Artificial Intelligence · arXiv:2603.01209 v3 · submitted 1 Mar 2026
Evaluating Agents Across Runtime Contracts: When Mismatch Costs Efficiency or Quality
The paper says
“Evaluations of trace-fine-tuned agents should specify episode limits, test runtime transfer across caps, and report cost alongside quality; otherwise, the same mismatch may appear as redundant computation or near-total quality loss.”
from the abstract · picked out as the main claim by a classifier
Reviews
Reviews by people
No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
In CodeAct, language-model agents write Python that calls tools and use execution feedback to choose actions. Persistent runtimes preserve Python variables between actions; stateless runtimes clear them without resetting task progress. Training traces demonstrate task-solving strategies and runtime-specific ways to store and recover intermediate results. We study runtime transfer: whether agents trained under one contract remain effective under the other, a dependence that fixed-runtime evaluations can conceal. Across three tasks requiring a working record built from tool feedback, we fine-tune separate Qwen3-8B agents per task and runtime on instance-paired persistent and stateless traces and evaluate all four training-deployment combinations. We vary the per-turn tool-call cap: how many calls one Python action may execute. In our primary task, agents inspect hidden item attributes and select a high-value subset under a weight limit. At 80 calls per turn, a persistent-trained agent deployed statelessly scores 0.61 of optimal value against 0.77 under matched persistent deployment, but uses 4.4 times as many tokens rebuilding its record. Tightening the cap to 25 reduces mismatched quality to 0.07 while matched quality remains 0.66. Trace replay shows that the cap interrupts inspection, the reset erases the partial record, and rebuilding consumes calls needed for progress. The widening gap replicates across training seeds, a second rollout, and two additional base models. Across the settings studied, cap tightening produces a mismatch-specific collapse only where interruptions are frequent and matched execution resumes while mismatched execution restarts. Evaluations of trace-fine-tuned agents should specify episode limits, test runtime transfer across caps, and report cost alongside quality; otherwise, the same mismatch may appear as redundant computation or near-total quality loss.
The paper's own sentences, arranged. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-01 · no language-model review
How the main claim is stated: hedged possibility 0.62
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.66 · human participants 0.03 · controlled comparison 0.89 · randomized assignment 0.04 · states quantitative result 0.99 · compares with prior results 0.60 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v3). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| In CodeAct, language-model agents write Python that calls tools and use execution feedback to choose actions. | background 1.00 | 0.07 | 0.01 | asserted 1.00 |
| Persistent runtimes preserve Python variables between actions; stateless runtimes clear them without resetting task progress. | background 0.99 | 0.12 | 0.02 | asserted 1.00 |
| Training traces demonstrate task-solving strategies and runtime-specific ways to store and recover intermediate results. | background 0.73 | 0.22 | 0.02 | asserted 1.00 |
| We study runtime transfer: whether agents trained under one contract remain effective under the other, a dependence that fixed-runtime evaluations can conceal. | gap 0.54 | 0.19 | 0.10 | hedged possibility 0.52 |
| Across three tasks requiring a working record built from tool feedback, we fine-tune separate Qwen3-8B agents per task and runtime on instance-paired persistent and stateless traces and evaluate all four training-deployment combinations. | method 0.87 | 0.97 | 0.04 | asserted 0.90 |
| We vary the per-turn tool-call cap: how many calls one Python action may execute. | method 0.97 | 0.85 | 0.02 | asserted 0.68 |
| In our primary task, agents inspect hidden item attributes and select a high-value subset under a weight limit. | method 0.95 | 0.71 | 0.02 | asserted 0.95 |
| At 80 calls per turn, a persistent-trained agent deployed statelessly scores 0.61 of optimal value against 0.77 under matched persistent deployment, but uses 4.4 times as many tokens rebuilding its record. | result 1.00 | 0.96 | 0.11 | asserted 1.00 |
| Tightening the cap to 25 reduces mismatched quality to 0.07 while matched quality remains 0.66. | result 1.00 | 0.95 | 0.09 | asserted 1.00 |
| Trace replay shows that the cap interrupts inspection, the reset erases the partial record, and rebuilding consumes calls needed for progress. | interpretation 0.96 | 0.92 | 0.06 | asserted 1.00 |
| The widening gap replicates across training seeds, a second rollout, and two additional base models. | result 0.98 | 0.95 | 0.07 | asserted 1.00 |
| Across the settings studied, cap tightening produces a mismatch-specific collapse only where interruptions are frequent and matched execution resumes while mismatched execution restarts. | interpretation 0.73 | 0.90 | 0.45 | asserted 0.99 |
| Evaluations of trace-fine-tuned agents should specify episode limits, test runtime transfer across caps, and report cost alongside quality; otherwise, the same mismatch may appear as redundant computation or near-total quality loss. | future work 0.78 | 0.42 | 0.81 | hedged possibility 0.62 |
What kind of paper the classifier reports this is: empirical study 1.00 · evidence from observation or experiment 0.96
What was examined, and what was not
Sources read
- arXiv listing abstract v3
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-01: 1 request, 9,512 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Challenges and discussion
Know something about this paper? Leave a comment.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment