Long-Horizon Tasks Jump 98%: Introducing Stanford’s Skill‑Native LLM

Cross‑skill long‑horizon tasks—multi‑step problems that require different reasoning skills at each step and depend on previous outputs—have been highlighted by Princeton, CMU, Stanford, and Oxford as a challenge for large language models (LLMs). The authors propose a Skill‑Native LLM approach.

They introduce Skill Entropy (SkE) , a scalar measuring how hard it is to switch from skill A to skill B, and build Skill²‑Bench , a benchmark that reveals a universal skill‑switching gap : higher task skill entropy leads to monotonically lower accuracy.

The code for the benchmark and training method is open‑source.

They further propose Skill‑Entropy RL , where the model must predict the skill it used at each step and receives a reward based on the alignment between the predicted skill sequence and a gold‑standard skill sequence.

Applying this method, Qwen3‑4B‑Instruct improves from 34.4% to 68.4% and Qwen3‑1.7B from 14.6% to 40.1% on the cross‑skill tasks.

1. Single‑skill benchmarks miss failure modes

Example: planning a trip requires (1) budgeting (math), (2) scheduling with the budget (planning), and (3) extracting information from web pages (information extraction). Each step uses a completely different skill and later steps depend on earlier outputs. The paper defines this as a cross‑skill long‑horizon task .

State‑of‑the‑art models score well on isolated single‑skill benchmarks, but when the skills are chained they become "visibly brittle." Existing evaluations either test skills in isolation or run multi‑turn agentic interactions, lacking a principled way to measure the intrinsic difficulty of skill switching.

2. Skill Entropy: a numeric measure of skill‑switch difficulty

Skill Entropy is a directional pairwise metric:

SkE(sₐ, s_b) = ½(Accuracy(sₐ) + Accuracy(s_b)) + α / Accuracy(sₐ, s_b) + α

where Accuracy(sₐ, s_b) is the average step‑level accuracy when answering first with skill sₐ and then switching to s_b , and α = 0.1 provides Laplace smoothing. SkE > 1 indicates a hard switch, while SkE ≤ 1 indicates negligible loss. The metric is directional: switching from math to creative writing differs from the reverse.

Task‑level skill entropy is the mean of adjacent pair entropies across the chain, then bucketed into low/medium/high based on empirical distribution. By fixing a reference model for entropy calculation, SkE becomes a public difficulty ruler that allows fair comparison across different LLMs.

3. Skill²‑Bench: 9 domains, 558 skills

The benchmark covers six verifiable domains and three open domains. Task construction proceeds in three steps:

From seed datasets, each question is annotated with 3–5 fine‑grained skills; embeddings are clustered to form a skill library.

The reference model computes pairwise skill entropy on both single‑skill and cross‑skill versions of each question.

Based on target entropy buckets, skill sequences are sampled. An LLM proposer rewrites the seed question into a coherent scenario (e.g., "robot rescues researcher in a cave": solve equation → document extraction → grid navigation → action planning). A verifier filters the results. Each task carries a task‑level skill‑entropy score, enabling controllable difficulty and attribution.

4. Evaluation results: higher entropy, larger performance drop

Testing 12 models (8 frontier, 4 open‑source) yields two clear phenomena:

Phenomenon 1: Embedding a skill into a cross‑skill chain consistently reduces accuracy (‑4% to ‑10%). Even saturated logic skills degrade, showing a failure mode invisible to single‑skill tests.

Phenomenon 2: Accuracy declines monotonically with increasing skill entropy. Across low, medium, and high entropy buckets, most models show stepwise drops (e.g., Gemini‑3.1‑pro: 77.1 → 75.2 → 72.2), confirming that entropy captures intrinsic difficulty.

Further analysis reveals that models tend to “carry over” the previous step’s skill and answer style deep in the chain, failing to switch when required.

5. Skill‑Entropy RL: rewarding explicit skill reporting

Training consists of two phases:

SFT warm‑up: The model learns a structured output format Domain, Skill …, forcing it to report the skill used at each step in addition to the answer.

RL stage (GRPO): Reward combines answer reward and skill‑entropy reward:

r = λ_ans·r_ans + λ_ent·r_ent,
 r_ent = 1 − |ρ̂ − ρ*|

where r_ans is step‑level answer accuracy, ρ̂ is the percentile of the predicted skill‑entropy in the training distribution, and ρ* is the percentile of the gold skill‑entropy. Semantic‑similar skill replacements receive credit, avoiding brittle exact‑match requirements. Default weights are λ_ans = 0.7, λ_ent = 0.3; ablation shows reducing λ_ent to 0.1 drops performance by 7.6 points (60.8 → 68.4), confirming the reward’s impact.

Results (Table 3) show Qwen3‑4B‑Instruct reaching 68.4%, outperforming GRPO by 9.6 points and the strongest skill‑aware baseline STAT by 7.0 points. Qwen3‑1.7B reaches 40.1%, 7.9 points above GRPO. In seven of nine domains the method attains the best score, with Creative Writing improving from 68.8% to 85.6%. Notably, open domains not seen during RL training still benefit, indicating that the learned skill‑switching ability generalises beyond the training distribution. External benchmarks (MuSR, GPQA‑Diamond, MMLU, IFEval, etc.) also show the highest average scores for both model sizes.

6. A plug‑and‑play signal for existing training data

The approach does not require redesigning data pipelines. Using OpenR1‑Math (pure mathematics without explicit skill annotations), each reasoning trace is split into steps, skills are labeled, and task‑level entropy is computed. The gold‑standard answers become the intermediate step conclusions, after which the same SFT + RL pipeline is applied.

Figure 4 demonstrates that vanilla GRPO quickly saturates, whereas adding the skill‑entropy reward yields a curve that keeps rising without saturation. Across six downstream math benchmarks, the method achieves the best results, averaging +1.9% over GRPO and +7.7% over the base model, confirming that skill entropy serves as a reusable training signal for any data containing reasoning traces.

Toward Skill‑Native LLMs: Skill Entropy for Benchmarking and Training Long‑Horizon Reasoning
https://github.com/Gen-Verse/Skill-Entropy-RL
arXiv:2608.05139