The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks

Hugging Face Daily Papers Papers
taste-benchmark llm-agents long-horizon-tasks decision-making agent-training

Summary

This paper introduces Taste-Bench, a benchmark for measuring taste in LLM agents' long-horizon decisions, finding that frontier models have low accuracy and that taste can be improved through distillation training.

LLM agents increasingly work on long-horizon tasks, and the decisions they make along the way, such as which hypothesis to test or which implementation to build on, determine the outcome of the whole run. Making these decisions well is becoming a key capability for both engineering and research agents. We refer to the ability to make good long-horizon decisions as the taste of an agent. While existing benchmarks measure the end-to-end success of agents on long-horizon tasks, none of them measures the taste of an agent. To address this problem, we build Taste-Bench, a benchmark of taste questions constructed automatically from trajectories that agents produced in engineering and research tasks. Each question presents a decision fork, a point in a trajectory where multiple directions are available and one of them leads to a better outcome, and the evaluated model chooses among these directions without seeing what happens after the fork. We mine these forks automatically from parallel attempts at the same task and from detours inside a single trajectory, without needing human annotation. We evaluate frontier models on Taste-Bench and find that the best model answers only 59.7% of the questions correctly. We further find that forks whose deciding evidence appears later in the trajectory are much harder for every model, and that a larger reasoning budget does not improve the accuracy. Finally, we show that taste can be trained. We distill the judgment of a teacher that has seen the outcome into a student model, and the student makes better decisions on unseen tasks and improves end-to-end success on held-out SWE-bench Pro tasks.
Original Article
View Cached Full Text

Cached at: 09/23/26, 07:31 AM

Paper page - The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks

Source: https://huggingface.co/papers/2609.25804

Abstract

LLMagentsincreasinglyworkonlong-horizontasks,andthedecisionstheymakealongtheway,suchaswhichhypothesistotestorwhichimplementationtobuildon,determinetheoutcomeofthewholerun.Makingthesedecisionswellisbecomingakeycapabilityforbothengineeringandresearchagents.Werefertotheabilitytomakegoodlong-horizondecisionsasthetasteofanagent.Whileexistingbenchmarksmeasuretheend-to-endsuccessofagentsonlong-horizontasks,noneofthemmeasuresthetasteofanagent.Toaddressthisproblem,webuildTaste-Bench,abenchmarkoftastequestionsconstructedautomaticallyfromtrajectoriesthatagentsproducedinengineeringandresearchtasks.Eachquestionpresentsadecisionfork,apointinatrajectorywheremultipledirectionsareavailableandoneofthemleadstoabetteroutcome,andtheevaluatedmodelchoosesamongthesedirectionswithoutseeingwhathappensafterthefork.Weminetheseforksautomaticallyfromparallelattemptsatthesametaskandfromdetoursinsideasingletrajectory,withoutneedinghumanannotation.WeevaluatefrontiermodelsonTaste-Benchandfindthatthebestmodelanswersonly59.7%ofthequestionscorrectly.Wefurtherfindthatforkswhosedecidingevidenceappearslaterinthetrajectoryaremuchharderforeverymodel,andthatalargerreasoningbudgetdoesnotimprovetheaccuracy.Finally,weshowthattastecanbetrained.Wedistillthejudgmentofateacherthathasseentheoutcomeintoastudentmodel,andthestudentmakesbetterdecisionsonunseentasksandimprovesend-to-endsuccessonheld-outSWE-benchProtasks.

View arXiv pageView PDFGitHub1Add to collection

Get this paper in your agent:

hf papers read 2609\.25804

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.25804 in a model README.md to link it from this page.

Datasets citing this paper1

#### wenbopan/taste-bench Viewer• Updatedabout 4 hours ago • 502 • 7

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.25804 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

A Matter of TASTE: Improving Coverage and Difficulty of Agent Benchmarks

Hugging Face Daily Papers

TASTE is an automated method for generating challenging agent benchmarks with broader tool-use coverage by evolving tool sequences through adaptive contrastive n-gram modeling and iterative difficulty refinement. The resulting τ^c-Bench reveals that models nearly saturating existing benchmarks suffer severe performance drops, indicating saturation rather than robust skill.