PaperlayerView exact revision on arXiv

Source-linked AI summary

Budget-Constrained Agentic Large Language Models: Intention-Based Planning for Costly Tool Use

Hanbing Liu, Chunhao Tian, Nan An, Ziyuan Wang, Pinyan Lu, Changyuan Yu, Qi Qi

arXiv:2602.11541v1cs.AIcs.LG

TL;DR

Budget-constrained tool use remains underexplored for agents operating with monetized, stochastic tools and hard monetary limits. INTENT uses intention-aware inference-time planning to estimate future costs and enforce budgets, improving task success across StableToolBench budgets and market settings while approaching the empirical upper bound.

  • Problem

    Budget-constrained tool use is underexplored despite agents needing to trade off information gain against expenditure in dynamic markets with hard monetary constraints.

  • Method

    INTENT uses intention-based inference-time planning with a learned world model and calibrated Monte Carlo lookahead to estimate future costs and guide budget-aware tool use.

  • Results

    INTENT achieves the highest pass rate across tested backbones while strictly respecting budgets, with moderate inference-time overhead across diverse budgets and market settings.

  • Takeaways & Limitations

    Budget-aware planning is essential for reliable and deployable agents that use external tools in open, dynamic markets.

  • Takeaways & Limitations

    Single-sample Monte Carlo cost estimates have high variance and can underestimate expected costs, causing budget overruns in deployment.

Abstract

from arXiv · show

We study budget-constrained tool-augmented agents, where a large language model must solve multi-step tasks by invoking external tools under a strict monetary budget. We formalize this setting as sequential decision making in context space with priced and stochastic tool executions, making direct planning intractable due to massive state-action spaces, high variance of outcomes and prohibitive exploration cost. To address these challenges, we propose INTENT, an inference-time planning framework that leverages an intention-aware hierarchical world model to anticipate future tool usage, risk-calibrated cost, and guide decisions online. Across cost-augmented StableToolBench, INTENT strictly enforces hard budget feasibility while substantially improving task success over baselines, and remains robust under dynamic market shifts such as tool price changes and varying budgets.

1. Introduction

The paper formulates tool-augmented agent operation as budget-constrained sequential decision making under stochastic outcomes, heterogeneous prices, and dynamic tool markets. It proposes INTENT, a lightweight inference-time planner that estimates future costs through intention-aware world modeling and calibrated lookahead without retraining or environment interaction.

  • Motivation: Agents access thousands of heterogeneous APIs, expanding the action space while exposing economically consequential resources with real monetary costs.Examples include real-time financial feeds, high-resolution satellite imagery, and irrevocable blockchain state changes.
  • Problem: Explicit budget feedback remains insufficient: strong models frequently exceed hard budgets through repetitive retries and unproductive exploration.The setting is instantiated on StableToolBench with hard task budgets and diverse per-call tool prices.
  • Method: INTENT combines a learned language world model with calibrated Monte Carlo lookahead and intention-based decomposition to estimate future costs during inference.The decomposition separates whether a tool call satisfies the agent’s semantic intention from the concrete output content.
  • Results: INTENT improves StableToolBench performance across diverse budgets and market settings, approaching the empirical upper bound with mild overhead.The framework enables budget-aware decisions without retraining or environment interaction.

2. Model

The paper models tool-augmented agents as sequential decision makers operating over growing textual histories and dynamic tool markets. Agents alternate reasoning, tool calls, stochastic observations, and termination while maximizing answer quality under a hard budget.

  • Context Space: The agent state is its full interaction history, represented as a monotonically growing concatenation of serialized tokens.This uniform context space embeds tool specifications, feedback, reasoning traces, calls, and answers.
  • Environment and Tasks: Task instances combine a user query, hard budget constraint, and market snapshot listing currently accessible tools with per-call costs.The market can vary by retrieval, permissions, or provider status.
  • Sequential Interaction: At each step, the policy generates reasoning, selects either a structured tool call or final answer, and appends stochastic tool observations to history.Tool execution is governed by an external environment and incurs the selected tool’s per-call cost.
  • Reward and Objective: The objective maximizes expected answer-quality reward while enforcing hard budget feasibility, assigning zero reward to trajectories that exceed the budget.Answer quality is measured by J(aK, q) ∈ [0, 1].
  • Planning Motivation: Offline optimization is impractical because free-form tool arguments create huge exploration and credit-assignment costs, while non-stationary markets prevent reliable inference-time adaptation.These challenges motivate lightweight online planning with a learned world model, without modifying agent parameters.

3. Methodology

The methodology introduces INTENT, an inference-time planner that uses lightweight world-model simulations to enforce hard budgets while minimally intervening in a capable pretrained agent. It improves on single-rollout budget checking by modeling tool-call intention, ideal trajectories, and risk-adjusted future cost.

  • Framework Overview: INTENT performs lightweight online lookahead with a learned world model to anticipate future tool usage while dynamically enforcing hard budget constraints.The framework is designed for non-stationary tool markets and aims to minimally intervene in a strong pretrained agent.
  • World Model: The world model predicts tool-execution outcomes, emphasizing output format and structure rather than perfectly reproducing factual tool responses.This structural prediction is intended to elicit the agent’s latent planning logic.
  • MCO: Unlike exhaustive tree search, MCO uses a single trajectory rollout that alternates the world model and agent policy, then accepts actions only when projected cost fits the remaining budget.Exhaustive search is infeasible because free-form tool arguments create an unbounded action space and agent inference is costly.
  • MCO Limitation: MCO’s single-sample cost estimate can have high variance under stochastic tools, allowing underestimated costs and potential deployment-time budget overruns.Costly retry loops caused by irrelevant tool results are an example of this risk.
  • INTENT: INTENT introduces an intention-aware decomposition that estimates whether a proposed tool call satisfies the reasoning-encoded intention and conditionally generates its observation.It uses a binary satisfaction variable z_t and separates intention prediction from conditional observation generation.
  • Risk-Calibrated Planning: INTENT simulates an ideal trajectory with intention satisfaction forced at every future step, then calibrates expected cost pessimistically and applies a risk preference before accepting actions.Rejected actions receive predicted success probabilities to identify high-risk bottlenecks for targeted replanning.

4. Experiments

Experiments evaluate INTENT on cost-augmented StableToolBench under hard budget constraints and dynamic tool-market conditions. INTENT achieves the strongest overall balance of task performance, budget compliance, and inference efficiency while remaining robust to market changes.

  • Experimental goals: Experiments assess hard-budget compliance, inference-time strategy effectiveness, and robustness to price changes, new tools, and varying budgets.The evaluation covers budget-constrained and dynamic tool-market settings.
  • Experimental setup: StableToolBench is augmented with independently sampled tool prices under a fixed budget B = 50, using 20 retrieved tools per query and 765 test instances.Each tool cost is sampled from U(5, 50) to form an instance-specific market snapshot.
  • Evaluation metrics: Evaluation reports PERFORMANCE, COST-AWARENESS, and EFFICIENCY through pass, budget-optimal pass, win, feasible-rate, cost, price, time, latency, and token-consumption metrics.Efficiency values are reported relative to RAW, while performance is judged using final-answer quality and tool-call traces.
  • Main results: 32.8% of GPT 4.1 mini PROMPT tasks violate budgets, showing that instruction-based cost awareness cannot reliably enforce hard constraints.PROMPT improves pass rate over RAW but remains distant from the budget-optimal frontier.
  • Main results: INTENT consistently achieves the highest pass rate across non-reasoning and reasoning backbones while strictly respecting budgets with only moderate inference-time overhead.Enforce baselines guarantee feasibility but trade off performance or efficiency, with BATS incurring particularly high overhead for reasoning models.
  • Dynamic-market robustness: INTENT remains robust under price perturbations and varying budgets, adapts across resource regimes, and requires only periodic lightweight-oracle updates without changing the underlying language model.Its tight-budget performance can match PROMPT under substantially larger budgets.

5. Related Work

Prior agentic-AI research expands tool-use capabilities and efficiency, but largely overlooks hard monetary constraints and dynamic costs. Language world models support planning, yet INTENT addresses stochastic tools and costly retries through intention-level outcome abstraction for budget-aware planning.

  • Agentic AI: Agentic AI equips frontier LLMs with reasoning and tool-use capabilities for complex, multi-step tasks across research, GUI control, and software engineering.These systems have demonstrated strong performance in deep research, information synthesis, GUI control, and software engineering.
  • Agentic AI: Most existing agents expand capabilities through diverse tools, long-horizon trajectories, or multi-agent coordination while optimizing unconstrained task success.Such methods typically permit repeated tool calls until sufficient information is obtained.
  • Economic Constraints: Existing efficiency methods mainly reduce computational cost or latency through token-efficient reasoning and fewer or approximated tool calls, rather than monetary expenditure.The related work distinguishes reasoning compression and speculative decoding from approaches that reduce or approximate tool calls.
  • Economic Constraints: Existing agents do not model monetized tools, irreversible retry costs, or dynamic information-gain–expenditure trade-offs, motivating hard budget-constrained sequential decision making.The paper explicitly formalizes budget-constrained tool use under hard monetary constraints.
  • Language World Models: Language world models simulate environment dynamics for planning, but INTENT abstracts stochastic tool outcomes by whether calls satisfy intentions, enabling reliable cost estimation and budget-aware planning.Prior inference-time applications mainly target deterministic environments, whereas the proposed setting includes high stochasticity and costly retry loops.

6. Conclusions

The paper frames budget-aware tool use as essential for reliable agentic AI and proposes INTENT, an intention-based inference-time planner for enforcing hard resource constraints without retraining or heavy search.

  • Conclusions: Budget-aware tool use is presented as a fundamental challenge for reliable and deployable agentic AI operating through external tools in open, dynamic markets.Agents increasingly rely on external tools to create real-world impact while reasoning under hard resource constraints.
  • Conclusions: INTENT enforces budgets by anticipating whether future tool interactions satisfy high-level intent rather than predicting exact tool outcomes.This intention-based abstraction enables effective budget enforcement with minimal intervention, without retraining or heavy search.
  • Conclusions: The study argues that budget awareness should be a first-class objective, with intention-level reasoning connecting stochastic environments to resource-constrained decisions.The authors call for further exploration of lightweight, inference-time control mechanisms for real-world agentic systems.

A. Oracle Implementations … B. Implementation Details

The paper implements INTENT as an inference-time, oracle-guided budget-constrained agent using learned world, intention-generation, and calibrated intention-prediction models. Its algorithms combine hard budget checks with rollout reuse, repeated-action handling, and low-success tool exclusion to improve practical efficiency.

  • A.1. Training: Training uses naturally collected ReAct-style user-conversation trajectories without additional data collection or synthetic trajectories.Each task instance contains a query, budget, market, and interaction trajectory.
  • A.1. Training: The language world model predicts tool responses from tool-interaction records, while the conditional generator models responses from positive-intention samples.Both components are implemented as fine-tuned large language models trained by next-token prediction.
  • A.1. Training: The intention predictor estimates calibrated tool-call success probabilities from reasoning, context, and action inputs.It is an encoder-only Transformer with a classification head, trained with cross-entropy and post-hoc temperature scaling; the calibrated score supports geometric cost estimation.
  • A.2. Additional Mechanisms: Implementation adds lightweight mechanisms that reduce redundant rollouts and unnecessary exploration without changing the oracle’s underlying decision rules.The mechanisms are incorporated into the main algorithm for computational efficiency in practical deployments.
  • A.2. Additional Mechanisms: The Rollout Cache reuses an accepted ideal-trajectory action sequence when the next proposal approximately matches and remains budget feasible; cache misses trigger replacement by a new rollout.The Last Call Cache directly accepts a repeated rejected action when it is budget feasible, then clears the cache.
  • A.2. Additional Mechanisms: During rollout, tools with predicted success probability ˜ρt < δ are blacklisted and permanently removed from the candidate action space for the current task.This mechanism excludes low-probability tools from subsequent decisions within the task instance.
  • A.3. Algorithms: The budgeted-constrained agent proposes actions through a ReAct loop, while an inference-time oracle decides whether each costly tool call is accepted for real execution or rejected with feedback.Accepted actions execute in the environment and deduct COST(at) from the current budget; rejected actions provide oracle output, and budget exhaustion returns FAILURE.
  • A.3. Algorithms: The Monte Carlo Oracle simulates a future trajectory with the language world model and agent policy, accepts when estimated total cost is within budget, and otherwise returns predicted future actions as hints.INTENT additionally maintains a risk factor, rollout cache, last rejected action, blacklist, threshold, and market snapshot.

B.1. Training and Inference Details

INTENT’s components are trained on StableToolBench-derived interaction data with intention-satisfaction annotations, using Qwen-based architectures and standardized train/development/test splits. Training uses AdamW with component-specific hyperparameters and BF16 full-parameter fine-tuning, while experiments run on a single RTX Pro 6000 GPU.

  • Data: The Language World Model and Conditional Generator use approximately 100k MirrorAPI-Cache tool-interaction trajectories, retaining 28k trajectories annotated for intention satisfaction.GPT-4.1-mini provides the intention-satisfaction annotations.
  • Data: The Intention Predictor is trained on 86k annotated (r, a, o) triples, with all datasets split 8:1:1 and development data used for checkpoint selection and probability calibration.GPT-4.1-mini annotates whether each tool outcome satisfies the agent’s expressed intention.
  • Base Models: The Language World Model and Conditional Generator use Qwen2.5-3B-Instruct, while the Intention Predictor uses Qwen3-0.6B-Embedding with an additional classification head.
  • Optimization: All components use AdamW and BF16 full-parameter fine-tuning for 2 epochs, with batch size 64 and learning rate 7 × 10−5 for two components versus 32 and 5 × 10−5 for the Intention Predictor.The first two components use 5% linear warmup followed by cosine decay.
  • Infrastructure: Experiments run on a single NVIDIA RTX Pro 6000 GPU, using LLaMA-Factory, Transformers, scikit-learn, FlashAttention-2, vLLM, and FastAPI across training, calibration, and serving.

B.2. Algorithm Implementation

The experiments implement budget-aware versions of DFSDT, BTP, and BATS alongside MCO and INTENT, using fixed search, inference, caching, and risk-control settings. DFSDT prunes actions exceeding the remaining budget, while other baselines are adapted or configured for the unified budget setting.

  • DFSDT: DFSDT prunes any branch whose proposed action exceeds the remaining monetary budget.This augments the original implementation, which does not explicitly account for monetary budgets.
  • DFSDT: The search width is fixed to 10 in all experiments.
  • BTP: BTP uses DFSDT, reproduction data, GPT-4.1-mini tool-score evaluation, Qwen3-0.6B-Embedding similarity, and blacklist threshold τ = 0.15.The blacklist mechanism follows the original setting.
  • BATS: BATS is adapted from per-tool budgets to a unified global budget and capped at K = 5 Self-varification Module calls per iteration.The original BATS pipeline and prompts are faithfully reproduced before adaptation.
  • INTENT: INTENT enables rollout, last-call, and blacklist caches with fixed γ = 0.5 and δ = 0.1, without task-specific hyperparameter tuning.Its Conditional Generator sampling temperature is set to 0.3.

B.3. Evaluation Details

Evaluation uses budget-aware automatic pass and win rates, with GPT-4.1-mini judging and GPT-5-mini generating WIN RATE references. An empirical budget-optimal upper bound is estimated through exhaustive DFSDT search for valid tool-use solutions.

  • Automatic Evaluation: PASS RATE and WIN RATE retain StableToolBench’s SoPR and SoWR logic while requiring solutions to satisfy the budget constraint.The evaluator model is upgraded to GPT-4.1-mini.
  • Automatic Evaluation: WIN RATE uses a reference solution generated by the PROMPT baseline with GPT-5-mini.
  • Budget-Optimal Pass Rate: DFSDT with width w = 10 and GPT-5-nano exhaustively searches trajectories to estimate the BUDGET-OPTIMAL PASS RATE upper bound.Each query is searched until five unique budget-satisfying solutions are collected.
  • Budget-Optimal Pass Rate: A query is marked solvable under the budget when at least one enumerated solution is judged correct by the evaluator.

C. Group-wise Breakdown of Pass Rate · D. Further Experiments

The paper reports group-wise pass rates on cost-augmented StableToolBench for completeness, despite the groups being largely orthogonal to cost-sensitive decisions. Further experiments analyze INTENT through ablations, efficiency mechanisms, and sensitivity analysis.

  • C. Group-wise Breakdown of Pass Rate: StableToolBench partitions tasks into predefined groups that are largely orthogonal to the paper’s cost-sensitive decision setting.The authors therefore treat the grouping as separate from their modeling assumptions.
  • C. Group-wise Breakdown of Pass Rate: Table 2 reports disaggregated Pass Rate across six benchmark-defined task groups.The breakdown is conducted on cost-augmented StableToolBench.
  • C. Group-wise Breakdown of Pass Rate: The group-wise PASS RATE breakdown is reported for completeness rather than as a direct evaluation of the modeling assumptions.The passage explicitly distinguishes the benchmark grouping from the paper’s cost-sensitive setting.
  • C. Group-wise Breakdown of Pass Rate: Oracle-based methods consistently outperform other baselines across all task groups.This cross-group ordering is stated as a general result of the breakdown.
  • C. Group-wise Breakdown of Pass Rate: With the GPT 5 nano backbone, INTENT performs uniformly strongly across groups and closely tracks the empirical upper bound.The empirical upper bound is marked in gray in the reported breakdown.
  • D. Further Experiments: The additional experiments analyze INTENT through component ablations, efficiency mechanisms, and sensitivity analysis.These experiments are organized in Appendices D.1, D.2, and D.3.

D.1. Ablation of Oracle Components · D.2. Ablation of Additional Mechanisms

INTENT’s calibrated intention prediction and ideal-trajectory simulation are both necessary for stable, effective budget-aware planning. Additional caching and blacklisting mechanisms primarily improve efficiency, while task performance remains largely stable.

  • D.1. Ablation of Oracle Components: Fine-grained ablations selectively modify the Intention Predictor and Conditional Generator while keeping the underlying agent and algorithmic logic unchanged.Results are summarized in Table 3.
  • D.1. Ablation of Oracle Components: The Intention Predictor ablation compares OPTIMISTIC, UNCALIBRATED, and EMPIRICAL MEAN variants for estimating future tool-call success probabilities.OPTIMISTIC assumes all future tool calls succeed with probability one; UNCALIBRATED uses raw predictor probabilities; EMPIRICAL MEAN uses training-data success frequencies while ignoring contextual reasoning and arguments.
  • D.1. Ablation of Oracle Components: Figure 6 shows that post-hoc calibration substantially reduces predicted-probability misalignment and improves expected calibration error without collapsing probabilities to degenerate extremes.The figure compares predicted probabilities with empirical success rates before and after calibration and shows improved probability spreading.
  • D.2. Ablation of Additional Mechanisms: Table 4 evaluates the additional mechanisms used in INTENT by removing rollout caching, last-call caching, and blacklisting.These mechanisms are tested separately for their effects on efficiency and task performance.
  • D.1. Ablation of Oracle Components: The Conditional Generator ablation evaluates STOCHASTIC, NO ROLLOUT, and NO TRACE alternatives that remove or weaken latent-plan elicitation and future-trajectory guidance.STOCHASTIC samples stochastically from the language world model; NO ROLLOUT disables lookahead and checks only immediate risk-adjusted cost feasibility; NO TRACE removes the oracle-predicted trajectory upon rejection.
  • D.1. Ablation of Oracle Components: Consistent performance degradation after ablating either component confirms that calibrated intention prediction and ideal-trajectory simulation are essential for stable, effective budget-aware planning.Table 3 reports overall PASS RATE for INTENT oracle component ablations, with removing or simplifying either module degrading performance.
  • D.2. Ablation of Additional Mechanisms: Removing any additional mechanism increases end-to-end time and latency, while overall task performance and Pass Rate remain largely stable.The results indicate that rollout caching, last-call caching, and blacklisting primarily improve efficiency without altering core decision logic.

D.3. Sensitivity Analysis … Action 6 (ANSWER)

INTENT balances task success and cost through risk-sensitive planning, with intermediate γ values offering a broad high-performance, lower-cost region. In the case study, intention-aware lookahead redirects tool selection toward a successful Google financial-analysis trajectory that remains within the 50-credit budget.

  • D.3. Sensitivity Analysis: As γ increases, average cost decreases while pass rate gradually declines, producing a smooth performance–cost trade-off without abrupt degradation.Larger γ penalizes uncertain future tool usage more heavily and encourages conservative behavior.
  • D.3. Sensitivity Analysis: At γ ≈0.5–0.7, INTENT reaches a broad sweet spot with high pass rate and substantially reduced average cost, limiting sensitivity to precise tuning.This intermediate range provides flexible control over the performance–cost trade-off.
  • E. Case Study: The representative task asks for Google’s quarterly financial data under a 50-credit budget, with 30 heterogeneous-cost tools creating many feasible tool chains.The requested outputs include cash flow, balance sheet, income statement, ratios, net income, total assets, and revenue.
  • E.1. Failure of Direct Prompt: PROMPT fails despite explicit budget and tool-cost information because sequentially retrieving expensive data pushes expenditure beyond the allowed budget.Its trajectory spends 38 credits on cash flow and then reaches 61 credits after the balance-sheet call, exceeding the budget by 11 credits.
  • Action 1 (TOOL CALL): INTENT rejects an initial plan whose risk-adjusted cost is 60.36 > 50 because the cash-flow call has predicted success probability 0.35.The oracle exposes anticipated actions, prices, success probabilities, and expected costs while withholding predicted arguments and simulated observations.
  • Action 2 (TOOL CALL): The agent then switches to cheaper alternatives, but INTENT blocks the Freedom Finance plan because its low predicted intention-satisfaction probability could trigger retries and exhaust the budget.Optional BLACKLIST can temporarily remove extremely low-probability tools from the candidate set to stabilize subsequent planning.
  • E.2. Solution Trajectory of INTENT: INTENT approves a Holistic Finance plan after rollout shows lower projected cost and higher predicted intention-satisfaction probability, with risk-adjusted expenditure within the remaining budget.The approved cash-flow call costs 9 credits, leaving 41 credits, and later income-statement and balance-sheet calls leave 30 and 7 credits respectively.
  • Action 6 (ANSWER): The final INTENT answer successfully reports Google’s requested financial information, including a current ratio of approximately 0.73, while completing the task within budget.The case demonstrates that INTENT reshapes search using anticipated-plan risk and then reuses validated rollouts for smooth execution with minimal intervention.
Loading 2602.11541v1…