-aware scaling regimes for LLM agents (Liu et al., 21 Nov 2025).

2. Principal Mechanisms: Control-Tokens and Budget Trackers

Various architectures have been developed for implementing budget-aware interoception, differing in the way the resource state is exposed and processed:

• Control-token insertion: BudgetThinker injects specialized control tokens $c_k$ into the generation stream at pre-specified intervals. Each token encodes the proportion of budget expended, transforming the leftover budget into an internal, proprioceptive “ping” that updates the hidden state and modulates ongoing inference (Wen et al., 24 Aug 2025).
• Prompt-level budget blocks: Budget-Aware Tool-Use (via the Budget Tracker plug-in) maintains per-tool counters for used and remaining calls. After each tool action, a standardized budget-status block is appended to the LLM context, providing both live numerical counters and qualitative policy advice. This allows the agent to “see” and respond to its actual resource state at each decision point (Liu et al., 21 Nov 2025).
• Tag-based projections: ROI-Reasoning encodes predicted budget decisions (e.g., “Level-0”, “Level-1”, etc.) as structured tags, teaching the LLM to anticipate and report its intended cost before committing to inference. This supports explicit solve/skip decision-making and forms a meta-cognitive interoceptive signal (Zhao et al., 7 Jan 2026).

These mechanistic variants all realize an internal sensory channel, allowing the agent to introspect and condition its behavior on live budgetary signals.

3. Training Paradigms for Budget Awareness

Budget-aware interoception is generally established through a multi-stage training protocol:

• Supervised fine-tuning (SFT): Models are exposed to budget-annotated data streams, such as interleaved control tokens at predefined intervals, budget-status markup blocks, or predicted effort-level tags. The objective is to ground the semantics of the internal budget signal and familiarize the model with its functional role (Wen et al., 24 Aug 2025, Liu et al., 21 Nov 2025, Zhao et al., 7 Jan 2026).
• Reinforcement learning (RL): Fine-tuned models undergo curriculum-based RL with reward signals that strongly penalize budget overruns, reward strict adherence, and incentivize optimal task performance per unit cost. For instance, BudgetThinker’s RL phase uses a length-aware reward combining accuracy, budget utilization, and severe penalties for exceeding $B$ (Wen et al., 24 Aug 2025). ROI-Reasoning employs PPO-style RL (Dr. GRPO), optimizing for long-horizon allocation under a hard global constraint (Zhao et al., 7 Jan 2026).
• Meta-cognitive cost prediction: ROI-Reasoning’s Meta-Cognitive Fine-Tuning (MFT) explicitly teaches the model to predict both anticipated cost and expected utility before reasoning, an approach that improves both budget adherence and accuracy under constrained conditions (Zhao et al., 7 Jan 2026).

4. Planning, Adaptation, and Verification Enabled by Interoception

Budget-aware interoception enables new agent behaviors and planning strategies not accessible to myopic (budget-oblivious) models:

• Dynamic planning and resource allocation: In the BATS (Budget-Aware Test-time Scaling) framework, agents decompose tasks into exploration and verification phases, adaptively widen or prune search tree branches according to remaining budgets, and maintain explicit inventories of tool consumption per subtask (Liu et al., 21 Nov 2025).
• Self-verification and strategic halting: Agents use verification modules to audit candidate solutions against the trajectory under the current budget, deciding to “dig deeper,” “pivot,” or “terminate” based on interoceptive signals and constraint satisfaction (Liu et al., 21 Nov 2025). Early stopping behaviors emerge, saving cost and preventing overthinking when the internal margin for exploration is exhausted.
• Solve-or-skip inference: ROI-Reasoning enables LLMs to predict expected ROI for each problem and make explicit skip decisions when anticipated benefit does not justify the expenditure, thereby maximizing global reward under budgeted multi-problem settings (Zhao et al., 7 Jan 2026).

