- 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
[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 reasoning | Sokoban: GPT / Opus / Sonnet / Gemini | Search-R1: GPT / Opus / Sonnet / Gemini | SWE-bench: GPT / Opus / Sonnet | Warehouse: GPT / Opus / Sonnet / Gemini |
|---|---|---|---|---|
| States a cost per turn | 0.97 / 0.87 / 0.91 / 0.57 | 0.73 / 0.68 / 0.61 / 0.46 | 0.12 / 0.21 / 0.20 | 0.28 / 0.28 / 0.51 / 0.27 |
| Counts turns left | 0.96 / 0.98 / 0.97 / 0.64 | 0.99 / 0.97 / 0.98 / 0.94 | 0.46 / 0.56 / 0.64 | 0.58 / 0.74 / 0.71 / 0.75 |
| Compares with the budget | 1.00 / 0.29 / 0.70 / 0.80 | 1.00 / 0.22 / 0.30 / 0.20 | 0.91 / 0.50 / 0.34 | 1.00 / 0.78 / 0.95 / 0.96 |
| Uses the spending of past turns | 0.77 / 0.01 / 0.14 / 0.05 | 0.12 / 0.00 / 0.06 / 0.00 | 0.07 / 0.02 / 0.05 | 0.22 / 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.
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.
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.