acceptodds
Under review as a conference paper at ICLR 2027

Towards Budget-Aware Agents: Do Agents Know What They Will Spend?

Abstract

Foundation-model agents operate under growing resource constraints yet rarely know how much budget they will spend. We call this capability **budget awareness** and formalize it as **progressive interval estimation**: mid-execution, can the agent provide an interval that covers its remaining budget and predict when its continuation will not complete within budget? We score this with a rollout-replay protocol that re-queries the agent on every trajectory prefix, decomposing estimation into feasibility prediction, early failure detection, and interval estimation. We evaluate five frontier models across four environments, including internal token budgets (Sokoban, Search-R1, SWE-bench) and external multi-dimensional budgets (Warehouse), and train Qwen-7B estimators with SFT and RL. Task success is only a loose proxy for estimation quality: frontier models differ significantly as estimators, but not in order of success. Estimation fails in structured ways: agents underestimate remaining cost on search and coding, recognize failure late, and widen intervals with task scale rather than continuation uncertainty. At a matched loss of solved tasks, the models’ infeasibility signal saves more tokens than a hindsight-tuned spending rule in 9 of 15 model–environment cells. SFT makes binary feasibility learnable in-domain, whereas interval estimation stays task-specific.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Budget-Aware Agents, Agent Evaluation, Resource-Aware Planning, Agent Reliability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YMm1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper studies the problem of whether LLM agents completing multi-step tasks can estimate how much they will spend, both in terms of "internal budget" (tokens) and "external budget" (costs in the environment). This paper's contribution, relative to previous work, is to elicit budget estimates at intermediate steps (rather than just at the beginning). This allows it to diagnose the phenomenon in a more fine-grained way, for example by looking at how late in the rollout LLMs realize they are running over budget ("Hypothesis 4"). They test 5 LLMs, four frontier-level LLMs and also Qwen 3 235B. They also do a fine-tuning experiment on Qwen 2.5 7B to see whether budget estimation capabilities can be improved with SFT + RL. The estimation experiments are in an offline setting, in which an estimator LLM in a separate query is given a partial transcript and asked to estimate the remaining spend, rather than an online setting, in which the LLM agent is estimating its spend as it completes the task.

Critical strengths

I think the experimental design is good. Multiple established datasets, and also multiple reasonable metrics, are considered. I also think Section 5 ("Why does budget estimation fail?") is reasonably well-executed.

Critical weaknesses

W1. I think the current setup is too stylized and not super interesting. Section 5 ("Why does budget estimation fail?") mostly reports on-path summary statistics about how the LLMs estimate their budget. For example, whether they update their beliefs based on how the task is going, or how late in the task they realize they aren't going to hit the budget. I think this analysis does a good job of explaining what is going on. But I would push back on the statement that this is explaining why budget estimation fails. I think within the current offline estimation framework I would at least want to see an analysis of what the estimator LLM's reasoning process looks like. But really I think there is so much more insight to be gained from the online, "closed-loop stopping" setting. The kinds of questions it can answer are so much richer, for example: To what extent are the LLMs reasoning carefully about how many tokens they think each step is going to take? To what extent do they look back and do a post hoc analysis? To what extent do they change their approach to the task to fit within the budget, as opposed to just doing what they would have done anyway and keeping track of budget? I think answers to these questions would actually shed light on how LLMs reason about the costs they incur.

W2. The prose is obscure and difficult to understand. The results in particular are not well-explained. They read like Claude Code outputs. I think some human attention here would improve the paper.

Questions for authors

See the questions under W1.

2 comments

We thank the reviewer for the careful reading, and for describing the experimental design as "good," with "multiple established datasets, and also multiple reasonable metrics," and Section 5 as "reasonably well-executed."

[W1] Online stopping

Used online, the infeasibility prediction stops failing runs early and wrongly stops fewer solved runs than a token cap. We ran Opus 5.5 online on 32 Sokoban levels (seeds 123–154) under the paper's 2,500-token budget. In condition C, after every turn the same model estimates its remaining cost and predicts whether it can finish within budget; the estimate is shown to the agent, and the run stops when the model predicts it cannot finish. A is the same agent without estimates, and A2 repeats A to show run-to-run noise. B stops a run after 1,400 tokens, a cap fit on offline Opus 4.7 data. E applies only the stop rule to A's runs, without showing the estimate to the agent.

Table R3: Online Sokoban, 32 levels, one run per level per condition.

ConditionLevels solvedAgent tokens per levelEstimation tokens per levelSolved levels wrongly stopped (of the 26 A solves)
A (no estimate)26/321,8450-
A2 (repeat of A)24/321,89400
B (token cap, 1,400)14/321,456012
E (stop rule only)16/321,2182,96110
C (estimate shown + stop rule)18/321,3223,0588

C stops 5 of the 6 levels that A fails and wrongly stops 8 of the 26 levels that A solves, against 12 for the 1,400-token cap. To compare at the same number of wrong stops, as Table 2 does, we sweep the cap on A's log in hindsight. The best cap that wrongly stops 8 solved levels (1,670 tokens) saves 0.30 of the tokens on the 6 levels A fails, and the best cap that wrongly stops 10 (1,490 tokens) saves 0.35, against 0.48 for C and 0.59 for E. A and A2 differ by 2 levels without any intervention, so we read these as counts; they point the same way as Table 2, where the infeasibility prediction saves more tokens than a hindsight-tuned cap in 9 of 15 model–environment cells. The online run uses a standalone harness on its own level set (seeds 123–154), so its success rates are not comparable with Table R2.

This setting measures the value of the signal at its highest cost. Here the same frontier model estimates at every turn, so estimation adds about 3,000 tokens per level and the total cost of C is about 2.4 times that of A. Section 5.1 shows that SFT raises a 7B model's balanced feasibility accuracy from 0.255 to 0.911, so the signal can come from a small estimator; reducing its cost, with a small estimator or by estimating every few turns, is a future direction outside the scope of this paper.

Budget awareness does not yet change how the agent works. After seeing its estimate, the agent's output per turn shrinks by the same amount as without the estimate (difference +44 tokens per turn, 95% CI [−4, +98], 22 levels), and 0 of 54 turns mention budget or tokens. Using the estimate inside the agent's own decisions is therefore a separate capability, and the online run gives the baseline for it.

[W1] What the estimates rest on

In budget estimation, models mostly count "turns left × tokens per turn" and rarely use their own token spending record. We labeled the reasoning text of 7,780 estimates from four models across the four environments by what it does, with Sonnet 5.5 as labeler; Gemini has too few SWE-bench reasoning texts to label. On 35 estimates we checked by hand, the labels agree on 343 of 385 decisions (0.89).

Table R4: Share of estimates whose reasoning does each step.

Step in the reasoningSokoban: GPT / Opus / Sonnet / GeminiSearch-R1: GPT / Opus / Sonnet / GeminiSWE-bench: GPT / Opus / SonnetWarehouse: GPT / Opus / Sonnet / Gemini
States a cost per turn0.97 / 0.87 / 0.91 / 0.570.73 / 0.68 / 0.61 / 0.460.12 / 0.21 / 0.200.28 / 0.28 / 0.51 / 0.27
Counts turns left0.96 / 0.98 / 0.97 / 0.640.99 / 0.97 / 0.98 / 0.940.46 / 0.56 / 0.640.58 / 0.74 / 0.71 / 0.75
Compares with the budget1.00 / 0.29 / 0.70 / 0.801.00 / 0.22 / 0.30 / 0.200.91 / 0.50 / 0.341.00 / 0.78 / 0.95 / 0.96
Uses the spending of past turns0.77 / 0.01 / 0.14 / 0.050.12 / 0.00 / 0.06 / 0.000.07 / 0.02 / 0.050.22 / 0.16 / 0.47 / 0.12

On how carefully models reason about tokens per step: in Sokoban and Search-R1 most estimates state a cost per turn and count turns left, and in SWE-bench 0.12–0.21 state a cost per turn. On whether they look back: only 0.15 of the estimates in the token environments use the spending of past turns (485 of 3,235), and outside GPT-5.2 on Sokoban the share is at most 0.14. In Warehouse, where the budget is money and inventory over weeks, more estimates refer to past periods (0.12–0.47) and nearly all compare with the budget (0.78–1.00). The paper's control in Section 4.1 reaches the same conclusion from the other side: removing turn count and per-turn usage from the prompt leaves GPT-5.2's feasibility F1 unchanged (0.406 vs. 0.406). Table R4 shows the same property in the models' own reasoning.

[W2] Prose

We rewrite the hardest results paragraphs so that each one reads on its own. In the revised PDF, the rewritten paragraphs in Sections 4–6 are marked in blue. Each states the comparison first, then the number, then what it means for budget awareness. The new experiments, including the online run and the labeling scheme with its hand check, are in Appendix L, also in blue. We will also use "why" in Section 5 only for statements the data supports.