5. Unified Cost Metrics and Empirical Scaling Analysis

Budget-aware interoception is evaluated using unified cost metrics and cost-performance scaling curves:

• Unified economic cost metric: Budget-Aware Tool-Use employs $$C_\text{unified}(x;\pi) = c_\text{token}(x;\pi) + \sum_{i=1}^K c_i(x;\pi)P_i$$, where $c_\text{token}$ is the token cost and $c_i$ the number of tool calls, each with price $P_i$ (Liu et al., 21 Nov 2025).
• Empirical results:
  • Budget Tracker alone improves efficiency, achieving 12.8% accuracy under tight tool budgets (vs. 10.3% for baseline) while using up to 40% fewer search actions at comparable accuracy (Liu et al., 21 Nov 2025).
  • BATS framework extends scaling curves, reaching 24.6% accuracy at 100 calls vs. 12.6% for standard approaches, and maintaining monotonic improvement at higher budgets (where non-interoceptive models saturate) (Liu et al., 21 Nov 2025).
  • BudgetThinker attains budget-following ratios up to 86.8% at $B=2000$ (vs. 47% for originals), and closely matches actual output length to allotted budget, with incremental performance gains in mathematical reasoning tasks (Wen et al., 24 Aug 2025).
  • ROI-Reasoning achieves near-optimal global reward (0.97 under budget 512, 1.13 under 1024), minimizing regret to as low as 0.02 in mixed-difficulty, multi-problem settings (Zhao et al., 7 Jan 2026).

Framework	Interoceptive Signal	Key Result at Tight Budget
Budget Tracker	Prompt-level status	+2.5 pp accuracy, -40% tool calls
BATS	Dynamic plan + verify	24.6% accuracy at 100 calls
BudgetThinker	Control tokens in gen	86.8% budget-following at $B=2000$
ROI-Reasoning	Predicted cost/utility	0.97 score (regret 0.11 at 512)

6. Theoretical Implications and Deployment Considerations

Instrumenting LLMs and agents with live resource counters and exposing this information as an interoceptive feature fundamentally alters computational behavior:

• Self-monitoring: The agent develops a continual sense of its internal computational margin, supporting adaptive allocation and “rational spending” behavior (Liu et al., 21 Nov 2025, Zhao et al., 7 Jan 2026).
• Cost-performance trade-off optimization: Budget-aware interoception allows models to trace the cost-performance Pareto frontier across varying budgets and task regimes, maintaining output quality efficiently as budgets tighten or loosen (Liu et al., 21 Nov 2025, Wen et al., 24 Aug 2025).
• Predictable latency: The ability to hard-stop generation at a prescribed budget without slack is critical for real-time and embedded deployment (e.g., autonomous systems), with no overrun risk (Wen et al., 24 Aug 2025).

This suggests that budget-aware interoception forms a general recipe for resource-constrained agent design: by supplying live, granular budget signals and enforcing policy adaptation via planning and self-verification, agent performance can scale more favorably and predictably under both tight and generous resource regimes.

7. Extensions and Future Directions

Current approaches primarily utilize discrete, quantized budget representations (e.g., control-token bins, budget tiers). Future directions indicated by the literature include:

• Extending to continuous budget signals, such as injecting exact remaining counts or using soft, learned internal representations (Wen et al., 24 Aug 2025).
• Generalizing to heterogeneous, multi-dimensional budgets (e.g., wall-clock time, transformer FLOPs, API quotas, memory).
• Scaling beyond fixed-task batches to variable-length, continually arriving workloads and non-uniform resource constraints (Zhao et al., 7 Jan 2026).
• Tight integration of interoceptive signals with hierarchical reasoning, multi-agent orchestration, and real-world tool ecosystems.

Budget-aware interoception thus represents a critical step in the progression toward LLMs and autonomous agents that reason, plan, and act efficiently under real-world, bounded-resource conditions, matching the strategic self-monitoring found in skilled human problem-solvers.