Sign InOpen Brain
arXivPaperNeeds Review

Toward Skill-Native LLMs: Skill Entropy for Benchmarking and Training Long-Horizon Reasoning

Skill²-Bench measures whether models can switch reasoning skills across dependent steps. Its results expose a failure mode relevant to agents whose plans mix coding, retrieval, calculation and tool use.

arXiv · Aug 5, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Skill²-Bench covers **558 skills across 9 domains** and assigns cross-skill tasks an entropy-based difficulty level. Tests of **8 frontier and 4 open-source models** found accuracy falling as skill-switching difficulty rose.

Practical Implication

Evaluate long-running agents on transitions between skills, not only isolated task scores. The proposed RL method makes models predict each step's skill and improved Qwen3-4B-Instruct from **34.4% to 68.4%** and Qwen3-1.7B from **14.6% to 40.1%**.

Agent-Ready Context
Skill²-Bench covers **558 skills across 9 domains** and assigns cross-skill tasks an entropy-based difficulty level. Tests of **8 frontier and 4 open-source models** found accuracy falling as skill-switching difficulty rose.

Evaluate long-running agents on transitions between skills, not only isolated task scores. The proposed RL method makes models predict each step's skill and improved Qwen3-4B-Instruct from **34.4% to 68.4%** and Qwen3-1.7B from **14.6% to 40.1%**.

Those training gains are shown on two Qwen models and depend on gold skill sequences used by the reward. Whether the taxonomy transfers cleanly to messy, tool-driven software work remains an open question.
Connected Context · Feed7 Judgment

This adds skill transitions as a measurable source of long-horizon failure, complementing benchmarks that score final outcomes or sustained trajectories without explicitly modeling switches between capabilities. The reported training gains suggest step-level skill identification can help two Qwen models, but dependence on gold skill sequences and a fixed taxonomy limits claims about messy tool-driven agents.

AgentHPOBench: A Benchmark For Evaluating LLM Agents as Sequential Hyperparameter OptimizersAgentHPOBench evaluates whether agents improve across sequential experiments; Skill²-Bench adds a possible explanatory dimension by measuring how transitions between required skills increase difficulty.DeepSWE: A Contamination-Resistant Coding Benchmark — James Shi, DatacurveDeepSWE reinforces the need for long-horizon behavioral evaluation, while Skill²-Bench narrows that behavior into explicit skill-switching difficulty rather than repository completion alone.AdvancedMathBench: A Benchmark Suite for Advanced Mathematical Proof Generation and VerificationAdvancedMathBench exposes unreliable verification within rigorous reasoning, whereas Skill²-Bench isolates switching between skills; together they show that long-horizon reliability needs more than one aggregate reasoning score.
Context Map
benchmark#reasoning#agent-evals#agent-reliability
Uncertainty
Those training gains are shown on two Qwen models and depend on gold skill sequences used by the reward. Whether the taxonomy transfers cleanly to messy, tool-driven software work remains an open question.