Reviewer_uH3jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether an agent can estimate its remaining resource use from an execution prefix. It records budget-hidden rollouts, then separately asks the same model for a remaining-cost interval or an infeasibility prediction. Experiments cover Sokoban, Search-R1, SWE-bench, and a multi-resource Warehouse simulation. Reported failures include biased estimates, late alarms, and interval widths that track task scale more than continuation spread. SFT and GRPO probes test learnability, while offline stopping analysis compares the infeasibility signal with a spending threshold. The appendices add rollout-clustered uncertainty, simple baselines, format controls, and branching experiments.

Critical strengths
  1. Asking for estimates at intermediate prefixes is a useful way to study agent self-monitoring. The paper goes beyond correlating success with token use: it evaluates alarms and interval estimates separately, compares with simple feasibility rules, and examines multiple continuations from shared prefixes. These controls make the observed weaknesses worth investigating.

  2. Several negative results are informative. A one-feature spending rule matches or beats zero-shot feasibility predictions in every internal-budget cell, and strong in-domain binary learning does not produce general interval estimation. The paper also reports false aborts alongside offline savings and distinguishes task-bootstrap uncertainty from the much smaller training study.

Critical weaknesses
  1. Interval quality lacks a fixed statistical target. The prompt requests a range that is both tight and covering, with no nominal coverage level; Eq. 4 then rewards a narrow interval that happens to contain a single sampled continuation. Consequently, low hit rate alone cannot establish poor calibration or an inability to estimate uncertainty, and model hit-rate rankings can reflect different width choices.

  2. The model coverage is dated for the paper's claims about frontier agents. The main evaluation includes Claude Opus 4.7 and Sonnet 4.6, although newer generations such as Opus 5 and Sonnet 5 were available before submission. Since the paper draws broad conclusions about agents' budget-estimation limitations, it is important to establish whether the reported optimism, late failure detection, and weak interval estimates persist in more recent models. The current results do not distinguish persistent limitations from behavior specific to the evaluated model generations.

Questions for authors

Please see weakness 2

1 comment

We thank the reviewer for the detailed reading, and for describing intermediate-prefix estimation as "a useful way to study agent self-monitoring" and noting that "several negative results are informative."

[W1] Stated probability for the cost interval

Telling the model how sure to be has a limited effect on budget awareness. The original prompt asks for a tight cost interval without a probability. We add "give an interval that you believe contains the true remaining tokens with 80% probability" and rerun the same steps (Sokoban: 188–215 steps per model; Search-R1: 52–201). Hit rate is the share of cost intervals that contain the true remaining tokens.

Table R1: Hit rate with and without the 80% request.

EnvModelHit rate, original promptHit rate, 80% promptHow many times wider the interval must be to reach 80%
SokobanGPT-5.20.360.383.5
SokobanOpus 4.70.470.502.4
SokobanSonnet 4.60.450.392.8
Search-R1GPT-5.20.210.284.8
Search-R1Opus 4.70.230.332.6
Search-R1Sonnet 4.60.370.463.1

With the 80% request, hit rate stays at 0.28–0.50, and the models widen their intervals only slightly; keeping each model's midpoint, the interval would have to be 2.4–4.8 times wider to reach 80%. The model ranking does not depend on this choice: on Search-R1 the order of the three models is the same under both prompts, and on Sokoban Opus 4.7 has the highest hit rate under both. The midpoint itself misses by 0.32–0.55 of the true remaining cost, an error that does not depend on width. Removing "as tight as possible" as well gives Sokoban hit rates of 0.45, 0.50 and 0.37. The calibration gap the paper reports therefore remains when the target is stated, and Table R1 is added as Appendix L.1.

[W2] Newer models

We ran gpt-6.1-sol and Opus 5.5 on the same 128 Sokoban levels under the same 2,500-token budget; each model estimates its own runs, scored as in Table 1. We use Sokoban because the paper's branching, training and online analyses all run there. Table R2 gives the three properties the reviewer lists.

Table R2: The three properties on Sokoban, older and newer models.

PropertyOlder frontier modelsgpt-6.1-solOpus 5.5On newer models
Late failure detection: failed runs flagged before 60% of the budget0.04–0.230.730.93Improves
Interval width grows with task size: correlation0.44–0.670.860.51Persists
Direction of interval misses: share conservative (first 40 levels)0.79–0.940.940.84Same as older models

Late failure detection is a property of the generation in the paper, and newer models improve on it. Interval width still grows with task size. The direction of misses still depends on the environment: Section 5.2 reports optimistic misses on search and coding and conservative misses on Sokoban, with optimism in 15 of 20 model–environment pairs across five models from four providers, and the newer models keep the Sokoban pattern. After the rebuttal we will give more comparison on more models. In the revised paper, Section 5.4 states late recognition for the models in Table 1 and reports the newer models next to it (Appendix L.2).

Per-model numbers are in Table R6 (reply to Fuj2) and Appendix L.2. Opus 5.5 solves 0.554 of the levels within budget (0.976 when the budget is ignored) and has the lowest hit rate among the frontier models (0.228). It refuses prompts that ask for <think> tags, so it uses an answer-only prompt; the per-turn usage shown to its estimator includes the hidden reasoning tokens, which count toward the budget, and 54 of its 57 failing runs solve the level but go over the budget, so part of its early flagging may come from reading how fast it spends. gpt-6.1-sol solves 0.914 of the levels within budget, so 11 of its 128 runs fail. All models play the same levels, and we do not add harder levels for any single model, since that would change the task distribution and break comparison across rows. The paper reports budget-awareness properties for the models it evaluates, and showing how newer generations differ is what our benchmark is for.

Reviewer_Fuj2OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces an evaluation framework for measuring how well existing models estimate the remaining budget in multi-step agentic tasks. At each step, the model is prompted with the past steps and its usage so far and asked to estimate what remains. The authors test both closed and open-source models and present their findings.

Critical strengths

Budget estimation is an important problem, especially given the high token usage of complex agentic tasks. A framework and analysis that study this problem are therefore valuable to the community. The framework and metrics the authors propose in Sections 3.1-3.3 are reasonable.

Critical weaknesses

I personally think the key value of an analysis paper is to reveal insights that generalize beyond its experimental setup and point toward possible solutions, which I think this paper lacks. The early-stopping and fine-tuning experiments are limited to offline simulation and a single small-scale probe. Budget estimation is known to be difficult, as shown in prior cost estimation work, and the paper's analysis seems mainly to confirm this.

From a methodological point of view, the main evaluation relies on prompting (Appendix J), which also raises concerns about how useful the insights are. Prompted behavior depends heavily on the specific model and version (e.g., Opus 5.5 behaves differently from Opus 5, and Gemini differs from Opus), and on its own it says little about the models' internal behavior. As a result, the analysis reads as a study of the chosen models that offers limited generalizability to future models.

Questions for authors

See weakness

1 comment

We thank the reviewer for the careful reading, and for noting that budget estimation is "an important problem," that a framework and analysis for it "are therefore valuable to the community," and that the framework and metrics "are reasonable."

[W1] Findings beyond known difficulty, and beyond offline simulation

The paper reports four findings that prior cost-estimation work does not. (1) Among frontier models, estimation quality separates from task success: success and hit rate correlate at r = 0.81 when the weakest model is included and r = 0.33 among the four frontier models (Section 4.1). (2) The infeasibility signal saves more tokens than a hindsight-tuned spending rule in 9 of 15 model–environment cells (Table 2). (3) Interval width grows with task cost and does not follow continuation uncertainty (Appendix G.8). (4) Binary feasibility is learnable, with balanced accuracy rising from 0.255 to 0.911 after SFT, while intervals stay task-specific. The training result is not a single probe: Appendix I extends SFT and GRPO to three tasks (Search-R1, Sokoban, SWE-bench) and three model families (Qwen2.5-7B, Qwen3-4B, Llama-3.1-8B), and in-task training raises the reward in every setting.

The infeasibility signal also works when the agent runs online. Table R5 gives the online run (full table and setup in the reply to YMm1).

Table R5: Stop rules, online Sokoban, 32 levels.

ConditionLevels solvedAgent tokens per levelEstimation tokens per levelSolved levels wrongly stopped (of the 26 A solves)
A (no estimate)26/321,8450-
B (token cap, 1,400)14/321,456012
E (stop rule only)16/321,2182,96110
C (estimate shown + stop rule)18/321,3223,0588

Stopping on the infeasibility prediction wrongly stops 8 or 10 of the 26 solved levels, against 12 for the token cap. At the same number of wrong stops, it saves more tokens on failing levels than the best cap chosen in hindsight (0.48 and 0.59, against 0.30 and 0.35), as in the offline result in Table 2. The estimation cost here is an upper bound, because the same frontier model estimates at every turn; the paper's training results show that a small model can learn the signal.

[W2] Generalization across models

A benchmark is needed because budget awareness varies across models. Table R6 puts all seven models we tested on Sokoban side by side. Our protocol needs only a model's stated cost interval and infeasibility prediction, so it can be rerun on any future model. The model ranking holds under the prompt change we tested: with the 80% request, the order of models by hit rate is unchanged on Search-R1 and the top model is unchanged on Sokoban (Table R1, reply to uH3j). We evaluate the stated estimate by design, because a deployed agent can act only on what it states and most frontier agents are available only through APIs; probing internal representations is a valuable future direction outside the scope of this benchmark.

Table R6: Sokoban, all seven models.

ModelHit rateFailed runs flagged before 60% of the budgetCorrelation of interval width with task size
GPT-5.20.3620.230.44
Opus 4.70.4660.040.67
Sonnet 4.60.4510.180.58
Gemini 3.10.2390.180.03
Qwen3-235B0.1600.07-0.09
gpt-6.1-sol0.6250.730.86
Opus 5.50.2280.930.51

Hit rate ranges from 0.160 to 0.625 and early flagging from 0.04 to 0.93, so a model's budget awareness cannot be read off from another model. Interval width grows with task size in five of the seven models (correlation 0.44–0.86). Our benchmark separates all seven models on these measures, which is the information a practitioner needs when choosing a model to run under a budget. Setup details for the newer models are in the reply to uH3j.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Summary and Additional Experiments

We thank the reviewers for describing the experimental design as "good" with "multiple established datasets, and also multiple reasonable metrics" (YMm1), for describing intermediate-prefix estimation as "a useful way to study agent self-monitoring" with "several negative results" that are "informative" (uH3j), and for calling budget estimation "an important problem" with a "reasonable" framework and metrics (Fuj2). The main results of the paper already rest on experiments, and the reviewers' suggestions are all valuable. We used them to add experiments that make the paper better.

Summary

Most prior budget-aware agent work treats the budget as a fixed condition (e.g., Liu et al., 2026), and adaptive stopping decides when to halt without saying how much more is needed (Zhou et al., 2026). We treat the budget as something the agent must be aware of: at every step, the agent estimates how much it still needs to spend (as a cost interval, e.g., 800–1,200 more tokens) and predicts whether it can finish within budget (the infeasibility prediction). With this estimate, an agent can stop early or ask for more budget. We find that:

  • Solving a task does not mean the model is budget-aware.
  • The models in the paper cannot yet manage a budget with their own estimates: they recognize failure late, and their cost intervals grow wider as the task gets larger without following how uncertain this run is (Appendix G.8). Newer models flag failure earlier, and their intervals still grow with task size (Table R2).
  • Their "cannot finish" signal is already useful for cutting losses: when the agent says midway that it cannot finish within budget and we stop the run there, we save more tokens than a "stop after X tokens" rule at the same number of lost solved tasks, even though X is tuned in hindsight (Table 2).

Following the reviewers' suggestions, we ran four new experiments:

  1. Telling the model how sure to be has a limited effect on budget awareness. When the prompt asks for a cost interval that contains the true remaining cost with 80% probability, the interval still contains it only 3 to 5 times out of 10 (Table R1, response to uH3j). The calibration gap the paper reports therefore remains when the target is stated.
  2. Our benchmark tracks progress across model generations. gpt-6.1-sol and Opus 5.5 flag failing runs earlier, and their cost intervals still grow with task size (Table R2, response to uH3j; Table R6, response to Fuj2). Opus 5.5 solves 0.554 of the levels within budget (0.976 when the budget is ignored) and has the lowest hit rate among the frontier models.
  3. The infeasibility signal is worth extracting during execution. When the agent runs online and is stopped once it predicts it cannot finish, it stops 5 of the 6 levels that would fail, and at the same number of solved levels wrongly stopped it saves more tokens on failing levels than the best token cap chosen in hindsight (0.48 against 0.30), as in Table 2 (Table R3, response to YMm1; Table R5, response to Fuj2). The paper already shows that a 7B model can learn this signal (Section 5.1).
  4. Models rarely use their own token spending record when they estimate. Their reasoning mostly counts "turns left × tokens per turn", and only 0.15 of estimates in the token environments refer to the spending of past turns (Table R4, response to YMm1). This matches the paper's control in Section 4.1, where removing the spending history leaves feasibility F1 unchanged.

The new results show that the calibration gap is not a prompt effect, that our benchmark tracks progress across model generations, and that the infeasibility signal is useful when the agent runs online. We demonstrate this in more detail in each reviewer's individual responses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.